CONTRA is a training‑free method that discovers and qualifies behavior‑changing questions for selective clarification in large language model (LLM) code generation. It first generates candidate questions, filters out those unrelated to required behavior or already resolved, then creates programs conditioned on two plausible answers to check for stable behavioral differences on shared inputs. Experiments on ClarifyCodeBench show that CONTRA achieves the highest F1 across four coding agents, outperforming baselines by 13.88 percentage points, and it is also implemented as a Claude Code plugin for practical use.
By Zheng Fang, Yongmin Li, Yichang Zhang, Dongming Jin, Haoyu Wang, Shuai Wang, Zhi Jin, Ge Li
The paper investigates how a long-running assistant should summarize its memory when it must serve multiple user goals simultaneously. Three summarization policies are compared: a neutral summary with no goal in mind, a single summary covering all goals, and separate per‑goal summaries. Experiments show that per‑goal summaries outperform the others in relevance, completeness, and accuracy, while the all‑goal summary performs poorly even compared to a neutral summary with a smaller budget.
By Albert Sadowski, Jaros{\l}aw A. Chudziak
Wayfarer is a domain‑agnostic, online deep RL agent that discovers options via Laplacian representation learning from high‑dimensional observations and uses them for control. The discovered options improve exploration, accelerate credit assignment, and generalise to unseen settings, leading to faster learning of complex policies. Wayfarer achieves state‑of‑the‑art performance among single‑stream agents on the most challenging Atari 2600 games, especially those requiring long‑horizon exploration such as Montezuma's Revenge and Private Eye.
By Erik M. Lintunen, Marlos C. Machado
The paper introduces Generative Test-Driven Development (GTDD), a method where a separate testing agent continuously generates new inputs after each coding agent’s implementation, guided by a human-specified behavioral contract and prior feedback. A trusted evaluator validates these inputs, provides reduced counterexamples, and stores them for regression testing, ensuring that development repeatedly confronts failures beyond the initial examples. Experimental results on a stateful key-value store show that policies regenerating tests during development achieve lower mean failure rates than a single-generation policy, while providing the tester with the candidate’s source code does not yield additional improvement.
By Masahiro Kato
SafeStreamingFlow is a goal‑conditioned planner that aligns flow sampling dynamics with execution dynamics by sequentially integrating a learned state vector field with hierarchical state prediction. It enforces safety constraints only for the executed step using high‑order control barrier functions, reducing planning latency and improving safety compared to prior safe diffusion/flow planners. The method demonstrates competitive goal‑reaching success across navigation, racing, and locomotion benchmarks.
By Seunghwan Jang, Jeongyong Yang, Siddharth Ancha, SooJean Han
LEAP: Learning Efficient Action Proposals For LLM Agents proposes a method to speed up large language model agents by training a small 0.6B drafter to predict target actions accurately. The approach uses a latency framework that balances drafting, verification, and execution costs, achieving up to 60% faster end‑to‑end wall clock time without reducing task success. LEAP can be online trained, eliminating the need for prior trace collection and making it practical for real‑world deployment.
By Zhen Xu, Qizheng Zhang, Gerry Wan, Shang Zhu, Ce Zhang
The paper introduces AgentSecGraph, a static analysis framework that builds a Security-Aware Agent Dependency Graph (Security-ADG) for each security-sensitive operation in large language model (LLM) agents. It enriches operation identity with agent relevance, source and dependency evidence, trust-boundary context, guard evidence, and external-effect semantics. The authors also present AgentSecBench, a corpus of 67 real-world LLM-agent repositories, and demonstrate that their analyzer identifies thousands of operation candidates, recovers substantial dependency and guard evidence, and distinguishes guarded behaviors from vulnerabilities with high accuracy.
By Hang Cui
The paper introduces Turnover-Orthogonal Credit Assignment (TOCA), a method for open-team multi-agent reinforcement learning that separates the effects of agent actions, pure turnover, and action–turnover interactions in the team return. TOCA uses a permutation‑invariant centralized critic with event tokens to provide counterfactual per‑agent credit signals and a softly weighted interaction variant, TOCA‑β, for high‑variance environments. Experiments show that TOCA outperforms event‑aware MAPPO critics and that removing interaction credit degrades performance, especially in high‑turnover settings such as the Dynamic Spread benchmark.
