Socio-Foundation: A Model for Generalizable Individual Behavior Simulation via Hierarchical Capability Distillation

Liang Wang; Wenxuan Xie; Xinyi Mou; Yixin Luo; Zhongyu Wei

cs.AIOct 6, 2026

Socio-Foundation: A Model for Generalizable Individual Behavior Simulation via Hierarchical Capability Distillation

Authors: Liang Wang, Wenxuan Xie, Xinyi Mou, Yixin Luo, Zhongyu Wei

Organizations: School of Data Science, Fudan University · Shanghai Innovation Institute · Tencent

Abstract

Simulating individual behavior requires large language models (LLMs) to preserve persona traits while adapting to dynamic social contexts. However, general-purpose LLMs often flatten distinct personas, while task-specific tuning suffers from fragmentation and generalization. To overcome these challenges, we organize individual simulation into the \textbf{FONTS Taxonomy}, comprising five complementary capability dimensions: \emph{persona fidelity} (\textbf{F}), \emph{outcome realization} (\textbf{O}), \emph{behavioral naturalness} (\textbf{N}), \emph{trajectory coherence} (\textbf{T}), and \emph{social grounding} (\textbf{S}). Grounded in this taxonomy, we curate a standardized training corpus library of approximately 10 million instances across 14 representative datasets and present \textbf{Socio-Foundation}. Socio-Foundation decouples specialization from integration via a three-stage pipeline: learning task experts via DAPO, consolidating them into capability experts via off-policy distillation, and unifying them via multi-teacher on-policy distillation (MOPD). We also establish \textbf{IndiEval}, consolidating 29 metrics across the FONTS dimensions. Experiments show that Socio-Foundation outperforms its \textit{Qwen3-8B} base by 11.0 points and approaches frontier models such as \textit{GLM-5.2}, with ablations and out-of-distribution evaluations further demonstrating the effectiveness and generalization of our model.

Figures & tables

Figure 1: Three approaches to individual simulation. (a) General-purpose models may flatten distinct personas. (b) Task-driven post-training yields fragmented skills. (c) Individual simulation Foundation model integrates complementary capabilities.

Figure 2: The overall framework of Socio-Foundation. (a) Framework overview: We organize individual simulation into five capability dimensions (F,O,N,T,S), consolidate task experts into capability experts, and integrate them into a unified simulator. IndiEval evaluates the resulting model using task-native metrics grouped by the same dimensions. (b) Training details: Task-specific DAPO trains LoRA task experts; off-policy distillation trains capability experts on fixed offline prefixes; and MOPD supervises the unified student on its own rollouts using weighted forward KL.

Frontier models
	F									O							N							T			S			Avg.
Model	LifeChoices Acc. $\uparrow$	AlignX Choice acc. $\uparrow$	HumanLLM Hit@5 $\uparrow$	BehaviorChain Node $\uparrow$	CoSER Char. fid. $\uparrow$	HUMANUAL Response $\uparrow$	HUMANUAL State $\uparrow$	UserLM-S3 Intent adh. $\uparrow$	SOTOPIA Believ. $\uparrow$	SOTOPIA Goal $\uparrow$	SOTOPIA Relation. $\uparrow$	SOTOPIA Knowledge $\uparrow$	SOTOPIA Benefits $\uparrow$	AgentSense Goal maj. $\uparrow$	UserLM-LiC Intent cov. $\uparrow$	UserLM-LiC Task score $\uparrow$	UserLM-S3 Term. F1 $\uparrow$	MirrorBench GTEval $\uparrow$	CoSER Anthrop. $\uparrow$	UserLM-S3 Human lik. $\uparrow$	UserLM-S3 Diversity $\uparrow$	UserLM-S3 Inv. overlap $\uparrow$	UserLM-S3 Role adh. $\uparrow$	CoSER Story cons. $\uparrow$	BehaviorChain CumScore $\uparrow$	CoSER Story qual. $\uparrow$	FANToM Item acc. $\uparrow$	Social-R1 Acc. $\uparrow$	AgentSense Private acc. $\uparrow$	Average $\uparrow$
GPT-5.5	92.0	61.0	64.1	89.7	49.6	21.8	44.3	98.0	84.7	59.0	63.7	29.3	55.0	96.7	99.0	94.4	0.0	41.3	41.9	51.2	91.6	75.6	32.0	62.6	58.7	65.6	79.3	69.0	82.4	67.2
Open-weight general-purpose models
DeepSeek-V4-Flash	81.8	59.6	62.0	69.3	43.2	15.5	39.5	96.0	82.7	54.5	60.8	27.4	51.9	95.5	93.3	57.9	34.1	48.5	44.5	29.8	86.8	87.5	5.0	62.6	26.3	75.7	61.5	51.0	82.4	59.7
GLM-5.2	84.2	62.2	63.4	76.7	43.9	20.0	45.5	94.0	86.1	55.5	72.9	37.2	50.5	96.7	97.3	76.8	0.0	43.2	40.8	79.7	91.0	78.4	6.0	59.7	34.1	68.0	74.3	69.0	80.8	63.2

