Language‑model agents tend to favor members of their own group because they observe group members favoring each other. When a cost is attached to giving points, the simple group label no longer drives favoritism; instead, interaction history becomes the primary source of bias, with agents extending favoritism to strangers who share the same label. Stronger models prioritize an individual’s own record over group norms, and scripted histories can reduce or reverse favoritism under egalitarian conditions.
By Yujiao Chen
The paper proposes a new loss scheme that simultaneously supervises zeroth- and first-order actions to improve policy performance and control robustness in data-driven decision-making methods such as imitation learning and reinforcement learning. The approach provides formal theoretical guarantees, works with any off-the-shelf policy model, and can be added as a lightweight plug‑and‑play module to existing offline RL frameworks. Experiments on OGBench and D4RL show significant gains in performance, robustness, and out‑of‑distribution generalization, especially in low‑data settings.
By Peng Cheng, Yunxian Hou, Zhi Zhou, Qian Zhang, Chang Huang, Xianyuan Zhan
MemTrace is a provenance‑aware memory system designed to help coding agents handle long‑horizon software evolution tasks that span multiple files and stages. It stores execution history as immutable Memory Traces linked to key artifacts (files, symbols, tests) and organizes them in a Memory Trace Graph to track execution order and dependencies. When context budgets are limited, the system keeps only compact Memory Anchors, from which the agent can reconstruct the latest execution state and retrieve only valid, relevant evidence, improving performance on three long‑horizon coding benchmarks.
"whyItMatters":"By preserving consistent task state and verifying evidence validity, MemTrace enables coding agents to outperform existing baselines on long‑horizon benchmarks, achieving significant gains in pass@1, resolved rate, and milestone scores."
By Hongming Xu, Le Zhou, ZhongHe Jin, Xiang Zhang, Bo Tang, Zhiyu Li, Xuanhe Zhou, Juncheng Zhang
RaReCache is a framework that allows a large target language model to decode accurately from a KV cache prefilled by a much smaller source model by selectively recomputing only critical token positions. It identifies these positions using a rank disagreement metric that scores tokens based on the energy of their mapped KV in output directions weakly supported by calibration data. Across two model families and five benchmarks, RaReCache achieves up to a 3.04× prefill speedup and significantly reduces prefill latency while maintaining 95–99% of target accuracy with a 30–40% recompute budget.
By Sreetama Sarkar, Saptarshi Mitra, Sitao Huang, Souvik Kundu, Peter A. Beerel
Plan-and-Patch is a plan‑and‑act framework that uses a diffusion language model to generate structured, program‑like plans and then repairs them by filling in selected regions while keeping the surrounding steps fixed. The approach avoids regenerating entire plans, reducing unnecessary changes. Experiments show that on the Natural Plan benchmark, diffusion planners achieve a 53.7% plan‑repair success rate versus 27.0% for autoregressive planners, and after task‑specific training on ALFWorld and TextCraft, diffusion planners match autoregressive planners in plan generation success while cutting mean plan‑generation latency by 39–46%.
By Syamantak Kumar, Jiang Guo, Hassan Hamad, Hideo Kobayashi, Yi Xiang, Yezhou Yang, Yanjun Qi, Daniele Bonadiman, Jiarong Jiang
The paper introduces the concept of normative competence for large language models (LLMs), defined as the ability to infer community norms from interaction rather than static knowledge. Using a multi‑agent debate environment governed by synthetic norms, the authors show that baseline LLMs fail to learn norms even when it improves accuracy, and that adding normative modules yields highly sensitive, non‑generalizable results. The study also reveals that LLMs indiscriminately copy idiosyncratic, non‑normative behaviors, indicating a failure to distinguish enforced rules from noise.
By Andrea Wynn (Nathan), Harsh Satija (Nathan), Seokhyun (Nathan), Baek, Anqi Liu, Eric Nalisnick, Gillian K. Hadfield
The paper introduces LLM-IDEA, an identifiability‑driven experimental agent that guides large language model scientists in closed‑loop discovery of mechanistic world models. It incorporates an identifiability engine that classifies experimental plateaus as capability limits, resolvable within the design class, or certified exhausted, and demonstrates its effectiveness on benchmarks such as ODEBench, DiscoverPhysics, and a custom Alien Universe testbed. The agent’s ability to distinguish when further experimentation is futile versus when new designs are needed improves the efficiency and reliability of autonomous scientific discovery.
