Reproduce “Exploring the generalization of LLM truth directions on conversational formats” 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
Timour Ichmoukhamedov, David Martens
Ichmoukhamedov and Martens investigate how well "truth directions" in Large Language Models generalize across varied conversational settings. They find that while linear probes effectively detect truthfulness in short, simple formats, performance significantly degrades in longer conversations where falsehoods appear earlier; adding a specific key phrase at the conversation's end restores high accuracy, improving detection from ~50-60% to ~80-90%.
Key insights
- The "truth direction" in LLMs is context-dependent and does not universally generalize across different conversational lengths or lie positions.
- LLMs may de-emphasize or "forget" truthfulness information in longer conversations unless explicitly prompted to maintain focus on it.
- Explicitly asking about truthfulness via a specific key phrase can effectively re-align the LLM's internal representations, making truthfulness discernible again.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.