Do LLMs Benefit From Their Own Words?
when a model's own earlier response carries errors, hallucinations or style into later turns
Jenny Y. Huang, Leshem Choshen, Ramon Astudillo, Tamara Broderick, Jacob Andreas · 2026
In one sentence
Deleting all prior assistant responses from multi-turn chat histories on WildChat and ShareLM often leaves response quality unchanged while cutting cumulative context by 5-10x, because 36.4% of in-the-wild turns are self-contained and past responses can pollute later ones.
Abstract
In multi-turn conversations, large language models typically condition on the full conversation history: both past user prompts and assistant responses. We revisit this design choice by comparing full-context prompting to four alternative, substantially-reduced context configurations. Analyzing in-the-wild multi-turn conversations across three open reasoning and one state-of-the-art model, we find that response quality is largely preserved under aggressive context filtering: replacing all prior assistant turns with one-sentence summaries or keeping only the most recent user--assistant exchange often matches storing full context in performance while using roughly 8x less context. To understand this result, we observe that a substantial fraction of user turns (36.4%) in multi-turn conversations are self-contained and that many follow-up turns can be addressed by seeing only the immediately preceding user--assistant exchange. Furthermore, we find that when models condition on their own past responses, this can lead to context pollution, a phenomenon in which reasoning errors, hallucinations, or stylistic artifacts propagate across turns. Motivated by these findings, we design a context-filtering approach that selectively omits the assistant-side history. Taken together, these findings suggest moving away from storing full dialogue transcripts and instead retaining only what is relevant.
Questions this paper answers
- If you delete a chatbot's own earlier replies from the conversation history, do its later answers get worse?
- How does assistant-turn omission in the dialogue history affect judged response quality and on-topic rate across instruction-tuned LLMs?
- How do I trim assistant turns out of a multi-turn prompt without degrading the next response?
- Should I keep past model replies in my chat history, or send only the user turns?
- Omitting all prior assistant responses preserves average response quality for DeepSeek-R1-Distill-Llama-8B and GPT-OSS-20B, while Qwen3-4B and GPT-5.2 lose some quality relative to full context. Responses stay on-topic under omission for all four models.
Holds for: 300 real-world multi-turn conversations sampled from WildChat and ShareLM, 5-10 rounds each, filtered to English technical chats; GPT-5 as pairwise LLM judge seeing prior user and assistant turns.
- When the GPT-5 judge is shown only prior user-side turns, assistant-omitted responses match or beat full-context responses on both quality and on-topic dimensions for all four models. That reverses the full-context judge's preference for Qwen3-4B and GPT-5.2.
Holds for: Same 300 WildChat and ShareLM conversations and same four models (Qwen3-4B, DeepSeek-R1-Distill-Llama-8B, GPT-OSS-20B, GPT-5.2); a judge without assistant history cannot resolve prompts that explicitly reference an earlier response.
- How many characters of prompt do you save in a long chat by keeping only what the user wrote?
- What context-length reduction does user-turn-only prompting yield over full dialogue history by round 8?
- How do I cut token spend on a growing chat history without truncating or summarizing?
- Will dropping assistant turns actually make my multi-turn requests meaningfully cheaper?
- Full-context histories grow to roughly 25,000-55,000 characters by round 8, whereas user-turn-only context stays near-constant at 5,000-10,000 characters, a 5 to 10x reduction in context usage.
Holds for: Character counts over 5-10 round WildChat and ShareLM technical conversations; measured in characters rather than tokens, and does not include generated response length.
- In real conversations with a chatbot, how often is a person's next message a brand-new request rather than a follow-up?
- What proportion of non-initial user turns in in-the-wild human-LLM logs are self-contained versus dependent on prior assistant output?
- How do I tell whether a multi-turn chat corpus really tests long-context multi-turn reasoning?
- Can I trust WildChat-style logs as a benchmark for multi-turn context dependence?
- In real-world multi-turn chats, 36.4% of non-initial user turns are self-contained new asks, 30.5% are follow-ups with concrete feedback, and 33.1% are follow-ups referencing an earlier turn without actionable feedback.
