Research map / Large language models / Language, safety and interpretability

Mutable Transcripts: Mitigating Context Pollution through Editable Conversation State

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes mutable transcripts, where users revise chat history via natural-language edit requests that rewrite the full transcript, and evaluates it with a prototype on Gemini 2.5, a 17-person within-subjects user study, and a small token-count transcript analysis.

Problem

Immutable, append-only chat transcripts accumulate outdated or contradictory turns (context pollution) that keep influencing LLM responses after user intent has changed, and current interfaces offer no way to revise earlier history beyond editing the last message or branching.

Approach

Treat the conversation transcript as editable state: a natural-language edit instruction is applied to the full history by an LLM, which rewrites all user and assistant turns, replacing the displayed and future-context transcript instead of appending corrections.

Key results

• In the 17-participant within-subjects study, the mutable transcript condition was rated significantly better on all five Likert measures (all p<0.001), with 88.2% saying they would use an Edit History feature.
• In the transcript analysis, mutable transcripts reduced average turns from 8 to 4 (-50%) and tokens from 581 to 238 (-59%) across three scenarios.
• Baseline chat histories contained on average 71% obsolete tokens, versus none in the mutable transcripts.

Surprising findings

• Some participants reported that useful information was lost when earlier turns were overwritten during revision.

Limitations

• Small sample (n=17), mostly technology professionals, and short scripted tasks; no branching-baseline comparison.
• Full-transcript rewriting adds latency and cost that scale with conversation length, and overwriting can discard useful context.
• Security and accountability risks from retroactive edits are not addressed.

Experimental setup

Benchmarks: user study (n=17) on three tasks transcript token analysis

Metrics: Likert transcript cleanliness state clarity confidence restart intent ease of use turns tokens obsolete token percentage

Baselines: standard append-only chat ChatGPT Claude Gemini Duck.ai

Builds on: ChatGPT branching editing

Closest papers by meaning

• Exploring Conversational Adaptability: Assessing the Proficiency of Large Language Models in Dynamic Alignment with Updated User Intent (AAAI 2025 · similarity 0.86)
• LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (ICLR 2025 · similarity 0.86)
• Aligning Spoken Dialogue Models from User Interactions (ICML 2025 · similarity 0.85)
• SayNext-Bench: Why Do LLMs Struggle with Next-Utterance Anticipation? (NeurIPS 2026 · similarity 0.85)
• Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap (NeurIPS 2026 · similarity 0.85)
• Beyond Prompts: Dynamic Conversational Benchmarking of Large Language Models (NeurIPS 2024 · similarity 0.85)