Table 1: Main results on IndiEval. Scores are scaled to 0–100 (higher is better). Avg. equally weights the five dataset-balanced capability scores. Bold and underline indicate the best and second-best distinct scores in each column across all listed models, including ties.

Figure 3: Operationalization of IndiEval across the FONTS dimensions. The five capability dimensions connect to 11 evaluation targets and 29 native metrics from 13 benchmark entries. Colored links trace the correspondence between dimensions, evaluation targets, metrics, and benchmark task settings. Metric tags identify rule/reference-based scoring (R), LLM-as-judge scoring (J), and detector-based model scoring (M). For each benchmark,

n

denotes sampled units before expansion into evaluation cases.

Figure 4: Evaluation-set composition and coverage of IndiEval. (A) Distribution of 6,297 evaluation cases across 13 benchmark entries. (B) Scoring-type coverage and (C) capability coverage, weighted by evaluation-case counts. For benchmarks spanning multiple scoring types or capability dimensions, their case counts are allocated equally across the corresponding categories. These coverage proportions are independent of the weights used to compute IndiEval Avg. Additional OOD benchmarks outside IndiEval are excluded.

	F									O							N							T			S			Avg.
Variant	LifeChoices Acc. $\uparrow$	AlignX Choice acc. $\uparrow$	HumanLLM Hit@5 $\uparrow$	BehaviorChain Node $\uparrow$	CoSER Char. fid. $\uparrow$	HUMANUAL Response $\uparrow$	HUMANUAL State $\uparrow$	UserLM-S3 Intent adh. $\uparrow$	SOTOPIA Believ. $\uparrow$	SOTOPIA Goal $\uparrow$	SOTOPIA Relation. $\uparrow$	SOTOPIA Knowledge $\uparrow$	SOTOPIA Benefits $\uparrow$	AgentSense Goal maj. $\uparrow$	UserLM-LiC Intent cov. $\uparrow$	UserLM-LiC Task score $\uparrow$	UserLM-S3 Term. F1 $\uparrow$	MirrorBench GTEval $\uparrow$	CoSER Anthrop. $\uparrow$	UserLM-S3 Human lik. $\uparrow$	UserLM-S3 Diversity $\uparrow$	UserLM-S3 Inv. overlap $\uparrow$	UserLM-S3 Role adh. $\uparrow$	CoSER Story cons. $\uparrow$	BehaviorChain CumScore $\uparrow$	CoSER Story qual. $\uparrow$	FANToM Item acc. $\uparrow$	Social-R1 Acc. $\uparrow$	AgentSense Private acc. $\uparrow$	Average $\uparrow$
Task-expert SFT	72.8	58.6	58.6	83.9	26.3	14.5	41.2	100.0	77.6	45.3	55.6	28.3	48.7	91.2	70.0	31.6	45.3	53.9	30.1	34.0	80.5	97.9	100.0	53.4	44.2	48.1	78.2	61.0	79.6	59.5
Task-expert OPD	70.8	52.6	56.6	77.4	19.1	13.2	35.5	99.0	72.6	44.4	54.8	23.0	50.0	91.4	74.6	36.8	0.0	42.8	21.1	34.7	85.4	79.2	91.0	51.7	43.2	38.8	60.2	48.0	82.1	54.2
Task-expert averaging	68.2	57.2	55.8	76.5	20.6	12.2	35.0	97.0	77.4	45.7	55.5	27.1	50.7	89.5	72.0	32.3	12.1	40.2	25.4	26.4	86.1	74.5	71.0	49.6	44.8	42.2	63.9	50.0	83.5	54.5
Capability-expert averaging	70.2	55.4	58.2	77.6	20.4	12.6	35.4	100.0	76.5	46.5	55.9	25.3	50.5	89.1	79.2	37.9	58.1	42.1	26.5	27.1	82.9	92.2	100.0	53.7	45.3	44.2	65.8	51.0	79.6	56.8
Capability-expert off-policy distill.	75.2	55.0	53.6	87.3	23.6	13.2	38.1	100.0	70.3	45.5	52.4	25.0	49.8	89.5	55.5	20.4	54.8	50.6	29.1	40.4	79.8	96.1	100.0	56.0	54.1	47.7	72.3	65.0	85.1	59.3