Holds for: GPT-5 as automated annotator over the sampled WildChat and ShareLM English technical conversations of 5-10 rounds; the 33.1% is an upper bound on assistant-dependence.
- "Do LLMs Benefit From Their Own Words?" argues that in-the-wild multi-turn chat logs are weak benchmarks for long-context multi-turn reasoning, and calls for corpora curated for genuine multi-turn dependence. The reason is that a large share of turns in real chat logs do not depend on earlier assistant responses.
Holds for: Argued from English technical (math and coding keyword) conversations of 5-10 rounds in WildChat and ShareLM; agentic settings with tool outputs and scratchpads are discussed but not measured.
- Which kinds of user messages still need the chatbot's earlier answer in the history, and which do not?
- Does the benefit of retaining assistant history differ between new-ask turns and follow-up turns?
- How do I decide per turn whether a user message needs the assistant's prior response?
- If most of my users ask follow-up questions, is stripping assistant history still safe for me?
- For Qwen3-4B and GPT-5.2, full-context and assistant-omitted prompting perform comparably on new-ask turns, while full context helps most on follow-up turns; assistant-omitted context still wins roughly 40% of follow-up comparisons for Qwen3-4B and 30% for GPT-5.2.
Holds for: The two models for which uniform omission lowered overall quality; win rates averaged over the quality and on-topic dimensions, ties handled as in the pairwise judge protocol.
- In real-world multi-turn chats, 36.4% of non-initial user turns are self-contained new asks, 30.5% are follow-ups with concrete feedback, and 33.1% are follow-ups referencing an earlier turn without actionable feedback.
Holds for: GPT-5 as automated annotator over the sampled WildChat and ShareLM English technical conversations of 5-10 rounds; the 33.1% is an upper bound on assistant-dependence.
- Can a chatbot's own earlier mistakes in the conversation get repeated into later answers?
- What does context pollution from retained assistant turns look like in multi-turn LLM conversations?
- How do I stop a wrong assumption or hallucination from an earlier reply carrying into later turns?
- Is keeping my model's previous outputs in the prompt risking repeated hallucinations and stale code?
- 5 conversations in which assistant-omitted responses scored far above full context exhibit context pollution. The cases are UMAP arguments carried into t-SNE code, hallucinated book titles repeated across turns, a misattributed NBER citation, tutorial style overriding a reflection request, and a reversed temperature formula.
Holds for: Cases surfaced by sorting rounds by 1-10 judge score gap (AO minus FC) and reviewing the largest positive gaps; illustrative, not a frequency estimate, and includes GPT-5.2 as well as smaller open models.
- Can a small model learn when to keep the chatbot's earlier replies and when to throw them away?
- Can a per-turn classifier select between full and assistant-omitted context, and is judge preference predictable from turn features?
- How do I build a per-turn policy that drops assistant history only when it is safe?
- Is adaptive per-turn context filtering worth implementing, or will I lose quality against always sending full history?
- A per-turn classifier choosing between full and assistant-omitted context retains over 95% of full-context-only win-or-tie performance, and matches full-context-only at about 70% of the context consumption. It beats a heuristic that omits assistant history on all new-ask turns.
Holds for: GPT-5.2 only; L1-regularized logistic regression over round metadata, prompt category and PCA-reduced text-embedding-3-large embeddings; ties counted as wins, and the heuristic baseline is measured on a 20% held-out subset so it may improve with more data.
- Predicting whether the judge will prefer full context over assistant-omitted context is hard: the L1-regularized logistic regression reaches only a 5-fold cross-validated F1 of 0.6106 ± 0.0119. None of the top 20 features are significant at the 5% level.
Holds for: GPT-5.2 responses on the sampled WildChat and ShareLM conversations; features are round metadata, prompt category, and 20 principal components each of prompt and history embeddings explaining 38.0% and 51.5% of variance.
- Is it better to replace a chatbot's past answers with one-sentence summaries than to keep them in full?