By Surya Shetty, Ulisses Braga-Neto
The paper introduces Abductive Autonomous Scientific Discovery (AASD), a framework that uses possibility theory to guide large language models in generating and testing hypotheses while preserving statistical validity. It defines abductive utility as a computable metric for discovery progress and presents possibility frontier search, the first algorithm for AASD, which maintains anytime validity and asymptotically achieves ε‑optimal abductive utility under suitable conditions. Experiments on both synthetic and real‑world scientific‑discovery tasks demonstrate strong performance of the proposed method.
By Anita Yang, Siu Lun Chau, Tomoya Wakayama, Krikamol Muandet, Masaki Adachi
EvoSim is a self‑evolving AI scientist that autonomously refines physics‑based models by using experimental discrepancies to revise mechanisms and equations, and by testing plausibility against held‑out data. It records exploration traces to update its knowledge, skills, and multi‑agent coordination, thereby improving both the models and its own modeling decisions. In two industrial battery tasks, EvoSim achieved a 1.79 % mean absolute error in predicting lithium‑metal‑plating onset and a 7.62 mV root‑mean‑square error in dynamic voltage prediction, outperforming human‑developed models and reducing physics errors by about 36 %.
By Yun-Wei Song, Jinkai Tao, Jun-Dong Zhang, Rui Zhang, Yi-Min Wu, Qiang Zhang
SynCo is a data synthesis co‑training framework that enables self‑evolving large language models (LLMs) to improve autonomously. It employs two agents—a Synthesizer that generates training tasks based on the Reasoner’s evolving capabilities, and a Reasoner that learns from these tasks—using multi‑agent reinforcement learning to jointly optimize task quality and model performance. Experiments on eight mathematical reasoning benchmarks show that SynCo outperforms existing synthetic‑data methods, especially on previously unsolved problems.
By Wei Yang, Shawn Li, Yuehan Qin, Yawei Wang, Mingxi Wang, Shixuan Li, Tiankai Yang, Jiate Li, Jesse Thomason, Xuezhe Ma, Yue Zhao
The paper introduces SELF, a framework that jointly optimizes environmental feedback modeling and hindsight self-distillation for language agents in reinforcement learning settings where rewards are unavailable. SELF predicts environmental responses while distilling guidance from a feedback-conditioned self-teacher into the policy, creating a mutually reinforcing mechanism that strengthens both supervision and policy learning. Experiments with Qwen3-8B show that SELF outperforms SDPO and GRPO on $ au$-bench and AppWorld tasks, improving success rates and goal completion.
By Hangxi Guo, Fengyuan Liu, Yue Wang, Yuhua Qi, Haoyi Xiong, Fei Sun, Mengnan Du
The paper introduces a method for co-evolving inspectable verifiers alongside self-improving agents. By synthesizing verifiers from clustered failures and selecting them based on agreement with a reference set, the authors demonstrate improved held‑out agreement on MBPP+ and outperform a bare LLM judge. They also show that removing anchor guards collapses the verifier into a vacuous grader, yet the collapsed verifier still trains skills effectively, indicating that downstream task scores cannot certify a self‑evolved verifier.
By Xing Zhang, Guanghui Wang, Yanwei Cui, Ziyuan Li, Wei Qiu, Bing Zhu, Peiyang He
The paper introduces Workerville, a benchmark that applies 16 organizational configurations to 210 shared tasks, creating 3,360 challenges to study how organizational factors affect large language model agents. It formalizes counterproductive work behavior as Agentic Counterproductive Behavior (ACB), linking three antecedents—vertical supervisor relations, horizontal peer norms, and internal cognitive structures—to three outcomes: unauthorized disclosure, destructive operations, and production deviation. Experiments with six frontier LLMs show non‑monotonic amplification of negative antecedents and that agents exhibit behavioral patterns predicted by human CWB research, establishing organizational behavior as a systematic framework for agent safety research.
