The paper introduces CORSA, a router‑aware attack framework that optimizes skill injections for both retrieval and execution in AI agents that use modular third‑party skills. It demonstrates that existing evaluations overestimate attack success because they assume the malicious skill is already selected, whereas realistic multi‑skill environments reduce success rates by 87‑97%. CORSA improves retrieval and end‑to‑end attack success while maintaining user utility and shows transferability across different router architectures and LLM backbones.
By Haneen Najjar, Luca Scionis, Haritz Puerto, Sahar Abdelnabi
arXiv:2610. 08339v1 Announce Type: new Abstract: Camera-based 3D perception for autonomous driving relies heavily on large annotated datasets, and deploying such a system to a new target region typically requires data collection and annotation.
By Hojun Lim, Hyeongseok Jeon, Donghyun Kim, Soonyoung Jung, Heecheol Yoo
The paper introduces the concept of a visual orchestration tax in agentic vision‑language model (VLM) pipelines, where the same static images are repeatedly processed by multiple agents, creating redundancy. It proposes a measurement‑to‑certification framework that audits visual evidence reuse using metrics M1_trace and M2, and certifies reuse through a SharedVisCache contract. Experiments on SeeingEye and MAMMQA show high redundancy rates (66.8‑75.6%) and demonstrate that certified reuse can reduce vision‑API calls while preserving output fidelity.
By Lingteng Zeng
The paper identifies a representational bottleneck in 3D CT embeddings that limits pathology coverage in automated radiology reports. It introduces AdaRAG-CT, an adaptive retrieval-augmented framework that supplements visual features with controlled textual retrieval to improve report generation. On the CT-RATE benchmark, AdaRAG-CT achieves state‑of‑the‑art clinical efficacy, raising Clinical F1 from 0.420 to 0.480.
By Renjie Liang, Yiling Ma, Yang Xing, Zhengkang Fan, Jinqian Pan, Chengkun Sun, Li Li, Kuang Gong, Jie Xu
The paper addresses the problem of setting optimal non‑discriminatory prices in a network of strategic agents with private information and potential misreporting. It shows that the optimal prices can be decomposed according to each agent’s squared centrality in a preference‑weighted network, leading to the Poll algorithm. Poll samples one agent per round, briefly explores its neighborhood, and updates prices locally, achieving significant computational, statistical, and economic efficiency gains over exact or other distributed methods.
By Chenyu Zhang, Rohit Parasnis, Saurabh Amin
The paper introduces scalable regularized estimation techniques for logarithmic multiplicative error models (log‑vMEM) applied to multivariate positive‑valued financial time series. It employs hierarchical lag structures and various group penalties, using a blockwise coordinate descent algorithm with Gauss‑Seidel updates to efficiently estimate parameters. GPU‑accelerated quadrature integration is used to overcome the main computational bottleneck, and the methods are evaluated through extensive simulations and applied to intraday realized volatility data for Microsoft.
By Rohan Hemant Chhatre, Chiranjit Dutta, Nalini Ravishanker, Sumanta Basu
The paper introduces PAIR, a pipeline that creates user-conditioned application states to evaluate mobile GUI agents across different users, and RePAIR, a reinforcement learning method that incorporates personalization-aware interaction rewards. Experiments with six agents reveal significant variability in task success and subgoal achievement across users, especially when targeting personal content. RePAIR improves success rates over supervised fine-tuning, suggesting that learning from cross-user differences can enhance GUI-agent reliability.
By Yeji Park, Jaeyun Shim, Taesik Gong
SIGMA is a self‑improving alignment framework that uses a model’s own reasoning to generate diverse alignment dilemma scenarios from a simple "Model Spec" and then fine‑tunes the model with self‑judged rewards. By performing spec‑guided task synthesis and rubric‑based reinforcement learning, SIGMA improves safety alignment in multi‑turn agentic settings, reducing harmfulness and misalignment scores while preserving general capability. The approach demonstrates that a model can autonomously strengthen its own safety reasoning without external supervision.