- How does one-sentence self-summarization of assistant turns compare with full history and with full omission on Lost-in-Conversation and WildChat?
- How do I compress assistant turns in a chat history instead of deleting them outright?
- Should I summarize my model's earlier replies rather than dropping them or keeping them verbatim?
- Replacing each assistant response with a one-sentence self-summary improves response quality over full context for both DeepSeek-R1-Distill-Llama-8B and Qwen3-4B on both Lost-in-Conversation and WildChat. Fully omitting assistant turns helps on Lost-in-Conversation but is mixed on WildChat.
Holds for: 2 models and 2 datasets, using an earlier pairwise judge pipeline that compared final responses only, so numbers are not directly comparable to the main-text 1-10 scoring runs.
- How much can you trust a big model's scoring of which chat answer is better, and does what it sees change its verdict?
- How well does the GPT-5 judge agree with manual annotation, and does restricting the judge to user-side turns flip the preference between context conditions?
- How do I set up an LLM judge for comparing responses under different context conditions without biasing it toward the longer history?
- Can I rely on LLM-as-judge results comparing full-context and assistant-omitted responses?
- The GPT-5 judge agreed with an author's manual verdict on 54 of 60 quality judgments (90.0%) and 55 of 60 on-topic judgments (91.7%).
Holds for: 15 sampled judgments per model across the 4 models, annotated by one of the authors; no larger-scale or multi-annotator human study was run.
- When the GPT-5 judge is shown only prior user-side turns, assistant-omitted responses match or beat full-context responses on both quality and on-topic dimensions for all four models. That reverses the full-context judge's preference for Qwen3-4B and GPT-5.2.
Holds for: Same 300 WildChat and ShareLM conversations and same four models (Qwen3-4B, DeepSeek-R1-Distill-Llama-8B, GPT-OSS-20B, GPT-5.2); a judge without assistant history cannot resolve prompts that explicitly reference an earlier response.
- Is there a study that asks whether keeping a chatbot's own past replies in the prompt is actually worth it?
- What work evaluates assistant-history retention on in-the-wild human-LLM logs rather than synthetic multi-turn dialogues?
- What should I read before designing context management for a multi-turn chat or agent system?
- Which paper should I cite if I want to argue that real chat logs are weak multi-turn benchmarks?
- Huang et al.'s "Do LLMs Benefit From Their Own Words?" questions the default assumption of multi-turn chat and agentic context management that retaining a model's own past responses reliably helps. It tests that assumption on in-the-wild human-LLM logs rather than synthetic dialogues.
Holds for: As of the February 2026 preprint; prior turn-level context-editing work such as ERGO evaluated on synthetic conversations, and earlier conversational-QA findings about irrelevant turns concerned human-human histories.
- "Do LLMs Benefit From Their Own Words?" argues that in-the-wild multi-turn chat logs are weak benchmarks for long-context multi-turn reasoning, and calls for corpora curated for genuine multi-turn dependence. The reason is that a large share of turns in real chat logs do not depend on earlier assistant responses.
Holds for: Argued from English technical (math and coding keyword) conversations of 5-10 rounds in WildChat and ShareLM; agentic settings with tool outputs and scratchpads are discussed but not measured.
- Do big commercial chatbots get thrown off by their own earlier replies, or is that only a small-model problem?
- Which models were evaluated for assistant-turn omission, and does over-conditioning on prior assistant output appear in frontier models too?
- How do I check whether the model I use is being misled by its own earlier turns?
- I run a frontier model in production, is over-conditioning on its own past answers something I have to worry about?
- Omitting all prior assistant responses preserves average response quality for DeepSeek-R1-Distill-Llama-8B and GPT-OSS-20B, while Qwen3-4B and GPT-5.2 lose some quality relative to full context. Responses stay on-topic under omission for all four models.
Holds for: 300 real-world multi-turn conversations sampled from WildChat and ShareLM, 5-10 rounds each, filtered to English technical chats; GPT-5 as pairwise LLM judge seeing prior user and assistant turns.
