# Multi-Turn Degradation

> Multi-turn degradation is the drop in model performance when a task's requirements are revealed across several conversation turns instead of all at once, because models commit to early assumptions and fail to recover.

- Identifier: PTL-0174
- Category: Failure Modes & Evaluation
- Canonical URL: https://protologue.com/t/multi-turn-degradation/
- Also known as: lost in conversation, multi-turn performance drop
- Introduced: 2025

## Description

A large 2025 study found that models that performed well on fully specified single-turn instructions did markedly worse in multi-turn settings, mainly through increased unreliability rather than a loss of capability. The practical advice is to consolidate requirements into a fresh prompt when a conversation goes off track.

## Related terms

- [Instruction Following](https://protologue.com/t/instruction-following/)
- [Context Compaction](https://protologue.com/t/context-compaction/)
- [Crescendo Attack](https://protologue.com/t/crescendo-attack/)
- [Context Rot](https://protologue.com/t/context-rot/)

## Sources

- Laban et al. (2025). LLMs Get Lost In Multi-Turn Conversation. https://arxiv.org/abs/2505.06120

## Cite this entry

Protologue. (2026). Multi-Turn Degradation. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0174). https://protologue.com/t/multi-turn-degradation/

License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)