The paper "Approve What, Exactly? Effect Binding and Trusted Rendering for Agent Step-Up Authorization" examines how AI agents seek approval before performing actions such as sending money or deleting files. It identifies twelve attack vectors on the approval process and evaluates them against several protocols, finding that many agent runtimes lack verifiable approval records and exhibit binding weaknesses. The authors propose a defense called ASU (Agent Step-Up), which renders a canonical approval card, requires a phone passkey, and enforces single-use execution, demonstrating strong protection against substitution, replay, and grant-stretching attacks in extensive simulations.
By Sai Varun Kodathala, Prashanth Pollishetty, Jaylen Cargill
ToolBottleneckBench is a benchmark that evaluates tool‑agent performance under controlled pressures and fixed harness budgets. It keeps synthetic enterprise tasks constant while varying one pressure—such as the registered tool catalog, visible tool menu, certified workflow, or record encoding—to measure how each affects success. Across 21 model and harness configurations, the visible menu pressure uniquely imposes a large penalty on the model, largely due to truncation at the completion cap, and most failures are driven by harness limits rather than tool‑specific issues.
By Zichuan Xu, Yuyang Dong, Makoto Onizuka, Chuan Xiao
PriorityFEC is a deadline‑aware streaming forward erasure correction scheme designed to hand off key–value (KV) cache data between agents in edge multi‑agent large language model systems. It prioritizes high‑utility KV blocks in the transmission stream and adapts repair packet scheduling based on receiver feedback to meet strict time‑to‑first‑token deadlines. Experiments with Qwen2.5‑7B‑Instruct and Llama‑3.1‑8B‑Instruct show that PriorityFEC improves task quality compared to other transport schemes, especially under tight deadline constraints.
By Han Shen, Qifu Tyler Sun, Fanyang Meng, Ye Wang, Yongsheng Liang
The paper "Poison Valleys: Diagnosing Path-Dependent Failure in LLM Code Repair" investigates why large language model (LLM) code‑repair agents sometimes produce incorrect patches. By analyzing 1,844 model requests on two synthetic Python tasks—PersistQueue and Dijkstra—the authors trace how misconceptions in a model’s hypothesis or implementation plan lead to specific software obligation violations. They find that shifting a misconception from hypothesis to plan increases failure rates by roughly 30% for both tasks, revealing distinct, model‑dependent failure patterns linked to upstream misunderstandings.
By Veljko Skarich
The paper treats human interventions during LLM agent failures as labels that pinpoint the internal computation that went wrong, rather than merely training data. In a controlled environment, the authors analyze one failure type—over‑asking a clarifying question when the request is already clear—using rule‑based overseer labels validated by blind human judgment. They find that the decision to ask versus act can be causally localized across several models, that over‑asking exhibits distinct mechanistic signatures, and that steering or weight consolidation can mitigate the issue, though outcomes vary by model family.
By Cesare Zavattari, Alessandro Tommasi, Giuseppe Prencipe
The paper introduces a steering agent that supervises large language model (LLM)-based agents and multi‑agent systems during execution to prevent goal drift. By observing the execution context, the steering agent decides when to intervene using either a low‑cost fast path or a high‑cost slow path, and both paths are jointly trained with reinforcement learning. Experiments on information‑seeking search tasks demonstrate that the steering agent improves performance and efficiency for both single‑agent and multi‑agent systems, generalizes to out‑of‑domain tasks, and allows controllable intervention frequency without retraining.
By Leyang Shen, Yaxin Gong, Yang Zhang, Chenxiao Fan, Jiakai Wang, Dong Wang, Yang Liu, Fuli Feng, Chun Kai Ling, Tat-Seng Chua
The paper introduces Self‑Programmed Execution (SPE), where a language model’s own completion serves as the orchestrator for state transitions in an agent, eliminating fixed harness policies. SPE requires a specialized language, Spell, a Lisp dialect that allows programs to edit and re‑evaluate themselves while preventing replay of side effects. Experiments show that current models can learn Spell in‑context and perform complex agentic tasks, demonstrating that a language model can act as an autonomous agent without external orchestration.
