Microsoft paper: LLMs abandon correct answers when challenged wrong; Self-Disagreement training lifts hold rate to 89%

rohanpaul_ai · x · 2026-11-19

A Microsoft Research paper shows LLMs often give up correct answers under wrong counter-arguments — a distinct capability (dialectical competence) that accuracy training, more sampling, or self-reflection cannot fix. Their Self-Disagreement method trains the model to state an answer, write its strongest counter-case, rule on it, and answer again, rewarded by the same answer checker (pure RLVR, no reward model or human data). Base Qwen3-8B held correct answers 53.3% of the time against wrong challengers, Gemma-2 9B just 25.1%; after training both hit 89%, while accuracy-only training added nothing. Takeaway: if your model faces pushback from users or agents, test answer-holding separately — accuracy scores won't reveal this gap.

Related event: Microsoft Paper: LLMs Abandon Correct Answers When Wrongly Challenged(2 posts)→

Original post →

More from coding & agent

coding & agent channel →