Microsoft Research introduces PDD Kit, a tool that formally verifies code generated by AI agents for full‑stack web applications. By translating developer intent into Verus specifications, PDD Kit proves that the AI‑written code satisfies these specifications for all inputs, supplementing proofs with tests when necessary. This approach allows developers to review high‑level intent rather than inspecting every line of code.
By Alex Lavaee, Matthai Philipose
Le Chonk provides advanced model control for Mistral Studio, enabling users to build and deploy enterprise agents, workflows, and applications more quickly. The integration offers full control over models, data, and infrastructure, streamlining the development process for businesses.
OpenAI introduces GPT‑6 Cyber, a new cybersecurity model designed to identify exploitable vulnerabilities and develop corresponding fixes. The model is specifically trained to operate within authorized scopes, ensuring it remains aligned with user permissions and security protocols.
Learning POMDP World Models from Observations with Language-Model Priors proposes Pinductor, a method that uses a large language model to generate and refine candidate POMDP models from limited observation-action trajectories. Pinductor achieves performance comparable to privileged LLM-based methods while being up to 330 times more sample‑efficient than standard baselines like DreamerV3, and its effectiveness scales with LLM capability. The approach demonstrates that language‑model priors can substantially reduce the interaction cost required for learning internal world models under partial observability.
By Valentin Six, Frederik Panse, Mathis Fajeau, Lancelot Da Costa, Mridul Sharma, Alfonso Amayuelas, Tim Z. Xiao, David Hyland, Philipp Hennig, Bernhard Sch\"olkopf
The paper investigates how a reviewer in multi‑agent language‑model systems can be authorized to override an upstream agent’s decision. It compares unrestricted override authority with an evidence‑gated approach that requires a valid record of a binding‑constraint violation and a compliant repair before replacement is allowed. Experiments on two open‑weight models across 96 spatial‑policy tasks show that the evidence gate sharply reduces corruption of correct proposals while largely preserving legitimate repairs, though the effectiveness varies with reviewer transitions and upstream states.
By Fuad Ali
The paper introduces ReScopeBench, a benchmark designed to evaluate how agents manage the lifecycle of experience in evolving environments. It differentiates between structural experience—knowledge about how an environment is organized—and rule experience—principles that guide decisions across tasks. Using an experience-first construction approach, the study tracks how agents acquire, reuse, and revise experience across task streams, revealing that while retained experience boosts performance, agents often fail to keep it current when environments change.
By Shijue Huang, Hangyu Guo, Hengheng Zhang, Guo An, Zhitao He, Qi Tian, Yiren Feng
The paper introduces Assurance-Conditioned Continuous Authorization (ACCA), a framework that maintains operational authority for autonomous AI agents by separating evidence validity from contextual applicability. ACCA derives assurance claims from eligible evidence, resolves a multidimensional effective-authority envelope, and uses a typed dependency graph for selective re‑evaluation of authorization‑relevant changes. The authors demonstrate the approach with controlled scenario tests in a trading domain, showing implementation‑level evidence for the method’s effectiveness.
By Mehdi Shajari
The paper proposes a new reporting standard for attributing agent failures to their history, insisting on intervention‑verified evidence rather than post‑hoc taxonomies. It outlines a method to freeze an executable mid‑trajectory state, alter only the agent’s perceived history, and continue execution to obtain a measured replay certificate. The authors review existing manipulation classes and highlight that while several methods perform interventions, none currently report a quantified replay certificate, and they provide two cautionary examples illustrating the pitfalls of lacking such verification.
By Egor Pakhomov, Erik Nijkamp
The paper proposes a method for combining expert knowledge using mixtures of feedback fuzzy cognitive maps (FCMs) guided by a relevance criterion adapted from legal evidence theory. An answer is deemed relevant only if it is both probative and material to the query, and this relevance is used to weight and mix expert FCMs into a query‑specific knowledge network. The authors introduce a Bayes‑like de‑mixing theorem to estimate posterior‑like FCMs and demonstrate the approach on synthetic FCMs derived from a 13‑node expert model of ischemic‑stroke risk.
By Olaoluwa Adigun, Akash Kumar Panda, Bart Kosko
PersonaCodeBench is a new benchmark that evaluates coding agents in multi‑turn, persona‑driven sessions, turning user characteristics into verifiable obligations. The benchmark’s persona pool covers roles, technical abilities, interaction preferences, and practical constraints, generating task‑grounded, evolving requirements for implementation, permissions, testing, explanation, and delivery. Evaluation uses a controller to manage user simulator responses, online requirement evaluation, and separate verifiers for native correctness and persona fulfillment, with a joint pass condition requiring adherence to budget and validity conditions.
By Haixin Wang, Jianheng Hou, Hongyu Shen, Wenzhe Fan, Wei-Chieh Huang, Qianyu Zhu, Weichen Li, Taoran Li, Yuhang Yao, Ruohan Wang, Zhejian Peng, Wanting Jiang, Yifan Chen, Shushu Liang, Haoming Li, Defu Cao, Xu Hu, Siyang Wu, Chen Yang, Haodi Hu, Haichao Zhang, Cong Wang, Zhiyuan Yu, Weizhi Zhang, Yifan Wang, Jintao Huang, Yizhou Sun, Yijun Wang, Yuexing Hao, Xiaomin Li
The paper evaluates FlyHash, a locality‑sensitive hashing scheme inspired by the fruit fly’s olfactory circuit, for use in large language model (LLM) agent memory retrieval. It demonstrates that FlyHash maintains a quality advantage over SimHash when bits are matched, and that this advantage persists with real agent‑memory relevance labels. However, FlyHash’s larger, slower codes and the dominance of dense‑vector indexing (HNSW) limit its practical benefits in terms of memory and latency savings.