- 5 conversations in which assistant-omitted responses scored far above full context exhibit context pollution. The cases are UMAP arguments carried into t-SNE code, hallucinated book titles repeated across turns, a misattributed NBER citation, tutorial style overriding a reflection request, and a reversed temperature formula.
Holds for: Cases surfaced by sorting rounds by 1-10 judge score gap (AO minus FC) and reviewing the largest positive gaps; illustrative, not a frequency estimate, and includes GPT-5.2 as well as smaller open models.
Claims and scope
- Omitting all prior assistant responses preserves average response quality for DeepSeek-R1-Distill-Llama-8B and GPT-OSS-20B, while Qwen3-4B and GPT-5.2 lose some quality relative to full context. Responses stay on-topic under omission for all four models. (Figure 2)
Scope: 300 real-world multi-turn conversations sampled from WildChat and ShareLM, 5-10 rounds each, filtered to English technical chats; GPT-5 as pairwise LLM judge seeing prior user and assistant turns.
- When the GPT-5 judge is shown only prior user-side turns, assistant-omitted responses match or beat full-context responses on both quality and on-topic dimensions for all four models. That reverses the full-context judge's preference for Qwen3-4B and GPT-5.2. (Figure 9)
Scope: Same 300 WildChat and ShareLM conversations and same four models (Qwen3-4B, DeepSeek-R1-Distill-Llama-8B, GPT-OSS-20B, GPT-5.2); a judge without assistant history cannot resolve prompts that explicitly reference an earlier response.
- Full-context histories grow to roughly 25,000-55,000 characters by round 8, whereas user-turn-only context stays near-constant at 5,000-10,000 characters, a 5 to 10x reduction in context usage. (Figure 6 (Section 2.2))
Scope: Character counts over 5-10 round WildChat and ShareLM technical conversations; measured in characters rather than tokens, and does not include generated response length.
- In real-world multi-turn chats, 36.4% of non-initial user turns are self-contained new asks, 30.5% are follow-ups with concrete feedback, and 33.1% are follow-ups referencing an earlier turn without actionable feedback. (Section 2.3)
Scope: GPT-5 as automated annotator over the sampled WildChat and ShareLM English technical conversations of 5-10 rounds; the 33.1% is an upper bound on assistant-dependence.
- For Qwen3-4B and GPT-5.2, full-context and assistant-omitted prompting perform comparably on new-ask turns, while full context helps most on follow-up turns; assistant-omitted context still wins roughly 40% of follow-up comparisons for Qwen3-4B and 30% for GPT-5.2. (Figure 3 (Section 2.4))
Scope: The two models for which uniform omission lowered overall quality; win rates averaged over the quality and on-topic dimensions, ties handled as in the pairwise judge protocol.
- 5 conversations in which assistant-omitted responses scored far above full context exhibit context pollution. The cases are UMAP arguments carried into t-SNE code, hallucinated book titles repeated across turns, a misattributed NBER citation, tutorial style overriding a reflection request, and a reversed temperature formula. (Table 1 (full annotations in Appendix A.12))
Scope: Cases surfaced by sorting rounds by 1-10 judge score gap (AO minus FC) and reviewing the largest positive gaps; illustrative, not a frequency estimate, and includes GPT-5.2 as well as smaller open models.
- A per-turn classifier choosing between full and assistant-omitted context retains over 95% of full-context-only win-or-tie performance, and matches full-context-only at about 70% of the context consumption. It beats a heuristic that omits assistant history on all new-ask turns. (Figure 5 (Section 3.2))
Scope: GPT-5.2 only; L1-regularized logistic regression over round metadata, prompt category and PCA-reduced text-embedding-3-large embeddings; ties counted as wins, and the heuristic baseline is measured on a 20% held-out subset so it may improve with more data.
