2026· Annual Meeting of the Association for Computational Linguistics· pp. 877-887· 0 citations· 31 references
Computer Science
TL;DR
This thesis proposes a framework for trajectory-level confidence analysis in the tool-use agent setting and expects the framework to expose early warning signals for agent failure and offer interpretable diagnostic tools for understanding when and why LLM agents lose confidence.
Abstract
Large language model (LLM) agents that invoke external tools must make sequences of interdependent decisions, yet existing uncertainty quantification (UQ) methods treat each step in isolation, ignoring how confidence evolves and compounds across a full task trajectory. We propose a framework for trajectory-level confidence analysis in the tool-use agent setting. The thesis pursues three aims: (1) estimating action-level confidence by adapting step-wise UQ to the heterogeneous think-act-observe cycles of tool-using agents; (2) aggregating the diverse action space into semantically coherent action types to enable meaningful trajectory-level analysis; and (3) discovering temporal patterns in the resulting confidence trajectories that reliably predict task success or failure. We ground the work in standard tool-use benchmarks and expect the framework to expose early warning signals for agent failure and offer interpretable diagnostic tools for understanding when and why LLM agents lose confidence, with improved calibration of multi-step agentic pipelines as a secondary benefit.
UQ methods developed for single generations should be revalidated at the trajectory level, with careful attention to the consistency measurement, aggregator choice, and computational budget.
Experimental results show that RUPA consistently outperforms existing UQ methods by providing more accurate uncertainty estimates, enabling earlier failure detection, and improving uncertainty-guided agent execution across diverse agent tasks.
Reliable web-agent monitoring is difficult when model-internal uncertainty signals such as token logits are unavailable. In this work, we study prefix-level risk prediction for web agents using observable trajectory signals: given an evolving prefix, estimate whether the current execution remains on track or is tending toward failure. We derive two observable trajectory representations: Macro features summarize cross-step agent--environment behavior and feedback, while Micro features measure the consistency of intention, action, and anticipated state change through repeated black-box queries. Instead of inheriting the final result label, we label the first critical error that remains uncorrected in the observed continuation and is associated with final failure as a key-step boundary, preserving valid early prefixes of failed trajectories as on track. Across WebArena-Lite and Online Mind2Web web agent benchmarks with five open- and closed-source backbones, observable trajectory signals are competitive with internal-signal baselines. The resulting predictors also support early intervention under fixed false-cut budgets and transfer across held-out website categories. These findings show that observable trajectory signals support valuable risk prediction abilities.
Si-Tong Pan, Yi-Peng Shen, Yilin Lu et al.· 0 citations
OODA-Tool, a typed closed-loop policy designed to mitigate state preservation from action realization, consistently improves task success across model sizes, with larger gains on smaller models and on tasks whose actions depend strongly on information accumulated across turns and prior tool results.
Rongfeng Guo, Yin-Xuan Huang, Yusen Wu et al.· 0 citations
As large language models (LLMs) are deployed as autonomous agents, safety failures increasingly involve consequential actions. We study agentic misalignment, where agents take harmful actions under goal conflicts and pressures. Using chain-of-thought (CoT) monitoring, we find that harmful execution is often preceded by intent signals in reasoning. However, post-hoc CoT labels are too coarse to show how intent changes during generation. We introduce INTENT-AS-A-TOOL, an approach that adds intent-targeted tools to give the model a dedicated channel for expressing commitment to a target behavior. The probability of calling an intent tool provides a judge-free, fine-grained signal of the model's tendency to pursue that behavior. Our results show that INTENT-AS-A-TOOL complements CoT monitoring, expands post-hoc CoT labels into dense trajectories, and identifies critical steps for online intervention. These findings suggest that action preferences are useful for tracking agentic misalignment during reasoning. Our code and data are accessible: https://github.com/RebeccaZhang22/intent-as-a-tool.
Yutong Zhang, Jianshuo Dong, Peng Xu et al.· 0 citations
This work introduces the \method (\methodshort), a self-evolving critic framework in which an LLM critic accumulates evidence from its own past judgments and their observed consequences, reducing ECE by up to $54\% relative to the strongest training-free baseline.
Yaopei Zeng, Congchao Wang, JianHang Chen et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