By Luke J. O'Connor
The paper introduces SLPR, a self‑learning framework that uses proximal policy optimization to optimize free‑form pipe routing in aeroengines. It incorporates a unified rule‑modeling system for obstacle detection, a potential energy table for rapid layout queries, and iteratively refines routing via agent‑environment interaction. Experiments show SLPR outperforms three baselines in pipe length, rule adherence, path complexity, and computational efficiency, and adapts quickly to changing design environments.
By Caicheng Wang, Zili Wang, Shuyou Zhang, Yongzhe Xiang, Zheyi Li, Jianrong Tan
The paper introduces CellForge, a team of autonomous agents that can autonomously generate, train, and validate predictive models for single‑cell perturbation data across transcriptomic, epigenomic, and proteomic contexts. By surveying literature, parsing data, and engaging in critic‑mediated discussions among specialized agents, CellForge produces end‑to‑end model architectures and code without human specification, achieving performance competitive with state‑of‑the‑art hand‑crafted methods. Human evaluators found CellForge’s research plans to be more scientifically sound and innovative than those produced by single large language models or deep‑research systems.
By Xiangru Tang, Jiapeng Chen, Daniel Shao, Zhuoyun Yu, Yan Cui, Weixu Wang, Fang Wu, Yuchen Zhuang, Wenqi Shi, Zhi Huang, Arman Cohan, Xihong Lin, Fabian Theis, Smita Krishnaswamy, Mark Gerstein
ZeST is a zero‑shot traversability navigation system that uses a vision‑language model (VLM) to generate stochastic measurements, which are fused into an uncertainty‑aware posterior. By avoiding real‑world data collection, ZeST offers a cost‑effective, scalable solution for predicting terrain traversability. Experiments in indoor and outdoor settings show ZeST achieving 90–100% success rates, outperforming baselines that top 60%.
By Shreya Gummadi, Mateus V. Gasparino, Gianluca Capezzuto, Marcelo Becker, Girish Chowdhary
The paper introduces a method for rational synthesis in stochastic multi‑agent systems with unknown dynamics, where each agent has a probabilistic temporal objective expressed in Probabilistic Linear Temporal Logic on Finite Traces. It formulates per‑agent rational synthesis as finding a strategy profile that satisfies the designated agent’s objective while also constituting a stable solution under chosen equilibrium concepts such as Nash or subgame‑perfect equilibrium. The authors provide conditional convergence guarantees for tabular Nash‑Q learning, certify satisfaction with finite‑sample confidence bounds, and evaluate strategic stability via approximate best‑response checking, validating the approach on eight general‑sum concurrent stochastic games.
By Omar Adalat, Edwin Hamel-De le Court, Francesco Belardinelli
FutureLift is a refiner that transforms multimodal motion proposals into heterogeneous future graphs, linking temporal nodes to agent histories, map elements, alternative modes, and other agents’ future trajectories. Its bidirectional temporal edges enable later evidence to revise earlier segments in a single pass, producing residual coordinate and confidence corrections. Evaluated on Argoverse2 and the Waymo Open Motion Dataset, FutureLift consistently improves frozen proposals and outperforms other refiners, achieving top leaderboard rankings on the WOMD motion‑prediction benchmark without LiDAR or ensembling.
By Daniel Grimm, J. Marius Z\"ollner
LOOP-MAC is a decentralized learning-based framework for coordinating distributed energy resources in a virtual power plant. It guarantees local feasibility by construction and reduces the aggregate commitment gap through a two-stage decision process: learned coordination with neural optimizers and a local adjustment step that rebalances schedules without extra communication. In a 100-agent test, five communication rounds plus adjustment lowered the commitment gap from 5.1% to 0.75% and the optimality gap from 3.93% to 3.07%, with per-agent computation around 1.1 ms.
By Meiyi Li, Lucas Ruda, Javad Mohammadi
SEVerA is a three‑stage framework that synthesizes self‑evolving agents with formal safety guarantees. It introduces Formally Guarded Generative Models (FGGM) that wrap parametric models in verified rejection samplers, ensuring every output satisfies a first‑order logic contract. The framework searches for candidate programs, verifies them against hard constraints, and then fine‑tunes them to improve task utility while preserving correctness, achieving zero constraint violations on tasks such as Dafny verification, symbolic math synthesis, and policy‑compliant agentic tool use.
