BIG5-CHAT: Shaping LLM Personalities Through Training on...

Trait induction and control2024OpenReviewApproved editorial review

Original title: BIG5-CHAT: Shaping LLM Personalities Through Training on Human-Grounded Data

Authors: Wenkai Li, Jiarui Liu, Andy Liu, Xuhui Zhou, Mona Diab, Maarten Sap

Keywords: Computation and Language, Personality traits, Human-grounded data

Source: Open primary source (opens in a new tab)

6
Authors
7
Findings
7
Limitations
6
Evidence

Editorial summary

English

The final ACL 2025 version of BIG5-CHAT presents a generation and training pipeline for expressing high or low levels of each Big Five trait in LLM responses. Its empirical basis is not a final human-written dialogue corpus. It starts from PsychGenerator, 846,304 Facebook posts linked to their authors' personality scores, and 10,000 SODA scenarios, themselves originally generated with GPT-3.5. The authors fine-tune five LLaMA-3-8B-Instruct expert generators, one per trait, on the upper and lower ends of PsychGenerator. They then combine expert logits with LLaMA-3-70B-Instruct through a DExperts-style method and produce 100,000 synthetic single-turn dialogues: 20,000 per trait, balanced between high and low levels. “Human-grounded” therefore refers to personality signals inherited from human social-media posts, not human authorship of BIG5-CHAT. A RoBERTa-large classifier trained on PsychGenerator reaches 93.8% on its binarized test set and assigns the intended level to 80.4% of expert-generator outputs, versus 59.2% for a GPT-4o-mini baseline. This evaluator shares the generator's source domain, and appendix experiments with Big Five Essay classifiers reach only about 50–60%, exposing weak cross-domain generalization. LLaMA-3-8B-Instruct and 70B-Instruct are aligned on BIG5-CHAT with LoRA-based SFT and DPO and compared with direct inference and instruction- or demonstration-based prompting. Across BFI-44 and IPIP-NEO-120, repeated five times at temperature 0.6, all interventions produce some high–low separation. SFT and DPO usually push scores closer to the scale extremes, especially for low-trait conditions, but DPO has no consistent overall advantage over SFT. Appendix baselines using keywords, MPI items, or LLM-written descriptions can match or exceed some fine-tuning results, although the authors exclude them from the main comparison because of lexical overlap with the questionnaires. SFT best approximates the selected human within-trait correlation matrix, with Frobenius distance 1.55 versus 2.10 for prompting and 2.06 for DPO, while notable discrepancies remain, especially for Neuroticism. Two graduate annotators compare 200 response pairs: BIG5-CHAT beats a prompt baseline on expressiveness in 50.3% of cases and realism in 47.8%, with 39.8% and 42.3% ties and kappa values of 0.50 and 0.55. The most favorable reasoning result is 70B SFT, which improves average social, mathematical, and commonsense performance, but does not uniformly beat direct inference on TruthfulQA or general knowledge. Results for 8B are far less stable: DPO collapses HumanEval in several conditions, SFT sharply reduces GSM8K, and Agreeableness and Neuroticism effects are weak or contradictory. The study therefore supports controllable output patterns under one pipeline, not a causal claim that personality improves reasoning. Training method and data remain confounded. Its hallucination discussion uses selected qualitative examples rather than a systematic metric, and the work does not test multi-trait combinations, long conversations, temporal stability, other languages, or user effects.

