LLMs Get Lost In Multi-Turn Conversation

LLMs Get Lost In Multi-Turn Conversation asks whether a model that solves a complete one-shot instruction remains dependable when the same requirements arrive over several turns. Read it to see how instruction sharding holds the intended task fixed, why best-case aptitude hides a 112% rise in unreliability, which recap and temperature interventions fall short, and where an idealized simulated user limits the result.

Reading focus: How Full, Concat, and Sharded delivery separate instruction rewriting from multi-turn information timing. Why repeated simulations need both an aptitude measure and an unreliability range instead of one average score. How to read the 39% average degradation, partial mitigations, episodic translation control, and simulator limits without turning them into claims about natural user behavior.

ICLR 2026 Outstanding Paper. Philippe Laban et al.. 36 min read, easy difficulty.