News without the noise
Digests of notable academic papers, primarily from arXiv
212 articles · page 4 of 5
Chen and colleagues introduce PeakBench, a benchmark of executable multi-tool workflows annotated with dependencies and measured resource demands. Their evaluation shows that agents capable of sound logical planning do not necessarily schedule tools safely or efficiently, while access to resource information can reduce avoidable capacity overflows.
Kim et al. built Simthesizer, a framework in which a harnessed coding agent modifies a shared simulator of LLM serving workflows rather than constructing a separate simulator for each new mechanism. The authors report 2.51% average throughput error against a vLLM-based system and simulation speeds up to 284.96 times those of competing simulators.
The authors trained a guard model to identify risky actions at individual steps in an agent’s trajectory, including before a tool call is executed. On AgentDojo and AgentDyn, StepGuard reduced mean attack success by 77.3% relative to unguarded agents while lowering mean utility by 2.8 percentage points.
Hazra et al. introduce Crase, a scholarly search system that explores a predefined citation neighborhood rather than letting an agent search iteratively without a fixed stopping rule. Across two benchmarks on a 500,000-paper arXiv corpus, the authors report up to three times the recall@50 of proprietary-model deep research agents at roughly one-third of their cost.
Huang et al. replace per-query iterative action optimization with a learned trajectory prior conditioned on current and goal images. Across four goal-conditioned pixel-control benchmarks, the authors report higher success rates and an order-of-magnitude reduction in planning time.
Ganz et al. tested coding agents that switched between lower-cost, lower-capability models and higher-cost, higher-capability models during long-running tasks. They find a “handoff tax”: giving a stronger model the weaker model’s full trajectory recovers less than half the quality gap between them while adding substantial cost.
Tam et al. jointly trained a 4-billion-parameter Qwen3 model on tool creation and tool use across 13 procedural reasoning tasks. The resulting model achieved 79.8% macro-average accuracy on held-out tasks and transferred to tabular and visual question answering without training on either data type.
The authors developed a search framework that iteratively modifies an agent’s surrounding harness using tasks selected according to baseline failure patterns. Across three enterprise benchmarks, evolved harnesses improved full-benchmark performance by 20–35 percentage points, with gains extending to held-out tasks and transferring across GPT and Qwen model families.
The strongest signal this week is that agent performance increasingly depends on choosing what to inspect, retrieve, or reward—not simply allocating more context or evaluation. Separate results also show that soft guidance remains unreliable when systems must obey hard operational or generative constraints.
Luo and colleagues built a controlled financial-compliance environment and a monitoring benchmark to test whether language-model agents ground their actions in executable rules. Across experiments with DeepSeek V4 Pro and Gemini 3.5 Flash, displaying rules reduced but did not eliminate rejected orders, while independent monitors could be misled by trader rationales when enforcement evidence was withheld.
Chen et al. formulate the choice of reusable skill documents for LLM agents as a budget-constrained optimization problem rather than an independent relevance-ranking task. Their Best Prefix Selection algorithm achieved 0.73 task success on a contamination-controlled BigCodeBench variant, compared with 0.20–0.52 for existing selection methods, while using fewer tokens than the strongest released router.
Han et al. analyze preference optimization for flow-matching generative models, showing that reward-driven updates can push outputs away from the pretrained data manifold when the induced displacement has a component normal to that manifold. Their proposed ThermoDPO-weighted objective improves both a toy benchmark and text-rendering metrics for Stable Diffusion 3.5 Medium.
Suwannapichat et al. add reward-guided fine-tuning to a model that generates communication graphs for LLM-based multi-agent systems. They report matching ARG-Designer’s task accuracy while reducing average token consumption by 20.5%.
Miyai, Aizawa and Yamasaki propose a method for improving the code and instructions surrounding an LLM agent without changing its model weights. Their approach adaptively evaluates candidate harnesses on tasks where past candidates disagreed, substantially reducing evaluation costs while preserving final performance in the reported experiments.