By Jingyu Zhang, Shruti Palaskar, Daniel Khashabi, Benjamin Van Durme, Leon A. Gatys, Joseph Yitan Cheng
EIO-Agents introduces an open specification for interoperable AI agent evaluation, consisting of two layers: the Evaluation Intelligence Ontology (EIO) and the Portable Evaluation Record (PER). The EIO layer supplies a semantic framework that defines typed evidence, behavioral predicates, evidence contracts, claims, witness rules, proof status, recurrence, and computable derivations for metrics, findings, controls, and decisions such as PASS, REVIEW, or BLOCK. The PER layer offers a canonical, content‑addressed record of an evaluation that preserves the evidence‑to‑decision chain, enabling re‑derivation, explanation, and verification of the evaluation outcome.
By Fouad Bousetouane
POLAR is a guardrail framework designed for small tool‑calling language‑model agents that evaluates the reversibility of each action using a two‑layer ontology. By assigning graded reversibility scores and pruning calls that fall below a threshold, POLAR aims to prevent operational risks before they occur. In experiments on the τ²‑bench across six agent models, POLAR increased mean task reward by 0.11 to 0.18 points on airline tasks for four of six agents, though improvements were limited to eight of eighteen model‑domain cells, with some domains showing regression.
By Yunju Kang, Seonghyeon Cho, Irene Li, Yeo-Chan Yoon, Chanjun Park
The paper introduces Hindsight Meta-Experience Distillation (HMED), a method that lets self‑improving agents evaluate how revisions to their Meta‑Skills affect skill discovery by re‑executing both the original and revised Meta‑Skills from the same initial state. By comparing these outcomes, HMED creates structured Meta‑Experiences that provide learning signals even for revisions that are not ultimately kept. Experiments on three interactive benchmarks show that HMED consistently outperforms strong baselines, improving skill discovery by focusing on the consequences of the improvement process rather than just branch outcomes.
By Qianhan Feng, Zhongzhen Huang, Yakun Zhu, Xiaofan Zhang, Qi Dou
The paper presents a continuous Dyna loop that trains a multi‑agent world model for a ten‑hero MOBA game and evaluates the resulting policy in the real game. By using imagined episodes and real‑game data to continuously refine the model, the policy achieves a 70.2% win rate as radiant on unseen seeds, a dramatic improvement over dream‑only training. The study highlights that internal dream metrics can miss model exploitation, that the world model must be repaired on the policy’s own games, and that the policy’s success stems from macro‑level pressure rather than precise micro‑tactics.
By Jordy Kieto
The paper introduces a repeat‑run evaluation framework for AI agents that process large unstructured datasets, focusing on theme churn and volume disagreement as key metrics. Experiments on three customer‑feedback tasks across eight frontier models show that a taxonomy‑grounded agent (TGA) dramatically reduces theme churn by 86–88% and eliminates volume disagreement compared to raw generation and hierarchical decomposition. The framework is applicable beyond customer feedback, to any recurring synthesis of unstructured corpora such as financial reports, legal documents, incident records, and scientific literature.
By Viraj Bagal, Raviraja Ganta, Prabhath Chellingi
The paper introduces VisualNoiseQA, a benchmark that tests large language models (LLMs) on active reasoning tasks involving noisy visual feedback. In this setup, a text-only LLM must solve visual question answering (VQA) problems by repeatedly querying an off-the-shelf vision-language model (VLM) that acts as a stochastic sensor. Each query generates multiple samples, and the resulting empirical uncertainty is exposed via self‑consistency, allowing the LLM to decide what to ask next and when to stop. The benchmark is automatically constructed from diverse VQA sources and two noisy VLMs, retaining only questions where the sensor is inconsistent yet solvable by humans, and evaluates several LLM reasoners on 1,000 instances covering perception, chart understanding, and knowledge-intensive reasoning.
