Mutable Transcripts: Mitigating Context Pollution through Editable Conversation State

Dan Barry; Andrew Hines

cs.AISep 25, 2026

Mutable Transcripts: Mitigating Context Pollution through Editable Conversation State

Authors: Dan Barry, Andrew Hines

Organizations: School of Computer Science University College Dublin Dublin, Ireland

Abstract

Contemporary large language model (LLM) chat systems treat conversation history as an immutable sequence of turns that defines the model's working context. However, user intent in real interactions is not static: it evolves through correction, refinement, and shifting constraints. This mismatch between dynamic intent and static transcripts can result in context pollution, where outdated or irrelevant information persists and continues to influence subsequent responses. We introduce mutable transcripts, a new interaction paradigm that enables users to revise prior turns through natural language edit requests, allowing the conversation history itself to be updated rather than appended. This reframes the transcript from a passive record into an editable representation of conversational state. We present a working prototype that integrates transcript-level revision into a standard chat interface and evaluate its feasibility through a controlled user study (n=17) and an illustrative transcript analysis of representative interaction scenarios. Participants significantly preferred mutable transcripts over standard chat across measures of clarity, confidence, and ease of use, with reduced intent to restart conversations. Transcript analysis of representative user study conversations shows that mutable transcripts can reduce conversation length and eliminate obsolete retained context. These findings provide initial evidence that user-driven revision of conversational history can improve interaction quality and help maintain a more current representation of user intent. The source code and prototype can be accessed at https://github.com/QxLabIreland/ReChat

Figures & tables

Figure 1: Interface design: (A) At any point in the conversation, the user can engage “Edit History” mode and enter a prompt to modify the chat history. (B) The UI refreshes and shows that previous conversational turns are updating. (C) The UI refreshes with the updated conversation constraint.

Figure 2: System overview of mutable transcripts operation. In normal chat mode, user prompts are appended to an immutable JSON transcript. In edit mode, natural language edit instructions are applied to the full transcript, which is rewritten and replaced before being re-rendered in the chat interface.

Figure 3: Condition A vs. B combined preference for five evaluation metrics: transcript cleanliness, state clarity, confidence, restart intent, and ease of use. We see no impact due to order bias in A vs. B compared to B vs. A. The combined results show statistically significant preference for Condition B (mutable) vs. A across all questions, including a lower restart preference. Error bars represent 95% confidence intervals over participant responses.

Scenario	Method	Turns	Tokens	Obsolete Tokens	Obsolete %
Retroactive Correction	Baseline	8	422	236	56%
	Mutable	4	169	–	–
Constraint Injection	Baseline	6	608	353	58%
	Mutable	4	479	–	–
Context Pruning	Baseline	10	712	655	92%
	Mutable	4	66	–	–

Table 1: Transcript efficiency comparison between standard immutable chat histories (Baseline) and mutable transcripts (Mutable) across three representative revision scenarios. Mutable transcripts reduce conversation length and eliminate obsolete retained context arising from superseded corrections and constraints.

Figure 4: Current correction methods used by respondents (blue) and the percentage of respondents who would use a mutable transcript feature if it was available (orange). The majority of people (94.1%) send a follow-up message to perform corrections and 88.2% of people said they would use “Edit History” if it existed in a real chat interface.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Profession Type	Count	Percent
Researcher	9	52.9
Data Science / Machine Learning	4	23.5
Software Development	2	11.8
Other (Specified "Lecturer")	1	5.9
Business/Management	1	5.9

Appendix

Table 2: Participant professional or academic field (

N=17

Response	Count	Percent
Multiple times a day	9	52.9
Several times a week	5	29.4
Once a day	2	11.8
Less than once a week	1	5.9

Appendix

Table 3: Frequency of LLM chatbot use for work, study, or personal tasks (

N=17

Option	Count	Percent
Gemini	16	94.1
ChatGPT (OpenAI)	12	70.6
Claude (Anthropic)	3	17.6
Microsoft Copilot	3	17.6
Other	3	17.6

Appendix

Table 4: LLM chatbots respondents were most familiar with or used most often (

N=17

respondents; multi-select).

Response	Count	Percent
Sometimes	8	47.1
Frequently	6	35.3
Never	2	11.8
Always or almost always	1	5.9

Appendix

Table 5: How often respondents felt the need to edit or modify a previous chatbot message or prompt (

N=17

Option	Count	Percent
Edit the previous message directly (if the feature is available)	6	35.3
Send a new follow-up message with the correction	16	94.1
Copy the previous prompt, modify it, and resend it as a new prompt	12	70.6
Start a completely new chat session	4	23.5
I usually don’t correct previous inputs	1	5.9

Appendix

Table 6: Current actions taken when changing a previous chatbot input (

N=17

respondents; multi-select).

Response	Count	Percent
Sometimes	6	35.3
Very often	5	29.4
Often	4	23.5
Rarely	2	11.8

Appendix

Table 7: Whether respondents had encountered similar situations in real chatbot use (

N=17

Response	Count	Percent
Likely	9	52.9
Very Likely	6	35.3
Neutral	1	5.9
Unlikely	1	5.9

Appendix

Table 8: Likelihood of using an “Edit History” feature if available in real chatbots (

N=17

LLM Context Management

Large language model context management studies how to select, compress, retrieve, and organize information within bounded model inputs.

See all 251 papers

Do LLMs Benefit From Their Own Words?
Feb 27, 2026Jenny Y. Huang, Leshem Choshen, Wei Sun +5LLM Context Management Context Compression
Department of Electrical Engineering and Computer Science, MIT · MIT–IBM Watson AI Lab · IBM Research
What if LLMs Ate Their Words: Causal History Effects in Multi-Turn Interaction
Sep 5, 2026Jinnan Li, Zheren Fu, Yue Wang +3LLM Context Management Causal Interventions in Language Models
School of Artificial Intelligence Jilin University Changchun, China · School of Cyber Science and Technology, University of Science and Technology of China Hefei, China · School of Information and Library Science, University of North Carolina at Chapel Hill Chapel Hill, North Carolina, USA
MT-OSC: Path for LLMs that Get Lost in Multi-Turn Conversation
Apr 9, 2026Jyotika Singh, Fang Tu, Miguel Ballesteros +6LLM Context Management
Oracle AI

Feb 27, 2026J/K move · Enter open · S save

Do LLMs Benefit From Their Own Words?

Jenny Y. Huang, Leshem Choshen, Wei Sun, Omar Khattab+4

Department of Electrical Engineering and Computer Science, MIT · MIT–IBM Watson AI Lab · IBM Research

LLM Context Management Context Compression

Open paper AskPDF arXiv

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.

Abstract

LLM Context Management

Explore similar work

Do LLMs Benefit From Their Own Words?

What if LLMs Ate Their Words: Causal History Effects in Multi-Turn Interaction

MT-OSC: Path for LLMs that Get Lost in Multi-Turn Conversation