Large language models (LLMs) excel at single‑turn tasks but often falter in extended multi‑turn collaborations, for example by proposing solutions before fully grasping user intent or by failing to amend mistaken assumptions after feedback. The Collaborative Theory of Mind (CToM) framework tackles these issues by explicitly training LLMs to anticipate the user’s next reply and to use that prediction to revise its own output before responding.
CToM operates in four steps: draft a response, predict the user’s likely reply, perform an internal revision guided by the prediction, and finally present the revised answer. During training, the actual user reply to each draft serves as the target; the model minimizes the negative log‑likelihood of that reply. This prediction objective is combined with Group‑Relative Policy Optimization (GRPO) to jointly improve task success and interaction efficiency.
Across mathematical reasoning, coding, and writing tasks, CToM outperforms all baselines. On a collaborative math reasoning benchmark, final accuracy rises from 74.1 % to 79.6 % compared with a naïve GRPO baseline. Ablation studies show that prediction supervision reduces the number of user replies by 8 % in coding and 17 % in writing while still boosting performance under the same reward signal.
Behavioral analysis reveals that prediction accuracy improves after training, and higher prediction fidelity correlates with larger gains in collaborative performance through self‑revision.
Review