By Bach Nguyen, Zhaonan Li, Mau Son Nguyen, Sanika Chavan, Nilay Kumar, Hong Anh Nguyen, Khoa Vo, Ben Zhou
SpeedrunBench is a new benchmark that tests large language model agents on their ability to develop and refine strategies for completing nine different video games as quickly as possible. The benchmark requires agents to repeatedly improve, reflect, and reason over long action horizons to beat both themselves and other players. Experiments show that while agents can approach human records on simple platformers, they lag behind humans on longer, more complex games within realistic resource limits.
By Yoshinari Fujinuma, Keisuke Kamahori, Ryuto Koike, Abdelrahman Madkour, Varun Prashant Gangal, Monty Bichouna, Martyna Markiewicz, Shivani Jain, Duncan Curtis, Rebecca Qian, Anand Kannappan
The study investigates how post‑training quantization (8‑bit vs. 4‑bit) affects language‑model agents’ ability to recover from temporary tool failures. Using Llama‑3.1‑8B‑Instruct and Qwen2.5‑7B‑Instruct on twenty deterministic tool‑use tasks across five prompts, the authors find that recovery performance varies with prompt, evaluation target, and scoring policy, sometimes favoring 8‑bit, sometimes 4‑bit, and even reversing direction under different conditions. The results highlight that a single prompt, task set, or scoring rule cannot reliably determine the robustness of quantized agents.
whyItMatters":"The findings emphasize that evaluation design choices critically influence conclusions about quantized‑agent robustness, underscoring the need for comprehensive, matched‑task assessments before deployment decisions are made."
By Yuhe Hu
The paper introduces STEPGATE, an uncertainty‑aware handoff framework that evaluates each step of a small language model (SLM) agent and selectively escalates difficult steps to a stronger model. On a 52‑task single‑step benchmark, the Qwen2.5‑1.5B/7B pair achieved 82.7% task success with only 30.8% escalation, outperforming local‑only and random escalation baselines. In multi‑turn tests, STEPGATE reached 69.0% trajectory success and 84.0% action success while using only 30.0% cloud actions, demonstrating that step‑level escalation can close much of the performance gap to a stronger backend with fewer remote tokens.
By Abolfazl Younesi
RA-MoWE introduces workflow‑affinity embeddings to cluster queries and guide the creation of reusable expert workflows for large language models. Each embedding captures how well a set of reference workflows solves a query, revealing common reasoning strategies. The framework uses cluster embeddings to initialize and refine specialized workflows, and an encoder predicts embeddings from query text, enabling efficient expert selection without executing reference workflows.
By Qi Cheng, Shengyu Chen, Wei Cheng, Yiqun Xie, Haoyu Wang, Haifeng Chen, Xiaowei Jia
WorkflowOps is a multi‑agent workflow orchestration framework that learns collaboration priors from historical workflows. It uses a transition probability matrix to guide DAG construction, a sufficiency‑driven loop to create new agents on demand via an LLM, and a layered semantic matching strategy to reduce LLM calls. Experiments show that WorkflowOps improves end‑to‑end pass rates, especially on structured, decomposable tasks.
By Qi Cheng, Shengyu Chen, Wei Cheng, Zhengzhang Chen, Xiaowei Jia, Haoyu Wang, Haifeng Chen
GameGo is a framework that converts short game ideas into detailed Product Requirements Documents using industry practices, enabling coding agents to generate complete games from sparse user queries. It employs dynamic compression to keep essential gameplay constraints while allowing design flexibility. The authors built GameGoData with over 55,000 development trajectories and GameGoBench with 124 game queries, training GameGoCoder to outperform baselines and match leading models on gamedev benchmarks.
By Haoyue Yang, Jingyao Li, Zhengfan Wu, Jing Liu, Xuanle Zhao, Kang Liu