The paper introduces a two‑phase pipeline that mines reproducible design principles for large language model multi‑agent systems by reusing all intermediate search trajectories from an MCTS optimizer. Using six composable operators and 26 Enigmata tasks, the authors generate a corpus of 28,568 executable workflows and apply an AutoResearch loop to discover and validate orchestration patterns. Key findings include that Star‑first workflows underperform, SharedMemory outperforms debate, and prompt specificity and placeholder‑based routing provide comparable gains to communication‑structure selection.
By Xavier Cadet, Junyan Cheng, Peter Chin, Kyle Richardson
AgentTrace is an active framework that fingerprints black‑box large language model agents by injecting ambiguous probes into tool interfaces, causing tokenization‑sensitive serialization failures. The initial failures separate model families, while the sequence of repair operations—modeled with lightweight Markov chains—reveals fine‑grained, model‑specific recovery behaviors. This approach achieves 96.0% attribution accuracy even under adversarial identity spoofing and can distinguish intra‑family variants that share identical tokenizers, outperforming unordered behavioral statistics and resisting prompt obfuscation and temperature scaling.
By Gejian Zhao, Hanzhou Wu, Xinpeng Zhang
The article discusses a new approach to diagnosing vehicle faults by treating diagnostic trouble codes (DTCs) as a high‑cardinality language. It proposes Transformer‑based models for predicting event sequences, scalable causal discovery frameworks, and a multi‑agent system that automatically generates Boolean error‑pattern rules. These methods aim to replace manual, error‑prone grouping of DTCs with automated, scalable fault diagnostics.
By Hugo Math
The paper introduces the Symbolic Guidance Taxonomy (SGT) to regulate the autonomy of large language model agents in distributed multi‑agent coordination. By providing symbolic guidance—ranging from open‑ended reasoning to fully prescriptive algorithmic steps—the authors find that intermediate levels of autonomy consistently outperform both unguided agents and fully prescriptive specifications. These results suggest that carefully controlling agent autonomy is a crucial design principle for effective LLM‑based coordination.
By Ben Rachmut, Ning Zhang, Yevgeniy Vorobeychik, William Yeoh
The paper introduces ID‑TAP, a neural dueling‑bandit framework that personalizes large‑language‑model agents at test time by selecting personas through binary comparisons of candidate responses. It combines a query‑conditioned neural utility model with a lightweight Bradley‑Terry explorer, and uses similarity‑weighted warm starts to transfer evidence across related queries. Experiments on seven user‑conditioned generation benchmarks show that ID‑TAP outperforms baselines within a ten‑round comparison budget, demonstrating the effectiveness of information‑directed challenger selection for persona adaptation.
By Jie Mao, Pingchen Lu, Xiangyi Wang, Hengquan Huang, Shuang Qiu, Min Zhang, Zhongxiang Dai
The paper presents a multimodal agent for adaptive control of scientific instruments whose responses change over time. The agent combines deterministic algorithms that extract symbolic physical quantities and visual evidence from waveforms, a two‑layer memory that stores trajectories and procedural knowledge, and a multimodal large language model that proposes actions, all verified by an independent feasibility check. In experiments on 455 shots from four EXL‑50U neutral beam injection ion sources, the system achieved high classification performance and a beam‑extraction rate of 86.80% across 409 shots, demonstrating its engineering feasibility for supervised sequential tuning.
By Guoyang Shi, Hongwen Li, Yiding Yu, Zhuang Fu, Siqi Ding, Zhengyuan Chen, JianGuo Chen, Hanyue Zhao, Tianyuan Liu
TGR‑EvoDAG is a DAG‑based multi‑agent framework that automates code evolution for generative recommendation systems, addressing cost blindness and short‑sighted evaluation issues in existing recursive self‑improvement methods. It incorporates stage‑aware cost planning and active downstream co‑adaptation to manage GPU‑hour budgets and navigate temporary regressions, enabling autonomous evolution of a complete generative recommendation pipeline. Experiments on Amazon datasets and deployment on Tencent’s media platform show significant offline metric gains and a measurable increase in online click‑through rate.
