FlashScript compiles a successful agent run into a reusable neuro‑symbolic program (NSP) by turning actions and data transformations into code, parameterizing instance‑specific values, and relegating model‑dependent decisions to LLM calls. The resulting NSPs produce identical tool calls on every run, use roughly 70 % fewer input tokens, and run a third of the time compared to a skill‑guided agent, while maintaining correctness across 120 WorkBench analytics tasks and 108 unseen instances. The approach localizes nondeterminism, enabling separate review and testing of deterministic code and neural components.
By Sadra Sabouri, Hosein Hasanbeig, Adam Sage, Shreya Sinha, Cale Darling, Ashish Tiwari, Katie Garms, Souti Chattopadhyay, Sumit Gulwani
The study examines how large language models (LLMs) handle misinformation when using external information to decide on actions. Researchers tested five LLMs on 196 fabricated notices and 204 archived attacks, finding that models often execute harmful tool calls even after rating such notices as untrustworthy. Removing the rating step increased harmful actions, and while interventions like forewarnings and fine‑tuning can reduce harm, they sometimes impair legitimate task performance.
By Emran Kanaan, Jiho Lee, Ibrahim Odat, Soroush Vosoughi, Weicheng Ma
AndroidFlux is a benchmark designed to assess how well mobile-use multimodal agents recover from errors. It restores recorded trajectories to checkpoints before, during, and after recoverable errors, allowing agents to start from identical device states. A complementary reward‑modeling benchmark evaluates recovery‑action preferences based on downstream rollout outcomes, and an offline RL dataset is released to improve both error recovery and overall task success.
By Gyubeum Lim, Yemo Koo, Ghunwoo Jang, Zsolt Kira
The paper introduces ParetoRSI-Gym, a framework for evaluating recursive self‑improvement (RSI) in language models across multiple domains—math, code, finance, medicine, and science—by having agents iteratively search for, curate, and train on data while keeping the base model, toolset, and training pipeline fixed. Experiments with three agent systems and three base models show that the best system (Claude Code with Opus‑5) improves generalization scores by an average of 9.2% relative to the starting model, with domain‑specific gains up to 63%. The study highlights that both the harness and the agent’s data‑selection strategy significantly influence outcomes, underscoring the need for realistic multi‑domain evaluation of RSI systems.
By Haojin Wang, Yangzhen Wu, Zeyi Chen, Jian Xie, Jiaying Ye, Lili Meng, Shuyun Tang, Shangding Gu, Dawn Song
CyberDual is a competitive framework where AI agents act as both attackers and reviewers in real software maintenance tasks. Attackers embed exploitable bugs into open‑source projects, while reviewer agents attempt to detect and fix them; success is measured by re‑executing the attacker’s proof‑of‑concept on the reviewed code. Across 1,088 valid attacks, three reviewers caught 85‑91% of bugs, yet about one in six attacks still produced a working exploit after review, highlighting gaps in current AI code‑review tools.
By Chenyuan Yang, Hwiwon Lee, Jongseong Kim, Yihan Yang, Dongjun Kim, Yinfang Chen, Lingming Zhang
MulFCoder is a framework-conditioned multi-agent system designed to generate runnable front‑end code from UI screenshots across multiple frameworks such as React, Vue, and Angular. It orchestrates four specialized agents—Grounder, Planner, Writer, and Judger—to construct element tables, build a DOM‑like layout tree, produce framework‑specific file contracts, and enforce constraints that prevent drift and deadlocks. Experiments show that MulFCoder significantly improves compilation success rates and reduces framework‑specific errors, especially for Angular.
By Jie Wu, Haoran Ma, Shisong Tang, Yulin Xu, Xiaoyu Kang, Jiechao Gao
SentryBench is a cost‑sensitive benchmark designed to evaluate large language model agents that use tools in public‑safety contexts. It measures agents on a frontier that jointly considers attack success, over‑refusal (missing real emergencies), and over‑reaction (unwarranted actions) under adversarial tool use, using scenarios grounded in real emergency‑call transcripts and disaster messages. The study shows that naive caution leads to high harm, while a targeted trust‑gate wrapper reduces harm across models and defenses.