By Mohan Bheemadolu, Chrisogonas O. Odhiambo
The paper surveys how graphs are increasingly used in large language model systems to support reasoning, knowledge grounding, retrieval, memory, execution, and multi‑agent coordination. It introduces an operation‑level taxonomy that distinguishes the function of each graph operation, the component that computes it, and how its output is subsequently used. The authors compare graph observations with external computations, analyze memory updates, and examine execution dependencies, then propose an evaluation framework that organizes benchmarks, diagnostics, and system costs, concluding with open challenges in adaptive computation, memory maintenance, and dependency‑aware recovery.
By Haihui Liu, Yang Chen, Zice Qiu, Yun Zhang, Jialun Zhong, Wenhua Wang, Yanzeng Li, Tian Wang
ViBench 2.0 is a new benchmark for long‑horizon software building, featuring 17 full‑stack web products that evolve through 241 feature requests. The benchmark evaluates models only when the entire product is finished, using 544 test plans per product to ensure a strict grading protocol. Initial results show a reliability gap: top models pass most test plans but rarely deliver fully working products, though recent releases have improved performance significantly.
By Daniel Furman, Preeya Kirani, Ibrahim Sheikh, Zhen Li, Peter Zhong, Michele Catasta
The paper introduces a formal definition of catastrophic performance in AI, distinguishing it from merely suboptimal outcomes. It uses two reference values: a baseline value representing the state if no agent is deployed, and an orthogonal value representing a trajectory unrelated to human preferences, typically much worse than the baseline. Catastrophic performance is defined as an AI agent undoing benefits previously achieved by human activity, thereby causing significant harm.
By Henrik Marklund, Alex Infanger, Benjamin Van Roy
HarnessFlow is a framework that automates the optimization of executable harnesses for large language model agents using a Monte Carlo Tree Search policy to balance exploration and exploitation. It refines harness candidates through a coding agent that leverages archived code, execution feedback, and reusable mechanisms, operating within a limited evaluation budget. Across six benchmarks covering agentic tasks, coding, and web interaction, HarnessFlow achieves the highest mean test scores, improving average performance by 10.89 percentage points over initial harnesses and 3.95 over Meta-Harness, while also reducing per-task execution costs.
By Zhi Hong, Qianlong Wang, Hongchen Wei, Aidan Wu, Daohong Ying, Mingze Kong, Bei Liu, Chong Luo
LegalEvolver is a query‑optimization framework for legal judgment prediction that generates legally informative queries to improve retrieval quality. It uses heuristic query seeding, summarizes rollout experiences into rubrics, and applies rubric‑guided reinforcement learning for iterative self‑evolution. Experiments on large benchmarks show that evolving rubrics outperform fixed ones and that rubric guidance boosts performance beyond direct supervised fine‑tuning plus RL.
By Xinyu Yang, Jiejun Tan, Chenlong Deng, Chenghao Zhang, Tongyu Wen, Zhicheng Dou
MA-Mesh is a framework that adds a role‑based self‑healing layer to multi‑agent collaboration patterns. It generates resilience‑aware system specifications and synthesizes them into executable policy code. The framework introduces cross‑agent reflection and role‑based failover to enable agents to evaluate peer performance and dynamically reassign roles, improving task performance and robustness in uncertain environments.
By JunYong Lee, Woo Kyung Kim, Honguk Woo
SUMONE is a lightweight summarizer designed to compress long interaction histories of language‑model agents. It is trained via masked reconstruction, where a frozen reconstructor predicts randomly masked parts of the original trajectory; reconstruction quality plus a length penalty yields a task‑agnostic reward for optimizing the summarizer with GRPO. At inference, only the summarizer remains, enabling plug‑and‑play context compression that reduces inference cost by roughly 78 % on DeepSearchQA and 75 % on SWE‑bench without affecting the underlying agent.
By Jizhe Zhang, Ang Li, Chenheng Zhang, Yisen Wang
ServerBench is a benchmark that lets large language model (LLM) coding agents design and test specialized LLM serving systems. The benchmark evaluates agents on a single objective—maximizing latency‑qualified token value per dollar of GPU cost—using stateful replay to preserve session dependencies and independent verification to confirm system correctness. In experiments with 11 coding‑agent models, Claude Sonnet 5.5 achieved 2.97 dollars per dollar on Gemma E4B, outperforming the vLLM reference by 4.87×, and the study highlights key design insights such as coordinating capacity with load selection, interpreting performance feedback through relevant experiments, and verifying execution paths for reliable delivery.
By Xinrui Zhong, Zhiming Wang, Yongqi Liu, Zhaohui Wang, Shanglin Yang, Zikun Li, Zhuoming Chen, Beidi Chen
Jev-AD combines a typed decision model, Jev, with a language world model to predict how nearby vehicles will move over the next two seconds for each candidate action. By comparing these predictions to the decision model’s outputs, Jev-AD reduces collision rates dramatically in both simulated highway merging (from 26.0% to 2.4%) and real‑world NGSIM episodes (from 48.2% to 2.8%). The system also achieves passenger intent in 93% of free‑form command trials and foresees 93% of dangerous actions in real traffic.
By Yansong Qu, Jiancong Chen, Samuel Labi