AirLLM introduces a hierarchical diffusion policy framework that adapts Low‑Rank Adaptation (LoRA) rank configurations for remote fine‑tuning of large language models over wireless channels. The system uses a Proximal Policy Optimization agent to make coarse decisions based on wireless and linguistic cues, then refines these decisions with Denoising Diffusion Implicit Models to produce task‑ and channel‑specific rank vectors. Experiments across different signal‑to‑noise ratios show that AirLLM improves fine‑tuning performance while substantially lowering transmission costs.
By Shiyi Yang, Xiaoxue Yu, Rongpeng Li, Jianhang Zhu, Zhifeng Zhao, Honggang Zhang
The paper introduces a novel LLM‑driven multi‑agent pipeline that converts relational databases into graph databases by standardizing table and column names and iteratively refining the graph schema through ETL, Analyzer, and Graph agents. The resulting graph database meets accuracy, groundedness, and faithfulness criteria and shows significant performance gains, achieving 85.6% Q&A accuracy—12.12% higher than an SQL agent on PostgreSQL—and reducing latency by roughly threefold on a BFSI dataset. This demonstrates an efficient, automated method for transforming tabular data into a more intuitive and faster‑executing graph format.
By Dinh-Khanh Pham, Quy-Anh Dang, Lam Mai Thanh, Khanh Bui, Truong-Son Hy
The paper demonstrates that safety mechanisms for autonomous large language model agents fail to compose across iterative loops, as trajectory‑scoped monitors cannot detect attacks whose evidence is spread over multiple iterations. It introduces LoopHarness, a system that maintains a persistent, non‑decaying safety state across loops, bounding unauthorized actions with a constant that does not grow with the number of iterations. The authors provide a comprehensive evaluation protocol, including attacks that require cross‑iteration evidence, module ablations, and adaptive white‑box red‑team testing.
By Chenhao Wu, Haoxuan Jia, Yang Liu, Yingguang Yang, Yuhan Lin, Chongyang Zhang, Hao Zheng, Yulin Huang, Jianshen Zhang, Yongzhi Qi, Shang Luo, Kefu Xu, Jifeng Zhu, Bin Chong
Diff Mining is a framework that identifies what a finetuned language model has learned by comparing its logits to those of its base model. It extracts per-context logit differences on a reference corpus and aggregates them into an interpretable token set using either a Top‑K frequency method or Non‑negative Matrix Factorization. The approach outperforms existing model‑diffing methods in domain detection and bias identification, and it requires only access to output logits, making it scalable to large models.
By Greg Kocher, Robert West, Cl\'ement Dumas, Julian Minder
GraphMemix introduces a combinatorial‑optimization graph memory framework that constructs query‑aware evidence forests for long‑term multimodal agent memory. It expands seed memories via schema and semantic relations, decouples memory support from relation verification to reduce redundancy, and optimizes a forest‑format context within a maximum evidence budget. Experiments on four benchmarks show significant accuracy gains and a new Pareto frontier between accuracy and lifecycle cost.
By Geng Li, Yuhao Wang, Dong Li, Jianye Hao, Yuxin Peng
GRAIN is a single-agent reinforcement learning framework that improves large language models’ robustness to real‑world shifts in node identifiers and task formulations by treating reasoning as a semantic parsing and tool‑execution pipeline. It introduces a Structure Invariance Reward that validates intermediate graphs against ground‑truth topologies, encouraging the model to learn genuine text‑to‑structure mappings instead of overfitting to surface patterns. On the new GRIT benchmark, GRAIN surpasses multi‑agent baselines by 16.45% in accuracy, reduces latency by about 24%, and halves the out‑of‑distribution gap of fine‑tuned models while remaining robust on large‑scale graphs beyond the training distribution.
By Zike Yuan, Han Zhang, Jianzhi Yan, Le Liu, Cai Ke, Huozhi Zhou, Jian Xie, Jiran Yin, Yukun Cao, Yue Yu, Hui Wang, Ming Liu, Bing Qin
The paper introduces a Latent World Model (LWM) for robot navigation that predicts action‑conditioned latent feature compatibility instead of reconstructing future observations. By exploiting the correlation between spatial proximity and latent feature similarity, the model evaluates action consequences directly in latent space and supports counterfactual training using sampled action sequences. The learned world model can supervise policy learning from unlabeled video and further improve policies via reinforcement learning entirely within the model, eliminating the need for action annotations and additional environment interaction.
