Mutable Transcripts: Mitigating Context Pollution through Editable Conversation State

TL;DR

The paper proposes mutable transcripts, a new chat interaction model where users can edit past conversation turns using natural language, allowing the conversation history to evolve dynamically. This approach addresses context pollution by enabling the removal or correction of outdated information, improving alignment with shifting user intent. A prototype was evaluated in a user study and analysis showed improved clarity, confidence, and reduced need to restart conversations.

Why It Matters

Current chat systems suffer from context pollution due to static conversation histories, which can mislead models and degrade interaction quality. Mutable transcripts offer a practical way to maintain accurate, up-to-date conversational state, enhancing usability and reducing cognitive load for users. This could lead to more natural, efficient, and reliable LLM interactions in real-world applications.

Key Findings

• Users significantly preferred mutable transcripts over standard chat in terms of clarity, confidence, and ease of use.
• Participants reported a lower intent to restart conversations when using mutable transcripts.
• Analysis of conversation transcripts showed reduced length and elimination of obsolete context when edits were applied.