Table 2: Pipeline ablations on IndiEval. We compare the full pipeline with SFT on task-expert rollouts, Task-expert OPD, LoRA averaging of task or capability experts, final-stage off-policy distillation from capability experts, SFT on pooled task data, and OPD with reverse KL. Bold and underline indicate the best and second-best distinct scores in each column.

Variant	F	O	N	T	S	Avg.
Socio-Foundation	63.93	63.85	54.41	50.50	75.77	61.69
w/o F	60.72	66.99	53.33	49.38	72.81	60.65
w/o O	62.57	63.28	52.46	49.32	73.95	60.32
w/o N	61.66	64.12	46.66	49.19	74.59	59.24
w/o T	62.44	65.20	51.57	49.36	74.11	60.54
w/o S	62.94	66.03	51.87	48.93	68.85	59.72

Table 3: Capability-expert ablations. w/o X denotes using other 4 experts to MOPD.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Cluster	Theme / dimension	Capability requirements
C1	Preference / Value Fidelity ( F )	A01 Preference alignment; A02 Viewpoint and value fidelity
C2	Persona / Expression Fidelity ( F )	A03 Persona maintenance; A04 Individual expression-style fidelity
C3	Behavioral / Decision Fidelity ( F )	A05 Individual response and choice fidelity; A06 User-intent fidelity; A10 Individual cognitive and decision-pattern fidelity
C4	Goal / Outcome Realization ( O )	A07 Individual goal realization
C5	Human-like Expression / Behavior ( N )	A08 Natural dialogue behavior; A09 Human-like behavior
C6	Trajectory / Narrative Coherence ( T )	A11 Behavior-sequence and cross-turn consistency; A12 Narrative progression and response consistency

Appendix

Table 4: The eight capability clusters and their 19 requirement labels. F/O/N/T/S denote Persona Fidelity, Outcome Realization, Behavioral Naturalness, Trajectory Coherence, and Social Grounding.

Figure 8: Capability-structured organization of the surveyed data and supplementary benchmarks. The five colored regions contain eight clusters. Each chip identifies a dataset/task entry, and an entry may appear in multiple clusters. The diagram summarizes the annotations in Table 5 ; the stated task and role interpretations apply to the displayed memberships. Table 6 separately records the dimensions measured by the selected main-table metrics.

Dataset / task	Task interpretation	Requirements	Clusters	Dimensions
Surveyed dataset/task entries
alignx_v2 †	Preference-conditioned response selection.	A01	C1	F
convokit_casino-corpus	Negotiation with private item preferences and individual objectives.	A01, A07	C1, C4	F, O
convokit_chromium-corpus	Participants in software-review discussions.	A08, A16	C5, C8	N, S
convokit_emotional-support	Emotional-support conversations.	A15, A18	C8	S
convokit_mediasum-corpus	Dialogue simulation from interview transcripts.	A08	C5	N

Appendix

Table 5: Dataset/task inventory and capability annotations. A-labels are defined in Table 4 ; C-labels identify clusters. The final two columns report cluster membership and FONTS dimensions separately. Supplementary entries complete coverage of the paper’s benchmarks and protocols.

