In this work, we view multi-agent reasoning as a process of seeking a collective equilibrium of implicit belief dynamics and introduce Stream of Belief (SoB), where agents maintain continuous latent beliefs and iteratively refine them through communication and reflection. Our method seeks a collective belief state of a multi-agent system where further communication induces no meaningful information propagation. Through extensive experiments, we demonstrate the generalization of SoB across tasks and model families. More Broadly, our results suggest a new perspective on scaling multi-agent reasoning from belief dynamics.
Mechanistic Interpretability (MI) has made significant progress in explaining the internal computations of Large Language Models (LLMs); however, these advances have yet to demonstrate comparable success in explaining the decision-making processes of LLM-based agents. This paper examines the challenges MI tools face when applied to realistic agentic failures. We focus on one such failure: web agents hallucinating actions that are not available on a website. We apply a series of mechanistic analyzes to Qwen2.5-7B-Instruct as it interacts with live websites. We find that a linear probe trained on the last prompt token can predict hallucinations before generation begins (0.80 AUC) and steering along a probe direction corrects 85\% of hallucinations, but disrupts 7\% of valid actions. However, these results only provide a partial explanation of why the agent hallucinates. The probe direction is sensitive to prompt formatting, losing much of its steering effect after a minor change to how actions are represented. Additionally, circuit identification can be difficult as the computation is distributed across multiple components and positions, and the information the agent relies on varies across examples. We discuss how these observations challenge the assumptions underlying standard MI methods for circuit discovery and feature identification, and outline future directions for making these methods better suited to explain failures in LLM-based agents.
Neural networks are likely to be among the most consequential artefacts in human history. As these systems become more capable and autonomous, their potential to cause societal harm also grows, making AI safety one of the defining challenges of our time. Yet we continue to build increasingly powerful systems without fully understanding how they work, how they will behave in unfamiliar situations, or how to reliably control them. Mechanistic interpretability offers a path towards understanding these systems by studying their internal representations and computations, with the aim of evaluating, predicting, and controlling behaviour. We survey mechanistic interpretability for AI safety in order to answer the question: Does understanding how models work actually make them safer? We place mechanistic interpretability in the broader context of interpretability research and compare its mechanistic evidence with behavioural evaluation. We introduce the foundations of mechanistic interpretability, including features, polysemanticity, transformer computation, and circuits, and survey methods for feature extraction and analysis, circuit discovery and causal verification, explaining intermediate representations in natural language, learning dynamics, and training-data attribution. Beyond surveying existing methods, we propose a unified framework for evaluating mechanistic interpretability claims and characterise six research orientations by their goals, assumptions, and trade-offs. Finally, we examine how mechanistic understanding can support AI safety, including knowledge editing, steering, and monitoring, and assess its limitations, failure modes, open problems, and relationship to behavioural safety measures. We also release a repository of code notebooks with worked examples that reproduce key findings from surveyed methods, allowing readers to build intuition through hands-on experimentation.
Dingyu Yao, Chenxu Yang, Zhengyang Tong Institute of Information Engineering, Chinese Academy of Sciences, Zheng Lin, Wei Liu, Jian Luan Xiaomi Corporation, Weiping Wang IIE
原文摘要
The Key-Value (KV) cache introduces substantial memory overhead during large language model (LLM) inference. Although existing vector quantization (VQ) methods reduce KV cache usage and provide flexible representational capacity across bit-widths, they suffer severe performance degradation at ultra-low bit-widths due to key cache outliers that hinder effective codebook utilization. To address this challenge, we propose VecInfer, a novel VQ method for aggressive KV cache compression while enabling efficient inference. By applying smooth and Hadamard transformations, VecInfer suppresses outliers in the key cache, enabling the codebook to comprehensively cover the original data distribution and thereby reducing quantization difficulty. To facilitate efficient deployment, we design an optimized CUDA kernel that fuses computation with dequantization to minimize memory access overhead. Extensive evaluations demonstrate that VecInfer consistently outperforms existing quantization baselines across both long-context understanding and mathematical reasoning tasks. With only 2-bit quantization, VecInfer achieves performance comparable to full precision, while delivering up to \mathbf{2.7\times} speedup in large-batch self-attention computation and \mathbf{8.3\times} reduction in single-batch end-to-end latency on Llama-3.1-8B with a 196k sequence length.
