Knowledge After Collapse: Context and Wording Shape Factual Accessibility in Language Models

Impact 7 arXiv cs.AI

Summary

  • Researchers introduce Knowledge Answer Divergence (KAD), an evaluation protocol measuring how language models handle false corrections during factual discussions.
  • KAD freezes conversations at the first adoption of a false target and branches probes to evaluate persistence, recovery, and abstention.
  • Evaluating eight local checkpoints and six hosted endpoints shows that post-adoption behavior heavily depends on conversational context and wording.
  • For example, a neutral-history branch yielded 42/46 correct responses in Llama 3.2 Instruct 3B compared to 16/46 under full history.

Why it matters

Understanding how language models persist in or recover from false-answer adoption helps developers design better conversational interventions and improve factual reliability.

Who should care