Fisch et al. model AI routing as a search problem in which inspecting each candidate model’s likely value is itself costly. Across multi-LLM routing, retrieval-augmented specialists, and variable reasoning budgets, their centralized policy reportedly preserves the routing quality of exhaustive estimation while invoking expensive estimators much less often.
Qian et al. propose MileGPO, a policy-optimization method for agents that receive supervision mainly from final task outcomes. On ALFWorld and WebShop, the authors report that its locally calibrated intermediate credits improve performance without auxiliary models or additional environment interactions.
Xu et al. developed a group-image generation framework that represents each requested person with an addressed identity token, predicts an identity–layout plan, and uses that plan to guide synthesis. On an identity-disjoint benchmark, the method improved face similarity over GPT-Image-2 while covering 97.3% of requested identities with a 2.8% duplicate rate.
Kimi Team introduces Kimi K3, a 2.8T parameter Mixture-of-Experts model with 104B activated parameters, native vision, and a 1M-token context window. The authors demonstrate that innovations in attention (Kimi Delta Attention, Attention Residuals) and routing (Stable LatentMoE) yield roughly 2.5× scaling efficiency improvement over Kimi K2, enabling frontier-level performance across long-horizon coding, agentic, reasoning, and vision tasks.
The authors introduce Skill Self-Play (Skill-SP), a framework where LLMs generate and solve challenging tasks by dynamically routing through learned skills. In a reinforcement learning loop, the system co-evolves a proposer, solver, and skill controller, showing consistent improvements on tool-use and reasoning benchmarks, even turning around initially misaligned models.
The authors propose WorldWeaver, a streaming multi-agent video diffusion model that augments autoregressive rollout with cross-agent world state registers: learnable tokens that maintain shared world information, track individual agent status, and are updated after each generated chunk. The registers are grounded with supervision from agent status, global bird's-eye views, and scene text. Experiments in two-agent Minecraft show that explicit world-state modeling improves logical consistency and generation quality over baselines that only carry forward observation history.
The authors present OpenForgeRL, an open-source framework that trains stateful, harness-based AI agents end-to-end using standard reinforcement learning libraries. It wraps the harness's model calls via a proxy for data recording and orchestrates rollouts in remote Kubernetes containers, yielding competitive results on tool-use, browser, and computer-use benchmarks with relatively few training tasks.
The authors introduce SANA-Video 2.0, a video diffusion transformer that uses a hybrid of linear and softmax attention to achieve efficient long-sequence processing. The model, scaled to 5B and 14B parameters, generates 720p video on a single H100 GPU and achieves a VBench score of 84.30 at 480p in 13.2 seconds, matching the quality of larger softmax-based models while being significantly faster.
The authors propose Agentic Designer, a multi-agent system that generates interior furniture layouts by breaking the task into stages of proposal, evaluation, and refinement. Compared to one-shot methods, their approach enforces geometric constraints at each step, achieving superior structural adherence and functional coherence on a new benchmark, InStruct.
The authors introduce NVIDIA Object-Oriented Agents (NOOA), a model-agnostic Python framework where any Python object can act as an AI agent: methods define available actions, fields store state, docstrings serve as prompts, and type annotations act as contracts. Methods whose body consists only of '...' are completed at runtime by an LLM-driven loop, while normal methods remain deterministic. The framework is demonstrated on SWE-bench Verified, Terminal-Bench 2.0, and ARC-AGI-3, showing that current models use this interface effectively.
Hu et al. introduce Molt, a PyTorch-native training framework for agentic reinforcement learning that prioritizes researcher readability and end-to-end algorithm traceability. In benchmarked asynchronous training runs, Molt achieved statistically comparable performance to a state-of-the-art Megatron-based stack, demonstrating that leanness does not compromise scalability.
The authors formalize the trade-off between retaining raw memory and consolidating it via compression, showing that consolidation improves accuracy by up to 48% under tight context budgets, while retention is preferable when budgets are generous. They propose OAS, a lightweight learner that selects the optimal operator per query.
