Megadose The AI curve, compressed.

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

· ArXiv · AI/CL/LG ·
Steering the agent’s internal activations corrected 85% of hallucinated web actions, but disrupted 7% of valid actions.

Researchers studied Qwen2.5-7B-Instruct interacting with live websites and attempting actions those sites did not offer. A linear probe predicted hallucinations before generation began, with an AUC of 0.80. A minor change to how actions were represented in the prompt removed much of the steering effect. Computation was spread across components and positions, and the information used varied between examples, leaving only a partial explanation of the failures.

ArXiv · AI/CL/LG's note

score 5

Categories: Research