Reproduce “Can Mechanistic Interpretability Help Explain LLM Agent Failures? A Case Study on Web Element Hallucination” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mohamed Aghzal, Mor Geva, Gregory J. Stein, Ziyu Yao
The authors use mechanistic interpretability to detect and steer hallucinated actions in Qwen2.5-7B-Instruct. A linear probe predicts unavailable actions before generation, reaching test AUC 0.80 at layer 18; a selector-start direction fixes 85.4% of confirmed hallucinations and breaks 7.0% of valid actions. Under a numbered-Markdown format change, its fix rate falls to 58.1%, limiting claims of transfer or explanation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.