Do LLMs Benefit From Their Own Words?

2026-02-27

AI Summary

This study investigates whether large language models genuinely require their own prior responses as part of the context in multi-turn dialogues. Through a systematic comparison between full-context prompting and a strategy that retains only user utterances, the authors find that most dialogue turns are self-contained and that including assistant history often introduces “contextual contamination.” To address this, they propose a selective filtering strategy that prunes redundant assistant responses from the dialogue history. Extensive experiments across multiple open-source and state-of-the-art models demonstrate that this approach reduces context length by up to 10× without compromising response quality, substantially lowering memory overhead while even improving generation performance.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• large language models
• multi-turn conversations
• context pollution
• assistant responses
• prompting

Innovation

• context filtering
• multi-turn dialogue
• assistant history omission
• context pollution
• prompt efficiency