By Zengmao Wang, Wei Gao, Shuhan Shen
RTNav is a new architecture for real‑time zero‑shot object navigation that explicitly accounts for inference latency, asynchronous environment stepping, and bounded compute. It addresses the performance drop seen in existing methods when deployed under realistic timing constraints. On real‑time variants of HM3D‑v1, HM3D‑v2, and HM3D‑OVON, RTNav boosts success rate by up to 11% and Success weighted by Completion Time by up to 5.1 points.
By Easop Lee, Lingyu Zhang, Boyuan Chen
MineExplorer is a benchmark designed to assess the open‑world exploration abilities of multimodal large language models (MLLMs) in Minecraft. It filters out tasks that rely heavily on Minecraft‑specific knowledge, organizes tasks into ReAct‑style capabilities, and composes atomic tasks into implicit multi‑hop challenges. A multi‑agent synthesis workflow creates reliable task graphs, sandbox scenes, and rule‑based milestone evaluators, and human evaluation confirms its superiority over a single‑agent baseline. Experiments show that while advanced MLLMs can handle many single‑hop tasks, they struggle with longer trajectories that require coordinating hidden prerequisites, and larger models or different thinking modes do not consistently improve performance.
By Tianjie Ju, Yueqing Sun, Zheng Wu, Wei Zhang, Yaqi Huo, Xi Su, Qi Gu, Xunliang Cai, Gongshen Liu, Zhuosheng Zhang
The paper introduces the Multi-Instruction Multi-Shot Long-Video Editing (MMLVE) task, aiming to enable consistent editing of long videos with multiple instructions. It proposes an agentic editing framework that combines Large Language Models and Vision-Language Models for shot-level decoupling and precise instruction parsing. The authors also present the MMLVE-Bench dataset and evaluation metrics, showing that their MMLVE-Agent outperforms existing state‑of‑the‑art methods by eliminating hallucinations and preserving temporal consistency.
By Chenyang Wu, Fuchen Long, Binyuan Huang, Xinlong Sun, Xi Chen, Chun-Le Guo, Chongyi Li
CIFQA is a deterministic, tool‑grounded multi‑agent framework that separates language understanding from numerical execution for financial question answering. It assigns specialized agents for interpretation, routing, parameter extraction, computation planning, and response generation, while deterministic Python tools perform the calculations. On a fixed‑deposit benchmark, CIFQA achieves 95.54% accuracy on calculation‑intensive queries and 90.87% overall, outperforming larger LLM baselines and showing that architecture, not scale, drives numerical reliability.
By Kunjesh Parekh, Anil Kumar Tiwari, Divya Saxena
The paper introduces CoVeMem, a Collaborative Vector Memory system that replaces text-based memory in agentic recommender systems with vectorized user and item states derived from a frozen LightGCN model. By retrieving relevant historical states at each decision and integrating them as soft tokens in the LLM’s context, CoVeMem enables contrastive alignment and listwise co‑training to learn how to read and rank these states, achieving performance on par with or better than existing text‑memory agents across multiple benchmarks without extra LLM calls for memory updates.
By Hanchong Chen, Xing Tang, Lingjie Li, Xiongfeng Shan, Xiuqiang He
FaulT-Bench is a new benchmark comprising 200 network troubleshooting scenarios across eight topologies, designed to test large‑language‑model agents on realistic, noisy user tickets that may contain false premises or incorrect fault claims. The benchmark includes 72 rewritten tickets that vary reporter confidence and detail while keeping the network state constant, allowing isolation of the impact of ticket wording on diagnosis. Evaluation of agents such as SADE, ReAct, and Claude Code shows they perform well on accurate tickets but degrade sharply on misleading or healthy‑network tickets, revealing differing failure modes and highlighting the importance of robust reasoning over unreliable input.
