JevSpawn is a new compositional policy that links natural language task specifications to finite probabilistic exploration, enabling LLM agents to generate actions more efficiently. It uses parallel action spawning, feedback‑driven branch selection, representation revision, and recovery from retained alternatives to adapt actions during interaction. Evaluations on eight benchmark tasks show that JevSpawn outperforms seven agent baselines and a TypeSafe Jev variant, improving task performance and speeding navigation.
By Haoyang Su, Weiran Huang
CompMat-Bench is a new benchmark comprising 94 tasks drawn from recent computational materials science studies, designed to evaluate AI agents on realistic research steps without requiring costly simulations during testing. The benchmark pre‑reproduces inputs and outputs to provide ground truth, allowing agents to be graded with fixed rules rather than an LLM judge. It supports both single tasks and multi‑step workflows, with varying levels of methodological guidance, and shows that while agents can achieve high pass rates on individual tasks, performance drops in longer workflows or with reduced guidance, often due to scientific rather than software errors.
By Chenmu Zhang, Levi Felix, Jun-Jie Zhang, Xingfu Li, Xuelian Jiang, Tao Jiang, Subhendu Mishra, Xixi Qin, Boris Yakobson
The paper introduces OSCAR, an LLM‑based framework that translates business descriptions into accurate optimization models while verifying and improving them through a simulator, coder, and reviewer. OSCAR uses a cost‑ordered escalation strategy to select among LLMs of varying price and capability, achieving 95–100% accuracy on benchmark problems with local, open‑weight models. The framework also provides competitive guarantees and token‑cost advantages over existing LLMs like Codex and Claude Code.
By Jinzhi Bu, Haixin Tang, Huanan Zhang
PG-SFT: Balancing Capability Acquisition and Retention in Offline Agent Fine-Tuning explores how to maintain a model’s existing abilities while teaching it new ones through supervised fine‑tuning on offline agent trajectories. The authors compare standard SFT, KL‑penalty, and update‑magnitude constraints, finding that these methods still degrade non‑target capabilities. They introduce Privilege‑Guided SFT (PG‑SFT), which uses turn‑level information gain to modulate supervision strength, achieving a better trade‑off between acquiring new skills and preserving existing ones, though with a slight drop in target‑task performance.
By Ronghua Li, Zi Liang, Zhishan Li, Shinan Liu
The study investigates how reputation, strategy, and emotional signals influence cooperation in generative AI models using the iterated prisoner's dilemma. Non‑reasoning models (Claude 3.5, Gemini 2.0 Flash, GPT‑4o) showed cooperation shaped by all three factors, while reasoning models (Claude 4.6, Gemini 3, GPT‑5.2) relied more on strategy and reputation, displayed reduced emotional influence, and exhibited varied end‑game behaviors. These results highlight the growing sophistication and heterogeneity of AI social behavior, suggesting the need for standardized cooperation benchmarks.
By Celso de Melo, Zishan Feng, James Hale, Kazunori Terada, Giorgio Coricelli, Jonathan Gratch
The paper introduces Revision‑Aware Independent Agent Graphs (RIAG) to address dynamic task routing, where an event stream continually revises task bindings and a system must select the correct document version at query time. By repurposing six benchmarks into over 31,000 dynamic episodes, the authors demonstrate that RIAG balances recomputation and reuse, achieving 54.24 % joint routing‑and‑answer accuracy with only 0.62 calls per query—substantially better than the strongest baseline. The study highlights the trade‑off between stale conclusions and wasted work in dynamic reasoning settings.
By Yan Luo, Selim-Antoine Lali, Jeremy Moebel, Iliass Khoutaibi, Ahmadou Aidara, Mengyu Wang
DAYJOB is a new benchmark comprising 130 long‑horizon professional tasks created by experts in healthcare (50 tasks) and finance (80 tasks). Each task is packaged as a containerized Harbor environment and evaluated against a detailed binary rubric, requiring agents to meet every criterion to pass. In tests across 30 model configurations, the best model (Claude Opus 5.5) achieves about 24% success in both domains, while the median model scores below 3%.
