Eight LLMs Lose 35–46% of Strategy Calibration in Free-Form Generation

Researchers at Meta Superintelligence Labs and the University of Southern California tested eight closed and open language models on a skill that every chatbot with memory needs. An assistant may correctly remember that a user likes salty food, and keep mentioning it even though the user has since said they've come to prefer sweet. The authors call the required skill adaptive personalization under epistemic uncertainty: the response strategy should change as the reliability of a stored preference changes.

To measure it, they use strategy calibration, the share of responses in which the model selects or generates a tactic that is acceptable for the preference's current state. When models choose from ready-made options, they perform well. In free-form generation, strategy calibration drops by 35–46%.

The authors call this the gap between recognition and generation.

Three Preference States and Five Strategies

The authors identify three states that knowledge about a user can be in. A preference is either established (stated explicitly and not contradicted), uncertain (the context is incomplete or ambiguous), or revised (new information has overridden the old). Each state calls for its own behavior: act directly, ask about what is missing, or correct course.

The set contains five strategies in total: act on the known preference, request missing information, double-check whether the preference is still current, offer options without forcing a choice, and update behavior after a contradiction.

Before building the test, the team ran a survey in English, Spanish, Portuguese, Hindi, and Indonesian. Three annotators rated each of 500 dialogues, yielding about 1500 ratings. In every state, a large majority of people chose a small set of strategies. To verify this, the authors went through all 15 ways of splitting the five strategies into two groups and compared each split against human choices using an agreement criterion. In each state, one split fit better than the rest. That split became the list of acceptable strategies.

How the Benchmark Works

The team built the test with a three-role pipeline of several LLMs. The foundation was 250 personas from PersonaHub, selected by farthest-point sampling (picking the most dissimilar points). Each persona received preferences from 20 life domains. The team then planned dialogues of 10–14 sessions, with scheduled transitions of a preference between the three states. A separate filter model checked dialogue coherence and label accuracy.

The result is 25,006 utterances in 2,939 multi-session dialogues in five languages. Each utterance comes with four candidate responses: two with acceptable strategies and two "traps." The traps are spread across seven error types: over-personalization, epistemic overreach, staleness, need mismatch, excessive questioning, style mismatch, and missing necessary questions.

Humans also validated the benchmark. Annotators rated 500 dialogues on five criteria: coherence, clarity of the epistemic state, correctness of the strategy label, trap type, and acceptability of the response. Annotator agreement with the LLM jury was substantial. Human accuracy in choosing an acceptable response from the options was 92%.

Why Models Fail and What Helps

The error is systematic. By default, models follow a "use what you know first" principle: they rely too heavily on earlier preferences regardless of how reliable those are. Yet in the multiple-choice test, they usually identify the right strategy correctly.

The authors checked whether prompting could fix this. Supplying preference context in the prompt did not close the gap. Probing of internal activations showed that the models encode the epistemic state in their hidden representations. However, this knowledge does not steer strategy choice during generation.

The researchers then trained the models with a reward function that pulls the predicted response strategy toward the human choice for a given preference state. This training narrowed the acceptability gap by a third for new personas and unfamiliar dialogue styles. The result is measured against standard fine-tuning.

The practical takeaway for developers of assistants with long-term memory: good retrieval over conversation history is not enough. Teams need to separately test whether the system knows how to ask again, double-check, or switch tactics when the data about the user has gone stale. The project's code is published on GitHub.