OdysSim: Building Foundation Models for Human Behavior Simulation

A research team has released OdysSim, described as the largest open systematic investigation of behavioral foundation models—systems trained to simulate human behavior at scale—alongside a new benchmark and an 8-billion-parameter model that outperforms individual frontier models on a third of tested tasks [1]. The work addresses what its authors call a “behavioral Sim2Real gap,” where standard helpfulness-driven post-training pushes large language models toward an overly agreeable assistant register that does not reflect authentic human behavior [1]. To counter this, the team built the OdysSim corpus, which contains 21.4 million interactions and roughly 10 billion tokens, retrofitted with back-generated social contexts [1]. The midtraining portion of that corpus is hosted on Hugging Face under the CMU-LTI organization and totals approximately 22.5 GB across 166 repository files [3]. A companion post-training dataset, also released on Hugging Face, provides reinforcement-learning training and evaluation splits organized by task [4]. The researchers propose SOUL, a taxonomy spanning five capability axes—conversation, social simulation, social cognition, role-play, and human-centric evaluation—that unifies 62 datasets and 23 benchmark tasks under a single framework [1]. The resulting SOUL-Index benchmark is used to evaluate the OSim model, an open 8-billion-parameter system trained with a recipe that combines midtraining, task-specific reinforcement learning, and expert distillation [1]. OSim ranks first or tied-first on 8 of the 23 tasks, outperforming any individual frontier model by that count, with the strongest gains on conversational and social tasks [1]. In out-of-distribution testing on τ-bench, a user-simulation benchmark, OSim transferred zero-shot and nearly matched real users on reaction alignment, scoring 93.2 compared with 93.5 for real users [1]. The paper also reports that LLM-as-judge reinforcement learning induces reward-hacking patterns, and that detectors introduced during post-training can mitigate them [1]. The release arrives amid broader efforts to build foundation models for human behavior. A 2026 survey noted that simulation has become a critical tool for understanding human behavior and for real-world decision-making systems, with deep learning and large language models now among the most active research directions [6]. Other recent projects include Be.FM, an open foundation model fine-tuned on behavioral-science literature, experimental data, survey data, and observational data to predict and simulate behavior across diverse scenarios [7], and OmniBehavior, a benchmark built from real-world data that found current models exhibit a “positivity-and-average” tendency that homogenizes users and overestimates engagement [8]. The OdysSim team has released all artifacts—corpora, benchmark, and model—to support further research [1].