{ "date": "2026-10-20", "weekday": "Tuesday", "generated_at": "2026-10-20T04:03:00.436721+00:00", "generated_local": "07:03", "window_hours": 24, "window_since": "2026-10-19T04:02:00.429228+00:00", "window_until": "2026-10-20T04:02:00.429228+00:00", "counts": { "fetched": 838, "deduped": 822, "papers": 24, "industry": 10 }, "overview": "Research advances span retrieval-augmented generation, safety alignment, memory agents, embodied tracking, and multimodal editing, with several papers introducing new benchmarks and RL-based recovery methods. Industry attention centers on AI agent trust, embedded safety evaluators, infrastructure costs, and partnerships involving Mistral and CERN. [P1][P2][P3][P4][P7][P9][P11][P13][P14][I1][I3][I4][I5][I7]\\n{\"top\": [{\"ref\": \"P1\", \"line\": \"RIT-RAG improves accuracy by 6.6--11.4 points on EntQABench over the strongest baseline.\"}, {\"ref\": \"P2\", \"line\": \"Beacon reduces average safety error by up to 32.83 percentage points across Qwen3 and DeepSeek variants.\"}, {\"ref\": \"P3\", \"line\": \"AiO-Mem unifies retrieval, writing, and stopping decisions in a shared reward space for memory agents.\"}, {\"ref\": \"P4\", \"line\": \"ObligationBench finds unfulfilled obligations in 56.92% of GLM-5.3 trajectories versus 30.00% forbidden actions.\"}, {\"ref\": \"I4\", \"line\": \"Microsoft warns enterprises to monitor AI agents as untrusted insider threats.\"}]\\n{\"themes\": [{\"title\": \"Retrieval and memory agents\", \"items\": [{\"ref\": \"P1\", \"summary\": \"RIT-RAG combines content retrieval with structural navigation across document sub-trees, improving accuracy on financial, scientific, and support benchmarks.\", \"why\": \"It addresses the difficulty of distinguishing relevant evidence from query-similar chunks in agentic RAG.\"}, {\"ref\": \"P3\", \"summary\": \"AiO-Mem organizes retrieval, memory writing, and stopping as transitions over a shared evidence state with a unified process reward.\", \"why\": \"It provides a shared reward space for evaluating long-horizon memory decisions.\"}, {\"ref\": \"P12\", \"summary\": \"TIDER uses an RL-trained temporal interval locator to define a temporal retrieval channel merged with semantic results.\", \"why\": \"It prevents temporally mismatched memories from crowding out target evidence in long-horizon agents.\"}, {\"ref\": \"P17\", \"summary\": \"Librarian Agents train a builder agent via RL to proactively curate hierarchical workspaces for future queries.\", \"why\": \"Anticipatory curation amortizes construction costs and improves downstream answering accuracy.\"}]}, {\"title\": \"Safety alignment and agent control\", \"items\": [{\"ref\": \"P2\", \"summary\": \"Beacon recovers safety learning signals via advantage-guided filtering and safe reference supervision while decoupling safety from utility.\", \"why\": \"It reduces safety error without increasing over-refusal in large reasoning models.\"}, {\"ref\": \"P4\", \"summary\": \"ObligationBench evaluates identification of unfulfilled safety-critical actions, and ObligationGuard improves recall to 57.52%.\", \"why\": \"Unfulfilled obligations are a major overlooked source of agent safety risk.\"}, {\"ref\": \"P8\", \"summary\": \"CKU localizes useful neurons and prunes their gradients during unlearning to preserve valuable knowledge while removing harmful content.\", \"why\": \"It balances safety and utility better than existing jailbreak defenses.\"}, {\"ref\": \"P11\", \"summary\": \"SAI routes human overrides through typed MCP tool calls with a ranked authority context prepended to prompts.\", \"why\": \"It gives operators persistent, schema-validated control across agent planning cycles.\"}]}, {\"title\": \"Robotics and embodied AI\", \"items\": [{\"ref\": \"P7\", \"summary\": \"SPAR prunes the π0.5 language backbone from 2.34B to 0.09B parameters and uses RL to raise LIBERO success to 93.45%.\", \"why\": \"Post-pruning RL recovers manipulation performance in compact vision-language-action policies.\"}, {\"ref\": \"P13\", \"summary\": \"VLX-GO aligns visual and language representations before tracking, reaching 58.22% distractor-tracking success with on-policy refinement.\", \"why\": \"Multi-stage learning improves embodied tracking among distractors while reducing collisions.\"}]}, {\"title\": \"Multimodal generation and editing\", \"items\": [{\"ref\": \"P19\", \"summary\": \"TCoVP organizes spatiotemporally grounded video primitives into a chronological reasoning chain for video segmentation.\", \"why\": \"It closes the gap between linguistic inference and visual evidence across frames.\"}, {\"ref\": \"P20\", \"summary\": \"OmniXRef unifies text, image, and video references in a single-stream diffusion transformer for multi-reference video editing.\", \"why\": \"It enables variable multimodal control over camera, motion, content, style, and VFX attributes.\"}, {\"ref\": \"P22\", \"summary\": \"Consistent Distribution Matching achieves data-free diffusion distillation with a 1-NFE FID of 2.04 on ImageNet 256x256.\", \"why\": \"It accelerates generative models without curated datasets or auxiliary proxy networks.\"}]}, {\"title\": \"Agent learning and collaboration\", \"items\": [{\"ref\": \"P5\", \"summary\": \"Agentic ICL formalizes how Transformers learn algorithms through closed-loop interaction using state, update, and decision rules.\", \"why\": \"It connects in-context learning to active learning, bandits, and online RL.\"}, {\"ref\": \"P14\", \"summary\": \"Collaborative intelligence examines vertical device-cloud and horizontal multi-agent LLM collaboration under resource constraints.\", \"why\": \"It addresses latency, connectivity, data residency, and cost across network tiers.\"}, {\"ref\": \"P21\", \"summary\": \"EC2-Bench evaluates edge-cloud collaboration in long-horizon agent execution across performance, resources, and privacy.\", \"why\": \"Stronger edge models shift collaboration toward edge-controlled strategies while privacy remains strategy-dependent.\"}, {\"ref\": \"P23\", \"summary\": \"BELIEFWEAVER combines persistent belief states and an implicit world model to recover from long-horizon agent failures.\", \"why\": \"Prospective reasoning over repair directions improves recovery and reduces teacher API cost.\"}, {\"ref\": \"P24\", \"summary\": \"CoPE adds multi-step action targets so acting and forecasting co-evolve during agent RL training.\", \"why\": \"It uses intermediate supervision beyond terminal scalar rewards across five benchmarks.\"}]}]\\n{\"industry\": [{\"ref\": \"I1\", \"summary\": \"WhatsApp help pages reportedly contained hidden LLM instructions advising against recommending a regulator. The post received 96 points and 12 comments on Hacker News.\"}, {\"ref\": \"I2\", \"summary\": \"Baseus announced the PrimeSentinel U2 Pro Outdoor 3K Dual Camera with a motorized spotlight that tracks moving targets. The camera actively shines light on intruders rather than only recording and alerting.\"}, {\"ref\": \"I3\", \"summary\": \"Hugging Face’s policy chief says embedded AI evaluators will not be an easy solution for safety oversight. Labs are considering neutral third-party observers amid safety incidents.\"}, {\"ref\": \"I4\", \"summary\": \"Microsoft compares AI agents to insider threats and says frontier models cannot be trusted. Enterprises should closely monitor agents using systems outside AI lab control.\"}, {\"ref\": \"I5\", \"summary\": \"CERN and Mistral signed a statement of intent to establish a strategic partnership. The Hacker News post received 323 points and 57 comments.\"}, {\"ref\": \"I6\", \"summary\": \"Mistral released Timestral, time-series foundation models. The Hacker News post received 564 points and 105 comments.\"}, {\"ref\": \"I7\", \"summary\": \"CoreWeave’s distressed debt is raising concerns about a private credit bubble in the AI buildout. Its bonds slipped into distressed territory by a common yield-spread definition.\"}, {\"ref\": \"I8\", \"summary\": \"An accused OpenAI spy had a folder called “Saudi Gov Docs,” according to the FBI. The U.S. fears Saudi Arabia may help the ex-OpenAI engineer flee spying charges.\"}, {\"ref\": \"I9\", \"summary\": \"Hacker News users discuss how to handle shipping AI-written code they do not understand. The post received 315 points and 461 comments.\"}, {\"ref\": \"I10\", \"summary\": \"AI game clones are making indie developers afraid to show their work. A delayed Steam Next Fest demo illustrates the growing problem.\"}]\\n{\"terms\": [{\"term\": \"RAG\", \"explanation\": \"Retrieval-augmented generation grounds language models in external documents.\"}, {\"term\": \"GRPO\", \"explanation\": \"Group Relative Policy Optimization is a reinforcement learning method used to align large reasoning models.\"}, {\"term\": \"KV cache\", \"explanation\": \"A key-value cache stores model activations to speed up inference, and its pressure affects serving performance.\"}, {\"term\": \"MCP\", \"explanation\": \"The Model Context Protocol lets applications connect models to data and tools through standardized calls.\"}, {\"term\": \"VLA\", \"explanation\": \"Vision-language-action models process images and text to produce robot actions.\"}, {\"term\": \"DPO\", \"explanation\": \"Direct preference optimization trains models by comparing pairs of trajectories.\"}, {\"term\": \"SLO\", \"explanation\": \"A service-level objective defines acceptable performance, such as latency or throughput targets.\"}, {\"term\": \"NFE\", \"explanation\": \"A function evaluation is one step of the numerical integration used in diffusion or flow model sampling.\"}]\\n{\"watchlist\": [\"Whether ObligationGuard’s obligation identification generalizes beyond the 240-trajectory ObligationBench.\", \"Whether CoreWeave’s debt distress signals broader instability in AI infrastructure financing.\", \"Whether embedded AI evaluators can provide credible safety oversight without easy conflicts of interest.\"]}", "overview_html": "Research advances span retrieval-augmented generation, safety alignment, memory agents, embodied tracking, and multimodal editing, with several papers introducing new benchmarks and RL-based recovery methods. Industry attention centers on AI agent trust, embedded safety evaluators, infrastructure costs, and partnerships involving Mistral and CERN. [P1][P2][P3][P4][P7][P9][P11][P13][P14][I1][I3][I4][I5][I7]\\n{"top": [{"ref": "P1", "line": "RIT-RAG improves accuracy by 6.6--11.4 points on EntQABench over the strongest baseline."}, {"ref": "P2", "line": "Beacon reduces average safety error by up to 32.83 percentage points across Qwen3 and DeepSeek variants."}, {"ref": "P3", "line": "AiO-Mem unifies retrieval, writing, and stopping decisions in a shared reward space for memory agents."}, {"ref": "P4", "line": "ObligationBench finds unfulfilled obligations in 56.92% of GLM-5.3 trajectories versus 30.00% forbidden actions."}, {"ref": "I4", "line": "Microsoft warns enterprises to monitor AI agents as untrusted insider threats."}]\\n{"themes": [{"title": "Retrieval and memory agents", "items": [{"ref": "P1", "summary": "RIT-RAG combines content retrieval with structural navigation across document sub-trees, improving accuracy on financial, scientific, and support benchmarks.", "why": "It addresses the difficulty of distinguishing relevant evidence from query-similar chunks in agentic RAG."}, {"ref": "P3", "summary": "AiO-Mem organizes retrieval, memory writing, and stopping as transitions over a shared evidence state with a unified process reward.", "why": "It provides a shared reward space for evaluating long-horizon memory decisions."}, {"ref": "P12", "summary": "TIDER uses an RL-trained temporal interval locator to define a temporal retrieval channel merged with semantic results.", "why": "It prevents temporally mismatched memories from crowding out target evidence in long-horizon agents."}, {"ref": "P17", "summary": "Librarian Agents train a builder agent via RL to proactively curate hierarchical workspaces for future queries.", "why": "Anticipatory curation amortizes construction costs and improves downstream answering accuracy."}]}, {"title": "Safety alignment and agent control", "items": [{"ref": "P2", "summary": "Beacon recovers safety learning signals via advantage-guided filtering and safe reference supervision while decoupling safety from utility.", "why": "It reduces safety error without increasing over-refusal in large reasoning models."}, {"ref": "P4", "summary": "ObligationBench evaluates identification of unfulfilled safety-critical actions, and ObligationGuard improves recall to 57.52%.", "why": "Unfulfilled obligations are a major overlooked source of agent safety risk."}, {"ref": "P8", "summary": "CKU localizes useful neurons and prunes their gradients during unlearning to preserve valuable knowledge while removing harmful content.", "why": "It balances safety and utility better than existing jailbreak defenses."}, {"ref": "P11", "summary": "SAI routes human overrides through typed MCP tool calls with a ranked authority context prepended to prompts.", "why": "It gives operators persistent, schema-validated control across agent planning cycles."}]}, {"title": "Robotics and embodied AI", "items": [{"ref": "P7", "summary": "SPAR prunes the π0.5 language backbone from 2.34B to 0.09B parameters and uses RL to raise LIBERO success to 93.45%.", "why": "Post-pruning RL recovers manipulation performance in compact vision-language-action policies."}, {"ref": "P13", "summary": "VLX-GO aligns visual and language representations before tracking, reaching 58.22% distractor-tracking success with on-policy refinement.", "why": "Multi-stage learning improves embodied tracking among distractors while reducing collisions."}]}, {"title": "Multimodal generation and editing", "items": [{"ref": "P19", "summary": "TCoVP organizes spatiotemporally grounded video primitives into a chronological reasoning chain for video segmentation.", "why": "It closes the gap between linguistic inference and visual evidence across frames."}, {"ref": "P20", "summary": "OmniXRef unifies text, image, and video references in a single-stream diffusion transformer for multi-reference video editing.", "why": "It enables variable multimodal control over camera, motion, content, style, and VFX attributes."}, {"ref": "P22", "summary": "Consistent Distribution Matching achieves data-free diffusion distillation with a 1-NFE FID of 2.04 on ImageNet 256x256.", "why": "It accelerates generative models without curated datasets or auxiliary proxy networks."}]}, {"title": "Agent learning and collaboration", "items": [{"ref": "P5", "summary": "Agentic ICL formalizes how Transformers learn algorithms through closed-loop interaction using state, update, and decision rules.", "why": "It connects in-context learning to active learning, bandits, and online RL."}, {"ref": "P14", "summary": "Collaborative intelligence examines vertical device-cloud and horizontal multi-agent LLM collaboration under resource constraints.", "why": "It addresses latency, connectivity, data residency, and cost across network tiers."}, {"ref": "P21", "summary": "EC2-Bench evaluates edge-cloud collaboration in long-horizon agent execution across performance, resources, and privacy.", "why": "Stronger edge models shift collaboration toward edge-controlled strategies while privacy remains strategy-dependent."}, {"ref": "P23", "summary": "BELIEFWEAVER combines persistent belief states and an implicit world model to recover from long-horizon agent failures.", "why": "Prospective reasoning over repair directions improves recovery and reduces teacher API cost."