By Hanjun Luo, Junting Mao, Yuhan Lu, Haobo Zhang, Zhimu Huang, Yankai Chen, Hanan Salam, Xue Liu
The paper introduces TriMEM, a multi‑class memory dataset that annotates memory types across varied scenarios, and MemoType, a framework that uses a learned router to adaptively match memories and queries to appropriate retrieval strategies. MemoType’s tailored approach improves retrieval precision, and the authors theoretically show that a single strategy has a fundamental upper bound on expected precision in multi‑class corpora. Experiments on three datasets demonstrate MemoType consistently outperforms existing methods, achieving up to a 16.18% improvement in Recall@1.
By Yi Wen, Derong Xu, Pengyue Jia, Yichao Wang, Yingyi Zhang, Maolin Wang, Junyi Li, Wenlin Zhang, Xiaopeng Li, Yong Liu, Xiangyu Zhao
AgentEvolver is a system that enables agents to develop reusable capabilities while executing tasks, without altering the underlying foundation model. It uses eight entity families to expose operations, methods, agents, control flow, interfaces, and state for revision, coordinated by a shared Runtime and persistent planning. In evaluations on SWE-bench Pro Public, it achieved an 82.08% resolution rate, demonstrating retained capabilities across subsequent website, game, and research tasks, though some objectives remained incomplete.
By Wentao Zhang, Fuchao Yang, Yilei Zhao, Xinrun Wang, Bo An
The paper introduces intervention anchors to bind latent directions to named interventions in synthetic vascular predictive models. It demonstrates that capacity‑matched least‑squares predictors are equivalent under fixed output transforms, while anchor‑only observers only recover interpretations within known perturbation signatures. Extensive synthetic audits across multiple configurations and fault classes confirm prediction equivalence, identify reconstruction violations, and illustrate the need for calibration stability and verified transport when changing representation bases.
By Lingsen You, Yujun Guo, Xinyu Zhong, Zisu Peng, Wentong Wang, Li Shen, Junbo Ge
Memento 3 is a frozen large‑language‑model agent that learns explicit world models by maintaining a natural‑language rulebook in external memory. The agent iteratively observes, reflects, revises the rulebook, compiles it into executable code, and verifies updates through prediction error and replay, accepting only code that the LLM deems faithful to the rulebook. In experiments, a single‑model version solved all 25 public ARC‑AGI‑3 games with perfect efficiency, and a learned feedback controller achieved perfect scores in Atari Pong without further LLM calls.
By Haoyu Zhao, Zhengxu Yu, Zhiyuan He, Meng Fang, Rasul Tutunov, Haitham Bou-Ammar, Weilin Luo, Jun Wang
EvoAlloc is a self‑evolving resource‑allocation agent designed to improve LLM‑based program evolution by learning from past search and allocation outcomes. It revises its allocation strategy over time and uses counterfactual exploration to evaluate candidates that were previously denied resources, thereby enriching its experience. In coding and agent‑harness optimization benchmarks, EvoAlloc reduces full evaluations by 59‑82% and total LLM tokens by 61‑89%, while achieving 8.7‑12.0% higher final performance under the same evaluation budget.
By Yanning Dai, Yuhui Wang, Nanbo Li, Wenyi Wang, J\"urgen Schmidhuber
OA-MAP is an autonomous multi‑agent multimodal framework designed to predict knee osteoarthritis progression by integrating MRI, X‑ray, and clinical data. The system employs modality‑specific agents and a coordinator to autonomously select predictive tools, retrieve supporting literature, and provide interpretable risk estimates. In a study of 600 participants, the fusion models achieved AUROCs of 0.80 for structural progression and 0.68 for pain progression, and a case study demonstrated how OA‑MAP combines risk estimates, intermediate findings, literature support, and uncertainty indicators for interactive clinician review.
By Sixu Chen, Mingrui Yang, Qiang Guan, Xiaojuan Li
OnTrack introduces a real‑time, streaming monitoring system for large language model agents that compares each step and its dependencies against recorded successful runs, enabling rapid alerts or blocking in roughly a millisecond per step. The method is evaluated across three data‑access regimes—from full historical data to only step logs—and demonstrates improved failure detection over content‑similarity baselines, achieving a 0.057 AUROC gain. In practical tests on SWE‑bench trajectories, OnTrack’s abort policy prevented 18% of wasted compute, correctly interrupting 83% of runs that would have failed.
By Babak Barazandeh, Connor Swanson, Chinmay Kulkarni, Nikhil Mungel