By Kuan-Hao Tseng, Niruth Bogahawatta, Yasod Ginige, Kunjan Patel, Kosta Dakic, Suranga Seneviratne
DataSTORM is an LLM‑based agentic system designed to conduct deep research over large‑scale structured databases and internet sources. It applies principles of Exploratory Data Analysis and Data Storytelling to frame research as a thesis‑driven analytical process, iteratively generating hypotheses, performing quantitative reasoning, and crafting coherent narratives. Evaluations on InsightBench and a new ACLED‑based dataset show that DataSTORM surpasses existing systems, achieving significant improvements in insight‑level recall and summary‑level scores.
By Shicheng Liu, Yucheng Jiang, Sajid Farook, Camila Nicollier Sanchez, David Fernando Castro Pena, Monica S. Lam
The article reviews how Large Models (LMs) based on Transformer architectures and self‑supervised pre‑training can address longstanding challenges in Battery Prognostics and Health Management (BPHM). It surveys LM applications across data scarcity, generalization, interpretability, and system automation, and outlines a roadmap for future research, including collaborative data ecosystems, validation, trustworthiness, and efficient deployment. The review aims to guide researchers and practitioners in developing next‑generation battery management systems that are safe, reliable, and autonomous throughout battery lifecycles.
By Jiale Liu, Huan Wang, Weicheng Wang, Rong Zhu, Qiqi Wang, Min Xie
The paper investigates whether large language models (LLMs) follow Occam's Razor when performing inductive and abductive reasoning. It introduces a synthetic framework for generating questions that require both types of reasoning and a new automated metric to evaluate the simplicity and correctness of generated hypotheses. Experiments show that while LLMs can handle simple scenarios, they struggle with complex world models and producing high‑quality, simplest hypotheses, even when using advanced reasoning techniques.
By Yunxin Sun, Abulhair Saparov
arXiv:2608. 27260v1 Announce Type: new Abstract: LLM agents increasingly rely on generated interaction data to learn how to interact with external environments.
By Xingshan Zeng, Zishan Xu, Boju Zhang, Yuzhou Wu, Lingzhi Wang, Jianghao Lin, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Weinan Zhang, Yong Yu, Qun Liu, Weiwen Liu
ContextEcho is a benchmark and harness designed to measure persona drift in large language models during long, tool‑using coding sessions. It includes a 25‑probe identity suite, a snapshot‑then‑probe protocol that preserves the main conversation, and both judged and judge‑free measurement surfaces. Across 23 frontier models and thousands of turns, the benchmark shows that persona drift is widespread, not limited to specific model families, and that simple in‑session compaction does not reset it, while a single‑shot anchor can restore the intended persona.
By Xianzhong Ding, Yangyang Yu, Changwei Liu, Bill Zhao, Le Chen, Tao Chen
DIANOIA introduces a diagnostic framework for multi‑agent large language model systems, decomposing reasoning gain into three measurable channels—coverage, fidelity, and synthesis. The protocol identifies bottleneck channels for a given task and implements a corresponding multi‑agent system with role‑diverse proposers, execution‑grounded verification, and iterative synthesis. Experiments on GSM8K, AIME‑2025, MBPP, and BFCL‑SP show that DIANOIA outperforms strong baselines, achieving significant token savings and accuracy gains while accurately pinpointing the critical channels.
By Yiming Yang, Zhuoyuan Li, Fanxiang Zeng, Hao Fu, Yue Liu
SWE-Prime introduces a two-stage supervised fine-tuning data selection process for large language models tackling software issues. The first stage filters entire trajectories by quality and representativeness, while the second stage selects meaningful semantic segments based on contribution, learnability, and risk. Experiments on SWE-Bench Pro and Verified demonstrate that training on just 10% of trajectories chosen by SWE-Prime surpasses full-dataset training, achieving up to 12.2% and 24.2% performance gains.
By Dewu Zheng, Ruizhe Ye, Yanlin Wang, Yang Ye, Hongyu Zhang, Ensheng Shi, Xilin Liu, Yuchi Ma, Jianxing Yu, Zibin Zheng