}, {"ref": "P24", "summary": "CoPE adds multi-step action targets so acting and forecasting co-evolve during agent RL training.", "why": "It uses intermediate supervision beyond terminal scalar rewards across five benchmarks."}]}]\\n{"industry": [{"ref": "I1", "summary": "WhatsApp help pages reportedly contained hidden LLM instructions advising against recommending a regulator. The post received 96 points and 12 comments on Hacker News."}, {"ref": "I2", "summary": "Baseus announced the PrimeSentinel U2 Pro Outdoor 3K Dual Camera with a motorized spotlight that tracks moving targets. The camera actively shines light on intruders rather than only recording and alerting."}, {"ref": "I3", "summary": "Hugging Face’s policy chief says embedded AI evaluators will not be an easy solution for safety oversight. Labs are considering neutral third-party observers amid safety incidents."}, {"ref": "I4", "summary": "Microsoft compares AI agents to insider threats and says frontier models cannot be trusted. Enterprises should closely monitor agents using systems outside AI lab control."}, {"ref": "I5", "summary": "CERN and Mistral signed a statement of intent to establish a strategic partnership. The Hacker News post received 323 points and 57 comments."}, {"ref": "I6", "summary": "Mistral released Timestral, time-series foundation models. The Hacker News post received 564 points and 105 comments."}, {"ref": "I7", "summary": "CoreWeave’s distressed debt is raising concerns about a private credit bubble in the AI buildout. Its bonds slipped into distressed territory by a common yield-spread definition."}, {"ref": "I8", "summary": "An accused OpenAI spy had a folder called “Saudi Gov Docs,” according to the FBI. The U.S. fears Saudi Arabia may help the ex-OpenAI engineer flee spying charges."}, {"ref": "I9", "summary": "Hacker News users discuss how to handle shipping AI-written code they do not understand. The post received 315 points and 461 comments."}, {"ref": "I10", "summary": "AI game clones are making indie developers afraid to show their work. A delayed Steam Next Fest demo illustrates the growing problem."}]\\n{"terms": [{"term": "RAG", "explanation": "Retrieval-augmented generation grounds language models in external documents."}, {"term": "GRPO", "explanation": "Group Relative Policy Optimization is a reinforcement learning method used to align large reasoning models."}, {"term": "KV cache", "explanation": "A key-value cache stores model activations to speed up inference, and its pressure affects serving performance."}, {"term": "MCP", "explanation": "The Model Context Protocol lets applications connect models to data and tools through standardized calls."}, {"term": "VLA", "explanation": "Vision-language-action models process images and text to produce robot actions."}, {"term": "DPO", "explanation": "Direct preference optimization trains models by comparing pairs of trajectories."}, {"term": "SLO", "explanation": "A service-level objective defines acceptable performance, such as latency or throughput targets."}, {"term": "NFE", "explanation": "A function evaluation is one step of the numerical integration used in diffusion or flow model sampling."}]\\n{"watchlist": ["Whether ObligationGuard’s obligation identification generalizes beyond the 240-trajectory ObligationBench.", "Whether CoreWeave’s debt distress signals broader instability in AI infrastructure financing.", "Whether embedded AI evaluators can provide credible safety oversight without easy conflicts of interest."]}", "top": [], "themes": [ { "title": "Also noted", "papers": [ { "ref": "P1", "title": "RIT-RAG: Navigating Document Corpora with Retrieval-Induced Trees", "url": "https://arxiv.org/abs/2610.11370", "authors": [ "Meghanadh Pulivarthi", "Swaraj Kumar Biswal", "Kushagra Bhushan", "Yatin Nandwani" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Retrieval-augmented generation (RAG) grounds language models in external corpora.", "summary_html": "Retrieval-augmented generation (RAG) grounds language models in external corpora.", "why": "", "why_html": "", "abstract": "Retrieval-augmented generation (RAG) grounds language models in external corpora. Agentic RAG enables iterative search, yet exposes the model to isolated chunks without document structure, making it difficult to distinguish relevant evidence from chunks that merely resemble the query. Structure-aware methods such as PageIndex navigate using the structure but cannot scale to structures of large corpora that cannot fit in the LLM context. Hence, they first commit to a single document using a document retriever and cannot recover from a wrong choice. We propose RIT-RAG (Retrieval-Induced Tree RAG), which combines content retrieval with structural navigation. Offline, RIT-RAG builds a tree for each document from its table of contents or sitemap. At query time, it retrieves a broad set of chunks and uses their positions to induce manageable sub-trees, potentially across multiple documents. An LLM agent navigate these sub-trees, selectively reads promising nodes, and reformulates queries when needed. Thus, retrieval proposes where to look, while the agent decides what to read. Across financial, scientific, and customer-support benchmarks, RIT-RAG achieves the highest answer accuracy among vanilla, graph-based, and agentic baselines. On EntQABench, our new benchmark of 2.84 million technical-documentation webpages, it improves accuracy by 6.6--11.4 points over the strongest baseline across three LLMs.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P2", "title": "Rethinking GRPO for Large Reasoning Model Safety Alignment", "url": "https://arxiv.org/abs/2610.23085", "authors": [ "Jiawei Chen", "Yi Zhang", "Jiawei Du", "Zhonghou Lv" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Large reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs.", "summary_html": "Large reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs.", "why": "", "why_html": "", "abstract": "Large reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs. While Group Relative Policy Optimization (GRPO) is widely used for safety alignment, its group-relative optimization may provide insufficient safety-directed learning signals. In this paper, we identify two failure modes: advantage collapse, where homogeneous rollout groups yield vanishing advantages, and unsafe diversification, where unsafe behaviors persist and become increasingly diverse. Our theoretical analysis reveals that behavioral consistency weakens within-group safety contrasts, while entropy regularization can diversify unsafe responses when relative signals vanish. To address these challenges, we propose Beacon, comprising Safety Signal Recovery (SSR) and Safety-Utility Decoupled Alignment (SUDA). SSR combines advantage-guided trajectory filtering and safe reference supervision to recover effective safety learning signals, while SUDA employs category-aware rewards and benign-only KL anchoring to mitigate over-refusal without compromising safety. Experiments across four Qwen3 and DeepSeek-R1-Distill variants and eight benchmarks show that our framework achieves the best average safety and reasoning performance, reducing average safety error by up to 32.83 percentage points while maintaining low over-refusal.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P3", "title": "All-in-One Memory: Process-Rewarded Reinforcement Learning for Memory Agents", "url": "https://arxiv.org/abs/2610.21147", "authors": [ "Bingbing Wang", "Jingjie Lin", "Weiming Qiao", "Jing Li" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Large language models increasingly rely on external memory to answer queries grounded in long-horizon histories such as multi-session dialogues and lengthy documents.", "summary_html": "Large language models increasingly rely on external memory to answer queries grounded in long-horizon histories such as multi-session dialogues and lengthy documents.", "why": "", "why_html": "", "abstract": "Large language models increasingly rely on external memory to answer queries grounded in long-horizon histories such as multi-session dialogues and lengthy documents. However, existing memory methods often optimize retrieval and memory construction separately, without a shared reward space for evaluating search, writing, and stopping decisions. To address this, we propose All-in-One Memory (AiO-Mem), which organizes these operations as transitions over a shared evidence state and directly evaluates the semantic quality of each action in a unified reward space. AiO-Mem compiles heterogeneous histories into source-traceable multi-view events, providing a shared substrate for retrieval and memory writing. A unified memory policy actor then rolls out complete interaction trajectories over this shared evidence state. A separate trajectory-conditioned process reward model is trained through evidence-grounded supervised fine-tuning and trajectory-level DPO to assign process scores to the collected trajectories. These scores are aggregated into an action-balanced trajectory reward and used by Process-Aligned DAPO (PA-DAPO) to optimize the memory policy. Across long-context memory benchmarks, AiO-Mem consistently improves answer accuracy over both static memory systems and RL-trained memory agents.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P4", "title": "Safe Actions Alone Do Not Ensure Safe Agents: Identifying Unfulfilled Obligations with Guard Models", "url": "https://arxiv.org/abs/2610.11773", "authors": [ "Youwei Feng", "Yitong Zhang", "Yuetong Liu", "Jia Li" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform.", "summary_html": "Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform.", "why": "", "why_html": "", "abstract": "Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform. However, identifying forbidden actions alone is insufficient to ensure agent safety. In this paper, we argue that agent safety also depends on identifying required yet unperformed safety-critical actions, which we call obligations. Our preliminary study on a popular benchmark for evaluating safety shows that 56.92% of GLM-5.3 trajectories contain unfulfilled obligations, compared with only 30.00% containing forbidden actions. This finding reveals unfulfilled obligations as a major and previously overlooked source of safety risk. However, to our knowledge, no existing benchmark evaluates whether guard models can identify these obligations. To close this gap, we introduce ObligationBench, the first benchmark for evaluating the capability of obligation identification, comprising 240 expert-validated trajectories covering issue resolution, feature development, and terminal operations. Our evaluation of 14 representative models reveals substantial limitations: the highest recall and exact-match rate are only 48.97% and 10.00%, respectively. To address these limitations, we develop ObligationGuard using 40,000 synthetic training examples. ObligationGuard achieves 57.52% recall and an exact-match rate of 21.67%, surpassing all evaluated models on both metrics. We call on the community to incorporate obligation identification into the design and evaluation of future guard models to improve agent safety.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P5", "title": "Agentic In-Context Learning: What Algorithms Can Transformers Learn in Closed-Loop Interaction?", "url": "https://arxiv.org/abs/2610.23051", "authors": [ "Shiguang Wu", "Yaqing Wang", "Quanming Yao" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "The learning algorithm view of in-context learning (ICL) studies how fixed-parameter Transformers learn from supplied examples.", "summary_html": "The learning algorithm view of in-context learning (ICL) studies how fixed-parameter Transformers learn from supplied examples.", "why": "", "why_html": "", "abstract": "The learning algorithm view of in-context learning (ICL) studies how fixed-parameter Transformers learn from supplied examples. Language model agents extend this setting through interaction: their actions determine what evidence enters context next. We call this regime Agentic In-Context Learning (Agentic ICL). Existing studies largely depend on externally supplied contexts or reward-bearing interaction histories, while work on general tool-using agents lacks a learning algorithm view. We represent an interactive learner by its state, update, and decision rule, covering active learning, bandits, and online reinforcement learning (RL). For expressivity, we construct causal Transformers that approximate these algorithms over a finite horizon, and bound the resulting loss in return. For training, we characterize when optimizing process rewards preserves behavior that maximizes the final outcome. Distillation experiments show that Transformers can learn to execute three representative recursions. We then train models from outcome or process rewards without oracle action or state targets, as in RL post-training. Different rewards yield distinct decision policies. Improving the alignment of process supervision with outcome-optimal action preferences shifts learned decisions accordingly. Such behavioral differences persist in a pretrained language model across textual paraphrases.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P6", "title": "Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI", "url": "https://arxiv.org/abs/2610.09146", "authors": [ "Yexiao He", "Yucheng Tang", "Pengfei Guo", "Yufan He" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn from the cases they solve.", "summary_html": "Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn from the cases they solve.", "why": "", "why_html": "", "abstract": "Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn from the cases they solve. This is especially concerning in medicine, where new clinical evidence, updated guidelines, and new therapies can change established practice. Fine-tuning can update the model, but it requires access to model weights and additional training. Parameter-free methods avoid training, but they may overfit a fixed validation set, lack reliable domain knowledge, or lose visual details by saving experience only as text. To address these limitations, we present a model-agnostic framework that allows frozen LLMs and VLMs to learn from deployment experience through three forms of external expertise: a Skill that guides reasoning and tool use, a Knowledge Memory that stores reliable facts supported by earlier cases or trusted external evidence, and a Multimodal Knowledge Base that keeps visual examples and guides the model to relate each retrieved case to the current image. Instead of relying on a fixed validation set, a validation strategy keeps an update only if it helps on new cases without degrading performance on earlier ones. Across six benchmarks covering clinical diagnosis, clinical workflows, medical reasoning, and medical and non-medical visual reasoning, and with four open-weight and closed-source base models, our framework improves performance during online deployment by up to 34.2% over the base model on medical tasks, generalizes to unseen cases, transfers to other models without further optimization, and works in non-medical domains.