Strong solo answers do not make a good collaborator
The paper starts from a gap between individual task skill and multi-turn collaboration. A model can solve math, write code, or draft an article on its own, yet still attempt a solution before understanding what the user needs, make assumptions that require repeated correction, or fail to recover once its first approach goes wrong. The authors illustrate this with a writing request: an assistant without a model of the user produces a long draft that misses the requested balance of benefits and risks, and the user has to correct it twice.
The proposed fix is Collaborative Theory of Mind (CToM), a training framework rather than a prompting trick. Its premise is that anticipating the user's next reply gives the assistant a usable signal about its own draft before that draft is exposed. The user's actual reply to each draft becomes the training target for that prediction, so the model learns from the feedback it would otherwise only receive after committing.
A draft, a predicted reply, then a hidden revision
Each turn begins with a short plan and an action decision: draft, ask, or commit. After a draft, the interaction loop opens a user-role slot marked as anticipated, and the policy writes what it thinks the user would say, in the user's first-person voice, in one to three sentences. A fixed classifier reads only that predicted reply and labels it satisfied or unsatisfied. If it is unsatisfied and the revision cap of three has not been reached, the assistant writes an improved draft; otherwise the current draft is shown.
The revision is conditioned only on the earlier draft and the model's own predicted reply, both of which stay in context. There is no natural-language instruction telling the model how to revise. Hidden drafts and predictions do not count toward interaction cost because the user never reads them, and at test time the same loop runs with no training signal, with the real user reply arriving only after a draft is shown.
Prediction loss is capped so task reward stays primary
Training combines two objectives. The prediction term minimizes the negative log-likelihood of the user's actual reply to each draft, which is a dense token-level signal. The task term is group-relative policy optimization (GRPO), which assigns a single scalar reward per conversation based on task success and user effort, with an effort weight of 0.01 and coefficients of 0.5 and 0.4 for reply count and token length.
Because a token-level loss could overwhelm a per-conversation reward, the authors scale the prediction contribution at each update to the largest value at which its gradient is at most half the gradient of the GRPO term. Setting the prediction coefficient to zero removes prediction supervision while keeping the same prompt and revision loop, which is how the ablation isolates its effect. Task scores differ by domain: exact boxed-answer correctness for math, fraction of unit tests passed for coding, and ROUGE-L F1 against a reference article for writing.
Math accuracy rises 5.5 points over naive GRPO
Across mathematical reasoning, coding, and writing, CToM reports higher collaborative task performance than all baselines. The clearest single comparison is collaborative math reasoning, where final accuracy goes from 74.1% under naive GRPO to 79.6% under CToM. In the reported action table, the math score is 79.62 with a standard error of 1.65, against 77.19 for the task-only variant that keeps the revision loop but drops prediction supervision.
That gap matters for reading the paper's claim. The task-only arm already includes the draft-predict-revise structure and the same prompts, so the difference between 79.62 and 77.19 isolates what prediction supervision adds on top of the loop. The authors also report that adding an action-entropy regularizer raises the ask rate in all six matched pairs and lowers task scores in all six, which is why the main configuration does not use it.
Prediction supervision cuts user replies at equal reward
The efficiency result comes from ablations rather than from the headline accuracy number. Under the same reward, adding prediction supervision reduces the number of user replies by 8% on coding and 17% on writing while improving task performance. These counts use only conversations that continue past the first assistant turn, and replies are the user's messages after the opening request, so the denominator excludes single-turn exchanges.
The paper's own example shows the mechanism in miniature: CToM asks about tone first, then internally predicts that its first draft's treatment of a risk example is too thin, revises, and has its first visible draft accepted. The user in that comparison reads three long drafts and writes two corrections without the model of the user, versus one accepted draft with it. This is an illustrative trajectory, not an average, so it should be read alongside the aggregate reply counts rather than instead of them.
Better predictions track larger self-revision gains
Behavioral analysis of interaction trajectories supports the proposed chain. User-response predictions become more accurate after training, and more accurate predictions are associated with larger improvements in collaborative task performance through self-revision. The association is correlational; the paper does not establish that accuracy alone causes the gain.
Prediction accuracy is measured two ways, and the two are not interchangeable. Lexical F1 compares word sets between predicted and observed replies, so it rewards shared wording; on math it rises from 11.91 at update 0 to 19.70 at the final checkpoint, while the no-prediction variant stays near 14. Because the math pairing convention changes the denominator, the authors note that part of the early math gain reflects the trained model reliably emitting a prediction at all. An LLM-judge score, which rates how well a prediction anticipates the substance of the reply, also puts every trained CToM checkpoint above the untrained model, with final scores of 46.6, 47.1, and 41.7 on math, coding, and writing versus 29.9, 30.4, and 33.0 without prediction.
The evidence comes from simulated users
All experiments use a simulated user, and the authors name a human-user study as the most important next step. Whether an assistant can predict how real users respond, whether CToM keeps its task performance with them, and whether people prefer working with it remain untested here. The satisfaction classifier is also a fixed set of lexical rules mined from simulator replies, not a learned model, so when revision triggers depends on what the assistant predicts the user will say rather than on any check of the answer itself.
The scope of the mechanism is narrower than the framing. CToM applies its user model only to the draft action; decisions about whether to ask, what to ask, and when to commit are not guided by the prediction. The authors also note that revising whenever the predicted reply signals dissatisfaction may not be the best use of a prediction for every task, and that other forms of user modeling, such as personality and preference modeling, are left open.
The gain sits in the revision loop, not the prompt
The paper's most decision-relevant comparison is internal: CToM against its task-only variant, which shares the prompts, the draft-predict-revise loop, and the GRPO objective but sets the prediction coefficient to zero. On math that is 79.62 versus 77.19, and the reply-count reductions of 8% on coding and 17% on writing are likewise attributed to adding prediction supervision at the same reward. A reader weighing whether to adopt this approach should therefore treat the prediction objective, not the multi-turn scaffolding alone, as the component the evidence supports.
The corresponding caution is that the strongest accuracy evidence is the math result, while the efficiency evidence is strongest on writing, where the reply reduction is largest. Coding sits between the two, with a smaller reply reduction and a task score of 51.05 for CToM against 48.11 for the task-only arm. The paper's own priority ordering follows from this: test with human users before assuming the simulated-user gains transfer, and expect the clearest benefit where a draft can be checked internally before it is shown.
Sources used in this answer
Collaborative Theory of Mind: Learning to Self-Revise by Anticipating User Responses
CToM trains LLMs to predict user replies and revise drafts internally, improving collaborative math accuracy from 74.1% to 79.6% over naive GRPO and cutting user replies by 8% on coding and 17% on writing at equal reward.