By Junlin He, Junzhang Zhu, Yudong Li, Yihong Tang, Xuefeng Jiang, Wei Ma, Chengxiang Zhuo, Zang Li
arXiv:2604.21907v2 Announce Type: replace-cross
Abstract: Equity Bias is a philosophical and practical framework for building more equitable AI systems. It treats AI as an interpretive agent that con...
By Mary Lockwood
OmniSpace is a plug‑and‑play framework that enhances multimodal large language models (MLLMs) with geometry awareness using only 2D observations. It introduces a Camera Pose Injector, a Multi‑view Epipolar Attention module, and a 3D Geometric Distillation objective to improve cross‑view correspondence and depth estimation. Experiments demonstrate that OmniSpace outperforms existing methods on planning, risk detection, language, and generalization benchmarks for autonomous vehicles.
By Hao Vo, Phu Loc Nguyen, Khoa Vo, Sieu Tran, Duc Minh Nguyen, Ngo Xuan Cuong, Nghi D. Q. Bui, Anh Nguyen, Duy Minh Ho Nguyen, Ngan Le
POSEIDON is an automated pipeline that generates diverse, verifiable training environments from scientific code repositories, scaling up the number of environments by nearly tenfold compared to previous methods. It decomposes scientific workflows into four task types—Reproduction, Subflow, Completion, and Correction—to target distinct scientific research capabilities, and uses a recursive improvement loop to refine and recover environments. Using Poseidon, the authors created Pearl, a collection of 97,659 executable and verifiable environments from 7,110 public repositories across ten scientific disciplines, the largest multidisciplinary set to date, which enabled Qwen3.5 models to achieve consistent gains on six scientific-agent benchmarks.
By Qifan Zhang, Yuhan Li, Nuo Chen, Bowen Liu, Xiao Hu, Miao Peng, Chenyi Zi, Lixin Liu, Lydell Huang, Jing Tang, Jia Li
The paper introduces SolveBind, a GUI‑agent framework that tackles verification barriers such as CAPTCHAs by combining hierarchical solving (Perceive, Ground, Solve, Act, Verify) with a runtime supervisor that grounds actions in evidence and manages recovery budgets. Across nine CAPTCHA families and multiple backbone models, SolveBind outperforms baselines by 13.3–34.5 percentage points, achieving the highest success rates on all tested AndroidWorld backbones and improving dynamic verification performance compared to MobileAgent. The study also quantifies the trade‑offs between evidence‑grounded execution control, success rates, latency, and token usage, highlighting the impact of binding and recovery mechanisms on overall efficiency.
By Xinhao Song, Haoxuan Peng, Su Su, Hongliang Wu, Sirui Song, Gongshen Liu, Linfeng Zhang, Dongrui Liu
The paper investigates whether large language model (LLM) agents maintain complementary consistency when reporting probabilities for a claim and its negation in opinion simulations. Using three small open models (Qwen3‑4B, Llama‑3.2‑3B, Gemma‑3‑4B), the authors find that paired reports often sum to more than one, even on real fact‑checking claims, and that this inconsistency persists across different simulation settings. Averaging a report with its complement restores consistency but only marginally improves predictive accuracy for one model.
By Anouar Imel, Jin Qi
The paper introduces a framework that compresses coastal hydrodynamic data—tide levels, non‑tidal residuals, and surface wind speeds—into compact token grids using vector‑quantized variational autoencoders. These tokens enable multi‑agent autonomous reasoning systems to traverse and analyze six months of coastal data in seconds, outperforming traditional raw archive access. The approach demonstrates that agents can perform complex spatiotemporal analyses and answer under‑specified queries efficiently on both on‑prem and off‑prem platforms.
By Jinpai Zhao, Albert Cerrone, Joannes Westerink, Clint Dawson
arXiv:2511.11628v2 Announce Type: replace-cross
Abstract: Modern operating system process schedulers typically enforce a static, ``one-policy-fits-all'' design, which inevitably compromises performan...