By Stephanie Finley, Liudas Panavas, Thomas Mikkelson, Cam Hinton, Stacey Ganss, Bradley Monton, Emily Kendall, Michelle Spradlin, Lydia Bye, Michael O'Brien, Lauren Ylvisaker, Derek Ray, Suhaas Garre, Sushant Mehta, Edwin Chen
AiSearch is a flexible multimodal retrieval framework that uses Vision Language Models (VLMs) to enable natural language search over images and videos. It supports interactive search refinement through user feedback, allowing results to be tailored to the user's intent in real time. The system also provides visual benchmarking across multiple VLMs, enabling users to choose the most suitable model for their specific task.
By Ali Koksal, Mei Chee Leong, Vicky Sintunata, Ching Ling Chin, Wee Teck Fong
Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States introduces PoS, an inference-time framework that builds and maintains explicit belief states to guide large language model agents. Each belief state combines an estimate of the current world with unresolved task requirements, making clear what the agent still needs to learn and accomplish. PoS validates consistency, monitors task progress to detect Belief Trapping, and tailors recovery to the trapping pattern and unresolved requirements, achieving top performance across four benchmarks with all three LLM backbones.
By Yu Luo, Jiamin Jiang, Yimin Zuo, Xidao Wen, Rongchen Gao, Yongqian Sun, Shenglin Zhang, Guiyang Liu, Cheng Zhang, Fang Situ, Qi Zhou, Dan Pei
NextMe-800 is an approximately 800‑hour first‑person video dataset collected from a single volunteer over 126 days, featuring 1 Hz images, gaze, and audio. The data are captioned at five hierarchical abstraction levels—from atomic actions to major activities—enabling personalized action anticipation as an open‑vocabulary K‑step sequence prediction task. The authors also introduce NextAct, a 1,500‑point benchmark that combines NextMe‑800 with the multi‑person EgoLife dataset, and evaluate models using an embedding‑based soft edit distance to assess how well personal behavior can be anticipated across abstraction levels and prediction horizons.
By Zhaoxu Meng, Yiming Sun, Mingyuan Gao, Jiachang Zhang, Zhuhan Dai, Yipeng Du, Zheng Lian, Jian-Qiao Zhu
The paper introduces MCRI, a four‑dimensional framework for analyzing and evaluating agent skills, and implements it as MCRI‑Eval, a large‑language‑model‑based evaluation method. Using 63,812 public skills from OpenClaw and 58,275 skill‑conditioned executions across BigCodeBench, BFCL‑Fundamental, and Mind2Web, MCRI‑Eval’s scores correlate with community popularity and outperform other methods in downstream ranking agreement. The evaluation also shows that MCRI‑Eval improves top‑1 skill selection by 17.7, 22.8, and 19.6 percentile points on the three benchmarks, offering a valuable pre‑execution signal for prioritizing promising skills before costly evaluation.
By Zongrui Yang, Li Xintong, Runchen Xu, Zhongsheng Wang, Zhedong Lin, Haoyuan Li, Jiamou Liu
The paper argues that evaluating agents as fixed models is insufficient, proposing instead to treat them as configurable systems. Using a new benchmark of four scientific tasks, the authors analyze how five configuration aspects—task information, reasoning, self‑verification, time budget, and backbone model—affect performance, noting that about 54% of outcome variance arises from run‑to‑run differences even with the same settings. The study finds that providing more task information has the strongest impact, while interactions among settings (e.g., extra time only helps with adequate information or model capability) and the choice of verification tools significantly shape agent behavior.
By Luis Wiedmann, Leander Girrbach, Cordelia Schmid, Zeynep Akata
Mingbird is a local‑first agent harness designed for small open‑weight language models (2–9 B) that run on ordinary laptops. It introduces ten mechanisms—such as a byte‑level net‑zero prefill budget, a finish gate that re‑reads the task before accepting completion, and signature‑level loop detection—to address common failure modes that arise from the harness rather than the model itself. In controlled experiments on the LRAB benchmark and the $ au^2$‑bench, Mingbird achieves higher overall scores (0.886 and 0.856 respectively) compared to other harnesses, and its ablation studies show that each mechanism contributes measurable performance gains.