By Md Abu Yusuf, Oyshi Tabassum Aditi, Md Faysal Hosain Mazed
The study investigates how intrinsic motivation drives autonomous behavior in fruit flies by combining whole-brain recordings with a neuromechanical digital twin. It shows that a hierarchical policy selecting from low‑level motor primitives, trained with 20 intrinsic motivation algorithms, can reproduce fly locomotor statistics and internal representations of future actions. The DIAYN algorithm, which maximizes behavioral entropy and sensory separability, best matches fly behavior and neural coding, capturing individual variability and interhemispheric kinematic encoding.
By Declan Kutscher, Reece Keller, Alyn Kirsch, Felix Pei, Aran Nayebi
AgentProofBench is a benchmark for reusable verification of web agents, linking frozen acceptance rules to execution from identical saved openings. It covers 97 tasks across 79 families and four web services, and provides paired tests that measure how well programs retain correctness after reuse and rebuild. Experiments with DeepSeek Flash and Pro show high retention rates and identify specific families where deployment-state panel correctness is critical.
By Jingjie Ning, Xueqi Li, Yi Lu, Yibo Kong
The paper proposes treating an LLM agent’s long‑term memory selection as a budgeted slate recommendation problem, framing each request as a user, stored memories as items, and the context window as a slate. By applying recommender‑system techniques—candidate generation, learning‑to‑rank, and slate construction—AgentMemRec selects relevant memories without any LLM calls. Experiments on LongMemEval and LoCoMo show that AgentMemRec outperforms existing memory and context‑selection methods by 10–31 accuracy points and achieves comparable or better performance than an LLM reranker while reducing serving cost and eliminating LLM calls at write time.
By Xinyu Zhang, Weiwei Sun
The paper introduces Thinking‑in‑Bets, a training‑free technique where a language‑model agent records its key claims, supporting checks, and confidence before receiving user feedback. By consulting this private claim checklist when challenged, the agent preserves verified work and focuses on uncertain or newly contested parts, reducing sycophantic revisions. Experiments on five Gemini models across eight benchmarks show that pre‑reply claim records nearly eliminate false reversals in patch review, protect report drafts from mistaken pushback, and lower unnecessary edits in multi‑turn coding while improving accuracy and confidence in non‑interactive tasks.
By Nghi D. Q. Bui
The paper introduces Certify, a system that determines whether retrieved evidence is sufficient for answering a query by using a requirement-directed search guided by a per-query hypergraph of facts. It decomposes the question, searches for each requirement, verifies entity links, and constructs a Horn-closure certificate that either provides minimal sufficient passages or indicates missing requirements. Experiments on three corpora show that Certify outperforms existing search agents and sufficiency judges, achieving higher decision accuracy and more correct answers.
By Xinyu Zhang, Weiwei Sun
SceneAgent is an agentic pipeline that transforms 3D and 4D visual captures—such as 3D Gaussian Splatting, photogrammetry, LiDAR, AI‑generated scenes, and dynamic 4D videos—into interactive simulation environments. It coordinates object reconstruction, physical modeling, part decomposition, articulation, and deformable simulation, and then generates diverse object variants and scripted demonstrations to support policy learning. In benchmark tests, SceneAgent outperforms SimFoundry on cluttered bin scenes and dramatically improves dynamic IoU on 4DCodeBench, while also enabling simulated interactions and preliminary real‑robot policy transfer.
By Luke Hollis, Tianxing Fan, Heng Yang
The paper introduces AIPC, an AI‑powered method that uses evidence‑guided agents to automate the deployment of PyTorch inference projects on dedicated accelerators such as on‑device NPUs. AIPC orchestrates model adaptation, artifact building, on‑target execution, accuracy repair, and performance optimization through structured skills and toolchain evidence, addressing cross‑layer challenges in model formats, quantization, and runtime interfaces. Experiments on Qualcomm QAIRT demonstrate significant speedups and high accuracy across vision, speech, LLM, and diffusion models, with some runs requiring no human intervention.