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Lingrui Mei, Shenghua Liu Institute of Computing Technology, Chinese Academy of Sciences, Yiwei Wang University of California, Merced, Yuyao Ge, Baolong Bi, Jiayu Yao, Jun Wan UBS Group AG, Ziling Yin 等 10 位作者
原文摘要
Long contexts break transformers: attention scores dilute across thousands of tokens, critical information gets lost in the middle, and the model cannot adapt to novel patterns at inference time. Recent work on test-time adaptation addresses this by maintaining a form of working memory—transient parameters updated on the current context—but existing approaches employ uniform write policies that waste computation on low-value regions and suffer from high gradient variance across semantically heterogeneous contexts. In this work, we reframe test-time adaptation as a budget-constrained memory consolidation problem, asking: given limited computational budget, which parts of the context should be consolidated into working memory? We propose GDWM (\textbf{G}ated \textbf{D}ifferentiable \textbf{W}orking \textbf{M}emory), a framework that introduces a Write Controller to gate the memory consolidation process. Our controller estimates Contextual Utility—an information-theoretic measure quantifying how much each region depends on long-range context—and allocates gradient steps accordingly, subject to a coverage constraint that ensures global representation. Theoretically, we prove that our chunk-restricted sampling strategy reduces gradient variance by eliminating inter-chunk variance via the Law of Total Variance. Experiments on ZeroSCROLLS and LongBench v2 benchmarks demonstrate that GDWM achieves comparable or superior performance with 4 ×fewer gradient steps compared to uniform baselines—excelling on sparse-information tasks (+6–13% on Qasper, +5–13% on GovReport for smaller models) while revealing principled trade-offs on dense-coverage tasks, establishing a new efficiency-performance Pareto frontier for test-time adaptation.
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Mohamed Aghzal, Gregory J. Stein, Ziyu Yao George Mason University
原文摘要
Large language model (LLM) web agents are increasingly used for web navigation but remain far from human reliability on realistic, long-horizon tasks. Existing evaluations focus primarily on end-to-end success, offering limited insight into where failures arise. We propose a hierarchical planning framework that analyzes web agents across three layers (i.e., high-level planning, low-level execution, and re-planning), enabling process-based evaluation of reasoning, grounding, and recovery. Our experiments show that structured Planning Domain Definition Language (PDDL) plans produce more concise and goal-directed strategies than natural language (NL) plans, but low-level execution remains the dominant bottleneck. These results indicate that improving perceptual grounding and adaptive control, not only high-level reasoning, is critical for achieving human-level reliability. This hierarchical perspective provides a principled foundation for diagnosing and advancing LLM web agents.
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Xingyu Yao, Zhendong Mao University of Science and Technology of China, Quan Wang Beijing University of Posts and Telecommunications
原文摘要
Detecting LLM-generated code is crucial for ensuring software provenance, security, reliability, and licensing compliance. Existing training-free detectors, mostly adapted from text-based methods, rely on global statistics of the Token Perplexity Sequence (TPS) and struggle with code. We reveal a key insight: despite the convergence of global statistics, LLM-generated and human-written code differ fundamentally in their local TPS dynamics: the former shows narrow transient spikes while the latter exhibits broad sustained fluctuations. To capture this distinction, we introduce CodeRipple, a novel training-free detection framework that employs wavelet analysis to characterize TPS morphology across scales. It jointly leverages the Stationary Wavelet Transform to model fluctuation shape and the Discrete Wavelet Transform to quantify cross-scale energy distribution. Evaluated on three challenging benchmarks spanning diverse programming languages, multiple generating LLMs, and various evasion strategies, CodeRipple consistently outperforms existing training-free methods, demonstrating its superior effectiveness and generalizability without any model training. Code available at: https://github.com/yaoxingyu77/CodeRipple.
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Xiaoting Wu, Yi Huang China Mobile Research Institute, Chunyang Gao, Mengfei Guo, Jingyu Yao China Mobile Communications Company Limited, Junlan Feng
原文摘要
Existing user simulators based on prompting to role-play or SFT are generally confined to imitating users’ textual utterances, without adequately considering the multi-faceted cognitive processes that underlie human decision-making during interactions. To facilitate better alignment with real human thinking patterns, we construct the LMSYS-UserThinking dataset, in which we augment 51k human–LLM conversations by reconstructing the user’s inner reasoning both during and at the end of each dialogue. Furthermore, to enhance controllability and situational coherence, we introduce scenario settings that describe the global context and user goals throughout multi-turn conversations. Using this dataset, we train user simulators called ThinkingUS on different base models. We evaluate our approach from both offline and online user simulation perspectives, ultimately demonstrating its effectiveness.
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao
原文摘要
While large language models (LLMs) have demonstrated impressive capabilities across tasks in language understanding and interactive decision making, their abilities for reasoning (e.g. chain-of-thought prompting) and acting (e.g. action plan generation) have primarily been studied as separate topics. In this paper, we explore the use of LLMs to generate both reasoning traces and task-specific actions in an interleaved manner, allowing for greater synergy between the two: reasoning traces help the model induce, track, and update action plans as well as handle exceptions, while actions allow it to interface with external sources, such as knowledge bases or environments, to gather additional information. We apply our approach, named ReAct, to a diverse set of language and decision making tasks and demonstrate its effectiveness over state-of-the-art baselines, as well as improved human interpretability and trustworthiness over methods without reasoning or acting components. Concretely, on question answering (HotpotQA) and fact verification (Fever), ReAct overcomes issues of hallucination and error propagation prevalent in chain-of-thought reasoning by interacting with a simple Wikipedia API, and generates human-like task-solving trajectories that are more interpretable than baselines without reasoning traces. On two interactive decision making benchmarks (ALFWorld and WebShop), ReAct outperforms imitation and reinforcement learning methods by an absolute success rate of 34% and 10% respectively, while being prompted with only one or two in-context examples. Project site with code: https://react-lm.github.io