New Persuasion Benchmark Pits LLMs Against Each Other — Grok Is the Hardest to Sway
A novel benchmark that tests how easily one LLM can persuade another to change its answers finds Grok 4.20 Beta nearly immovable, while Xiaomi and Gemini models flip under pressure.
A new LLM evaluation methodology is testing something benchmarks rarely measure: how easily a model can be talked out of its own answers. As @LechMazur detailed, the persuasion benchmark pits models against each other in adversarial dialogues, measuring both a model's ability to persuade and its resistance to being persuaded. The results are striking: Grok 4.20 Beta with reasoning enabled is "nearly immovable on average," while models from Xiaomi and Google's Gemini family proved far more susceptible to changing their positions under pressure.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.