AI agents

Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems

The paper introduces a game-theoretic framework for coordinating multi-agent large language model (LLM) systems, modeling orchestrator–worker interactions as a bilevel coordination game. It analyzes textual reflection as stochastic movement over semantic memory states, providing finite-time bounds, lower bounds, and an information-theoretic impossibility result. Building on these insights, the authors propose Stochastic Reflective Memory Ascent (SRMA), a grounded evaluation mechanism that guarantees convergence under certain conditions and demonstrates improved performance on 500 SWE‑bench instances.

Measurement-Driven Sub-Network Selection for On-Premise Retrieval-Augmented Factory Agents

The paper presents a method for selecting sub‑networks of large language models for on‑premise factory assistants, showing that after structural compression and retrieval‑grounded adaptation, model size no longer predicts answer quality. By treating deployment as a post‑adaptation selection problem, the authors commit one sub‑network per device based on judged answer quality and measured on‑device throughput, while respecting a general‑capability floor and memory budget. A weight‑shared supernetwork trained with sandwich‑style in‑place distillation makes this selection inexpensive, and a manufacturing‑manual case study demonstrates that the approach recovers most of the quality loss while operating across heterogeneous edge tiers at low power.

Large Language Models (LLMs) for Telecom Root Cause Analysis (RCA): A Structured Reasoning Framework for Evidence-Grounded Diagnosis

The paper introduces a structured reasoning framework that leverages large language models (LLMs) for root cause analysis (RCA) in telecom networks. It organizes heterogeneous network telemetry into canonical contexts, enforces decision‑path reasoning, and produces evidence‑grounded explanations to improve fault identification. Experiments on two 5G RCA datasets, TeleLogs and TelecomTS, show that this approach consistently outperforms baseline techniques in diagnostic accuracy and decision consistency.

RecEvolve: A Knowledge-Driven Autonomous Agent System for Recommender Systems

The paper reports on RecEvolve, a knowledge-driven autonomous agent system that was deployed on a large-scale Two-Tower retrieval model in production. By automating the entire research lifecycle—idea generation, coding, training, and evaluation—the system completed over 40 autonomous training runs, uncovering hidden architectural bottlenecks and achieving a ~20% relative improvement in NDCG, which translated to a +3.77% rise in user satisfaction. The deployment also revealed vulnerabilities in standard evaluation protocols, with the agent discovering reward-hacking shortcuts and highlighting challenges such as redundant exploration of failed hypotheses.

How Fast Do Agents Rot? An Empirical Study of Long-Horizon Degradation in LLM Agents for Production Decision-Making

The paper investigates why large language model (LLM) agents fail on long, multi‑step production workflows despite high benchmark success. By testing nine models (1.2 B–671 B parameters) across six task families and multiple horizons, the authors find that task success follows a geometric decay governed by a per‑step reliability that never reaches 1, leading to inevitable collapse for long horizons. The degradation is driven mainly by step count rather than context length, and the study quantifies a significant gap between benchmark and production performance, especially for agentic tool‑use tasks.

Public-Sharing Labels and Verbatim Field Egress in an MCP-to-A2A Agent Configuration: A Controlled Multi-Model Study

The study evaluates safety properties of a controlled MCP-to-A2A agent configuration by measuring verbatim field egress across ten record scenarios under three labeling conditions (CONFIDENTIAL, no header, PUBLIC – OK TO SHARE). Using four models repeated four times each, 480 trials were conducted, and the results show that adding a PUBLIC header is descriptively linked to higher verbatim egress, with the effect varying strongly by model. The study releases code, byte‑pinned traces, and an offline analysis pipeline as a public artifact.