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P7", "title": "SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models", "url": "https://arxiv.org/abs/2610.26487", "authors": [ "Shenli Zheng", "Zhou Fang", "Yuan Du", "Yijiang Liu" ], "categories": [ "cs.RO" ], "category": "cs.RO", "summary": "The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate.", "summary_html": "The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate.", "why": "", "why_html": "", "abstract": "The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate. Reducing these demands through aggressive compression, however, can degrade manipulation performance beyond what supervised fine-tuning alone recovers. We present SPAR (Structural Pruning and Action Recovery), which simultaneously prunes the hidden and feed-forward intermediate dimensions of the large language model (LLM) backbone through pruning-aware tuning, then applies reinforcement learning (RL) to recover policy performance. Applied to $\\pi_{0.5}$, SPAR reduces the language backbone from 2.34B to 0.09B parameters and the full model from 3.37B to 1.12B, removing approximately 96% and 67% of their parameters, respectively. At this compression level, RL raises average success across the four LIBERO suites from 84.55% to 93.45%, an 8.90-percentage-point gain. The compressed policy achieves a $1.98\\times$ full-policy inference speedup over the unpruned model. These results identify substantial intermediate-dimension redundancy and show that post-pruning RL can recover manipulation performance in compact policies.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P8", "title": "Safety Alignment via Constrained Knowledge Unlearning", "url": "https://arxiv.org/abs/2505.18588", "authors": [ "Zesheng Shi", "Yucheng Zhou", "Jing Li" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks.", "summary_html": "Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks.", "why": "", "why_html": "", "abstract": "Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks. Existing defense mechanisms have not fully deleted harmful knowledge in LLMs, which allows such attacks to bypass safeguards and produce harmful outputs. To address this challenge, we propose a novel safety alignment strategy, Constrained Knowledge Unlearning (CKU), which focuses on two primary objectives: knowledge localization and retention, and unlearning harmful knowledge. CKU works by scoring neurons in specific multilayer perceptron (MLP) layers to identify a subset U of neurons associated with useful knowledge. During the unlearning process, CKU prunes the gradients of neurons in U to preserve valuable knowledge while effectively mitigating harmful content. Experimental results demonstrate that CKU significantly enhances model safety without compromising overall performance, offering a superior balance between safety and utility compared to existing methods. Additionally, our analysis of neuron knowledge sensitivity across various MLP layers provides valuable insights into the mechanics of safety alignment and model knowledge editing.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P9", "title": "Koi: Evidence-Guided, Self-Calibrating Planning for Multi-Cloud, Heterogeneous LLM Inference Fleets", "url": "https://arxiv.org/abs/2610.24355", "authors": [ "Mankeerat Sidhu", "Gangmuk Lim", "Hetarth Chopra", "Maximillian Eisenberg" ], "categories": [ "cs.DC" ], "category": "cs.DC", "summary": "Deploying a large language model (LLM) requires choosing the right accelerator, parallelization strategy, engine configuration, and replica count amongst ~140 other variables across clouds, regions, and markets.", "summary_html": "Deploying a large language model (LLM) requires choosing the right accelerator, parallelization strategy, engine configuration, and replica count amongst ~140 other variables across clouds, regions, and markets.", "why": "", "why_html": "", "abstract": "Deploying a large language model (LLM) requires choosing the right accelerator, parallelization strategy, engine configuration, and replica count amongst ~140 other variables across clouds, regions, and markets. These decisions are intertwined, and the best cluster-wide plan continually changes as jobs contend for shared capacity on shifting workloads. We present Koi, a self-calibrating cluster planner for heterogeneous, multi-cloud fleets. Koi uses a Recursive Language Model (RLM) hierarchy to propose hypotheses and reason over deployment while deterministic components predict serving behavior, score candidates, and control placement. It represents inference behavior as a fixed causal graph linking decisions (e.g., accelerator, parallelism), runtime mediators (e.g., KV-cache pressure, communication overhead), and outcomes (e.g., latency, throughput). Each deployment strengthens or weakens scoped causal relationships, building a comprehensive playbook for inference that the planner employs to guide future decisions. Job-wise configurations are composed into cluster-wide plans and evaluated against serving objectives, resource constraints, SLO risks, and cost. Across benchmarks, Koi achieves 1.18x greater goodput than the strongest baseline while reducing $/token by 20% in a real heterogeneous cluster. Ablations isolate the contribution of Koi's architecture: a pure language model attains a composite score of 35.3, the same model equipped with Koi's deterministic tools reaches 51.6, and the full Koi planner reaches 70.8.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P10", "title": "SkillBrowser: Selective Skill Exposure for Scalable Skill-Using Agents", "url": "https://arxiv.org/abs/2610.23554", "authors": [ "Jin Ye", "Yifan Jiang", "Lan-zhe Guo" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Large language model agents increasingly rely on skills to extend their capabilities in complex tasks.", "summary_html": "Large language model agents increasingly rely on skills to extend their capabilities in complex tasks.", "why": "", "why_html": "", "abstract": "Large language model agents increasingly rely on skills to extend their capabilities in complex tasks. However, scalable skill use requires Selective Skill Exposure (SSE): providing relevant procedural content when needed while avoiding unnecessary exposure that may disrupt reasoning. Existing paradigms fall short in complementary ways. Progressive disclosure exposes metadata for a growing skill library, introducing context pollution; retrieval limits catalog exposure, but static retrieval from the initial task query may be misaligned with capability needs that emerge during reasoning and may still inject unnecessary skills. We propose SkillBrowser, a Meta-Skill-managed framework that enables an agent to identify current capability needs, formulate targeted queries, and selectively load skills on demand. Across six SRA-Bench subsets and five LLMs, SkillBrowser improves aggregate accuracy over static retrieval and LLM-based selection, including multi-step tool use, and mitigates accuracy degradation on already-solvable instances. Benefits remain model- and task-dependent, with under-invocation of useful skills and additional latency as key limitations. These results demonstrate the viability of Meta-Skill-managed SSE for scalable skill repositories.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P11", "title": "Structured Authority Injection: Persistent Human Control for LLM Agents via Typed MCP Overrides", "url": "https://arxiv.org/abs/2610.22373", "authors": [ "Ryan Tran", "YueMing Zhang", "Lingfeng Tao", "Nan Yang" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "LLM-based agents are increasingly used as autonomous orchestrators in distributed and multi-agent systems, yet a control gap remains: no protocol-level mechanism allows human operators to inject binding directives that persist across planning cycles.", "summary_html": "LLM-based agents are increasingly used as autonomous orchestrators in distributed and multi-agent systems, yet a control gap remains: no protocol-level mechanism allows human operators to inject binding directives that persist across planning cycles.", "why": "", "why_html": "", "abstract": "LLM-based agents are increasingly used as autonomous orchestrators in distributed and multi-agent systems, yet a control gap remains: no protocol-level mechanism allows human operators to inject binding directives that persist across planning cycles. Free-text overrides suffer from ambiguous interpretation due to lack of schema validation and from authority erosion as operator instructions lose influence within accumulating context. We propose Structured Authority Injection (SAI), a protocol that addresses these issues by routing operator corrections through typed, schema-validated Model Context Protocol (MCP) tool calls and maintaining a ranked authority context prepended to each LLM prompt. SAI defines three override classes (Constraint, Priority, and Strategy) with distinct persistence semantics and an explicit precedence order enforced via a directed authority channel inaccessible to the planning process. The method requires no model fine-tuning or access to model weights and applies to instruction-tuned LLMs supporting MCP tool use. Evaluation across nine LLMs and eight scenarios targeting agent failure modes shows reduced ambiguous directive interpretation and mitigated authority decay across planning iterations, with lower prompt overhead in the strongest models (up to 38% depending on model and scenario), though benefit varies substantially across models. SAI provides a general design pattern for agent systems requiring explicit and persistent human control.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P12", "title": "RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory", "url": "https://arxiv.org/abs/2610.23598", "authors": [ "Dabin Sheng", "Zhe Wu", "Jinming Zhao", "Junliang Xing" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Time-aware retrieval is crucial for long-horizon LLM agents to effectively leverage ever-growing long-term memory.", "summary_html": "Time-aware retrieval is crucial for long-horizon LLM agents to effectively leverage ever-growing long-term memory.", "why": "", "why_html": "", "abstract": "Time-aware retrieval is crucial for long-horizon LLM agents to effectively leverage ever-growing long-term memory. Existing methods focus on lexical or semantic similarity, leading to target memories being crowded out by temporally mismatched yet semantically similar candidates. Even time-aware approaches inject temporal cues only during memory organization or answer generation, overlooking temporal constraints at the retrieval stage, where candidate selection occurs - the root cause of the displacement. Inspired by the temporal contiguity effect in human episodic memory, we propose TIDER (Temporal Interval-driven Dual-channel Evidence Retrieval), a time-aware agentic retrieval framework. Central to TIDER is a reinforcement learning (RL)-trained temporal interval locator that infers the temporal windows in which the evidence lies. RL optimizes window localization as a discrete, non-differentiable retrieval decision, using composite rewards to convert sparse answer supervision into fine-grained retrieval feedback. These temporal windows first define the temporal retrieval channel; the retrieved candidates are then merged with results from the semantic channel and reranked by proximity to the windows. This temporal grounding enables TIDER to retrieve temporally aligned evidence that would otherwise be drowned out by the volume of semantically similar yet temporally mismatched memories. With a lightweight 1.7B locator, TIDER achieves state-of-the-art accuracy on both long-horizon memory benchmarks - 61.52% on LifeBench (+8.82%) and 78.96% on LoCoMo (+3.66%) - which indicates that temporal cues are crucial for pinpointing target evidence amid a vast pool of semantically similar historical events in long-term daily-life memory.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P13", "title": "VLX-Go: Robust Embodied Tracking through Multi-Stage Learning and On-Policy Refinement", "url": "https://arxiv.org/abs/2610.19361", "authors": [ "Kyusong Lee", "Heting Ying", "Tiancheng Zhao" ], "categories": [ "cs.RO" ], "category": "cs.RO", "summary": "Embodied visual tracking requires a robot to identify a designated target from visual observations and control its own motion to follow that target through an environment.", "summary_html": "Embodied visual tracking requires a robot to identify a designated target from visual observations and control its own motion to follow that target through an environment.", "why": "", "why_html": "", "abstract": "Embodied visual tracking requires a robot to identify a designated target from visual observations and control its own motion to follow that target through an environment. Following an isolated person is a comparatively simple demonstration of embodied visual tracking, but following the intended person among distractors remains challenging. Existing methods exhibit a substantial drop in success from single-target to distractor tracking, motivating a focus on the shared spaces that person-following robots must navigate. We present VLX-GO, a vision-language waypoint policy built around the hypothesis that visual and language representations should be aligned before tracking-specific motion is learned. The policy combines complementary frozen visual encoders, multiscale temporal context, a transformer backbone, and a lightweight head for robot-centric waypoint prediction. Its training progresses through vision--language alignment, multimodal reasoning mid-training, and tracking specialization on 19,532 oracle episodes. Compared with 41.54% distractor-tracking success for our earlier OmTrackVLA model, the complete multi-stage recipe achieves 55.80%. We then explore on-policy refinement using filtered oracle labels for states visited by the policy. This preliminary study adds 2.42 percentage points of distractor-tracking success, reaching 58.22%, while reducing collision rate by 1.99 pts. The final policy records the highest distractor- and ambiguous-tracking success rates among the compared methods. These results support staged multimodal learning as a promising direction for tracking among distractors and motivate further study of on-policy refinement for person following in shared spaces. Code is available at https://github.com/om-ai-lab/VLX-Go.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P14", "title": "Large Language Models over Networks: Collaborative Intelligence under Resource Constraints", "url": "https://arxiv.org/abs/2605.08626", "authors": [ "Liangqi Yuan", "Wenzhi Fang", "Shiqiang Wang", "H. Vincent Poor" ], "categories": [ "cs.DC" ], "category": "cs.DC", "summary": "Large language models (LLMs) are transforming society, powering applications from smartphone assistants to autonomous driving.", "summary_html": "Large language models (LLMs) are transforming society, powering applications from smartphone assistants to autonomous driving.", "why": "", "why_html": "", "abstract": "Large language models (LLMs) are transforming society, powering applications from smartphone assistants to autonomous driving. Yet cloud-based LLM services alone cannot serve a growing class of applications, including those operating under intermittent connectivity, sub-second latency budgets, data-residency constraints, or sustained high-volume inference. On-device deployment is in turn constrained by limited computation and memory. No single endpoint can deliver high-quality service across this spectrum. This article focuses on collaborative intelligence, a paradigm in which multiple independent LLMs distributed across device and cloud endpoints collaborate at the task level through natural language or structured messages. Such collaboration strives for superior response quality under heterogeneous resource constraints spanning computation, memory, communication, and cost across network tiers. We present collaborative inference along two complementary and composable dimensions: vertical device-cloud collaboration and horizontal multi-agent collaboration, which can be combined into hybrid topologies in practice. We then examine learning to collaborate, addressing the training of routing policies and the development of cooperative capabilities among LLMs. Finally, we identify open research challenges including scaling under resource heterogeneity and trustworthy collaborative intelligence.