Study probes why web agents hallucinate unavailable actions
A mechanistic interpretability case study finds a linear probe can predict Qwen2.5-7B-Instruct hallucinations before generation, but the explanation remains partial and fragile.
In plain words
Web agents sometimes try to click or use page elements that do not exist. Researchers tested whether interpretability tools, which inspect a model's internal computations, could explain these mistakes. A simple detector flagged errors early, but it did not fully explain why they happen.
What was announced
The arXiv paper examines one agentic failure: web agents hallucinating actions that are not available on a website. Applying a series of mechanistic analyses to Qwen2.5-7B-Instruct while it interacted with live websites, the authors report that a linear probe trained on the last prompt token could predict hallucinations before generation began, reaching 0.80 AUC. Steering along the probe direction corrected 85% of hallucinations, though it also disrupted 7% of valid actions.
The authors stress these results give only a partial explanation. The probe direction was sensitive to prompt formatting, losing much of its steering effect after a minor change in how actions were represented. Circuit identification also proved difficult because computation was spread across multiple components and positions, and the information the agent relied on varied between examples. The paper argues these observations challenge assumptions behind standard mechanistic interpretability methods for circuit discovery and feature identification, and outlines future directions for adapting such methods to LLM-based agent failures.
What it does not show
- The evidence is an abstract only; full methodology, datasets and caveats are not available for verification.
- Results cover a single model, Qwen2.5-7B-Instruct, and one failure type, so they may not generalize to other agents or errors.
- The authors themselves state the probe provides only a partial explanation, and its effectiveness depends on prompt formatting.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.