- Predicting whether the judge will prefer full context over assistant-omitted context is hard: the L1-regularized logistic regression reaches only a 5-fold cross-validated F1 of 0.6106 ± 0.0119. None of the top 20 features are significant at the 5% level. (Table 2 (Appendix A.13))
Scope: GPT-5.2 responses on the sampled WildChat and ShareLM conversations; features are round metadata, prompt category, and 20 principal components each of prompt and history embeddings explaining 38.0% and 51.5% of variance.
- Replacing each assistant response with a one-sentence self-summary improves response quality over full context for both DeepSeek-R1-Distill-Llama-8B and Qwen3-4B on both Lost-in-Conversation and WildChat. Fully omitting assistant turns helps on Lost-in-Conversation but is mixed on WildChat. (Figure 7)
Scope: 2 models and 2 datasets, using an earlier pairwise judge pipeline that compared final responses only, so numbers are not directly comparable to the main-text 1-10 scoring runs.
- The GPT-5 judge agreed with an author's manual verdict on 54 of 60 quality judgments (90.0%) and 55 of 60 on-topic judgments (91.7%). (Appendix A.6)
Scope: 15 sampled judgments per model across the 4 models, annotated by one of the authors; no larger-scale or multi-annotator human study was run.
- Huang et al.'s "Do LLMs Benefit From Their Own Words?" questions the default assumption of multi-turn chat and agentic context management that retaining a model's own past responses reliably helps. It tests that assumption on in-the-wild human-LLM logs rather than synthetic dialogues.
Scope: As of the February 2026 preprint; prior turn-level context-editing work such as ERGO evaluated on synthetic conversations, and earlier conversational-QA findings about irrelevant turns concerned human-human histories.
- "Do LLMs Benefit From Their Own Words?" argues that in-the-wild multi-turn chat logs are weak benchmarks for long-context multi-turn reasoning, and calls for corpora curated for genuine multi-turn dependence. The reason is that a large share of turns in real chat logs do not depend on earlier assistant responses.
Scope: Argued from English technical (math and coding keyword) conversations of 5-10 rounds in WildChat and ShareLM; agentic settings with tool outputs and scratchpads are discussed but not measured.
Common misreadings
- Omitting assistant history is not shown to be universally free: for Qwen3-4B and GPT-5.2, uniformly dropping assistant responses lowers average judged quality under a judge that sees the full history.
- The 36.4% figure counts non-initial user turns classified as self-contained new asks by a GPT-5 annotator on filtered English technical conversations from WildChat and ShareLM; it is not a claim about all chat traffic.
- The context-pollution cases in Table 1 were surfaced by sorting for the largest score gaps favouring assistant-omitted context, so they establish that the failure mode exists and reaches GPT-5.2, not how often it occurs.
- The adaptive omission strategy is not a strong predictive model: its 5-fold cross-validated F1 is 0.6106 and no individual feature is significant, so its context savings come from a weak signal rather than a reliable per-turn prediction.
- The 5 to 10x reduction is in cumulative context characters over conversation rounds, not a measured speedup or dollar cost saving.
Terminology in this paper
- Assistant-Omitted (AO) context
- A prompting configuration in which every past assistant response in a multi-turn conversation is replaced by the placeholder "[Response provided]", so the model conditions only on prior user turns while the alternating user/assistant structure is preserved.
- context pollution
- The phenomenon in which a language model over-conditions on its own earlier responses, so errors, hallucinations or stylistic artifacts introduced in one turn propagate into later turns.
- New Ask
- A non-initial user turn that introduces a fully self-contained request, understandable without any prior conversation round.
- Follow-up with Feedback
- A user turn that gives concrete, actionable feedback on a prior assistant response, such as "use Python instead of Java for the code example".
- Follow-up without Feedback
- A user turn that refers to an earlier conversation round without any concrete instruction for revision, such as "reflect on your response".
How to cite
@misc{huang2026llmsbenefitwords,
title={Do LLMs Benefit From Their Own Words?},
author={Jenny Y. Huang and Leshem Choshen and Ramon Astudillo and Tamara Broderick and Jacob Andreas},
year={2026},
eprint={2602.24287},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2602.24287},
}
References
See the full reference list in the paper.