The authors propose TRACE, a credit-assignment method for long-horizon agentic RL that derives per-action rewards from the log-probability of the gold answer under a frozen reference model, using temporal-difference learning. On the closed-web BrowseComp-Plus benchmark, TRACE raises Qwen3-4B from 7.2 to 35.6 and Qwen3-30B-A3B from 8.4 to 42.6, and enables pure RL training without cold-start supervised fine-tuning or live-web data.
The authors introduce Boogu-Image-0.1, a family of open-source multimodal models that unify understanding and generation. By strengthening the system's understanding capability through a better multimodal encoder, agentic prompt rewriting, and inference-time scaling, the model matches or surpasses existing open-source models and approaches closed-source systems, all with a theoretical training cost of roughly $400K.
Wang et al. revisit the evaluation of automatic harness evolution for LLM agents, comparing it with simple test-time scaling and discovery baselines under matched feedback and inference budgets. They find that harness evolution does not consistently outperform these baselines on Terminal-Bench 2.1 and shows limited generalization to held-out tasks, raising concerns about overfitting and the effectiveness of the approach.
The authors introduce STRACE, a framework that constructs high-signal optimization contexts for long-horizon agents by removing redundant and non-causal steps from execution traces. On a formal verification benchmark, STRACE boosted the success rate of human-expert designed agents from 42.5% to 58.5%, a 1.4× improvement over standard context-filtering baselines.
The authors present LingBot-World 2.0, an upgrade to their open-world simulator. Key advances include causal pretraining enabling infinite interaction horizons, a distilled real-time model running at 60fps for 720p video, and integration of an agentic harness with a pilot agent for character behavior and a director agent for procedural environment generation.
The authors introduce LLM-as-a-Verifier, a general-purpose verification framework that extracts continuous scores from LLMs by taking expectations over scoring token logits. This approach achieves state-of-the-art results on four benchmarks—Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%)—and provides fine-grained feedback that can be used for reinforcement learning and progress monitoring.
Li et al. propose CompactionRL, a reinforcement learning method that trains LLM agents to perform long-horizon tasks by compressing interaction histories. Applied to open models, it yields consistent gains on code agent benchmarks, with GLM-4.5-Air reaching 66.8% Pass@1 on SWE-bench Verified.
The authors introduce Untrusted Content Masking (UCM), a defense that prevents prompt injection attacks on web agents by masking untrusted content in the Document Object Model (DOM) before the agent processes it. They show that the DOM structure alone can distinguish trusted from untrusted page regions without reading their content, enabling agents to observe and interact safely with web pages while remaining isolated from adversarial content.
The authors introduce WhisperBench, a 108-case benchmark for stealth memory injection, and MemGhost, a one-shot payload generation framework. In experiments, MemGhost achieves 87.5% end-to-end success on OpenClaw with GPT-5.4 and 71.4% on Claude Code SDK with Sonnet 4.6, revealing that persistent memory can be a practical pathway for long-term agent compromise.
Hills et al. introduce Iterative VibeCoding, a benchmark for AI control in persistent codebases, and demonstrate that gradual attacks—distributing a covert side task across multiple pull requests—achieve high evasion (≥65%) across several attack models. They find that no single monitor is robust against both gradual and non-gradual attack strategies, but a stateful link-tracker monitor combined with trajectory monitors in a four-monitor ensemble reduces gradual-attack evasion from 93% under the weakest diff monitor to 47%.
Feng et al. present Vera, an end-to-end automated safety testing framework for LLM agents that uses a three-stage pipeline to discover risks, compose executable safety cases, and execute them in sandboxed environments with evidence-grounded verification. Evaluating on four production agent frameworks (OpenClaw, Hermes, Codex, and Claude Code), they find average attack success rates of 93.9% under multi-channel attacks, and release Vera-Bench with 1600 safety cases spanning 124 risk categories.