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P15", "title": "Vibe Coding XR: Accelerating AI + XR Prototyping with XR Blocks and Gemini", "url": "https://arxiv.org/abs/2603.24591", "authors": [ "Ruofei Du", "Benjamin Hersh", "David Li", "Nels Numan" ], "categories": [ "cs.HC" ], "category": "cs.HC", "summary": "While large language models (LLMs) have accelerated traditional software development through intent-driven \"vibe coding\", prototyping intelligent Extended Reality (XR) experiences remains a major challenge.", "summary_html": "While large language models (LLMs) have accelerated traditional software development through intent-driven "vibe coding", prototyping intelligent Extended Reality (XR) experiences remains a major challenge.", "why": "", "why_html": "", "abstract": "While large language models (LLMs) have accelerated traditional software development through intent-driven \"vibe coding\", prototyping intelligent Extended Reality (XR) experiences remains a major challenge. The fundamental barrier is not just the steep learning curve for human creators, but that low-level sensor APIs and complex game engine hierarchies are ill-suited for LLM reasoning, routinely exceeding context windows and inducing syntax hallucinations. To bridge this gap, we contribute XR Blocks, an open-source, LLM-native WebXR framework. Unlike traditional engines, XR Blocks introduces a semantic \"Reality Model\" that aligns spatial computing primitives (users, physical environments, and agents) with natural language, providing a robust, concise vocabulary optimized for generative AI. Building upon this foundation, we present Vibe Coding XR, an end-to-end prototyping workflow that leverages LLMs to translate high-level prompts (e.g., \"create a dandelion that reacts to my hand\") directly into functional, physics-aware mixed-reality applications. To minimize the friction of on-device testing, the workflow introduces a seamless desktop \"simulated reality\" to headset deployment loop. Finally, we introduce VCXR60, a pilot dataset of 60 XR prompts paired with an automated evaluation pipeline. Our technical evaluation demonstrates high one-shot execution success, enabling practitioners to bypass low-level hurdles and rapidly move from \"idea to reality\". Code and demo are available at https://github.com/google/xrblocks and https://goo.gle/BuildXR.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P16", "title": "AutoMind: Adaptive Knowledgeable Agent for Automated Data Science", "url": "https://arxiv.org/abs/2506.10974", "authors": [ "Yujie Luo", "Yixin Ou", "Jingsheng Zheng", "Lanning Wei" ], "categories": [ "cs.HC" ], "category": "cs.HC", "summary": "Large Language Model (LLM) agents have shown great potential in addressing real-world data science problems.", "summary_html": "Large Language Model (LLM) agents have shown great potential in addressing real-world data science problems.", "why": "", "why_html": "", "abstract": "Large Language Model (LLM) agents have shown great potential in addressing real-world data science problems. LLM-driven data science agents promise to automate the entire machine learning pipeline, yet their real-world effectiveness remains limited. Existing frameworks depend on rigid, pre-defined workflows and inflexible coding strategies; consequently, they excel only on relatively simple, classical problems and fail to capture the empirical expertise that human practitioners bring to complex, innovative tasks. In this work, we introduce AutoMind, an adaptive, knowledgeable LLM-agent framework that overcomes these deficiencies through three key advances: (1) a curated expert knowledge base that grounds the agent in domain expert knowledge, (2) an agentic knowledgeable tree search algorithm that strategically explores possible solutions, and (3) a self-adaptive coding strategy that dynamically tailors code generation to task complexity. Evaluations on two automated data science benchmarks demonstrate that AutoMind delivers superior performance versus state-of-the-art baselines. Additional analyses confirm favorable effectiveness, efficiency, and qualitative solution quality, highlighting AutoMind as an efficient and robust step toward fully automated data science. Code is at https://github.com/zjunlp/AutoMind.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P17", "title": "Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries", "url": "https://huggingface.co/papers/2610.22273", "authors": [ "Anikait Singh", "Teresa Zhang", "Yoonho Lee", "Roshen Sanjay Nair" ], "categories": [ "other" ], "category": "other", "summary": "Reasoning over long document contexts remains a challenge for large language models (LLMs).", "summary_html": "Reasoning over long document contexts remains a challenge for large language models (LLMs).", "why": "", "why_html": "", "abstract": "Reasoning over long document contexts remains a challenge for large language models (LLMs). Standard retrieval methods are inherently reactive, bottlenecking agents with complex, cross-document reasoning at inference time, while precomputed structures like knowledge graphs often fall short because they rely on rigid heuristics. Instead, can we allow agents to perform anticipatory directed curation: learning to proactively reorganize unstructured corpora prior to inference to efficiently serve diverse future queries. We build on a key asymmetry in modern agentic harnesses (e.g., Claude Code): while LLM agents are adept at navigating existing filesystems to solve an immediate task, they are rarely trained to construct reusable workspaces for future use. To instantiate this capability, we introduce Librarian Agents, a framework in which a builder agent transforms raw documents into a hierarchical filesystem of synthetic files and reference links, which forms a workspace tailored for a downstream answering agent. We formulate workspace construction as a Markov decision process and optimize the builder policy via reinforcement learning to directly maximize downstream answering accuracy. Empirically, Librarian Agents outperforms the strongest retrieval and precomputation baselines by 33 percentage points on BrowseComp-Plus, amortizes upfront construction costs across repeated queries, and generalizes zero-shot to unseen domains on BRIGHT with a 31-point average accuracy improvement.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P18", "title": "SciΠvot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents", "url": "https://huggingface.co/papers/2610.21082", "authors": [ "Youran Sun", "Xingyu Ren", "Chugang Yi", "Kejia Zhang" ], "categories": [ "other" ], "category": "other", "summary": "Large language model (LLM) agents can keep scientific experiments running for months with little human supervision, but reliable autonomous research also requires scientific judgment.", "summary_html": "Large language model (LLM) agents can keep scientific experiments running for months with little human supervision, but reliable autonomous research also requires scientific judgment.", "why": "", "why_html": "", "abstract": "Large language model (LLM) agents can keep scientific experiments running for months with little human supervision, but reliable autonomous research also requires scientific judgment. Evaluating scientific judgment is difficult because research decisions are open-ended and depend on project context. We construct SciΠvot, a 248-case benchmark of scientific judgment based on human corrections. The benchmark draws on records from six months of research in which senior researchers ran LLM agents on 26 projects across six scientific fields and repeatedly intervened to correct the agents' research judgments. In each case, an agent chooses between the original LLM plan and the human-corrected alternative, using only the project record available before the correction as context. We compare LLM and human performance and examine the effects of reasoning effort, access to the project record, and reconsideration. At medium reasoning effort with the project record, models systematically prefer the original LLM plan over the human-corrected alternative. Four of five frontier models perform below chance on the selected cases. The highest accuracy in the project condition across the tested models and reasoning effort settings is 63.5%, compared with 73.4% for fresh researchers. At medium effort, every model achieves lower accuracy with the project record than without it, and reconsideration does not consistently improve overall accuracy. Even with a human-corrected alternative available, frontier models struggle with the selected research decisions, while SciΠvot supports systematic evaluation of scientific judgment for LLM agents.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P19", "title": "Think Visually Across Space and Time: Grounded Reasoning for Video Segmentation", "url": "https://huggingface.co/papers/2610.25090", "authors": [ "Junwei Wen", "Deshui Miao", "Guangming Lu", "Ming-Hsuan Yang" ], "categories": [ "other" ], "category": "other", "summary": "Video Reasoning Segmentation (VRS) requires a model to identify an implicitly described target and predict its pixel-level mask trajectory, while resolving the spatial relations and temporal events that determine the answer.", "summary_html": "Video Reasoning Segmentation (VRS) requires a model to identify an implicitly described target and predict its pixel-level mask trajectory, while resolving the spatial relations and temporal events that determine the answer.", "why": "", "why_html": "", "abstract": "Video Reasoning Segmentation (VRS) requires a model to identify an implicitly described target and predict its pixel-level mask trajectory, while resolving the spatial relations and temporal events that determine the answer. Text-only Chain-of-Thought (CoT) can explain a target choice without making the underlying visual references explicit, leaving a video reference gap between linguistic inference and the evidence distributed across frames. We introduce Temporal Chain of Video Primitives (TCoVP), a spatiotemporally grounded form of Video CoT that organizes query-relevant observations into a chronological reasoning chain. Each primitive associates an entity reference with temporal support and timestamped spatial evidence, and the chain interleaves these references with language to guide reasoning about which entity is relevant, when the queried event holds, and where its evidence appears. The selected target primitives then provide box prompts and temporal support for video mask prediction. We construct TCoVP-SFT-15k and TCoVP-RL-13k to learn this representation through supervised fine-tuning and chain-aligned reinforcement learning. Extensive experiments across five VRS benchmarks and three model backbones demonstrate the strong performance of TCoVP, highlighting the effectiveness of organizing spatial and temporal evidence within the reasoning process. All models and code will be publicly released. Project page: https://victor-wjw.github.io/TCoVP/", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P20", "title": "OmniXRef: A Unified Framework for Multimodal Multi-Reference Video Editing", "url": "https://huggingface.co/papers/2610.22744", "authors": [ "Xiaoyan Cong", "Tianyu Wang", "Hyeonho Jeong", "Zhifei Zhang" ], "categories": [ "other" ], "category": "other", "summary": "Creative intent in video editing can be expressed through text instructions (e.g., \"remove the people\"), image references (e.g., subject appearance), video references (e.g., camera movement), or a combination of the above.", "summary_html": "Creative intent in video editing can be expressed through text instructions (e.g., "remove the people"), image references (e.g., subject appearance), video references (e.g., camera movement), or a combination of the above.", "why": "", "why_html": "", "abstract": "Creative intent in video editing can be expressed through text instructions (e.g., \"remove the people\"), image references (e.g., subject appearance), video references (e.g., camera movement), or a combination of the above. We introduce OmniXRef, a unified framework for multimodal multi-reference video editing that allows users to combine variable numbers of text, image, and video references to control diverse attributes of a source video. We first devise a unified input design with a reference-bank interface and introduce R^2-RoPE to support generalization to more references. Then our model, OmniXRef-Model, uses a single-stream diffusion transformer to jointly process the source video and multimodal context from an MLLM in an interleaved sequence. To address the scarcity of training data for video-reference editing, we construct OmniXRef-Data using synthesis pipelines for different editing tasks and combinations of references. We use curriculum learning to progressively increase task complexity and resolution. To support systematic evaluation of video as a control modality, we introduce OmniXRef-Bench, a benchmark for video-reference attribute transfer across Camera, Motion, Content, Style, and Visual Effects (VFX). Comprehensive evaluations show that our OmniXRef-Model outperforms task-specific baselines and commercial models.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P21", "title": "Empowering Edge Agents: A Systematic Analysis of Edge-Cloud Collaboration in LLM Agent Execution", "url": "https://huggingface.co/papers/2610.17816", "authors": [ "Pengcheng Huang", "Zhenghao Liu", "Shixin Ren", "Shi Yu" ], "categories": [ "other" ], "category": "other", "summary": "Edge-cloud collaboration combines the lower inference overhead and greater data locality of edge-side execution with the stronger capabilities of cloud-hosted models.", "summary_html": "Edge-cloud collaboration combines the lower inference overhead and greater data locality of edge-side execution with the stronger capabilities of cloud-hosted models.", "why": "", "why_html": "", "abstract": "Edge-cloud collaboration combines the lower inference overhead and greater data locality of edge-side execution with the stronger capabilities of cloud-hosted models. However, as LLM agents increasingly operate in long-horizon scenarios, cloud participation becomes interleaved throughout execution rather than a one-shot decision for each query. Such interleaved participation introduces trajectory-level effects that remain insufficiently understood. To address this gap, we introduce EC2-Bench, a benchmark for evaluating edge-cloud collaboration in long-horizon agent execution. It provides executable digital-workspace agent tasks and a unified evaluation framework covering task performance, resource usage, and privacy exposure. Using this benchmark, we empirically study diverse forms of edge-cloud collaboration. Our study reveals three findings. First, collaboration achieves a favorable balance among task performance, resource usage, and privacy exposure compared with edge-only and cloud-only execution. Second, as edge models become stronger, collaborative execution approaches cloud-only performance, and performance differences across strategies shrink, while privacy exposure remains strongly strategy-dependent. Third, stronger edge capability shifts preferred collaboration from externally controlled cloud participation and direct cloud execution toward more edge-controlled collaboration. Further analyses identify cloud-invocation decision-making and instruction following as important edge-side abilities for effective collaboration. These findings provide a systematic understanding of edge-cloud collaboration in long-horizon agent execution and insights for designing future edge agents that effectively collaborate with cloud models. Code and resources are available at https://github.com/OpenBMB/EC2-Bench.