Do LLMs Benefit From Their Own Words?

TL;DR

This study investigates whether large language models benefit from including their own previous responses in multi-turn conversations, finding that omitting assistant history often does not harm and can even improve response quality.

Contribution

The paper demonstrates that excluding assistant responses from context can reduce memory usage and sometimes enhance response quality, challenging standard multi-turn prompting practices.

Findings

Omitting assistant responses does not affect response quality in many cases.

Removing assistant history can reduce context length by up to 10x.

Context pollution from previous responses can negatively impact model outputs.

Abstract

Multi-turn interactions with large language models typically retain the assistant's own past responses in the conversation history. In this work, we revisit this design choice by asking whether large language models benefit from conditioning on their own prior responses. Using in-the-wild, multi-turn conversations, we compare standard (full-context) prompting with a user-turn-only prompting approach that omits all previous assistant responses, across three open reasoning models and one state-of-the-art model. To our surprise, we find that removing prior assistant responses does not affect response quality on a large fraction of turns. Omitting assistant-side history can reduce cumulative context lengths by up to 10x. To explain this result, we find that multi-turn conversations consist of a substantial proportion (36.4%) of self-contained prompts, and that many follow-up prompts provide sufficient instruction to be answered using only the current user turn and prior user turns. When analyzing cases where user-turn-only prompting substantially outperforms full context, we identify instances of context pollution, in which models over-condition on their previous responses, introducing errors, hallucinations, or stylistic artifacts that propagate across turns. Motivated by these findings, we design a context-filtering approach that selectively omits assistant-side context. Our findings suggest that selectively omitting assistant history can improve response quality while reducing memory consumption.