By Hao Wang, Ting Huang
ScholarCatalyst is a new benchmark that evaluates how well AI systems can retrieve research papers that inspire new work. The dataset was created by having 184 lead authors of 207 recent computer science papers annotate which earlier papers helped their projects, providing detailed rationales. The benchmark tests retrieval from the literature available at the start of a project, revealing that current agentic search and even advanced models like Claude Fable 5.1 perform only modestly better than simple embedding retrieval.
By Sohyeon Kim, Yoonho Lee, Bo Liu, Dayoon Ko, Rulin Shao, Seungone Kim, Graham Neubig, Pang Wei Koh, Aakanksha Chowdhery, Akari Asai, Omar Khattab, Yejin Choi, Gunhee Kim, Chelsea Finn
The paper introduces a bounded‑fidelity sim‑as‑demo‑stage design pattern that suppresses contact physics during object handoffs in simulators, using MuJoCo’s mocap‑body primitive and a lightweight Python adapter. This approach ensures audit‑chain stability, producing identical event‑log hashes across 1,000 replays per posture, whereas a contact‑force baseline yields significant divergence. The authors demonstrate that the pattern maintains reproducibility across various timesteps and sequential handoffs with minimal overhead, and they identify contexts where it should not be applied.
By Xue Qin, Simin Luan, Cong Yang, Zhijun Li
The paper investigates how inherited state affects sub-agent performance in multi-agent frameworks, comparing three inheritance policies—Reset, Selective, and Full—across a ladder of Qwen3 models. It finds that reliance on stale state decreases with model capability, but a mid-capability model (Qwen3‑1.7B) exhibits a statistically significant local minimum of net harm, defining a "danger band." Selective handoff consistently improves accuracy over Full, especially within the danger band, while a fixed-threshold router fails on other datasets.
By Jundong Hu, Shekar Ramachandran
The paper introduces a method for generating legible plans in arbitrary PDDL domains by extending prior legibility research to classical planning without custom planners. It incorporates a second‑order theory of mind to estimate the observer’s perspective, enabling robots to implicitly communicate goals in human‑robot teaming. Benchmark results show that increasing legibility typically trades off with plan efficiency, and a regularizing factor is needed to balance the two.
By Michele Persiani, Thomas Hellstr\"om
ShatterQuant is a hardware-software co-designed framework that enables mixed-precision quantization within individual tensors by assigning different bit-widths to blocks of a weight projection. It couples precision granularity with processing element configuration, allowing each precision to determine an effective block height. The framework includes a hardware-aware post-training method based on block-level standard deviation and weight sensitivity, a ShatterQuant Transformer Accelerator supporting 1/2/4/8-bit weight precision, precision-dependent PE configuration, block rescaling, and integrated softmax and piecewise-linear nonlinearities, and an evaluation showing 1.5 TOPS, 760 GOPS/$mm^2$ area efficiency, and 2.8 TOPS/W energy efficiency on a TSMC 16nm PDK implementation.
By Mikolaj Walczak, Edward Humes, Chao Fang, Marian Verhelst, Tinoosh Mohsenin
arXiv:2610.00385v1 Announce Type: cross
Abstract: Replay selectors often rank cached trajectories by format feedback, confidence, freshness, or response length, although cache-level correctness and d...
By Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Tianshu Fu, Daren Zha, Jun Xiao
arXiv:2610.00389v1 Announce Type: cross
Abstract: Open-ended query generation lacks standard answers, thus necessitating an effective reward mechanism. Pointwise scoring rubrics provide limited infor...
By Zihan Shen, Qi Liu, Zixuan Yang, Yiqun Chen, Chenglong Zhao, Xiaozhao Wang, Lei He