# Do LLMs Benefit From Their Own Words?

**Authors:** Jenny Y. Huang, Leshem Choshen, Ramon Astudillo, Tamara Broderick, Jacob Andreas  
**arXiv:** [2602.24287](https://arxiv.org/abs/2602.24287)  
**Date:** 2026-02-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2602.24287v1/

## Key points

- Authors: Jenny Y. Huang, Leshem Choshen, Ramon Astudillo, Tamara Broderick, Jacob Andreas
- Study compares traditional full-context prompting vs user-turn-only prompting in multi-turn interactions
- Removing prior assistant responses does not significantly affect response quality on many turns
- Reduction in cumulative context lengths by up to 10 times observed
- 36.4% of multi-turn conversations contain self-contained prompts
- User-turn-only prompting can outperform full context due to context pollution causing errors, hallucinations, or stylistic inconsistencies
- Proposal of a context-filtering approach to enhance response quality and decrease memory consumption

## Summary

The impact of conditioning large language models on their own prior responses in multi-turn interactions is investigated by authors Jenny Y. Huang, Leshem Choshen, Ramon Astudillo, Tamara Broderick, and Jacob Andreas. The study compares the traditional full-context prompting approach with a user-turn-only prompting method that excludes previous assistant responses across various open reasoning models and a state-of-the-art model. Interestingly, the results show that removing prior assistant responses does not significantly affect response quality on many turns. This leads to a reduction in cumulative context lengths by up to 10 times. The analysis reveals that a considerable portion (36.4%) of multi-turn conversations contain self-contained prompts and many follow-up prompts can be answered using only the current user turn and previous user turns. In cases where the user-turn-only prompting outperforms full context, it is attributed to instances of context pollution where models overly rely on their past responses resulting in errors, hallucinations or stylistic inconsistencies propagating through subsequent turns. Based on these findings, the researchers propose a context-filtering approach that selectively omits assistant-side history to enhance response quality while decreasing memory consumption. This study highlights how omitting assistant history can improve performance in large language models during multi-turn interactions and emphasizes the importance of considering contextual dependencies in dialogue systems for more effective communication outcomes.