Do LLMs Benefit From Their Own Words?

Abstract: Multi-turn interactions with LLMs typically retain the assistant's own past responses in the conversation history. In this work, we revisit this design choice by asking whether LLMs 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.

• The paper demonstrates that omitting assistant responses can maintain or even improve answer quality while significantly reducing context token usage.
• Methodical experiments on four LLMs revealed that adaptive context strategies achieve over 95% performance with only 70% of the full context usage.
• Empirical results indicate that selective retention helps mitigate error propagation and context pollution, enhancing efficiency and robustness in dialogue systems.

Analysis of "Do LLMs Benefit From Their Own Words?" (2602.24287)

Problem Formulation and Motivation

The paper examines a fundamental assumption in multi-turn conversational LLM applications: the necessity of retaining assistant-side turns in conversation history. Standard multi-turn interfaces typically concatenate past assistant and user turns, premised on the belief that full context improves downstream answer quality and continuity. However, empirical evidence for this assumption is limited, especially for LLM-generated (often error-prone or verbose) content. The authors interrogate whether LLMs' own previous outputs actually provide utility or can be detrimental for successive responses, especially with respect to computational efficiency, error propagation, and the context window.

Experimental Framework

A controlled analysis was performed on real-world, technical multi-turn conversations sourced from WildChat and ShareLM. Four LLMs were evaluated—Qwen3-4B, DeepSeek-R1-Distill-Llama-8B, GPT-OSS-20B, and GPT-5.2—under two prompting regimes:

• Full Context (FC): all prior user and assistant turns included
• Assistant-Omitted (AO): prior assistant turns replaced by placeholders, retaining structural alternation

Response quality and topic adherence were pairwise evaluated by a judge LLM (GPT-5), provided both full and AO context outputs as well as the conversation history.

Major Empirical Findings

1. Omission of Assistant-Side History

Removing previous assistant outputs has negligible or positive effects on response quality for a substantial fraction of turns. For open models, AO prompting preserved quality; for GPT-5.2 and Qwen3-4B, quality marginally decreased but remained acceptable. LLM-judges that themselves omit assistant context showed AO responses outperforming FC across all models. Cumulative context consumption is reduced by $5\times$–$10\times$ with AO strategy, mitigating inference slowdowns and context window saturation.

2. Self-Contained Turns and Prompt Structure

Multi-turn user prompts are often self-contained: 36.4% of prompts are classified as independent of assistant-side history. Moreover, most follow-up prompts either supply concrete feedback that can be satisfied using only user turns or reference prior instructions that are themselves actionable without assistant context. Only 33.1% reference prior assistant responses without actionable modification.

3. Context Pollution and Error Propagation

Models frequently over-condition on previous outputs, resulting in "context pollution." This manifests as propagation of hallucinated facts, erroneous code parameters, stylistic inertia, and misattributed citations. Cases are documented across domains and models, including frontier LLMs. Empirically, AO responses consistently outperform FC on rounds affected by context pollution.

4. Adaptive Context Management

An adaptive, per-turn context-strategy—selecting between FC and AO using logistic regression over prompt embeddings and metadata—retains 95% of FC-only performance with only 70% of the context usage. The classifier fuses prompt type, conversation embeddings, and round metadata. A simple heuristic ("Omit on New Ask") is inferior, further confirming the necessity for learned, fine-grained filtering. The paper highlights that the next logical advance is targeted inclusion of only referenced assistant turns, rather than binary FC/AO strategies.

Contradictory Claims and Numerical Outcomes

• Contradictory to assumptions: The paper directly counters the default belief that full-context always improves performance. Empirical win rates and judge scores show that indiscriminate retention of assistant turns can be neutral or harmful.
• Strong numerical claims:
  • Context reduction of up to $10\times$ with negligible impact on response quality for a majority of multi-turn prompts.
  • Adaptive context strategies achieve >95% FC-equivalent win rates at substantially reduced token usage.

Theoretical Implications

The findings challenge the necessity of universal context concatenation in dialogue models, demonstrating that multi-turn dependence is not inherent and is often overestimated. The study raises questions regarding the ability of LLM architectures to distinguish relevant from irrelevant context and the cognitive biases of LLMs toward their own outputs—a form of self-conditioning not present in human dialogue.

Broadly, this indicates that improved context management policies are critical, including automated pruning, targeted summarization, and relevance prediction. It also impacts research on long-context agents, tool-augmented reasoning, and conversational memory, suggesting that context policies should be dynamically aligned with conversation structure.

Practical Impact and Future Directions

• Efficiency: AO and adaptive strategies not only minimize context window requirements, but also enable more efficient deployment on limited-resources, reduce latency, and improve scalability for production chat systems.
• Robustness: Mitigating context pollution improves factual reliability—especially critical for technical, scientific, and agentic use cases.
• Agentic Systems: The paper notes that agent-based LLM systems (e.g., Cursor, Claude Code) already incorporate trajectory pruning and tool-output management. Incorporating more granular assistant-response filtering will further enhance agentic system robustness.
• Benchmark Recommendations: The research underscores the need for better real-world conversation benchmarks that accurately reflect genuine multi-turn dependence, facilitating advances in context-aware LLM engineering.

Future work should concentrate on learning selective retention policies at turn or artifact granularity, leveraging relevance models, and extending human evaluation coverage to validate automated judge reliability.

Conclusion

This research reevaluates foundational assumptions regarding context retention in multi-turn LLM conversations, establishing that assistant-side history is often non-essential and, under certain conditions, detrimental to downstream responses. The results substantiate the development of adaptive, selective context-management solutions that improve efficiency, minimize context pollution, and preserve quality across a wide spectrum of conversational interactivity. These findings inform both immediate system design and longer-term research on robust, memory-efficient, and contextually coherent LLM agents.