The paper introduces a pre‑action verification framework for large language model agents that emit shell commands or code edits. By running deterministic checks before execution, the system can catch invalid commands (95.8% success with a 10.0% false‑positive rate) and prevent silent failures in code edits, achieving high recall while minimizing false positives. The authors provide benchmarks, verifiers, and guards for both action modalities.
By Asaad Althoubi
ParaRecover is a new process-level benchmark designed to evaluate error localization and recovery in multi-turn parallel tool-use agents. It contains 10,626 instances across two difficulty levels, built on a taxonomy of 14 error types that cover planning dependencies, tool selection, and argument matching. The benchmark introduces the SDE rubric, which assesses structural integrity, diagnostic reasoning, and evolutionary strategy during agent execution, and demonstrates that it can guide improvements in agents’ reflective recovery capabilities.
By Bowen Guan, Zhentao Yin, Yanming Shen
TokenMapper is a framework that enables direct translation between different speech tokenizers, allowing heterogeneous speech models to communicate without converting tokens to waveform audio. It handles mismatched token spaces, including single and multi-codebook representations, while maintaining a shared effective token rate. Experiments on GLM-4-Voice, MiMi, and DualCodec show that TokenMapper achieves word error rates close to native reconstructions, comparable human MOS scores, and significantly reduces latency compared to waveform bridging.
By Tal Kozakov, Tal Rosenwein, Eliya Nachmani
The paper investigates learning Nash equilibria in partially observable Markov games (POMGs) where agents cannot fully observe the state. By focusing on a subclass with independent state transitions and a Markov potential game structure, the authors propose an independent learning algorithm that allows agents to converge to an approximate Nash equilibrium using only their own observations and actions, without communication. Under a filter stability assumption, finite‑history policies are shown to approximate the POMG sufficiently, enabling a surrogate near‑potential Markov game and yielding quasi‑polynomial sample and computational complexity.
By Philip Jordan, Maryam Kamgarpour
CoHyDE is an iterative co‑training framework that jointly trains a dense encoder and an LLM rewriter for tool retrieval from large API catalogs. The encoder is fine‑tuned with InfoNCE on catalog‑style hypothetical descriptions generated by the rewriter, while the rewriter is preference‑aligned via DPO against the encoder’s retrieval scores. On a 10k‑tool subset of ToolBench, three rounds of CoHyDE outperform the best single‑component baseline by 2.5 pp NDCG@5 on standard queries and 6.3 pp on vague queries, with the largest gains on the hardest vague tier.
By Vaishali Senthil, Ashutosh Hathidara, Sebastian Schreiber
The paper introduces RIPPLE, a method for adapting workflow-synthesizing agents through prompt-policy editing without retraining the underlying model. RIPPLE diagnoses failed execution trajectories, maps failures to specific policy segments, and restricts edits to those segments. It then evaluates candidate edits in isolation and replays only those that remain safe after composition, achieving up to a 23.1% improvement in validation success on a synthetic benchmark and positive gains on additional language‑model backbones.
By Manqing Mao, Hong Wang, Samson Koelle, Jie Yuan, Zhuoer Wang, James Feng, Yanjun Lin, Daniel Edmiston, Nikki Lijing Kuang, Zhecheng Sheng, Wei Niu
CueMem is a cue‑guided framework for long‑term conversational memory that reconstructs query‑relevant dialogue context from compressed memory records. Instead of treating memory units as self‑contained evidence, it extracts fine‑grained cues linked to their source turns and, at query time, expands from these cues over a turn graph to rebuild a compact evidence context. Experiments on LoCoMo and LongMemEval show that CueMem outperforms baseline memory methods, reduces input tokens and latency, and improves long‑term conversational question answering.
By Changjian Wang, Rongzhen Li, Weili Guan, Shuming Shi, Quan Lu, Ning Jiang
arXiv:2609.12655v1 Announce Type: new
Abstract: Large language model agents are expected to continuously adapt to new tasks and environments over their lifetime by reusing past experience. However, e...
By Yuli Qiu, Yutong Li, Wei Su, Zeming Liu, Wanxiang Che, Heyan Huang, Haifeng Wang, Yuang Guo
arXiv:2609.13141v1 Announce Type: new
Abstract: Post-training attention sparsification reduces the quadratic cumulative attention cost of pretrained Transformers by selecting a small set of context u...
By Zhiwei Li, Lei Zhu, Hao Gu, Xiang Hu, Yan Wang, Haitao Mi, Sirui Han, Leo Liang, Zhijiang Guo
arXiv:2609.12533v1 Announce Type: cross
Abstract: Real-world Earth observation (EO) agents must translate high-level scientific questions into executable workflows to acquire observations, prepare da...
By Zhutao Lv, Chenhao Dang, Yi Feng, Yanpei Gong, Xiaolei Wang, Junyan Ye, Conghui He, Weijia Li
arXiv:2609.04280v2 Announce Type: replace-cross
Abstract: Modern LLM-based agents operate through a harness of tools, reusable skills, and specialist agents that shapes what they observe and what the...
By Zixuan Ke, Vaidehi Patil, Haizhou Shi, Yang Li, Ye Liu, Sarath Shekkizhar, Anurag Koul, Jiayu Wang, Xuan Phi Nguyen, Semih Yavuz, Mohit Bansal, Shafiq Joty
arXiv:2609.12181v1 Announce Type: new
Abstract: Vision-language and language models increasingly interpret materials data, yet benchmarks report that they hallucinate invalid properties and violate p...