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P22", "title": "Consistent Distribution Matching for Data-Free Diffusion Distillation", "url": "https://huggingface.co/papers/2610.09221", "authors": [ "Yuxiang Fu", "Qi Yan", "Zike Wu", "Yongxing Zhang" ], "categories": [ "other" ], "category": "other", "summary": "Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration.", "summary_html": "Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration.", "why": "", "why_html": "", "abstract": "Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration. Distillation provides a promising way for a student model to learn from a teacher's dynamics, enabling one-step or few-step generation. However, existing methods often depend on curated distillation datasets, costly teacher rollouts, or auxiliary proxy networks, which complicate model training and scaling. In this work, we propose Consistent Distribution Matching, a simulation-free and data-free distillation method for accelerating diffusion and flow models while preserving strong generative capacity. Our key insight is to unify sample generation and score estimation with one student network. Thus, our framework uses only two models, a frozen teacher and a trainable student, and optimizes one objective. We prove that minimizing our objective indicates Wasserstein convergence of the student flow-map pushforwards to the teacher marginals. On ImageNet 256times256, our method attains an FID of 2.04 with a single function evaluation (1-NFE) and a 4-NFE FID of 1.37 within 40 epochs of training, surpassing the state-of-the-art distillation baselines without data. Our code code and model are available at https://consistentdmd.github.io/.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P23", "title": "Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents", "url": "https://huggingface.co/papers/2610.18082", "authors": [ "Xiyang Wu", "Zhichao Liu", "Shengxin zhang", "Zongxia Li" ], "categories": [ "other" ], "category": "other", "summary": "Long-horizon agents often struggle to recover from failures whose causes lie several steps earlier, beyond what local reflection or backtracking can resolve.", "summary_html": "Long-horizon agents often struggle to recover from failures whose causes lie several steps earlier, beyond what local reflection or backtracking can resolve.", "why": "", "why_html": "", "abstract": "Long-horizon agents often struggle to recover from failures whose causes lie several steps earlier, beyond what local reflection or backtracking can resolve. We introduce BELIEFWEAVER, a framework that combines persistent evidence across attempts with prospective reasoning over repair directions. It maintains a relational belief state over frontiers, attempts, observations, and checkpoints. Execution evidence updates related frontier priorities across attempts and rollback, while diagnostic evidence gates execution. An implicit world model (IWM) evaluates repair hypotheses, predicts progress, and identifies relations among frontiers to guide exploration. We further distill these recurring IWM reasoning components into a 9B student through supervised fine-tuning and on-policy distillation (OPD). Across three terminal-based benchmarks, BELIEFWEAVER improves aggregate final score by 5.1 percentage points over native Terminus-2 continuation, with positive trends in exploration efficiency. The distilled 9B student improves final score by 9.8 percentage points over the reported native baseline while reducing estimated teacher API cost by 54.8% relative to the matched teacher. These results show that persistent evidence and prospective reasoning can improve recovery in long-horizon agent tasks while substantially reducing reliance on teacher-model inference.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P24", "title": "Co-Evolving Acting and Forecasting for Language Agents", "url": "https://huggingface.co/papers/2610.20597", "authors": [ "Haofei Yu", "Yixing Chen", "Yining Zhao", "Xuyan Ye" ], "categories": [ "other" ], "category": "other", "summary": "Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused.", "summary_html": "Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused.", "why": "", "why_html": "", "abstract": "Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused. Prior work exploits this supervision through value targets that estimate per-step success probabilities or observation targets that fit environment response. Both are hard to construct well: (1) value targets come from the same single scalar signal for each step in the trajectory, and (2) observation targets, such as web pages and terminal logs, are long and noisy and reach only one step ahead. We propose CoPE, which adds a third kind of multi-step action targets. On successful on-policy trajectories, CoPE trains the policy backbone to generate the action sequences that follows each history. Each action is also supervised when forecast from the preceding histories, forming a temporal ensemble that the shared backbone absorbs during training, and this enables acting and forecasting to co-evolve. The policy still acts one step at a time at inference, with no overhead. Across five benchmarks (ALFWorld, ScienceWorld, WebShop, τ^2-bench, and AppWorld), CoPE (1) achieves stronger aggregate performance than GRPO and RL baselines with predictive auxiliary targets across simulated environments, while improving overall performance on both realistic applications; (2) delivers up to 1.9times and 2.0times higher sample efficiency than ECHO and GRPO, respectively, on ScienceWorld; and (3) achieves these gains with a single policy backbone, without additional models, separate pretraining, or inference overhead. The codebase is available at https://github.com/ulab-uiuc/CoPE.", "doi": null, "source": "HF Papers", "also": [] } ] } ], "papers": [ { "ref": "P1", "title": "RIT-RAG: Navigating Document Corpora with Retrieval-Induced Trees", "url": "https://arxiv.org/abs/2610.11370", "authors": [ "Meghanadh Pulivarthi", "Swaraj Kumar Biswal", "Kushagra Bhushan", "Yatin Nandwani" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Retrieval-augmented generation (RAG) grounds language models in external corpora.", "summary_html": "Retrieval-augmented generation (RAG) grounds language models in external corpora.", "why": "", "why_html": "", "abstract": "Retrieval-augmented generation (RAG) grounds language models in external corpora. Agentic RAG enables iterative search, yet exposes the model to isolated chunks without document structure, making it difficult to distinguish relevant evidence from chunks that merely resemble the query. Structure-aware methods such as PageIndex navigate using the structure but cannot scale to structures of large corpora that cannot fit in the LLM context. Hence, they first commit to a single document using a document retriever and cannot recover from a wrong choice. We propose RIT-RAG (Retrieval-Induced Tree RAG), which combines content retrieval with structural navigation. Offline, RIT-RAG builds a tree for each document from its table of contents or sitemap. At query time, it retrieves a broad set of chunks and uses their positions to induce manageable sub-trees, potentially across multiple documents. An LLM agent navigate these sub-trees, selectively reads promising nodes, and reformulates queries when needed. Thus, retrieval proposes where to look, while the agent decides what to read. Across financial, scientific, and customer-support benchmarks, RIT-RAG achieves the highest answer accuracy among vanilla, graph-based, and agentic baselines. On EntQABench, our new benchmark of 2.84 million technical-documentation webpages, it improves accuracy by 6.6--11.4 points over the strongest baseline across three LLMs.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P2", "title": "Rethinking GRPO for Large Reasoning Model Safety Alignment", "url": "https://arxiv.org/abs/2610.23085", "authors": [ "Jiawei Chen", "Yi Zhang", "Jiawei Du", "Zhonghou Lv" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Large reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs.", "summary_html": "Large reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs.", "why": "", "why_html": "", "abstract": "Large reasoning models (LRMs) exhibit strong reasoning capabilities but remain vulnerable to unsafe outputs. While Group Relative Policy Optimization (GRPO) is widely used for safety alignment, its group-relative optimization may provide insufficient safety-directed learning signals. In this paper, we identify two failure modes: advantage collapse, where homogeneous rollout groups yield vanishing advantages, and unsafe diversification, where unsafe behaviors persist and become increasingly diverse. Our theoretical analysis reveals that behavioral consistency weakens within-group safety contrasts, while entropy regularization can diversify unsafe responses when relative signals vanish. To address these challenges, we propose Beacon, comprising Safety Signal Recovery (SSR) and Safety-Utility Decoupled Alignment (SUDA). SSR combines advantage-guided trajectory filtering and safe reference supervision to recover effective safety learning signals, while SUDA employs category-aware rewards and benign-only KL anchoring to mitigate over-refusal without compromising safety. Experiments across four Qwen3 and DeepSeek-R1-Distill variants and eight benchmarks show that our framework achieves the best average safety and reasoning performance, reducing average safety error by up to 32.83 percentage points while maintaining low over-refusal.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P3", "title": "All-in-One Memory: Process-Rewarded Reinforcement Learning for Memory Agents", "url": "https://arxiv.org/abs/2610.21147", "authors": [ "Bingbing Wang", "Jingjie Lin", "Weiming Qiao", "Jing Li" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Large language models increasingly rely on external memory to answer queries grounded in long-horizon histories such as multi-session dialogues and lengthy documents.", "summary_html": "Large language models increasingly rely on external memory to answer queries grounded in long-horizon histories such as multi-session dialogues and lengthy documents.", "why": "", "why_html": "", "abstract": "Large language models increasingly rely on external memory to answer queries grounded in long-horizon histories such as multi-session dialogues and lengthy documents. However, existing memory methods often optimize retrieval and memory construction separately, without a shared reward space for evaluating search, writing, and stopping decisions. To address this, we propose All-in-One Memory (AiO-Mem), which organizes these operations as transitions over a shared evidence state and directly evaluates the semantic quality of each action in a unified reward space. AiO-Mem compiles heterogeneous histories into source-traceable multi-view events, providing a shared substrate for retrieval and memory writing. A unified memory policy actor then rolls out complete interaction trajectories over this shared evidence state. A separate trajectory-conditioned process reward model is trained through evidence-grounded supervised fine-tuning and trajectory-level DPO to assign process scores to the collected trajectories. These scores are aggregated into an action-balanced trajectory reward and used by Process-Aligned DAPO (PA-DAPO) to optimize the memory policy. Across long-context memory benchmarks, AiO-Mem consistently improves answer accuracy over both static memory systems and RL-trained memory agents.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P4", "title": "Safe Actions Alone Do Not Ensure Safe Agents: Identifying Unfulfilled Obligations with Guard Models", "url": "https://arxiv.org/abs/2610.11773", "authors": [ "Youwei Feng", "Yitong Zhang", "Yuetong Liu", "Jia Li" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform.", "summary_html": "Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform.", "why": "", "why_html": "", "abstract": "Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform. However, identifying forbidden actions alone is insufficient to ensure agent safety. In this paper, we argue that agent safety also depends on identifying required yet unperformed safety-critical actions, which we call obligations. Our preliminary study on a popular benchmark for evaluating safety shows that 56.92% of GLM-5.3 trajectories contain unfulfilled obligations, compared with only 30.00% containing forbidden actions. This finding reveals unfulfilled obligations as a major and previously overlooked source of safety risk. However, to our knowledge, no existing benchmark evaluates whether guard models can identify these obligations. To close this gap, we introduce ObligationBench, the first benchmark for evaluating the capability of obligation identification, comprising 240 expert-validated trajectories covering issue resolution, feature development, and terminal operations. Our evaluation of 14 representative models reveals substantial limitations: the highest recall and exact-match rate are only 48.97% and 10.00%, respectively. To address these limitations, we develop ObligationGuard using 40,000 synthetic training examples. ObligationGuard achieves 57.52% recall and an exact-match rate of 21.67%, surpassing all evaluated models on both metrics. We call on the community to incorporate obligation identification into the design and evaluation of future guard models to improve agent safety.