Collaborative Theory of Mind: Learning to Self-Revise by Anticipating User Responses
Large language models (LLMs) demonstrate strong individual task performance, yet still struggle to collaborate with users over extended, multi-turn interactions, for example by attempting solutions before understanding the user's needs or by failing to revise mistaken assumptions after receiving…
This article introduces Collaborative Theory of Mind (CToM), a framework designed to improve large language models' (LLMs) ability to collaborate with users during extended, multi-turn interactions by anticipating user responses and self-revising based on those predictions.
What happened
The article presents CToM, a novel approach to enhance LLM collaboration by training models to predict user feedback and adjust their responses accordingly. The method was tested across mathematical reasoning, coding, and writing tasks, showing measurable improvements in accuracy and interaction efficiency compared to baseline approaches.
Why it matters
For AI developers and technology professionals, this research addresses a critical challenge in human-AI collaboration: enabling models to adapt dynamically during extended interactions. Improved collaboration efficiency can lead to more effective AI tools for complex problem-solving and user-centric applications.
- Anticipating user responses enables LLMs to self-revise and improve collaboration
- Combining prediction objectives with policy optimization enhances both task performance and interaction efficiency
- Improved user response prediction directly correlates with better collaborative outcomes
- Efficiency gains in interactions could reduce user effort in AI-assisted workflows
- This approach addresses a key limitation in current LLM capabilities for extended dialogues
The bigger picture
The framework demonstrates practical benefits for AI systems requiring sustained user interaction, potentially reducing the number of back-and-forth exchanges needed to complete tasks while maintaining or improving accuracy.
CToM presents a significant advancement in AI collaboration by enabling LLMs to dynamically adapt through predictive self-revision, offering measurable improvements in both task accuracy and interaction efficiency across multiple domains.
Technical details
- User response prediction as a training objective to guide internal revisions in LLMs
- Group-relative policy optimization (GRPO) for balancing task performance and interaction efficiency
- Negative log-likelihood minimization for predicting user replies
- Multi-turn interaction frameworks for mathematical reasoning, coding, and writing tasks
FAQ
What is Collaborative Theory of Mind (CToM)?
CToM is a framework that trains large language models to anticipate user responses during interactions, using these predictions to guide internal revisions and improve collaborative performance.
How does CToM improve AI collaboration?
By predicting user feedback during interactions, CToM enables LLMs to self-revise before presenting responses, leading to more accurate and efficient collaboration compared to traditional approaches.
What are the key results of the CToM framework?
CToM achieved 79.6% accuracy in mathematical reasoning tasks (up from 74.1%), reduced user reply requirements by 8-17% in coding and writing tasks, and demonstrated improved user response prediction accuracy through training.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv - cs.AI on Oct 12, 2026.
Comments
Sign in to join the discussion.