Mechanistic interpretability partially predicts web agent hallucinations but has limits

Can Mechanistic Interpretability Help Explain LLM Agent Failures? A Case Study on Web Element Hallucination

Artificial IntelligenceComputation and Language

Summary

Sometimes language-based web agents try to do actions that don’t actually exist on a website, which is called hallucination. The authors studied how to explain and fix these hallucinations using tools that look inside the agent’s reasoning process. They found a way to predict when hallucinations will happen and reduce them by changing the agent’s internal signals. However, this method only works partially and is sensitive to small changes, showing that it’s hard to fully understand why the agent makes these mistakes.

What this means in practice

  • For web automation engineers: Predict and reduce hallucinated web actions by inspecting and steering internal agent signals during deployment on live sites.
  • For software reliability teams: Diagnose limitations of interpretability tools to improve debugging of decision errors in language-based automation agents.

Authors

Mohamed Aghzal, Mor Geva, Gregory J. Stein, Ziyu Yao

Abstract

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.