By Amit Thakur, Mukesh Singhal
The paper introduces cordial learning, a distributed training framework for agents with correlated data, where each agent’s label depends on inputs from other agents. By sharing only low‑dimensional outputs, agents train local models that extract informative signals from peers, forming a game‑theoretic loss that depends on others’ models. The authors prove convergence to a globally optimal solution for linear models and show empirical effectiveness on structured multi‑digit MNIST tasks, even in nonlinear settings.
By Sarah Shitrit, Ilai Bistritz
The paper introduces threat‑preserving representation sensitivity (TPRS) to assess how changes in agent‑visible wording affect attack success rates (ASR) while keeping the underlying task and evaluation criteria constant. Experiments on Agent Security Bench, MCPTox, and AgentDojo show that renaming threat‑related tools can significantly alter ASR—up to +13.21 percentage points on some models—yet may also degrade benign utility. The findings suggest that security scores based on a single representation may not generalize, urging robustness claims to be validated across multiple threat‑preserving representations.
By Neeraj Karamchandani, Piyush Nagasubramaniam, Xinhong Xie, Sencun Zhu, Dinghao Wu
The paper investigates whether adding a short classification instruction after a user’s prompt improves the ability of activation probes to detect malicious inputs in large language models. Across 13 safety benchmarks and three open‑weight model families, a classification suffix consistently boosts out‑of‑distribution detection (up to ~4 AUC points) compared to no suffix, and the benefit transfers to multi‑position pooling probes used in production. The improvement stems from the classification format itself rather than the specific content of the instruction, though the optimal suffix varies with the model and readout type.
By Elad David, Max Fomin
NeutronGym is the first executable environment that lets language‑model agents design neutron instruments, using tools that validate their builds, McStas ray‑tracing for simulation, and a level‑resolved grading ladder that evaluates syntax, runtime, structure, and science without an LLM judge. The platform provides procedural families of instrument layouts with held‑out parameter regimes and a curated set of 16 tasks from published instruments, called McStasBench, which includes memorization probes and a sandbox. Experiments show that while seven models can reproduce at most seven of the 16 tasks and none reaches a reference design, reinforcement learning can dramatically improve performance—e.g., Qwen3‑8B’s success rate jumps from 11% to 77% on held‑out instances—highlighting the ladder’s importance for partial credit and the potential of RL to match classical optimizers under realistic simulation budgets.
By Lijie Ding, Changwoo Do
ReFract is a new benchmark that tests whether language model agents can act appropriately based on a user’s role, a capability called Perspective Awareness. It contains 150 expert‑validated entries derived from anonymized domain support queries, and uses Text World Models to simulate the agents’ operating environments and generate perspective‑aware action trajectories. Current state‑of‑the‑art LLMs solve at most 69% of the tasks, with over half of their trajectories attempting perspective‑violating actions.
By Hainiu Xu, V\'{i}tor N. Louren\c{c}o, Mohnish Dubey, Yunfei Bai, Yulan He, Caroline Catmur, Aline Paes, Marco Caserta, Akash Chandrayan, Luca D'Angelo
The paper introduces an Embodied Neurocomputation framework that tackles the challenge of encoding and decoding between silicon computers and biological neural networks (BNNs). Using a large‑scale parameter sweep, the authors optimized over 1,300 encoding configurations for a BNN agent navigating an odor‑style gradient in a simulated grid‑world, running more than 4,000 hours of real‑time interactions. Twelve configurations consistently outperformed silicon‑based DQN agents within the same interaction budget, demonstrating the framework’s potential for scalable, goal‑oriented learning with BNNs.