Español

La versión final publicada en ACL 2025 de BIG5-CHAT presenta una cadena de generación y entrenamiento para expresar niveles altos o bajos de los cinco rasgos Big Five en respuestas de LLM. Su base empírica no es un corpus final escrito por personas: parte de PsychGenerator, 846.304 publicaciones de Facebook asociadas a puntuaciones de personalidad de sus autores, y de 10.000 escenarios de SODA, un conjunto de situaciones sociales generado originalmente con GPT-3.5. Los autores ajustan cinco generadores expertos LLaMA-3-8B-Instruct, uno por rasgo, con los extremos superior e inferior de PsychGenerator. Después combinan sus logits con LLaMA-3-70B-Instruct mediante una variante de DExperts y producen 100.000 diálogos sintéticos de un solo turno: 20.000 por rasgo, equilibrados entre nivel alto y bajo. Por ello, «human-grounded» describe la señal de personalidad heredada de publicaciones humanas, no la autoría humana de BIG5-CHAT. Un clasificador RoBERTa-large entrenado en PsychGenerator alcanza 93,8 % en su test binarizado y asigna correctamente el nivel objetivo al 80,4 % de las salidas del generador, frente a 59,2 % para una línea base GPT-4o-mini. Esta evaluación comparte dominio y datos de origen con el generador, y el apéndice muestra que clasificadores entrenados con Big Five Essay solo alcanzan aproximadamente 50–60 %, señal de escasa generalización entre dominios. Sobre BIG5-CHAT se ajustan LLaMA-3-8B-Instruct y 70B-Instruct mediante LoRA con SFT y DPO; se comparan con inferencia directa y prompts de instrucción o demostración. En BFI-44 e IPIP-NEO-120, repetidos cinco veces a temperatura 0,6, todas las intervenciones separan en alguna medida niveles altos y bajos. SFT y DPO suelen empujar con más fuerza los extremos, sobre todo las puntuaciones bajas, y en 70B se acercan a los límites 5 y 1 de las escalas; no aparece una ventaja sustancial y uniforme de DPO sobre SFT. Sin embargo, el apéndice muestra que prompts con palabras clave, ítems MPI o descripciones generadas por LLM pueden igualar o superar algunos resultados de ajuste, aunque los autores los excluyen de la comparación principal por su solapamiento léxico con los cuestionarios. Las correlaciones internas entre rasgos se acercan más a una referencia humana con SFT: la distancia de Frobenius es 1,55, frente a 2,10 para prompting y 2,06 para DPO, pero persisten discrepancias, especialmente en Neuroticism. Dos estudiantes de posgrado comparan 200 pares de respuestas; BIG5-CHAT gana a un baseline por prompt en expresividad el 50,3 % y en realismo el 47,8 %, con 39,8 % y 42,3 % de empates y kappas 0,50 y 0,55. En razonamiento, el resultado más favorable corresponde al SFT de 70B, con mejoras medias en razonamiento social, matemático y de sentido común; no supera de forma uniforme al modelo directo en TruthfulQA o conocimiento general. Los resultados de 8B son mucho más inestables: DPO colapsa HumanEval en varias condiciones, SFT reduce fuertemente GSM8K y los efectos de Agreeableness y Neuroticism son débiles o contradictorios. Por tanto, el estudio aporta evidencia de control de patrones de respuesta bajo un pipeline concreto, pero no demuestra que una personalidad cause mejor razonamiento. Las diferencias pueden proceder del método y los datos de ajuste. La discusión de alucinación se basa en ejemplos cualitativos escogidos, no en una métrica sistemática. Tampoco se estudian combinaciones de rasgos, diálogos prolongados, estabilidad temporal, otros idiomas o efectos sobre usuarios.

Research question

Can a conversational corpus based on human Big Five signals be constructed to align LLMs through SFT or DPO, to induce high and low levels more clearly than with prompting, and to relate those output profiles to reasoning task performance?

Method

Experimental pipeline in four stages. Five LLaMA-3-8B-Instruct models are fine-tuned as expert generators with PsychGenerator posts from the high and low thirds of each trait. Their logits are combined with LLaMA-3-70B-Instruct to rewrite responses in 10,000 SODA scenarios and create 100,000 single-turn synthetic dialogues. A RoBERTa-large trained on PsychGenerator and two human evaluations examine trait expression. LLaMA-3-8B-Instruct and 70B-Instruct are aligned with LoRA through SFT and DPO and compared with prompting using BFI-44, IPIP-NEO-120, correlation matrices, and reasoning benchmarks; each questionnaire is repeated five times at temperature 0.6.

Sample: One hundred thousand single-turn synthetic dialogues, 20,000 per trait and balanced by level, derived from 10,000 SODA scenarios and signals learned from 846,304 PsychGenerator posts. Two LLaMA-3 models are evaluated with five repetitions of two inventories. The human evaluation uses two graduate students and 200 comparisons per study; it does not include a representative sample of users or clinical psychologists.

Findings

  • The RoBERTa classifier trained on PsychGenerator obtains 93.8% on its binarized test, 80.4% when evaluating the expert generator, and 59.2% over the GPT-4o-mini baseline; this advantage is measured within the same data domain.
  • SFT and DPO produce more extreme high-low separations than the main baselines on BFI and IPIP-NEO, especially in LLaMA-3-70B, but there is no substantial general difference between SFT and DPO.
  • SFT better reproduces the human correlation matrix between traits according to Frobenius distance: 1.55 compared to 2.10 for prompting and 2.06 for DPO; the approximation remains imperfect, especially for Neuroticism.
  • Compared to prompting, BIG5-CHAT wins 50.3% of human comparisons on expressivity and 47.8% on realism; ties are 39.8% and 42.3%, with moderate agreement between two annotators.
  • In the 70B model, SFT achieves the best reasoning average and improves especially social, mathematical, and common sense tasks, but does not consistently outperform the direct model on TruthfulQA and general knowledge.
  • In 8B the effects are not robust: DPO severely degrades HumanEval under several conditions, SFT collapses GSM8K for many traits, and Agreeableness and Neuroticism do not follow a stable relationship with performance.
  • Validation with Big Five Essay drops to approximately 50-60%, which questions whether the classifier and the learned patterns transfer outside of PsychGenerator.