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P5", "title": "Agentic In-Context Learning: What Algorithms Can Transformers Learn in Closed-Loop Interaction?", "url": "https://arxiv.org/abs/2610.23051", "authors": [ "Shiguang Wu", "Yaqing Wang", "Quanming Yao" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "The learning algorithm view of in-context learning (ICL) studies how fixed-parameter Transformers learn from supplied examples.", "summary_html": "The learning algorithm view of in-context learning (ICL) studies how fixed-parameter Transformers learn from supplied examples.", "why": "", "why_html": "", "abstract": "The learning algorithm view of in-context learning (ICL) studies how fixed-parameter Transformers learn from supplied examples. Language model agents extend this setting through interaction: their actions determine what evidence enters context next. We call this regime Agentic In-Context Learning (Agentic ICL). Existing studies largely depend on externally supplied contexts or reward-bearing interaction histories, while work on general tool-using agents lacks a learning algorithm view. We represent an interactive learner by its state, update, and decision rule, covering active learning, bandits, and online reinforcement learning (RL). For expressivity, we construct causal Transformers that approximate these algorithms over a finite horizon, and bound the resulting loss in return. For training, we characterize when optimizing process rewards preserves behavior that maximizes the final outcome. Distillation experiments show that Transformers can learn to execute three representative recursions. We then train models from outcome or process rewards without oracle action or state targets, as in RL post-training. Different rewards yield distinct decision policies. Improving the alignment of process supervision with outcome-optimal action preferences shifts learned decisions accordingly. Such behavioral differences persist in a pretrained language model across textual paraphrases.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P6", "title": "Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI", "url": "https://arxiv.org/abs/2610.09146", "authors": [ "Yexiao He", "Yucheng Tang", "Pengfei Guo", "Yufan He" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn from the cases they solve.", "summary_html": "Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn from the cases they solve.", "why": "", "why_html": "", "abstract": "Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn from the cases they solve. This is especially concerning in medicine, where new clinical evidence, updated guidelines, and new therapies can change established practice. Fine-tuning can update the model, but it requires access to model weights and additional training. Parameter-free methods avoid training, but they may overfit a fixed validation set, lack reliable domain knowledge, or lose visual details by saving experience only as text. To address these limitations, we present a model-agnostic framework that allows frozen LLMs and VLMs to learn from deployment experience through three forms of external expertise: a Skill that guides reasoning and tool use, a Knowledge Memory that stores reliable facts supported by earlier cases or trusted external evidence, and a Multimodal Knowledge Base that keeps visual examples and guides the model to relate each retrieved case to the current image. Instead of relying on a fixed validation set, a validation strategy keeps an update only if it helps on new cases without degrading performance on earlier ones. Across six benchmarks covering clinical diagnosis, clinical workflows, medical reasoning, and medical and non-medical visual reasoning, and with four open-weight and closed-source base models, our framework improves performance during online deployment by up to 34.2% over the base model on medical tasks, generalizes to unseen cases, transfers to other models without further optimization, and works in non-medical domains.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P7", "title": "SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models", "url": "https://arxiv.org/abs/2610.26487", "authors": [ "Shenli Zheng", "Zhou Fang", "Yuan Du", "Yijiang Liu" ], "categories": [ "cs.RO" ], "category": "cs.RO", "summary": "The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate.", "summary_html": "The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate.", "why": "", "why_html": "", "abstract": "The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate. Reducing these demands through aggressive compression, however, can degrade manipulation performance beyond what supervised fine-tuning alone recovers. We present SPAR (Structural Pruning and Action Recovery), which simultaneously prunes the hidden and feed-forward intermediate dimensions of the large language model (LLM) backbone through pruning-aware tuning, then applies reinforcement learning (RL) to recover policy performance. Applied to $\\pi_{0.5}$, SPAR reduces the language backbone from 2.34B to 0.09B parameters and the full model from 3.37B to 1.12B, removing approximately 96% and 67% of their parameters, respectively. At this compression level, RL raises average success across the four LIBERO suites from 84.55% to 93.45%, an 8.90-percentage-point gain. The compressed policy achieves a $1.98\\times$ full-policy inference speedup over the unpruned model. These results identify substantial intermediate-dimension redundancy and show that post-pruning RL can recover manipulation performance in compact policies.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P8", "title": "Safety Alignment via Constrained Knowledge Unlearning", "url": "https://arxiv.org/abs/2505.18588", "authors": [ "Zesheng Shi", "Yucheng Zhou", "Jing Li" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks.", "summary_html": "Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks.", "why": "", "why_html": "", "abstract": "Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks. Existing defense mechanisms have not fully deleted harmful knowledge in LLMs, which allows such attacks to bypass safeguards and produce harmful outputs. To address this challenge, we propose a novel safety alignment strategy, Constrained Knowledge Unlearning (CKU), which focuses on two primary objectives: knowledge localization and retention, and unlearning harmful knowledge. CKU works by scoring neurons in specific multilayer perceptron (MLP) layers to identify a subset U of neurons associated with useful knowledge. During the unlearning process, CKU prunes the gradients of neurons in U to preserve valuable knowledge while effectively mitigating harmful content. Experimental results demonstrate that CKU significantly enhances model safety without compromising overall performance, offering a superior balance between safety and utility compared to existing methods. Additionally, our analysis of neuron knowledge sensitivity across various MLP layers provides valuable insights into the mechanics of safety alignment and model knowledge editing.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P9", "title": "Koi: Evidence-Guided, Self-Calibrating Planning for Multi-Cloud, Heterogeneous LLM Inference Fleets", "url": "https://arxiv.org/abs/2610.24355", "authors": [ "Mankeerat Sidhu", "Gangmuk Lim", "Hetarth Chopra", "Maximillian Eisenberg" ], "categories": [ "cs.DC" ], "category": "cs.DC", "summary": "Deploying a large language model (LLM) requires choosing the right accelerator, parallelization strategy, engine configuration, and replica count amongst ~140 other variables across clouds, regions, and markets.", "summary_html": "Deploying a large language model (LLM) requires choosing the right accelerator, parallelization strategy, engine configuration, and replica count amongst ~140 other variables across clouds, regions, and markets.", "why": "", "why_html": "", "abstract": "Deploying a large language model (LLM) requires choosing the right accelerator, parallelization strategy, engine configuration, and replica count amongst ~140 other variables across clouds, regions, and markets. These decisions are intertwined, and the best cluster-wide plan continually changes as jobs contend for shared capacity on shifting workloads. We present Koi, a self-calibrating cluster planner for heterogeneous, multi-cloud fleets. Koi uses a Recursive Language Model (RLM) hierarchy to propose hypotheses and reason over deployment while deterministic components predict serving behavior, score candidates, and control placement. It represents inference behavior as a fixed causal graph linking decisions (e.g., accelerator, parallelism), runtime mediators (e.g., KV-cache pressure, communication overhead), and outcomes (e.g., latency, throughput). Each deployment strengthens or weakens scoped causal relationships, building a comprehensive playbook for inference that the planner employs to guide future decisions. Job-wise configurations are composed into cluster-wide plans and evaluated against serving objectives, resource constraints, SLO risks, and cost. Across benchmarks, Koi achieves 1.18x greater goodput than the strongest baseline while reducing $/token by 20% in a real heterogeneous cluster. Ablations isolate the contribution of Koi's architecture: a pure language model attains a composite score of 35.3, the same model equipped with Koi's deterministic tools reaches 51.6, and the full Koi planner reaches 70.8.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P10", "title": "SkillBrowser: Selective Skill Exposure for Scalable Skill-Using Agents", "url": "https://arxiv.org/abs/2610.23554", "authors": [ "Jin Ye", "Yifan Jiang", "Lan-zhe Guo" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Large language model agents increasingly rely on skills to extend their capabilities in complex tasks.", "summary_html": "Large language model agents increasingly rely on skills to extend their capabilities in complex tasks.", "why": "", "why_html": "", "abstract": "Large language model agents increasingly rely on skills to extend their capabilities in complex tasks. However, scalable skill use requires Selective Skill Exposure (SSE): providing relevant procedural content when needed while avoiding unnecessary exposure that may disrupt reasoning. Existing paradigms fall short in complementary ways. Progressive disclosure exposes metadata for a growing skill library, introducing context pollution; retrieval limits catalog exposure, but static retrieval from the initial task query may be misaligned with capability needs that emerge during reasoning and may still inject unnecessary skills. We propose SkillBrowser, a Meta-Skill-managed framework that enables an agent to identify current capability needs, formulate targeted queries, and selectively load skills on demand. Across six SRA-Bench subsets and five LLMs, SkillBrowser improves aggregate accuracy over static retrieval and LLM-based selection, including multi-step tool use, and mitigates accuracy degradation on already-solvable instances. Benefits remain model- and task-dependent, with under-invocation of useful skills and additional latency as key limitations. These results demonstrate the viability of Meta-Skill-managed SSE for scalable skill repositories.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P11", "title": "Structured Authority Injection: Persistent Human Control for LLM Agents via Typed MCP Overrides", "url": "https://arxiv.org/abs/2610.22373", "authors": [ "Ryan Tran", "YueMing Zhang", "Lingfeng Tao", "Nan Yang" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "LLM-based agents are increasingly used as autonomous orchestrators in distributed and multi-agent systems, yet a control gap remains: no protocol-level mechanism allows human operators to inject binding directives that persist across planning cycles.", "summary_html": "LLM-based agents are increasingly used as autonomous orchestrators in distributed and multi-agent systems, yet a control gap remains: no protocol-level mechanism allows human operators to inject binding directives that persist across planning cycles.", "why": "", "why_html": "", "abstract": "LLM-based agents are increasingly used as autonomous orchestrators in distributed and multi-agent systems, yet a control gap remains: no protocol-level mechanism allows human operators to inject binding directives that persist across planning cycles. Free-text overrides suffer from ambiguous interpretation due to lack of schema validation and from authority erosion as operator instructions lose influence within accumulating context. We propose Structured Authority Injection (SAI), a protocol that addresses these issues by routing operator corrections through typed, schema-validated Model Context Protocol (MCP) tool calls and maintaining a ranked authority context prepended to each LLM prompt. SAI defines three override classes (Constraint, Priority, and Strategy) with distinct persistence semantics and an explicit precedence order enforced via a directed authority channel inaccessible to the planning process. The method requires no model fine-tuning or access to model weights and applies to instruction-tuned LLMs supporting MCP tool use. Evaluation across nine LLMs and eight scenarios targeting agent failure modes shows reduced ambiguous directive interpretation and mitigated authority decay across planning iterations, with lower prompt overhead in the strongest models (up to 38% depending on model and scenario), though benefit varies substantially across models. SAI provides a general design pattern for agent systems requiring explicit and persistent human control.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P12", "title": "RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory", "url": "https://arxiv.org/abs/2610.23598", "authors": [ "Dabin Sheng", "Zhe Wu", "Jinming Zhao", "Junliang Xing" ], "categories": [ "cs.AI" ], "category": "cs.AI", "summary": "Time-aware retrieval is crucial for long-horizon LLM agents to effectively leverage ever-growing long-term memory.", "summary_html": "Time-aware retrieval is crucial for long-horizon LLM agents to effectively leverage ever-growing long-term memory.", "why": "", "why_html": "", "abstract": "Time-aware retrieval is crucial for long-horizon LLM agents to effectively leverage ever-growing long-term memory. Existing methods focus on lexical or semantic similarity, leading to target memories being crowded out by temporally mismatched yet semantically similar candidates. Even time-aware approaches inject temporal cues only during memory organization or answer generation, overlooking temporal constraints at the retrieval stage, where candidate selection occurs - the root cause of the displacement. Inspired by the temporal contiguity effect in human episodic memory, we propose TIDER (Temporal Interval-driven Dual-channel Evidence Retrieval), a time-aware agentic retrieval framework. Central to TIDER is a reinforcement learning (RL)-trained temporal interval locator that infers the temporal windows in which the evidence lies. RL optimizes window localization as a discrete, non-differentiable retrieval decision, using composite rewards to convert sparse answer supervision into fine-grained retrieval feedback. These temporal windows first define the temporal retrieval channel; the retrieved candidates are then merged with results from the semantic channel and reranked by proximity to the windows. This temporal grounding enables TIDER to retrieve temporally aligned evidence that would otherwise be drowned out by the volume of semantically similar yet temporally mismatched memories. With a lightweight 1.7B locator, TIDER achieves state-of-the-art accuracy on both long-horizon memory benchmarks - 61.52% on LifeBench (+8.82%) and 78.96% on LoCoMo (+3.66%) - which indicates that temporal cues are crucial for pinpointing target evidence amid a vast pool of semantically similar historical events in long-term daily-life memory.