By Xinbo Wang, Shian Jia, Ziyang Huang, Jing Cao, Mingli Song
RSI-Opt introduces Recursive Self-Improvement Optimizer Agents that evolve their own training policies by accumulating knowledge from previous runs, including prompts, playbooks, optimizer code, and decision cadences. A meta‑agent rewrites and archives these policies, enabling reusable procedures and code that adapt to new tasks. Across multiple generations and tasks, RSI‑Opt improves optimizer performance, achieving higher z‑scores on evolution suites, better cross‑model transfer, and significant error reductions on PDE benchmarks compared to fixed‑prompt baselines.
By Binghang Lu, Yige Sun, Guang Lin
The paper presents a knowledge layer that reorganizes enterprise Retrieval-Augmented Generation (RAG) for multi‑agent systems by placing an ontology at the artifact level rather than on extracted entities. Using a seven‑agent reverse‑engineering pipeline, the authors processed 241 repositories across 18 services to generate thousands of artifacts that capture scope, journeys, rules, and acceptance criteria, each tagged with provenance metadata. The resulting self‑describing catalog, sparse metadata registry, and navigation hypergraph enable queries to resolve to artifact identifiers, thereby structurally disambiguating near‑identical rules and improving retrieval accuracy. The system was built in under a week and helped root‑cause over 100 production defects and attribute 35 release regressions to specific rules and services.
By Garima Agrawal, Prasun Das, Akshay Sarvade, Priyanka L, Minisha N, Sravani Joshna, Hemath Manivannan, Lavanya Korada, Sambhrit Pattanayak, Srashti Gupta, Siddu R, Prasad Kalyanasundaram
The paper reports on a self‑improving science agent that has admitted 252 scientific relations and uses Minimum Description Length (MDL) compression ratios as an unsupervised epistemic grader. The MDL‑ranked top discoveries match canonical physical laws (e.g., vis‑vis‑a, Balmer/Rydberg) with strong statistical enrichment and no counterexamples, while 86 % of the agent’s ledger entries compress poorly (cr≈1) and are identified as statistical regularities rather than true laws. The study introduces a dual‑axis bookkeeping system separating prediction value (R²) from compression value (cr), demonstrates its robustness, and presents falsifiable predictions.
By Wenbin Mao, Pei Zou
AI-for-Science agents, which combine large language models (LLMs) with domain-specific harnesses to formulate hypotheses, run experiments, and interpret evidence, may compromise model safety. The study evaluated five harnesses on 50 harmful chemistry requests, finding that harnesses can increase harmful compliance even with direct prompting and that adaptive attacks that fail against the base LLM succeed once the harness is added. These results indicate that the safety behavior of the underlying LLM is not durable when integrated into an agent harness.
By Michael Yew Chuan Lee, Manya Sriram, N. Benjamin Erichson
The paper presents a taxonomy of communication failures in multi-agent large language model systems, derived from 3,211 failed multi-round trajectories across diverse domains such as medical QA, coding, finance, multi-hop reasoning, ethics, and embodied text tasks. It analyzes how failures arise, noting that they are largely driven by recovery/correction and consensus/decision issues rather than isolated first-round errors, and highlights patterns like early convergence, minority-signal suppression, and wrong framing persistence. The authors recommend stronger disagreement handling, evidence verification, minority-signal preservation, and recovery checks before finalizing decisions.
By Priya Pitre, Jun Hou, Yi Fang, Naren Ramakrishnan, Xuan Wang
The paper investigates whether experience gained by large‑language‑model agents during molecular optimization in one protein pocket can be transferred to new, unseen pockets. By logging every edit and docking step, the authors build an experience base keyed on contact sites, the edit action, and the round, and find that this key sharply distinguishes improving from worsening edits. However, transfer is limited: only coarse, non‑specific descriptors of the edit and resulting molecule provide useful signals, while pocket‑specific information is largely irrelevant for predicting success.
By Zaifei Yang, Weiyu Chen, Yaqing Wang, James Kwok