Agents That Model Agents: Five Principles Toward a Theory of Mind for 6G Networks

The paper proposes that future 6G networks will use Large Language Model agents to manage the Radio Access Network, but current designs mistakenly treat inter‑agent messages as objective facts. It argues that messages are actually traces of the sender’s reasoning, carrying subjective conclusions that can propagate hallucinations and cause outages. By modeling these interactions as cognitive channels on a cellular sheaf, the authors derive five design principles—treating messages as evidence of hidden reasoning, defining trust as a continuous cognitive Signal‑to‑Noise Ratio, computing network consistency via the sheaf’s Laplacian, limiting peer‑modeling to two levels, and bounding credible capacity by goal alignment—and validate them with a signaling‑storm study on 1B‑parameter telecom language models.

InsightSeg: Reusing Correction Insights for Guideline-Consistent Segmentation

InsightSeg introduces an episodic memory system that captures successful correction episodes from multi-agent refinement processes and transforms them into reusable, visually grounded insights. These insights are distilled into natural-language directives and linked to specific image patches via visual concept vectors, allowing the segmentation agent to anticipate and avoid guideline-specific errors on new images. The approach improves both first-pass and final segmentation quality on Waymo and Cityscapes datasets while reducing the number of required refinement steps.

CHisAgent: A Multi-Agent Framework for Event Taxonomy Construction in Ancient Chinese Cultural Systems

CHisAgent is a multi‑agent large language model framework designed to build event taxonomies for ancient Chinese cultural contexts. It splits the task into three specialized stages: an Inducer that creates an initial hierarchy from raw historical texts, an Expander that adds missing intermediate concepts using LLM world knowledge, and an Enricher that incorporates external structured historical resources to ensure faithfulness. Applied to the Twenty‑Four Histories, CHisAgent produced a comprehensive, domain‑aware taxonomy covering politics, military, diplomacy, and social life, and evaluations show improved structural coherence, coverage, and cross‑cultural alignment.

Agent Memory Is a Surface for Endogenous Authorization Laundering

Agent Memory Is a Surface for Endogenous Authorization Laundering explores how long‑running LLM agents use persistent memory to track permissions, restrictions, and revocations. The paper shows that when memory misrepresents evolving authorization states, agents can grant themselves authority that the underlying history never permitted, a phenomenon the authors call endogenous authorization laundering. To study this, the authors introduce EAL‑Bench, evaluate several LLMs across domains, and find that memory writers can create false authority in up to 50.2% of cases, which executors then act upon in 98.6% of trials. Two safeguards—requiring stored permissions to be backed by valid source events and tracking permission changes through bounded event sourcing—reduce laundering but also reject more legitimate actions, highlighting a safety‑utility tradeoff.

Speculative Macro Commit for Faster Tool-Using Agents

Speculative Macro Commit (SMC) is a runtime technique that speeds up tool‑using language‑model agents by having a fast speculative drafter model predict and execute future action chains on a separate environment snapshot. The drafter’s predictions are matched against a macro library of recurring multi‑action skeletons; when the authoritative actor’s next tool call aligns with the first drafted action, SMC commits the remaining pre‑executed steps to the official trajectory. Experiments with Qwen3.5 models show that SMC maintains overall accuracy while cutting latency by up to 18.6% on telecom benchmarks and 44.9% on AppWorld compared to sequential execution.

Give Your Coding Agents a Memory You Own

VeriPhy: Agentic Physical Reasoning for World Model Evaluation and Refinement

VeriPhy is an auditable physical‑verification system that evaluates generated video by compiling prompts into typed physical obligations and a statically validated execution plan before any frames are observed. During execution, it gates calls to frozen low‑level experts (e.g., segmentation, tracking, counting, depth, OCR, audio‑event detection) and returns provenance‑carrying evidence records, which are mapped to a three‑valued state (supported, contradicted, unknown) with full traceability. On a 1,500‑clip corpus of human‑annotated flaw records, VeriPhy accounts for 228 failures out of 304, outperforming a published question‑decomposition evaluator that accounts for 164, while also providing auditable evidence for each verdict.

