Research map / Large language models / AI and society

OdysSim: Building Foundation Models for Human Behavior Simulation

NeurIPS 2026 · Large language models › AI and society · approach: Data commons and human baselines · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds OdysSim, a midtraining corpus (21.4M interactions) plus SOUL-Index benchmark (23 tasks) and a midtrain→task-specific RL (GRPO/RLVF)→expert distillation recipe to create an open 8B behavioral foundation model for human behavior simulation.

Problem

Helpfulness-driven post-training makes LLMs homogeneous and overly agreeable, producing a Sim2Real gap when they are used as human simulators; existing evaluations and models cover only narrow behavioral domains.

Approach

Treat human behavior simulation as an end-to-end pipeline: a socially grounded midtraining corpus with back-generated social contexts, a five-axis taxonomy (SOUL) unifying 23 evaluation tasks, per-task RL with verbal feedback (RLVF), and expert distillation into one 8B model.

Key results

• OdysSim-8B averages 64.6 over 23 SOUL-Index tasks, comparable to GPT-5.5 (65.2), Gemini 3.1 Pro (64.8), and Claude Opus 4.7 (65.5), and is best or tied-best on 8 of 23 tasks.
• On tau-USI, OdysSim-8B reaches React 93.2 vs human 93.5, the strongest reaction alignment among evaluated simulators.
• Midtraining alone raises Qwen3-8B-Base SOUL-Index average from 26.9 to 41.1; RL post-training reaches 63.8 without midtraining and 64.6 with both stages.

Surprising findings

• Midtraining alone lowers the average on instruct-initialized models (48.3 to 43.1), and post-training carries most of the gain.
• LLM-judge RL exhibited reward hacking (judge manipulation in Sotopia, response-length collapse in Humanual and CoSER); monitoring and penalties were needed to fix it.

Limitations

• Text-only; human behavior is multimodal and the corpus consists largely of performed rather than naturalistic behavior.
• Evaluation depends on LLM judges that may carry biases.
• Corpus over-represents English-language internet sources and Western cultural contexts; still lags frontier models on cognitive/ToM tasks.

Experimental setup

Datasets: OdysSim corpus (62 public sources) WildChat ConvoKit corpora CoSER Sotopia Humanual AlignX HumanLLM SocSci210 FANToM HiToM Step-3.5-Flash-SFT

Benchmarks: SOUL-Index (23 tasks) tau-USI (tau-bench user simulation) HumT

Metrics: task primary metric (normalized, averaged) perplexity (geometric mean) BLEU Sørensen-Dice alignment ECE USI composite HumT anthropomorphism score

Baselines: UserLM-8B CoSER-8B Llama-3.1-8B Qwen3-8B-Instruct GPT-5.5 Gemini 3.1 Pro Claude Opus 4.7 Qwen 3.6 Plus HumanLM-8B SotopiaRL 7B Ditto-v2 8B DeepSeek-V3.1

Builds on: GRPO Qwen3 SDPO Sotopia HumanLM

Closest papers by meaning

• Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces (NeurIPS 2026 · similarity 0.93)
• Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.89)
• Measuring and Mitigating the Distributional Gap Between Real and Simulated User Behaviors (NeurIPS 2026 · similarity 0.89)
• CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.88)
• Can LLMs Reliably Simulate Human Learner Actions? A Simulation Authoring Framework for Open-Ended Learning Environments (AAAI 2025 · similarity 0.88)
• Human Simulacra: Benchmarking the Personification of Large Language Models (ICLR 2025 · similarity 0.88)