MIT/Stanford Paper: Single-Feature Behavioral Profiles Best for LLM Human Simulation
sebkrier · x · 2026-11-05
A paper from MIT, Stanford and Google DeepMind asks what person descriptions actually help LLMs simulate human behavior under distribution shift.
- Method: an LLM-based profile generator builds "behavioral profiles" from observed actions, messages and surveys; forward feature selection keeps features only while they improve prediction.
- Evaluated across 616 public-goods games, 144 multi-party bargaining games, and 1,480 generalized 11–20 money-request designs with new participants and designs.
- Findings: leading one-feature profiles beat no-profile and general-purpose personas; more features add no clear benefit; cross-game-family profile transfer gives uneven gains and can hurt accuracy.
- Takeaway: choosing behavioral information becomes a testable part of simulation design.
More from Research
- ReasonCore to host "Dead Reckoning" post-training happy hour at NeurIPS 2026 Sydney — geoffwolfe · 2026-11-05
- Stanford paper traces ICL circuit: few-shot learning is gated label lookup, not learning — rohanpaul_ai · 2026-11-05
- Stanford Paper Maps the Circuit Behind In-Context Learning: Gated Key-Value Memory in Small Transformers — teortaxesTex · 2026-11-05
- Gym-Anything for Robotics uses coding agents to mass-generate realistic robot simulations — wellecks · 2026-11-05
- LeCun coauthors proof of Platonic Representation Hypothesis: JEPA views need 42x fewer samples — ylecun · 2026-11-05
- REINFORCE Zero: trainer-inference invariance keeps async LLM RL stable for 20,000 updates — teortaxesTex · 2026-11-05