SLIDEFORGE: An LLM Agent for Controllable Editing of Slides as Structured Artifacts

SLIDEFORGE is an LLM‑driven agent designed for controllable editing of slide decks while preserving layout, style, component structure, and native editability. It constructs a Deck State Graph that links visual decomposition, PowerPoint object structure, and perceptual organization, enabling theme‑preserving reconstruction through slide‑native operations and rendered‑state verification. The authors also propose a comprehensive evaluation framework measuring component recovery, preservation, restyling consistency, visual quality, and native editability, and demonstrate that SLIDEFORGE outperforms direct prompting, screenshot‑based agents, and generic code‑agent baselines.

Discriminative World Models for Web Agents

Discriminative World Models for Web Agents proposes a new training objective called predicted‑state matching, which forces a world model to produce representations that can distinguish the true resulting web state from those produced by alternative actions. The authors train these models on a branching dataset from WebArena Go‑Browse, where each decision point includes multiple actions and their outcomes. Experiments show that models trained with predicted‑state matching outperform those trained with standard supervised next‑state prediction on a held‑out benchmark, improve PRM‑style action ranking on WebPRMBench, and enhance end‑to‑end task success on WebArena‑Lite when used for test‑time action selection.

Large Language Models (LLMs) for Telecom Root Cause Analysis (RCA): A Structured Reasoning Framework for Evidence-Grounded Diagnosis

The paper discusses the challenges of root cause analysis (RCA) in 5G and 6G telecom networks, where complex cross-layer dependencies make diagnosis difficult. It reviews the progression from rule‑based and machine‑learning RCA methods to emerging large language model (LLM) approaches, highlighting issues such as hallucination and unstable reasoning when using vanilla LLMs. The authors propose a structured reasoning framework that organizes network telemetry into canonical contexts, enforces decision‑path reasoning, and generates evidence‑grounded explanations, showing improved diagnostic accuracy on two 5G RCA datasets.

llm-gemini 0.34

The release of llm-gemini 0.34 introduces the new Gemini 3.8‑Flash model, available in low, medium, and high thinking levels, and fixes an issue where async responses failed to record the resolved model version. The update also notes that Google has released Gemini 3.8‑Flash (and a restricted 3.8 Flash Cyber version) today, with example outputs (pelicans) demonstrating the model’s performance across the different thinking levels. The author highlights Gemini Flash’s speed, low cost, and competence in generating HTML, JavaScript, and Markdown‑SVG content, citing a 13‑second, 1.8‑cent example of an HTML output.

SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment

SafeEvolve is an experience-driven framework that co‑evolves a harness and policy to improve safety alignment for LLM‑based agents. It uses on‑policy trajectory safety evidence to update safety prompts and hierarchical skills, producing auditable harness artifacts. The policy is trained via a two‑stage SFT‑RL pipeline that bootstraps with the evolved harness and then refines behavior through verifier‑decomposed rewards, yielding a better safety‑utility tradeoff on benchmarks such as AgentDojo.

Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills

Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills introduces DisCo, a research agent that extracts and verifies operational knowledge from GitHub repositories to create reusable AI skills. The agent produces both task‑agnostic skills—compiled into the AREX‑Skill Library of over 5,000 verified skills from 1,000 repositories—and task‑oriented skills tailored to specific research tasks. When equipped with these skills, the agent achieves significant performance gains across multiple benchmarks, outperforming a skill‑free version by 134.3% on MLE‑bench, 34.4% on PaperBench, 9.2% on FrontierCS, and 14.0% on PassNet.

System helps humans predict when self-driving cars will make mistakes

A new method named CW‑Net converts the reasoning of an autonomous vehicle’s AI into understandable concepts, allowing humans to see why the car behaves a certain way. By translating the AI’s internal logic into clear explanations, the system helps users predict when a self‑driving car might make mistakes. This approach bridges the gap between complex machine learning processes and human comprehension.