Benchmark / entry	Cluster / dimension	Selected metrics and interpretation
LifeChoices	C1, C3 / F	Choice accuracy: persona-conditioned preference and choice fidelity.
AlignX	C1 / F	Direct-choice accuracy: personalized preference fidelity.
HumanLLM	C1 / F	Hit@5: personalized item selection; the story/scenario membership in the survey is outside this metric.
BehaviorChain	C3 / F; C6 / T	Node accuracy (F); chain-level CumScore (T).
CoSER	C2 / F; C5 / N; C6 / T	Character fidelity (F); anthropomorphism (N); storyline consistency and storyline quality (T).
HUMANUAL	C1–C3 / F	Response alignment and state alignment. The six domain entries supply the relevant individual-reference requirements.

Appendix

Table 6: Concordance between cluster memberships and the selected evaluation metrics. Each of the 29 main-table metrics appears once; this table introduces no new metric or score.

Dataset / source	Before	After	Removed	Transformed	Share (%)
alignx_v2	14,649,170	14,086,414	562,756	1,097	74.0663
socsci210	2,618,745	2,547,535	71,210	30,311	13.3949
human_llm	1,316,829	1,316,811	18	535	6.9238
wildchat	165,346	157,220	8,126	3,726	0.8267
coser	114,831	114,831	0	8	0.6038
lmsys	79,929	79,908	21	167	0.4202

Appendix

Table 7: Dataset cleaning statistics. Transformed counts repaired instances; Removed counts filtered-out instances. Share is the percentage of all retained rows. Transformation counts are reported separately and are not subtracted again from retained counts.

Figure 9: Task-expert training trajectories. DAPO, GRPO, and GSPO are compared on five tasks and their arithmetic mean. Dashed lines indicate the reported Socio-Foundation, GPT-5.5, and Qwen3-Next-80B reference scores.

Dataset	AlignX	BehaviorChain	CoSER	FANToM	HumanLLM	HUMANUAL-Book	HUMANUAL-Chat	HUMANUAL-Opinion	HUMANUAL-Politics	LifeChoices	MirrorBench	Social-R1	SOTOPIA	UserLM-S3
RL training instances	8,192	1,024	1,024	1,024	1,024	1,000	1,000	1,000	1,000	1,150	1,024	687	405	1,024

Appendix

Table 8: Training instances sampled for task-expert RL. Each column reports the number of instances sampled from the training corpus for DAPO on the corresponding dataset.

Dim.	Benchmark	Metric	Description
F	LifeChoices	Acc.	Accuracy of persona-conditioned choices against the reference decisions.
F	AlignX	Choice acc.	Accuracy of direct choices against the target individual’s reference preferences.
F	HumanLLM	Hit@5	Fraction of cases in which the reference item appears among the top five predicted items.
F	BehaviorChain	Node	Micro-averaged correctness of individual behavior predictions across chain nodes.
F	CoSER	Char. fid.	Scene-level assessment of consistency with the assigned character’s attributes and behavior.
F	HUMANUAL	Response	Alignment of generated responses with the target person’s reference responses.

Appendix

Table 9: Definitions and capability assignments of all main-table IndiEval metrics. Metric labels match the main table.

Dataset / benchmark	Sampling unit	Sampled units	Cases
IndiEval
FANToM	Complete question set (one per conversation)	100	1,468
Social-R1	Question	100	100
LifeChoices	Decision question (balanced sampling across books)	500	500
BehaviorChain	Complete persona behavior chain	100	1,529
AlignX	Core preference sample (five conditioning variants each)	100	500

Appendix

Table 10: Evaluation sampling statistics for IndiEval and

\tau

-USI. Sampled units follow each benchmark’s native structure; a unit may yield multiple evaluation cases.