Liu et al. present SimWorlds, a multi-agent framework that generates dynamic, editable 4D scenes from natural language descriptions. The system outperforms prior baselines in visual fidelity and physical consistency on a new benchmark, 4DBuildBench.
The authors introduce COMFYCLAW, an agentic harness that controls ComfyUI workflows for image generation. By representing workflow construction as typed graph editing, reverting invalid edits, and using a region-level vision-language model (VLM) verifier to provide actionable repair suggestions, COMFYCLAW evolves a skill library from past trajectories, errors, and verifier feedback. Across four benchmark splits, three agent models, and two image backbones, COMFYCLAW achieves the best average evaluation score in all six agent configurations, and human annotators prefer its outputs over variants without skill evolution.
The authors introduce AutoMem, a framework that treats memory management as a learnable skill for LLMs by promoting file-system operations to first-class actions. By automating the optimization of memory structure (prompts, schemas, action vocabulary) and model proficiency (training on good memory decisions), AutoMem improved base agent performance by approximately 2x to 4x across three long-horizon games without altering task-action behavior.
Zhang et al. propose a training paradigm that equips LLM agents with an internal world model for prospective reasoning. By first injecting predictive capabilities through a world model mid-training stage, then eliciting structured foresight via supervised fine-tuning, and finally calibrating with reinforcement learning, the authors demonstrate significant improvements on search and mathematical reasoning tasks over standard fine-tuning approaches.
The authors propose Qwen-Image-Agent to address the context gap where user requests for image generation are often underspecified, implicit, or depend on up-to-date knowledge. The framework uses Context-Aware Planning to identify missing details and Context Grounding to acquire them from reasoning, search, memory, and feedback. On the newly introduced IA-Bench benchmark and two other datasets, it achieves state-of-the-art performance against strong baselines.
The authors present CRAFT, a two-stage post-training recipe that enables enterprise coding agents to learn schema and tool-use behavior without exhaustive prompt-time schema injection. They show that CRAFT improves composite Agent Score by +9.6 percentage points, reduces input tokens by ~9x, and cuts schema-discovery loops by up to 5x compared to a schema-stuffed baseline.
The authors introduce SelfCompact, a scaffold that allows language model agents to dynamically decide when and how to compact their accumulated context. By combining a tool for summarization with a rubric that specifies when to invoke it, SelfCompact matches or exceeds fixed-interval baselines on math and agentic search benchmarks, using 30-70% fewer tokens.
The authors present Sol Video Inference Engine, a training-free framework that uses parallel agent modules to optimize cache, sparse attention, token pruning, quantization, and kernel fusion for a given model-hardware-configuration triplet. Across three models (64B Cosmos3-Super, 22B LTX-2.3, 2B SANA-Video), the full stack achieves over 2x end-to-end acceleration while maintaining near-lossless VBench quality, with minimal human effort.
The authors present Sakana Fugu, a family of orchestrator language models that generate dynamic, query-adaptive agentic scaffolds to harness the specialized capabilities of multiple LLMs. Fugu and Fugu-Ultra achieve state-of-the-art results on six challenging benchmarks, including SWE-Bench Pro, Terminal Bench, and GPQA-Diamond, surpassing any single publicly accessible model.
Kim et al. introduce Multi-Agent Transactive Memory (MATM), a framework for storing and retrieving agent-generated trajectories in a shared repository accessible to a population of LLM agents. In ALFWorld and WebArena environments, retrieving prior trajectories improves downstream task success rates and requires fewer interaction steps without any coordination or joint training.
The authors investigate whether heterogeneous LLMs can be aligned to directly transfer latent representations (KV-caches) between them, bypassing costly text decoding and re-encoding. They propose a method using a cross-model transformation and two-phase training (reconstruction then generation), achieving performance comparable to or better than text communication in context-aware settings at 2-3x lower compute, and remaining effective in context-unaware settings where prior heterogeneous methods collapsed.