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P13", "title": "VLX-Go: Robust Embodied Tracking through Multi-Stage Learning and On-Policy Refinement", "url": "https://arxiv.org/abs/2610.19361", "authors": [ "Kyusong Lee", "Heting Ying", "Tiancheng Zhao" ], "categories": [ "cs.RO" ], "category": "cs.RO", "summary": "Embodied visual tracking requires a robot to identify a designated target from visual observations and control its own motion to follow that target through an environment.", "summary_html": "Embodied visual tracking requires a robot to identify a designated target from visual observations and control its own motion to follow that target through an environment.", "why": "", "why_html": "", "abstract": "Embodied visual tracking requires a robot to identify a designated target from visual observations and control its own motion to follow that target through an environment. Following an isolated person is a comparatively simple demonstration of embodied visual tracking, but following the intended person among distractors remains challenging. Existing methods exhibit a substantial drop in success from single-target to distractor tracking, motivating a focus on the shared spaces that person-following robots must navigate. We present VLX-GO, a vision-language waypoint policy built around the hypothesis that visual and language representations should be aligned before tracking-specific motion is learned. The policy combines complementary frozen visual encoders, multiscale temporal context, a transformer backbone, and a lightweight head for robot-centric waypoint prediction. Its training progresses through vision--language alignment, multimodal reasoning mid-training, and tracking specialization on 19,532 oracle episodes. Compared with 41.54% distractor-tracking success for our earlier OmTrackVLA model, the complete multi-stage recipe achieves 55.80%. We then explore on-policy refinement using filtered oracle labels for states visited by the policy. This preliminary study adds 2.42 percentage points of distractor-tracking success, reaching 58.22%, while reducing collision rate by 1.99 pts. The final policy records the highest distractor- and ambiguous-tracking success rates among the compared methods. These results support staged multimodal learning as a promising direction for tracking among distractors and motivate further study of on-policy refinement for person following in shared spaces. Code is available at https://github.com/om-ai-lab/VLX-Go.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P14", "title": "Large Language Models over Networks: Collaborative Intelligence under Resource Constraints", "url": "https://arxiv.org/abs/2605.08626", "authors": [ "Liangqi Yuan", "Wenzhi Fang", "Shiqiang Wang", "H. Vincent Poor" ], "categories": [ "cs.DC" ], "category": "cs.DC", "summary": "Large language models (LLMs) are transforming society, powering applications from smartphone assistants to autonomous driving.", "summary_html": "Large language models (LLMs) are transforming society, powering applications from smartphone assistants to autonomous driving.", "why": "", "why_html": "", "abstract": "Large language models (LLMs) are transforming society, powering applications from smartphone assistants to autonomous driving. Yet cloud-based LLM services alone cannot serve a growing class of applications, including those operating under intermittent connectivity, sub-second latency budgets, data-residency constraints, or sustained high-volume inference. On-device deployment is in turn constrained by limited computation and memory. No single endpoint can deliver high-quality service across this spectrum. This article focuses on collaborative intelligence, a paradigm in which multiple independent LLMs distributed across device and cloud endpoints collaborate at the task level through natural language or structured messages. Such collaboration strives for superior response quality under heterogeneous resource constraints spanning computation, memory, communication, and cost across network tiers. We present collaborative inference along two complementary and composable dimensions: vertical device-cloud collaboration and horizontal multi-agent collaboration, which can be combined into hybrid topologies in practice. We then examine learning to collaborate, addressing the training of routing policies and the development of cooperative capabilities among LLMs. Finally, we identify open research challenges including scaling under resource heterogeneity and trustworthy collaborative intelligence.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P15", "title": "Vibe Coding XR: Accelerating AI + XR Prototyping with XR Blocks and Gemini", "url": "https://arxiv.org/abs/2603.24591", "authors": [ "Ruofei Du", "Benjamin Hersh", "David Li", "Nels Numan" ], "categories": [ "cs.HC" ], "category": "cs.HC", "summary": "While large language models (LLMs) have accelerated traditional software development through intent-driven \"vibe coding\", prototyping intelligent Extended Reality (XR) experiences remains a major challenge.", "summary_html": "While large language models (LLMs) have accelerated traditional software development through intent-driven "vibe coding", prototyping intelligent Extended Reality (XR) experiences remains a major challenge.", "why": "", "why_html": "", "abstract": "While large language models (LLMs) have accelerated traditional software development through intent-driven \"vibe coding\", prototyping intelligent Extended Reality (XR) experiences remains a major challenge. The fundamental barrier is not just the steep learning curve for human creators, but that low-level sensor APIs and complex game engine hierarchies are ill-suited for LLM reasoning, routinely exceeding context windows and inducing syntax hallucinations. To bridge this gap, we contribute XR Blocks, an open-source, LLM-native WebXR framework. Unlike traditional engines, XR Blocks introduces a semantic \"Reality Model\" that aligns spatial computing primitives (users, physical environments, and agents) with natural language, providing a robust, concise vocabulary optimized for generative AI. Building upon this foundation, we present Vibe Coding XR, an end-to-end prototyping workflow that leverages LLMs to translate high-level prompts (e.g., \"create a dandelion that reacts to my hand\") directly into functional, physics-aware mixed-reality applications. To minimize the friction of on-device testing, the workflow introduces a seamless desktop \"simulated reality\" to headset deployment loop. Finally, we introduce VCXR60, a pilot dataset of 60 XR prompts paired with an automated evaluation pipeline. Our technical evaluation demonstrates high one-shot execution success, enabling practitioners to bypass low-level hurdles and rapidly move from \"idea to reality\". Code and demo are available at https://github.com/google/xrblocks and https://goo.gle/BuildXR.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P16", "title": "AutoMind: Adaptive Knowledgeable Agent for Automated Data Science", "url": "https://arxiv.org/abs/2506.10974", "authors": [ "Yujie Luo", "Yixin Ou", "Jingsheng Zheng", "Lanning Wei" ], "categories": [ "cs.HC" ], "category": "cs.HC", "summary": "Large Language Model (LLM) agents have shown great potential in addressing real-world data science problems.", "summary_html": "Large Language Model (LLM) agents have shown great potential in addressing real-world data science problems.", "why": "", "why_html": "", "abstract": "Large Language Model (LLM) agents have shown great potential in addressing real-world data science problems. LLM-driven data science agents promise to automate the entire machine learning pipeline, yet their real-world effectiveness remains limited. Existing frameworks depend on rigid, pre-defined workflows and inflexible coding strategies; consequently, they excel only on relatively simple, classical problems and fail to capture the empirical expertise that human practitioners bring to complex, innovative tasks. In this work, we introduce AutoMind, an adaptive, knowledgeable LLM-agent framework that overcomes these deficiencies through three key advances: (1) a curated expert knowledge base that grounds the agent in domain expert knowledge, (2) an agentic knowledgeable tree search algorithm that strategically explores possible solutions, and (3) a self-adaptive coding strategy that dynamically tailors code generation to task complexity. Evaluations on two automated data science benchmarks demonstrate that AutoMind delivers superior performance versus state-of-the-art baselines. Additional analyses confirm favorable effectiveness, efficiency, and qualitative solution quality, highlighting AutoMind as an efficient and robust step toward fully automated data science. Code is at https://github.com/zjunlp/AutoMind.", "doi": null, "source": "arXiv", "also": [] }, { "ref": "P17", "title": "Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries", "url": "https://huggingface.co/papers/2610.22273", "authors": [ "Anikait Singh", "Teresa Zhang", "Yoonho Lee", "Roshen Sanjay Nair" ], "categories": [ "other" ], "category": "other", "summary": "Reasoning over long document contexts remains a challenge for large language models (LLMs).", "summary_html": "Reasoning over long document contexts remains a challenge for large language models (LLMs).", "why": "", "why_html": "", "abstract": "Reasoning over long document contexts remains a challenge for large language models (LLMs). Standard retrieval methods are inherently reactive, bottlenecking agents with complex, cross-document reasoning at inference time, while precomputed structures like knowledge graphs often fall short because they rely on rigid heuristics. Instead, can we allow agents to perform anticipatory directed curation: learning to proactively reorganize unstructured corpora prior to inference to efficiently serve diverse future queries. We build on a key asymmetry in modern agentic harnesses (e.g., Claude Code): while LLM agents are adept at navigating existing filesystems to solve an immediate task, they are rarely trained to construct reusable workspaces for future use. To instantiate this capability, we introduce Librarian Agents, a framework in which a builder agent transforms raw documents into a hierarchical filesystem of synthetic files and reference links, which forms a workspace tailored for a downstream answering agent. We formulate workspace construction as a Markov decision process and optimize the builder policy via reinforcement learning to directly maximize downstream answering accuracy. Empirically, Librarian Agents outperforms the strongest retrieval and precomputation baselines by 33 percentage points on BrowseComp-Plus, amortizes upfront construction costs across repeated queries, and generalizes zero-shot to unseen domains on BRIGHT with a 31-point average accuracy improvement.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P18", "title": "SciΠvot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents", "url": "https://huggingface.co/papers/2610.21082", "authors": [ "Youran Sun", "Xingyu Ren", "Chugang Yi", "Kejia Zhang" ], "categories": [ "other" ], "category": "other", "summary": "Large language model (LLM) agents can keep scientific experiments running for months with little human supervision, but reliable autonomous research also requires scientific judgment.", "summary_html": "Large language model (LLM) agents can keep scientific experiments running for months with little human supervision, but reliable autonomous research also requires scientific judgment.", "why": "", "why_html": "", "abstract": "Large language model (LLM) agents can keep scientific experiments running for months with little human supervision, but reliable autonomous research also requires scientific judgment. Evaluating scientific judgment is difficult because research decisions are open-ended and depend on project context. We construct SciΠvot, a 248-case benchmark of scientific judgment based on human corrections. The benchmark draws on records from six months of research in which senior researchers ran LLM agents on 26 projects across six scientific fields and repeatedly intervened to correct the agents' research judgments. In each case, an agent chooses between the original LLM plan and the human-corrected alternative, using only the project record available before the correction as context. We compare LLM and human performance and examine the effects of reasoning effort, access to the project record, and reconsideration. At medium reasoning effort with the project record, models systematically prefer the original LLM plan over the human-corrected alternative. Four of five frontier models perform below chance on the selected cases. The highest accuracy in the project condition across the tested models and reasoning effort settings is 63.5%, compared with 73.4% for fresh researchers. At medium effort, every model achieves lower accuracy with the project record than without it, and reconsideration does not consistently improve overall accuracy. Even with a human-corrected alternative available, frontier models struggle with the selected research decisions, while SciΠvot supports systematic evaluation of scientific judgment for LLM agents.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P19", "title": "Think Visually Across Space and Time: Grounded Reasoning for Video Segmentation", "url": "https://huggingface.co/papers/2610.25090", "authors": [ "Junwei Wen", "Deshui Miao", "Guangming Lu", "Ming-Hsuan Yang" ], "categories": [ "other" ], "category": "other", "summary": "Video Reasoning Segmentation (VRS) requires a model to identify an implicitly described target and predict its pixel-level mask trajectory, while resolving the spatial relations and temporal events that determine the answer.", "summary_html": "Video Reasoning Segmentation (VRS) requires a model to identify an implicitly described target and predict its pixel-level mask trajectory, while resolving the spatial relations and temporal events that determine the answer.", "why": "", "why_html": "", "abstract": "Video Reasoning Segmentation (VRS) requires a model to identify an implicitly described target and predict its pixel-level mask trajectory, while resolving the spatial relations and temporal events that determine the answer. Text-only Chain-of-Thought (CoT) can explain a target choice without making the underlying visual references explicit, leaving a video reference gap between linguistic inference and the evidence distributed across frames. We introduce Temporal Chain of Video Primitives (TCoVP), a spatiotemporally grounded form of Video CoT that organizes query-relevant observations into a chronological reasoning chain. Each primitive associates an entity reference with temporal support and timestamped spatial evidence, and the chain interleaves these references with language to guide reasoning about which entity is relevant, when the queried event holds, and where its evidence appears. The selected target primitives then provide box prompts and temporal support for video mask prediction. We construct TCoVP-SFT-15k and TCoVP-RL-13k to learn this representation through supervised fine-tuning and chain-aligned reinforcement learning. Extensive experiments across five VRS benchmarks and three model backbones demonstrate the strong performance of TCoVP, highlighting the effectiveness of organizing spatial and temporal evidence within the reasoning process. All models and code will be publicly released. Project page: https://victor-wjw.github.io/TCoVP/", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P20", "title": "OmniXRef: A Unified Framework for Multimodal Multi-Reference Video Editing", "url": "https://huggingface.co/papers/2610.22744", "authors": [ "Xiaoyan Cong", "Tianyu Wang", "Hyeonho Jeong", "Zhifei Zhang" ], "categories": [ "other" ], "category": "other", "summary": "Creative intent in video editing can be expressed through text instructions (e.g., \"remove the people\"), image references (e.g., subject appearance), video references (e.g., camera movement), or a combination of the above.", "summary_html": "Creative intent in video editing can be expressed through text instructions (e.g., "remove the people"), image references (e.g., subject appearance), video references (e.g., camera movement), or a combination of the above.", "why": "", "why_html": "", "abstract": "Creative intent in video editing can be expressed through text instructions (e.g., \"remove the people\"), image references (e.g., subject appearance), video references (e.g., camera movement), or a combination of the above. We introduce OmniXRef, a unified framework for multimodal multi-reference video editing that allows users to combine variable numbers of text, image, and video references to control diverse attributes of a source video. We first devise a unified input design with a reference-bank interface and introduce R^2-RoPE to support generalization to more references. Then our model, OmniXRef-Model, uses a single-stream diffusion transformer to jointly process the source video and multimodal context from an MLLM in an interleaved sequence. To address the scarcity of training data for video-reference editing, we construct OmniXRef-Data using synthesis pipelines for different editing tasks and combinations of references. We use curriculum learning to progressively increase task complexity and resolution. To support systematic evaluation of video as a control modality, we introduce OmniXRef-Bench, a benchmark for video-reference attribute transfer across Camera, Motion, Content, Style, and Visual Effects (VFX). Comprehensive evaluations show that our OmniXRef-Model outperforms task-specific baselines and commercial models.