Microsoft paper: LLMs abandon correct answers under wrong challenges, and accuracy training can't fix it
rohanpaul_ai · x · 2026-11-19
A new Microsoft paper by Soheil Abbasloo and Peng Cheng, "Benchmarks Don't Talk Back: Learning Through Self-Disagreement" (arXiv:2611.17563), shows LLMs frequently give up correct answers when challenged with wrong arguments.
Key findings:
- More accuracy training, more sampling, or self-reflection does not fix the problem
- The missing piece is a distinct capability the authors call "dialectical competence" — knowing when to hold a position and when to yield; it presupposes first-answer accuracy but doesn't follow from it
- Their fix, Self-Disagreement: the model states a position, generates the strongest case against it, rules on whether that case stands, and is trained on the ruling via RL from verifiable rewards — no second model, reward model, preference data, or human annotation
The capability must be trained rather than elicited, with direct implications for multi-agent debate and deployment scenarios far beyond benchmark conditions.
More from Models
- Nat Lambert rips White House plan to disclose 'Chinese model' use: where do you draw the line? — ivan_bezdomny · 2026-11-19
- Polymarket gives Google 79% odds of leading AI by year-end, OpenAI just 2.8% — Polymarket · 2026-11-19
- Heavy user pushes back on Astra 6.2 hate: big jump over 6.0, uses only 30% of quota — soumitrashukla9 · 2026-11-19
- Google buys bankrupt Spirit Airlines' internal workplace data for $10M to train AI models — Polymarket · 2026-11-19
- GPT-6.2 Astra tops Design Creativity Bench at 70.4, Sol close at 4.5x lower cost — sandersted · 2026-11-19
- Google's Logan Kilpatrick demos Gemini 4: games, Android apps and simulations from one prompt — Shruti_0810 · 2026-11-19