HACHIMI: Scalable and Controllable Student Persona Generation via Orchestrated Agents

Yilin Jiang; Fei Tan; Xuanyu Yin; Jing Leng; Aimin Zhou

cs.CLMar 5, 2026

HACHIMI: Scalable and Controllable Student Persona Generation via Orchestrated Agents

Authors: Yilin Jiang, Fei Tan, Xuanyu Yin, Jing Leng, Aimin Zhou

Organizations: East China Normal University, Shanghai, China · The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · Shanghai Innovation Institute, Shanghai, China

Abstract

Student Personas (SPs) are emerging as infrastructure for educational LLMs, yet prior work often relies on ad-hoc prompting or hand-crafted profiles with limited control over educational theory and population distributions. We formalize this as Theory-Aligned and Distribution-Controllable Persona Generation (TAD-PG) and introduce HACHIMI, a multi-agent Propose-Validate-Revise framework that generates theory-aligned, quota-controlled personas. HACHIMI factorizes each persona into a theory-anchored educational schema, enforces developmental and psychological constraints via a neuro-symbolic validator, and combines stratified sampling with semantic deduplication to reduce mode collapse. The resulting HACHIMI-1M corpus comprises 1 million personas for Grades 1-12. Intrinsic evaluation shows near-perfect schema validity, accurate quotas, and substantial diversity, while external evaluation instantiates personas as student agents answering CEPS and PISA 2022 surveys; across 16 cohorts, math and curiosity/growth constructs align strongly between humans and agents, whereas classroom-climate and well-being constructs are only moderately aligned, revealing a fidelity gradient. All personas are generated with Qwen2.5-72B, and HACHIMI provides a standardized synthetic student population for group-level benchmarking and social-science simulations. Resources available at https://github.com/ZeroLoss-Lab/HACHIMI

Figures & tables

Figure 1: HACHIMI pipeline overview. From target distributions (grade/gender/academic level), steps (1)–(5) produce the HACHIMI-1M corpus.

Panel of Figure 2 — Figure 2: Immersive role-playing prompt template used for HACHIMI student agents when answering CEPS- and PISA-based shadow surveys.

Figure 3: Pearson

r

and Spearman

\rho

between human and HACHIMI cohort means for each CEPS target.

Figure 4: Distribution of Pearson correlations between human and agent group means on PISA 2022.

Metric	Baseline	HACHIMI	$\Delta$
Hard error rate ( $\downarrow$ )	12.03%	0.00%	$-12.03$
Warning rate ( $\downarrow$ )	25.33%	0.82%	$-24.51$
Distinct-1 ( $\uparrow$ )	0.2328	0.3285	$+0.0957$
Distinct-2 ( $\uparrow$ )	0.4589	0.7893	$+0.3304$
Near-duplicate pairs ( $\downarrow$ )	157	0	$-157$

Table 1: Protocol-matched intrinsic comparison.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Field	Responsible Agent
id	System
Name	Enrollment & Development
Age	Enrollment & Development
Gender	Enrollment & Development
Grade	Enrollment & Development
Developmental Stage	Enrollment & Development

Appendix

Table 2: Field-to-agent responsibility mapping in the HACHIMI multi-agent system.

ID	Rule
R1	Age–grade norm: Grade 1–12 should approximately align with ages 6–18, allowing at most $\pm 1$ year deviation.
R2	Developmental-stage plausibility: Piaget, Erikson, and Kohlberg stage labels must remain broadly compatible with the student’s age band.
R3	Strong Subjects and Weak Subjects must both be non-empty and mutually disjoint.
R4	The eight creativity dimensions must show variation rather than identical levels; if feasibility is relatively low/low, proposing solutions cannot exceed a mid-level rating.
R5	If the values field presents stable positive physical/mental health, the mental-health field must not describe severe functional impairment or heavy clinical pathology.
R6	Agent-name regex: ^(?:[a-z]+[1-5]){1,2}_(?:[a-z]+[1-5]){1,3}$ .

Appendix

Table 3: Executable validator rules used in HACHIMI generation.

Field	Content
Name (pseudonym)	Shihan Wang
Agent ID	wang2_shi1han2
Age / Gender	13 / Female
Grade	Grade 7 (Junior Year 1)
Academic standing	Low; bottom 50% in school ranking
Strengths	Art

Appendix

Table 4: One example persona (English translation from the original Chinese record).

Grade	High	Medium	Low	Poor	Male	Female	Total
	Academic Level counts				Gender counts
1	21,154	20,942	21,257	21,182	42,197	42,338	84,535
2	21,298	21,026	21,150	21,165	42,192	42,447	84,639
3	21,116	21,329	21,111	21,240	42,306	42,490	84,796
4	21,234	21,093	20,909	20,749	41,962	42,023	83,985
5	21,267	20,632	21,226	21,249	42,237	42,137	84,374

Appendix

Table 5: Offline summary of HACHIMI-1M by grade, academic level, and gender.

Metric	Value	Notes
# personas	151,426	merged corpus size
Error rate	0.0000	fraction with $\geq 1$ hard error
Warning rate	0.0006	fraction with $\geq 1$ warning
Distinct-1	0.4044	over values/creativity/psychology
Distinct-2	0.8296	over values/creativity/psychology
Near-duplicate pairs	0	SimHash, Hamming $\leq T$

Appendix

Table 7: Summary of intrinsic metrics on 151,426 HACHIMI personas.

