Ji-An Li
Publications
Thinking Under Uncertainty: Evidence Use and Information-Seeking in Language Models
Inference-time thinking improves the performance of large language models, but aggregate outcomes do not reveal whether models use available evidence more effectively or seek information that could improve future decisions. We distinguish these responses by measuring action preference, thinking length, and reported confidence under matched uncertainty. Ten open-weight models completed matched horizon-style two-armed bandit trials in thinking and non-thinking modes. A cognitive model separated value-guided action and uncertainty-independent choice noise from two behavioral signatures of exploration: a UCB-like preference for the less-known arm and Thompson-like choice variability that increases with total uncertainty. On average, thinking strengthened value-guided action and reduced uncertainty-independent choice noise, without producing UCB-like exploration or strengthening Thompson-like exploration. Outside action, the information-imbalanced history condition, which also displayed more observations than the matched balanced condition, was associated with greater thinking length. Reported confidence became more sensitive to decision difficulty and more strongly associated with chosen task evidence. We interpret these thinking-length and reported-confidence patterns as consistent with metacognitive control and metacognitive monitoring, respectively, without establishing either process. Decoder sweeps, especially temperature, altered choice noise and thinking length but did not reproduce the joint cross-output pattern. In this controlled decision setting, thinking improved how models acted on current evidence, while neither measured signature supported a shift toward a more information-seeking policy.
Post-training makes large language models less human-like
Large language models (LLMs) are increasingly used as surrogates for human participants, but it remains unclear which models best capture human behavior and why. To address this, we introduce Psych-201, a novel dataset that enables us to measure behavioral alignment at scale. We find that post-training -- the stage that turns base models into useful assistants -- consistently reduces alignment with human behavior across model families, sizes, and objectives. Moreover, this misalignment widens in newer model generations even as base models continue to improve. Finally, we find that persona-induction -- a popular technique for eliciting human-like behavior by conditioning models on participant-specific information -- does not improve predictions at the level of individuals. Taken together, our results suggest that the very processes that are currently employed to turn LLMs into useful assistants also make them less accurate models of human behavior.
The World Leaks the Future: Harness Evolution for Future Prediction Agents
Many consequential decisions must be made before the relevant outcome is known. Such problems are commonly framed as future prediction, where an LLM agent must form a prediction for an unresolved question using only the public information available at the prediction time. The setting is difficult because public evidence evolves while useful supervision arrives only after the question is resolved, so most existing approaches still improve mainly from final outcomes. Yet final outcomes are too coarse to guide earlier factor tracking, evidence gathering and interpretation, or uncertainty handling. When the same unresolved question is revisited over time, temporal contrasts between earlier and later predictions can expose omissions in the earlier prediction process; we call this signal internal feedback. We introduce Milkyway, a self-evolving agent system that keeps the base model fixed and instead updates a persistent future prediction harness for factor tracking, evidence gathering and interpretation, and uncertainty handling. Across repeated predictions on the same unresolved question, Milkyway extracts internal feedback and writes reusable guidance back into the harness, so later predictions on that question can improve before the outcome is known. After the question is resolved, the final outcome provides a retrospective check before the updated harness is carried forward to subsequent questions. On FutureX and FutureWorld, Milkyway achieves the best overall score among the compared methods, improving FutureX from 44.07 to 60.90 and FutureWorld from 62.22 to 77.96.