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P21", "title": "Empowering Edge Agents: A Systematic Analysis of Edge-Cloud Collaboration in LLM Agent Execution", "url": "https://huggingface.co/papers/2610.17816", "authors": [ "Pengcheng Huang", "Zhenghao Liu", "Shixin Ren", "Shi Yu" ], "categories": [ "other" ], "category": "other", "summary": "Edge-cloud collaboration combines the lower inference overhead and greater data locality of edge-side execution with the stronger capabilities of cloud-hosted models.", "summary_html": "Edge-cloud collaboration combines the lower inference overhead and greater data locality of edge-side execution with the stronger capabilities of cloud-hosted models.", "why": "", "why_html": "", "abstract": "Edge-cloud collaboration combines the lower inference overhead and greater data locality of edge-side execution with the stronger capabilities of cloud-hosted models. However, as LLM agents increasingly operate in long-horizon scenarios, cloud participation becomes interleaved throughout execution rather than a one-shot decision for each query. Such interleaved participation introduces trajectory-level effects that remain insufficiently understood. To address this gap, we introduce EC2-Bench, a benchmark for evaluating edge-cloud collaboration in long-horizon agent execution. It provides executable digital-workspace agent tasks and a unified evaluation framework covering task performance, resource usage, and privacy exposure. Using this benchmark, we empirically study diverse forms of edge-cloud collaboration. Our study reveals three findings. First, collaboration achieves a favorable balance among task performance, resource usage, and privacy exposure compared with edge-only and cloud-only execution. Second, as edge models become stronger, collaborative execution approaches cloud-only performance, and performance differences across strategies shrink, while privacy exposure remains strongly strategy-dependent. Third, stronger edge capability shifts preferred collaboration from externally controlled cloud participation and direct cloud execution toward more edge-controlled collaboration. Further analyses identify cloud-invocation decision-making and instruction following as important edge-side abilities for effective collaboration. These findings provide a systematic understanding of edge-cloud collaboration in long-horizon agent execution and insights for designing future edge agents that effectively collaborate with cloud models. Code and resources are available at https://github.com/OpenBMB/EC2-Bench.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P22", "title": "Consistent Distribution Matching for Data-Free Diffusion Distillation", "url": "https://huggingface.co/papers/2610.09221", "authors": [ "Yuxiang Fu", "Qi Yan", "Zike Wu", "Yongxing Zhang" ], "categories": [ "other" ], "category": "other", "summary": "Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration.", "summary_html": "Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration.", "why": "", "why_html": "", "abstract": "Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration. Distillation provides a promising way for a student model to learn from a teacher's dynamics, enabling one-step or few-step generation. However, existing methods often depend on curated distillation datasets, costly teacher rollouts, or auxiliary proxy networks, which complicate model training and scaling. In this work, we propose Consistent Distribution Matching, a simulation-free and data-free distillation method for accelerating diffusion and flow models while preserving strong generative capacity. Our key insight is to unify sample generation and score estimation with one student network. Thus, our framework uses only two models, a frozen teacher and a trainable student, and optimizes one objective. We prove that minimizing our objective indicates Wasserstein convergence of the student flow-map pushforwards to the teacher marginals. On ImageNet 256times256, our method attains an FID of 2.04 with a single function evaluation (1-NFE) and a 4-NFE FID of 1.37 within 40 epochs of training, surpassing the state-of-the-art distillation baselines without data. Our code code and model are available at https://consistentdmd.github.io/.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P23", "title": "Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents", "url": "https://huggingface.co/papers/2610.18082", "authors": [ "Xiyang Wu", "Zhichao Liu", "Shengxin zhang", "Zongxia Li" ], "categories": [ "other" ], "category": "other", "summary": "Long-horizon agents often struggle to recover from failures whose causes lie several steps earlier, beyond what local reflection or backtracking can resolve.", "summary_html": "Long-horizon agents often struggle to recover from failures whose causes lie several steps earlier, beyond what local reflection or backtracking can resolve.", "why": "", "why_html": "", "abstract": "Long-horizon agents often struggle to recover from failures whose causes lie several steps earlier, beyond what local reflection or backtracking can resolve. We introduce BELIEFWEAVER, a framework that combines persistent evidence across attempts with prospective reasoning over repair directions. It maintains a relational belief state over frontiers, attempts, observations, and checkpoints. Execution evidence updates related frontier priorities across attempts and rollback, while diagnostic evidence gates execution. An implicit world model (IWM) evaluates repair hypotheses, predicts progress, and identifies relations among frontiers to guide exploration. We further distill these recurring IWM reasoning components into a 9B student through supervised fine-tuning and on-policy distillation (OPD). Across three terminal-based benchmarks, BELIEFWEAVER improves aggregate final score by 5.1 percentage points over native Terminus-2 continuation, with positive trends in exploration efficiency. The distilled 9B student improves final score by 9.8 percentage points over the reported native baseline while reducing estimated teacher API cost by 54.8% relative to the matched teacher. These results show that persistent evidence and prospective reasoning can improve recovery in long-horizon agent tasks while substantially reducing reliance on teacher-model inference.", "doi": null, "source": "HF Papers", "also": [] }, { "ref": "P24", "title": "Co-Evolving Acting and Forecasting for Language Agents", "url": "https://huggingface.co/papers/2610.20597", "authors": [ "Haofei Yu", "Yixing Chen", "Yining Zhao", "Xuyan Ye" ], "categories": [ "other" ], "category": "other", "summary": "Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused.", "summary_html": "Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused.", "why": "", "why_html": "", "abstract": "Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused. Prior work exploits this supervision through value targets that estimate per-step success probabilities or observation targets that fit environment response. Both are hard to construct well: (1) value targets come from the same single scalar signal for each step in the trajectory, and (2) observation targets, such as web pages and terminal logs, are long and noisy and reach only one step ahead. We propose CoPE, which adds a third kind of multi-step action targets. On successful on-policy trajectories, CoPE trains the policy backbone to generate the action sequences that follows each history. Each action is also supervised when forecast from the preceding histories, forming a temporal ensemble that the shared backbone absorbs during training, and this enables acting and forecasting to co-evolve. The policy still acts one step at a time at inference, with no overhead. Across five benchmarks (ALFWorld, ScienceWorld, WebShop, τ^2-bench, and AppWorld), CoPE (1) achieves stronger aggregate performance than GRPO and RL baselines with predictive auxiliary targets across simulated environments, while improving overall performance on both realistic applications; (2) delivers up to 1.9times and 2.0times higher sample efficiency than ECHO and GRPO, respectively, on ScienceWorld; and (3) achieves these gains with a single policy backbone, without additional models, separate pretraining, or inference overhead. The codebase is available at https://github.com/ulab-uiuc/CoPE.", "doi": null, "source": "HF Papers", "also": [] } ], "industry": [ { "ref": "I1", "title": "WhatsApp's help pages hid LLM instructions: dont recommend a regulator", "url": "https://redact.dev/blog/whatsapp-help-pages-hidden-ai-instructions", "source": "Hacker News", "summary": "96 points and 12 comments on Hacker News.", "summary_html": "96 points and 12 comments on Hacker News.", "points": 96 }, { "ref": "I2", "title": "Baseus’ new security camera has a spotlight that tracks and follows moving targets", "url": "https://www.theverge.com/tech/1005073/baseus-security-primesentinel-u2-pro-outdoor-3k-dual-camera-tracking-spotlight", "source": "The Verge", "summary": "Baseus announced a new addition to its lineup of AI-assisted security cameras that's more proactive about discouraging unwanted visitors instead of just recording and alerting you to their presence.", "summary_html": "Baseus announced a new addition to its lineup of AI-assisted security cameras that's more proactive about discouraging unwanted visitors instead of just recording and alerting you to their presence.", "points": null }, { "ref": "I3", "title": "Embedded AI evaluators won’t be an easy solution, Hugging Face’s policy chief says", "url": "https://techcrunch.com/2026/10/19/embedded-ai-evaluators-wont-be-an-easy-solution-hugging-faces-policy-chief-says/", "source": "TechCrunch", "summary": "The idea of embedded evaluators — in which supposedly \"neutral\" third-party observers are placed inside AI organizations — has been popularized over the last several months, as AI labs grapple with a rash of safety incidents that have spurred concerns over the potential harms of their technology.", "summary_html": "The idea of embedded evaluators — in which supposedly "neutral" third-party observers are placed inside AI organizations — has been popularized over the last several months, as AI labs grapple with a rash of safety incidents that have spurred concerns over the potential harms of their technology.", "points": null }, { "ref": "I4", "title": "Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’", "url": "https://techcrunch.com/2026/10/19/microsoft-says-ai-agents-are-like-insider-threats-as-frontier-models-cant-be-trusted/", "source": "TechCrunch", "summary": "Microsoft wants enterprises to closely watch the agents that employees use with systems not controlled by AI labs like OpenAI and Anthropic.", "summary_html": "Microsoft wants enterprises to closely watch the agents that employees use with systems not controlled by AI labs like OpenAI and Anthropic.", "points": null }, { "ref": "I5", "title": "CERN and Mistral sign a statement of intent to establish a strategic partnership", "url": "https://home.cern/cern-and-mistral-join-forces-to-accelerate-ai-for-science/", "source": "Hacker News", "summary": "323 points and 57 comments on Hacker News.", "summary_html": "323 points and 57 comments on Hacker News.", "points": 323 }, { "ref": "I6", "title": "Timestral: Mistral's time-series foundation models", "url": "https://mistral.ai/news/timestral/", "source": "Hacker News", "summary": "564 points and 105 comments on Hacker News.", "summary_html": "564 points and 105 comments on Hacker News.", "points": 564 }, { "ref": "I7", "title": "CoreWeave’s debt is becoming the whole AI industry’s problem", "url": "https://www.theverge.com/ai-artificial-intelligence/1014045/coreweave-stock-debt-sector-ai", "source": "The Verge", "summary": "CoreWeave had a bad week last week.", "summary_html": "CoreWeave had a bad week last week.", "points": null }, { "ref": "I8", "title": "Accused OpenAI spy had folder called “Saudi Gov Docs,” FBI says", "url": "https://arstechnica.com/tech-policy/2026/10/accused-openai-spy-had-folder-called-saudi-gov-docs-fbi-says/", "source": "Ars Technica", "summary": "US fears Saudi government will help ex-OpenAI engineer flee spying charges.", "summary_html": "US fears Saudi government will help ex-OpenAI engineer flee spying charges.", "points": null }, { "ref": "I9", "title": "Ask HN: How to deal with shipping AI-written code you don't understand?", "url": "https://news.ycombinator.com/item?id=50151036", "source": "Hacker News", "summary": "315 points and 461 comments on Hacker News.", "summary_html": "315 points and 461 comments on Hacker News.", "points": 315 }, { "ref": "I10", "title": "AI game clones are making indie developers afraid to show their work", "url": "https://arstechnica.com/gaming/2026/10/ai-game-clones-are-making-indie-developers-afraid-to-show-their-work/", "source": "Ars Technica", "summary": "Delayed Steam Next Fest demo is a sign of a growing problem for indie devs.", "summary_html": "Delayed Steam Next Fest demo is a sign of a growing problem for indie devs.", "points": null } ], "terms": [], "watchlist": [], "watchlist_html": [], "sources": [ { "name": "arXiv", "status": "ok", "count": 712 }, { "name": "HF Papers", "status": "ok", "count": 50 }, { "name": "OpenAI", "status": "ok", "count": 1 }, { "name": "Google Research", "status": "ok", "count": 0 }, { "name": "DeepMind", "status": "ok", "count": 0 }, { "name": "Google AI", "status": "ok", "count": 0 }, { "name": "TechCrunch", "status": "ok", "count": 19 }, { "name": "Ars Technica", "status": "ok", "count": 14 }, { "name": "The Verge", "status": "ok", "count": 10 }, { "name": "Hacker News", "status": "ok", "count": 32 } ], "model": "nvidia/nemotron-3.5-super-vl:free", "cost_usd": 0.0, "degraded": false, "note": null }