Multi-Turn Conversations Crash LLM Performance to 65%, Microsoft and Salesforce Paper Finds
A joint research paper tested leading LLMs in realistic multi-turn dialogues and found accuracy plummeted from 90% in single-turn to 65% — with compounding errors from forgotten context, false assumptions, and instruction drift.
A joint paper from Microsoft Research and Salesforce has quantified something that every developer building on LLMs has felt in production: models get dramatically worse as conversations get longer. According to @oliviscusAI, the study tested top-tier models across realistic multi-turn conversation scenarios and found that performance dropped from approximately 90% accuracy in single-turn exchanges to just 65% in multi-turn dialogues — a 25-percentage-point decline that the researchers attribute to compounding failures in context retention, assumption propagation, and instruction fidelity.
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.