Limitations

  • BIG5-CHAT is synthetic: SODA was generated with GPT-3.5 and the final responses come from LLMs; the human signal comes indirectly from Facebook posts, whose sample has demographic, cultural, and platform biases.
  • The generator and the evaluator classifier learn from PsychGenerator. The automatic evaluation may reward the same domain patterns rather than external psychological validity, and fails to generalize to Big Five Essay.
  • BFI and IPIP-NEO are scored as human measures without demonstrating invariance or metric equivalence for LLMs; saturation near 1 and 5 may reflect explicit label following and linguistic overlap.
  • Only two graduate students evaluate 200 pairs and are not experts in psychology; the size and annotation profile limit claims about human realism.
  • Personality, the adjustment method, the examples, and the explicit labels change at the same time, so reasoning differences do not identify a causal effect of the trait.
  • The analysis is limited to English, to isolated traits, and to single-turn dialogues; it does not examine Big Five combinations, longitudinal stability, prolonged interaction, or real effects on people.
  • The discussion on bias and hallucination is qualitative and uses selected examples; it does not provide a systematic rate or a safety evaluation.

What the study does not establish

  • It does not demonstrate that LLaMA possesses personality, mental states, or internal human traits; it demonstrates control of responses under specific instructions and fine-tuning.
  • It does not demonstrate that BIG5-CHAT is written by people or that its dialogues are natural observations; it is a synthetic corpus indirectly supported by human signals.
  • It does not prove that SFT or DPO are universally superior to prompting, because some additional prompts from the appendix match or exceed results and performance depends on the model and the instrument.
  • It does not establish that a personality trait causes better reasoning or that human patterns transfer to LLMs; the 8B results and several domains contradict a general rule.
  • It does not validate the PsychGenerator classifier as a personality measure outside its domain or BFI/IPIP scores as equivalent to human scores.
  • It does not demonstrate safety, absence of stereotypes, or suitability for personalization in real applications.

Traceability

Scope: Full text

Version: ACL 2025 Long Paper, pp. 20434–20471

Consulted source: https://aclanthology.org/2025.acl-long.999.pdf

Review: Codex editorial review, 2026-07-14

Approval: Codex fidelity pass, 2026-07-14

English translation: approved, 2026-07-18

Models evaluated

  • LLaMA-3-8B-Instruct
  • LLaMA-3-70B-Instruct
  • GPT-4o-mini
  • GPT-4o
  • RoBERTa-large personality classifier

Instruments and metrics

  • Big Five Inventory (44 items)
  • IPIP-NEO-120
  • PsychGenerator-derived five-head personality classifier
  • Pairwise human evaluation of expressiveness and realism
  • Within-trait correlation matrices and Frobenius distance
  • LIWC2015 linguistic analysis

Data used

  • PsychGenerator (846,304 Facebook posts)
  • SODA (10,000 sampled scenarios)
  • BIG5-CHAT (100,000 synthetic single-turn dialogues)
  • PAPI-120-600K human IPIP-NEO response data
  • Big Five Essay (2,468 essays)
  • SocialIQA, GSM8K, MathQA, TruthfulQA, CommonsenseQA, PIQA, MMLU, GPQA, HumanEval and MBPP

Evidence and location

  • Construction of expert generators and human/synthetic origin of the data: ACL 2025, pp. 20436-20439, sections 3-5.1 and Table 1
  • Questionnaire results, reasoning, and stated limits: ACL 2025, pp. 20439-20443, sections 5.2-8 and Tables 2-4
  • Human evaluations and inter-annotator agreement: ACL 2025, pp. 20454-20456, Appendix E.1-E.2 and Tables 10-11
  • Additional baselines, internal correlations, and full reasoning: ACL 2025, pp. 20457-20465, Appendices E.3-F.2 and Tables 12-15
  • Generalization to Big Five Essay and clarification of the synthetic pipeline: ACL 2025, pp. 20451-20453 and 20466-20467, Appendices D.1 and H; Figure 5
  • Bias, hallucination, and linguistic analysis: ACL 2025, pp. 20465-20471, Appendices G-I and Table 16