Variable	KL Divergence
Grade	0.0001
Gender	0.0000
Academic level	0.0000

Appendix

Table 8: KL divergence between target quotas and empirical distributions.

Figure 5: Distribution of paragraph lengths (in characters) across the three long-text components. The solid curves represent kernel density estimates (KDE), and the dashed lines indicate the means. The distributions show that the generation process exercises tight control over output lengths, centered around the target specifications.

Component	Pearson $r$
Values	0.7671
Creativity	0.8997
Mental health	0.8380

Appendix

Table 9: Text-derived alignment between academic level and component scores.

Figure 6: Distribution of Pearson correlations between human and agent group means on PISA 2022, summarized by construct family (each violin aggregates all constructs within a family across the five regions; individual construct points are jittered).

Figure 7: Heatmap of Pearson correlations between human and agent group means on PISA 2022. Constructs are ordered by latent family, highlighting the high-consistency math/curiosity block and the lower-consistency well-being and workload block.

Construct	Description	East Asia	S. Europe	Lat. Am.	Mid. East	W. Europe
Math Effort & Efficacy
MATHEFF	Mathematics self-efficacy: formal and applied mathematics - response options reversed in 2022 (WLE)	0.95	0.98	0.97	0.97	0.97
MATHEASE	Perception of Mathematics as easier than other subjects	0.83	0.81	0.72	0.91	0.85
MATHEF21	Mathematics self-efficacy: mathematical reasoning and 21st century skills (WLE)	0.77	0.77	0.90	0.89	0.78
MATHPREF	Preference of Math over other core subjects	0.83	0.92	0.60	0.90	0.85
MATHPERS	Effort and Persistence in Mathematics (WLE)	0.85	0.70	0.76	0.86	0.79

Appendix

Table 10: Cross-regional consistency of Pearson correlations (

r

) between human and agent group means on PISA 2022. Constructs are grouped by family. Bold indicates strong alignment (

r\geq 0.80

); gray indicates negative alignment. Descriptions follow official PISA 2022 variable definitions.

Metric	HACHIMI	Baseline	Notes
# personas	10,000	10,000	merged corpus size
Error rate	0.0000	0.1203	fraction with $\geq 1$ hard error
Warning rate	0.0082	0.2533	fraction with $\geq 1$ warning
Distinct-1	0.3285	0.2328	over values/creativity/psychology
Distinct-2	0.7893	0.4589	over values/creativity/psychology
Near-duplicate pairs	0	157	SimHash, Hamming $\leq T$

Appendix

Table 11: Summary of intrinsic metrics for HACHIMI personas vs. the matched one-shot single-model baseline.

Synthetic Data Generation for Language Models

Synthetic data generation for language models creates artificial text, tasks, and interaction traces to train or evaluate models.

See all 293 papers

MetaPersona: Task-Grounded Synthetic Populations from Empirical Social Science
Sep 29, 2026Jinyi Ye, Yuangang Li, Chenxiao Yu +7Personality Modeling in Language Models Synthetic Data Generation
University of Southern California · University of California, Irvine · 2077AI +1
Persona Matters: Effects of Activation Steering on Short Answer Generation and Scoring
Apr 8, 2026Yongchao Wu, Aron HenrikssonOpen-Ended Generation Language Model Steering
Stockholm University, NOD-huset, Borgarfjordsgatan 12, 16455 Stockholm, Sweden
Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM
Mar 19, 2026Zizhao Hu, Mohammad Rostami, Jesse ThomasonPersona-Conditioned Generation Personalized Language Model Alignment
University of Southern California Los Angeles, California, USA

Sep 29, 2026J/K move · Enter open · S save

MetaPersona: Task-Grounded Synthetic Populations from Empirical Social Science

Jinyi Ye, Yuangang Li, Chenxiao Yu, Preyashi Poddar+6

University of Southern California · University of California, Irvine · 2077AI+1

Personality Modeling in Language Models Synthetic Data Generation Computational Social Science Synthetic Population Generation Large Language Model-Based Social Simulation

Open paper AskPDF arXiv

Personas used to seed LLM social simulations face a cold-start problem: existing methods lack a principled basis for deciding which attributes to include and how to assign their values. As a result, synthetic populations may misrepresent the demographic composition, latent attributes, and dependency structure that shape downstream behavior. We introduce MetaPersona-DB, a dataset of 11,000+ empirical human-subjects studies annotated with task-relevant variables, reported relationships, and aggregate-level population statistics. Building on this resource, we propose MetaPersona, a framework that retrieves task-relevant evidence, constructs literature-derived persona dependency graphs, and samples synthetic populations from empirical priors linking demographics, latent attributes, and outcomes. Across three downstream case studies, three baselines, and three frontier models, results vary by task and model: MetaPersona performs strongly on misinformation belief and AI-tool sentiment, while results on income redistribution are mixed. It also reduces persona-construction cost to under $0.5 per task using GPT-5.2. Finally, we present MetaPersona-Studio, a prototype interactive interface for empirically grounded persona generation.

Abstract

Synthetic Data Generation for Language Models

Explore similar work

MetaPersona: Task-Grounded Synthetic Populations from Empirical Social Science

Persona Matters: Effects of Activation Steering on Short Answer Generation and Scoring

Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM