The brief decides the outcome (Microsoft Research)

The quality of an AI conversation is largely settled before the second turn begins.

A study from Microsoft Research and Salesforce, which won a best paper award at ICLR 2026, tested 15 language models across more than 200,000 simulated conversations and found that performance fell by an average of 39% when the details of a task arrived gradually through ordinary back-and-forth instead of in one complete brief. The models committed to early assumptions and later corrections did not bring them back, and the two reasoning models in the set failed the same way. The authors' summary is blunt: when a model takes a wrong turn in a conversation, it gets lost and does not recover.

That cuts against how most organisations picture supervising agents. The working assumption is that a person can steer as they go, correcting course when the output drifts. If this research holds, the steering is far weaker than it feels, and what was written at the start decides most of the outcome.

The caveats belong in the picture. These were simulated conversations rather than real ones, which are messier in ways that could make the effect larger or simply different, and model makers will try to train the problem away. The practical implication survives those caveats: a conversation that has gone wrong is usually better restarted with the full picture than patched.

For managers this moves the skill from reviewing outputs to writing instructions, and it places accountability with whoever wrote the brief. Who in your organisation has been taught to brief an AI system properly, and is anyone checking the quality of those briefs?

#AI #Productivity #Leadership

Source: Microsoft Research, "LLMs Get Lost In Multi-Turn Conversation" (ICLR 2026 best paper) https://lnkd.in/gyVEcUiK