Mutable Transcripts: Mitigating Context Pollution through Editable Conversation State

Source facts

The preprint proposes allowing users to revise prior conversation turns through natural-language edit requests, treating the transcript as an editable representation of conversational state rather than an immutable sequence. A prototype was evaluated in a controlled user study with 17 participants; the abstract reports preference for mutable transcripts, reduced intent to restart conversations, and shorter conversations in the analyzed scenarios. The arXiv record states that the paper was accepted at NeurIPS 2026 and links source code and a prototype.

Analysis

The concept addresses a practical weakness of append-only chat history: outdated constraints can persist and influence later responses. Editable state could make correction explicit, but it also raises provenance and audit questions about what was changed and when.

Uncertainty/Doubt

• This is an arXiv preprint with an acceptance statement in its record. The small user study, interaction effects, and long-term behavior were not independently audited here.