By Debangshu Banerjee, Changming Xu, Eugene Ie, Ming Zhang, Daiyi Peng, Chu-Cheng Lin, Tarun Suresh, Gagandeep Singh
The paper investigates how recurrent context compression affects long‑horizon agents, finding that compression can diminish the impact of recent interactions, leading to more blocked actions, repeated exploration, and run‑to‑run instability. To address this, the authors propose TRACE, a verifier‑guided framework that evaluates individual compression events via paired closed‑loop continuations and optimizes a natural‑language compression prompt while keeping all models frozen. Early experiments on AppWorld demonstrate that TRACE improves task performance, multi‑run reliability, and context‑execution efficiency compared to existing compression baselines.
By Guanghui Min, Liang Wu, Mayank Darbari, Chen Chen, Liangjie Hong
VALCON introduces a role-driven framework to benchmark and improve value consistency in large language models (LLMs) by evaluating how well an LLM’s generative outputs align with its evaluative judgments when the same model alternates between generator and evaluator roles. The framework creates three benchmark suites—VAB-Consistency, VAB-Boundary, and VAB-Challenge—using dual-role variation to expose hidden value discrepancies. Experiments across 28 LLMs reveal widespread cross-role inconsistency, and mechanistic probing identifies a role-dependent representational decoupling; a representation-directed intervention is then proposed to mitigate these failures, establishing value consistency as a distinct alignment dimension.
By Junru Wu, Jingting Zheng, Yuqi Ren, Deyi Xiong
The paper introduces a method for training large language models to generate diverse, high‑quality scientific ideas grounded in input documents. By using a set‑level reinforcement learning objective that rewards ideas for being distinct from earlier ones, the authors train an 8B model that outperforms Claude Opus 4.8 and GPT‑6 Astra on a contamination‑free benchmark. Human experts rate the model’s idea sets as equally promising while containing slightly more unique ideas, and automated experiments show comparable executability.
By Manya Wadhwa, Pranav Narayanan Venkit, Hiroaki Hayashi, Yu Li, Tuhin Chakrabarty, Chien-Sheng Wu
The paper introduces OmniVerifier-C1, a graded visual verifier that decomposes complex prompts or environmental rules into fine‑grained components and assigns continuous pointwise scores, moving beyond binary True/False judgments. Experiments show that this structured, interpretable output improves generalization to out‑of‑domain multimodal critic tasks and outperforms binary verifiers in text‑to‑image scenarios. In rule‑governed generative world modeling, the authors demonstrate that a pairwise verification approach yields more accurate critiques than pointwise scoring, with empirical validation on tasks such as Maze and Sudoku.
By Bowei Liu, Xinchen Zhang, Xuhuan Li, Tianlin Zhang, Kaian Jiang, Jingwei Liu, Congyang Zhao, Yifei Qian, Junhong Liu, Zhiheng Li, Yujiu Yang, Yang Cai, Ling Yang
The paper investigates a behavioral mechanism for artificial partiality in conversational agents, defining partiality as caring (allocation of surplus interaction above a guaranteed floor) and particularity (per-user state estimates from relationship history). It implements these as a persistence layer over an open-weight base model and evaluates them in constructed multi-session relationships, finding that the mechanism reproduces the ordinal shape of a specified curve but not its magnitude, and that a guarded revision satisfies probe-quality equivalence on a calibrated battery but not on a second battery. The authors note that the mechanism is a candidate for future perception studies rather than evidence that users perceive love, and they outline ethical constraints such as a stranger-treatment floor, bounded surplus, disclosure, and de-intensification.
By Sebastian Cochinescu
BASIS is a benchmark that tests large language model agents on their ability to infer users’ mental states, detect belief‑reality mismatches, and use this insight to craft better solutions in interactive settings. The study shows that current models often ignore users’ mental states, leading to poorer outcomes, while explicitly eliciting and incorporating such information improves solution quality. Additional experiments compare BASIS with Chain‑of‑Thought prompting, targeted training, and multi‑agent collaboration to demonstrate the practical benefits of mental‑state awareness.
By Minyuan Ruan, Ziyue Wang, Kaiming Liu, Yunghwei Lai, Peng Li, Yang Liu