By Johnson Zhou, Daniel Tanneberg, Forough Habibollahi, Alon Loeffler, Kiaran Lawson, Valentina Baccetti, Kwaku Dad Abu-Bonsrah, Candice Desouza, Finn Doensen, Bradley Watmuff, Daria Kornienko, Azin Azadi, Justin Leigh Bourke, Bernhard Sendhoff, Brett J. Kagan
The paper investigates whether a proximal policy optimisation (PPO) agent can learn to act as a broker in a continuous‑time broker‑trader game that has an analytically solved optimal strategy. By discretising the continuous‑time reward and validating the implementation, the authors show that a PPO agent with a feed‑forward neural network can approximate the reference action when there is no uninformed order flow, but struggles under stochastic uninformed flow, with critics failing to rank actions reliably. The study further demonstrates that a causal certainty‑equivalent controller based on observable history outperforms PPO under partial information, and that freezing the analytical policy and fine‑tuning PPO after a change in execution cost yields a measurable improvement.
whyItMatters:"The analytical solution serves both as a diagnostic benchmark for RL performance and as a practical starting policy that can be adapted to changing market conditions, illustrating how theoretical finance models can guide and improve reinforcement learning in complex financial control tasks."
By Siu Tung Wong (Institute of Finance and Technology, University College London), Carlo Campajola (Institute of Finance and Technology, University College London, UZH Blockchain Center)
FASTDIAR is a frame‑level speaker encoder designed for real‑time streaming diarization on a single CPU thread. It converts a state‑of‑the‑art speaker recognition architecture into a causal model that emits an embedding every 80 ms from a two‑second window, and pairs it with online clustering that updates only when the stream’s self‑similarity changes. Trained via distillation from an utterance‑level teacher, it achieves the highest accuracy on low‑overlap benchmarks, degrades less with more speakers, and runs five times faster than real time.
By Nikita Torgashov, Okan K\"op\"ukl\"u
HyperLogic is a forward‑construction benchmark for Chinese logical reasoning that decouples problem authoring from answer generation. It employs a multi‑agent workflow to harden undergraduate‑authored seeds, translate items into executable finite‑domain models, and adjudicate solver‑derived answers under human oversight. The benchmark includes a Base set of 195 items (922 sub‑questions) and a Hard set of 100 items, revealing significant performance gaps among frontier models and demonstrating the challenges of faithful formalization even with tool support.
By Ming Zhang, Qiyuan Peng, Yinxi Wei, Yujiong Shen, Kexin Tan, Yuhui Wang, Zhenghao Xiang, Junjie Ye, Zhangyue Yin, Zhiheng Xi, Shihan Dou, Weikang Wang, Yuhao Zhang, Tao Gui, Ruizhi Yang, Qi Zhang, Xuanjing Huang, Alex Chen, Maxm Pan
SEDIMA is a persistent hierarchical insight memory designed for evolutionary search agents that use large language models. It transforms raw search traces into natural‑language insights, clusters them by semantic similarity, and retrieves relevant guidance to inform future mutations, thereby accumulating transferable knowledge across runs and problems. When added as a drop‑in module, SEDIMA improves average final performance by 5.5% on AlgoTune and 6.6% on ALE‑Bench LITE, and reduces the number of iterations needed to reach baseline‑best performance by 32.3% on OpenEvolve.
By Amirhossein Abaskohi, Mahdi Mostajabdaveh, Zirui Zhou
The paper introduces Calibrated User Embeddings (CUE), a framework that encodes real user sessions into continuous representations and decodes them into persona commands to steer large language models (LLMs) as user simulators without additional training. CUE enables user-conditioned replay of past interactions and generates novel personas that better align with real-user success rates and failure patterns. Evaluations on the τ²‑Bench dataset show that CUE‑driven simulators commit fewer simulator‑attributed errors, more accurately reproduce real‑user failure modes, and maintain competitive user fidelity across various tasks and LLMs.
By Anjali Kantharuban, Jonas Mueller
The paper introduces a two‑stage framework for auditing false alarms in safety monitors for language‑model agents, treating the problem as a positive‑unlabeled (PU) ranking task. It adapts safe references toward the alarm domain, consolidates ordering from multiple PU models, and refines the ranking with structural guidance, all without requiring alarm safety labels. The method improves macro AUPRC to 0.6444 and recovers 33.3% more false alarms at a 5% review budget compared to leading PU baselines.
By Xichen Yan, Chongyang Gao, Kezhen Chen, Guangyi Zhang, Jiaqi Wu, Lixu Wang