By Hasan Kurban, Rasul Khanbayov, Mustafa Kurban
arXiv:2609.12517v1 Announce Type: new
Abstract: Long-Video Question Answering (LVQA) requires locating decisive evidence in hour-scale videos under a limited frame budget. Most training-free methods...
By Jian Hu, Zixu Cheng, Da Li, Wei Li, Ziquan Liu, Shaogang Gong
The paper introduces a contact tracing algorithm that incorporates differential privacy to protect users’ risk scores from adversarial inference. By releasing each risk score with an epsilon of 1, the algorithm achieves a two- to ten-fold reduction in infection rates across two leading COVID‑19 agent‑based simulators. This work claims to be the first to provide differential privacy guarantees for risk score disclosure in COVID‑19 contact tracing.
By Rob Romijnders, Christos Louizos, Yuki M. Asano, Max Welling
AMDKernelVault is an open HIP and Triton kernel corpus and training framework designed for AMD CDNA GPUs. It includes 62,153 verified HIP kernels, 39,893 Triton kernels, and 2,377 ROCm library QA entries, and introduces agent-driven pipelines (HIPKernelGen and TritonKernelGen) that convert PyTorch references into GPU kernels, compile, validate, and profile them on AMD hardware. The corpus was used to fine‑tune Qwen3-8B, achieving the highest correctness on several benchmarks such as PyTorch-to-HIP, TritonBench‑G, and ROCmBench under fixed evaluation budgets.
By Ji Liu, Saptarshi Majumder, Yiqing Huang, Wenwen Ouyang, Umang Pandey, Zeping Li, Chushi Chen, Zihao An, Puyuan Yang, Zekai Li, Sina Rafati, Ziqiong Liu, Pratik Prabhanjan Brahma, Dong Li, Zicheng Liu, Sharon Zhou, Emad Barsoum
DRS‑VPT is a feed‑forward transformer that performs image‑to‑scan registration by predicting the scan pose, point maps, and a coarse‑to‑fine feature pyramid for direct reprojective alignment. It unifies tasks like camera‑LiDAR calibration and indoor camera‑to‑map relocalization, achieving state‑of‑the‑art results in autonomous driving and competitive performance indoors without map‑specific training. The model also learns complex scan‑to‑image projection behaviors, such as occlusion of back‑facing points.
By Lanke Frank Tarimo Fu, Maurice Fallon
The paper introduces Adversarial Importance Sampling (Advis), a technique that leverages importance sampling over standard training trajectories to estimate and optimize worst‑case returns without extra environment interactions or auxiliary networks, thereby capturing long‑term robustness. It also presents advrl, a modular PyTorch library that consolidates existing robustness methods and adversarial attacks into single‑file implementations for easier prototyping and reproducible evaluation. Finally, the authors highlight that optimal adversarial hyperparameters do not transfer across agents, prompting evaluation against a broader set of attackers (6–14× more configurations) and demonstrate the effectiveness of their approach on continuous control tasks.
By Amine Andam, Jamal Bentahar, Mustapha Hedabou
The paper introduces AIM, a privacy‑aware memory framework that lets multi‑agent, multi‑user large language models manage both private and shared memory. AIM classifies data as private (user‑specific) or public (shared) and enforces index‑level access controls to protect sensitive information while enabling shared knowledge to improve coordination. The authors also present MUMBench, a new dataset for evaluating memory operations in multi‑user settings, and report high accuracy metrics for AIM on this benchmark.
By Zachary Johnson, Nigel Boachie Kumankumah, Somya Chatterjee, Tejas Sathyamurthi, Min Chen, Xinyi Alice Li, Xiao Wang, Emily Morgan Gelchie, Jessica Lin, Sadid A. Hasan, Sulaiman Vesal
The paper introduces LAST-CQ, a five-agent, training‑free, execution‑grounded framework for Text‑to‑Cypher that evaluates which components of an agentic pipeline contribute most to performance. Experiments on 2,471 live‑database queries across six backbones show that removing correction reduces execution‑BLEU by 3.1–12.3%, while substituting schema‑grounded feedback with raw error strings has negligible impact. Parallel sampling degrades quality by 10–11%, whereas failure detection and retry routing recover 91.7% of initially failed queries, highlighting that simple failure handling is more effective than sophisticated feedback or increased sampling.
By Ioannis Prokopiou, Athanasios Aidinis, Panagiotis-Christos Kyrmpatsos, Pantelis Vikatos
The paper presents a human‑annotated dataset of 100 Quran recitation recordings, identifying 348 scored units and 162 localized events across ten combined labels. An evaluator that scores both labels and word positions achieves a label‑aware F1 of 0.525 and a localization F1 of 0.826, with adapted production cleaner/alignment components yielding similar scores. A pilot experiment with eight 20‑minute runs across three coders and eight models shows wide variance in F1 (0.143–0.892) and highlights that most gold events are detected, leaving only span extent and label conventions as the remaining challenges.
By Mohamad Al Mdfaa, Nursultan Askarbekuly, Ahmed Helaly, Ubai Sandouk, Manuel Mazzara