On the Tip of the Tongue: First-Token Cueing as a Behavioral Probe of Recall Failures in LLMs

The paper argues that many closed-book factual errors in large language models are first-token recall failures rather than missing knowledge. The model commits to a wrong first token, but when given the correct first token it completes the answer. The authors propose first-token cueing as a behavioral probe for distinguishing these failures from genuine knowledge gaps. The supplied text is truncated, so the full methodology and quantitative results are not available here.

Why it matters

If many errors stem from recall initiation rather than absent knowledge, this could guide evaluation methods and mitigation strategies aimed at improving factual accuracy.

Read with care

Source text is truncated and results are not shown; claims are not independently checked.