Benchmark	Cases/run	Metric	Mean $\pm$ SD	Min.	Max.	Range
LifeChoices	500	Accuracy	$79.800\pm 0.583$	79.000	80.600	1.600
HUMANUAL	600	Response alignment	$12.486\pm 0.585$	11.716	13.325	1.609
		State alignment	$36.360\pm 0.587$	35.676	37.023	1.347
MirrorBench	200	GTEval	$54.425\pm 0.396$	53.975	54.925	0.950
		PI	$62.517\pm 1.690$	60.833	64.333	3.500
		RNR	$76.250\pm 1.723$	73.250	77.500	4.250

Appendix

Table 11: Five-run evaluation stability of a fixed unified OPD checkpoint. Bounded scores are multiplied by 100; for these

[0,1]

metrics, the displayed SD is also the dispersion as a percentage of the theoretical score range. LifeChoices values are accuracy percentages. MATTR retains its native z-score scale because it has no fixed finite bounds. Cases/run gives the number of test cases in each of the five repetitions. Range is the observed maximum minus minimum, computed before rounding; these extrema are not confidence bounds.

Explore similar work

OdysSim: Building Foundation Models for Human Behavior Simulation
Jun 12, 2026Xuhui Zhou, Weiwei Sun, Weihua Du +6User Simulation Behavioral Foundation Models
Carnegie Mellon University, Language Technologies Institute
How Well Do Large Language Models Capture Human Personality?
May 12, 2026Aanisha Bhattacharyya, Yaman Kumar Singla, Rajiv Ratn Shah +2Artificial Intelligence Personas Persona Consistency
Adobe Media and Data Science Research (MDSR) · IIIT-Delhi · SUNY at Buffalo
Will Scaling Improve Social Simulation with LLMs?
Jul 2, 2026Caleb Ziems, William Held, Su Doga Karaca +3Scaling Improve Social Simulation Sequential Scaling
Stanford University · Open Athena

Jun 12, 2026, cs.CLJ/K move · Enter open · S save

OdysSim: Building Foundation Models for Human Behavior Simulation

Xuhui Zhou, Weiwei Sun, Weihua Du, Jiarui Liu+5

Carnegie Mellon University, Language Technologies Institute

User Simulation Behavioral Foundation Models Agent Behavior Modeling Post-Training Sim-To-Real Gap Text Corpora

Open paper AskPDF arXiv

Large language models are increasingly deployed as human simulators for interactive evaluation and social simulation. Yet helpfulness-driven post-training pulls them toward a homogeneous, overly agreeable assistant register, creating a behavioral Sim2Real gap. We present OdysSim, the largest open systematic investigation of behavioral foundation models, i.e., models trained to simulate human behavior at scale. We propose SOUL, a taxonomy of five capability axes (CONV, SS, COG, ROLE, EVAL) that unifies 62 datasets and 23 benchmark tasks under one framework. Specifically, we curate the OdysSim corpus (21.4M interactions, 10B tokens, retrofitted with back-generated social contexts), construct the SOUL-Index benchmark, and develop an end-to-end training recipe combining midtraining, task-specific RL, and expert distillation. The resulting open 8B OSim model ranks first or tied-first on 8 of 23 tasks, outperforming any individual frontier model by this count, with the strongest gains on conversational and social tasks. Its outputs are also more human-like in length, formatting, and word choice, and it transfers zero-shot to out-of-distribution user simulation on

τ

-bench, nearly matching real users on reaction alignment (93.2 vs. 93.5). We further show that LLM-as-judge RL induces reward-hacking patterns, and that our detectors can mitigate them during post-training. Together, our findings suggest that behavioral foundation models require rethinking the LLM training paradigm. We release all artifacts to support future research.

Abstract

Explore similar work

OdysSim: Building Foundation Models for Human Behavior Simulation

How Well Do Large Language Models Capture Human Personality?

Will Scaling Improve Social Simulation with LLMs?