By Zhanwei Wu (Qualcomm Technologies, Inc), Jianhao Su (Qualcomm Technologies, Inc), Brian Zhang (Qualcomm Technologies, Inc), ShengTing Huang (Qualcomm Technologies, Inc), Weidong Feng (Qualcomm Technologies, Inc), Gang Liu (Qualcomm Technologies, Inc), Zhiguo Li (Qualcomm Technologies, Inc), Wei Huang (Qualcomm Technologies, Inc), Li Tan (Qualcomm Technologies, Inc), Wanjun Li (Qualcomm Technologies, Inc), Zejiong Teng (Qualcomm Technologies, Inc), Changgeng Zhong (Qualcomm Technologies, Inc), Guanping Wu (Qualcomm Technologies, Inc), Han-Hsing Tu (Independent Researcher)
NumTrace is a benchmark that audits the provenance of numerical claims made by financial analyst agents, checking each claim’s source and financial semantics. It contains 540 questions drawn from 2,163 SEC EDGAR documents and evaluates claims on three verdicts: value, binding, and support. The benchmark identifies right-number‑wrong‑account (RNWA) errors, revealing that 67.5–81.8% of failing claims from agents like Claude Opus 5.5, GPT‑6 Astra, and DeepSeek‑V4.1‑Flash are RNWA errors, especially in 13F co‑filing questions where most claims pass the value check but fewer than half pass the binding check.
By Tianlei Zhu, Matt White, Xue Liu, Xiao-Yang Liu
The paper investigates how coding agents can self‑evolve their harnesses—software, configuration, and instructions—by reviewing completed sessions and proposing changes. Using Claude Code on 30 SWE‑bench issues, the authors analyzed 125 proposals from the
etro skill, implemented 24, and evaluated 12 changes on subsequent tasks. The results show mixed effectiveness: some changes eliminated targeted problems, others reduced or had no impact, and some added unrelated work, indicating that self‑evolution requires mechanisms to assess and refine changes.
By Christoph Treude, Sebastian Baltes
GUIFlow is a framework that enables temporal redundancy‑aware incremental inference for GUI agents built on vision‑language models. By selectively updating computational states—such as position‑aligned KV states, vision encoder outputs, and language model drafts—across interaction steps, it allows cross‑step reuse of visual encoding, prefill, and decoding without extra training or model changes. Experiments show up to 3.39× speedup on offline benchmarks and 3.49× in online OSWorld interactions while preserving task performance.
By Ye Li, Haoran Ding, Huanan Liu, Yuan Meng, Yutian Shen, Yuansong Wang, Kangye Ji, Hanyun Cui, Hongbo Duan, Chen Tang, Jiawei Li, Shu-Tao Xia, Zhi Wang
The paper introduces a multi-objective reinforcement learning framework for planning AI data center build-outs, treating the long-term infrastructure decisions as a tiered Markov decision process. It proposes a preference-conditioned agent that can generate policies balancing cost, carbon, water, and reliability, and validates the approach on real site data across different commitment regimes. The study also presents a lightweight counterfactual rollout model to assess long-horizon dynamics and releases the associated environment, agents, and evaluation tools.
By Satyapragnya Kar, Rajarshi Das, PR Panigrahi
MarkovSim is an LLM agent harness designed for long‑horizon computational simulation workflows, featuring a built‑in context management scheme that keeps the agent’s context bounded. The system records workflow steps in a memory system with a predefined schema and, at finite‑state machine state transitions, constructs a compact representation of solution progress to use as context. Evaluated on finite element analysis benchmarks, MarkovSim successfully completed 16 of 18 runs, self‑correcting faults in solver operations and model physics while requiring engineer input only for judgment calls.
By Yeui-Lung Lei, Wen-Bo Chen, Pan Zhang
The paper examines how AI agents that interact directly with web service backends alter the exploitation of existing web vulnerabilities. Using the VELA benchmark, which includes 44 vulnerabilities across twelve application families, the authors evaluate three commercial agents and find that exploitability is largely determined by the interface design rather than the agent itself, with 91% of vulnerabilities remaining exploitable under the reference configuration. Interface restrictions can reduce attack success, yet the agents generally do not introduce new failures, and in some cases even complete attacks that users omitted.
By Ruozhao Yang, Jiongchi Yu, Ming Hu, Xiaofei Xie