HER: Human-like Reasoning and Reinforcement Learning for LLM Role-playing

Personas, identity, and agents2026ACL AnthologyApproved editorial review

Authors: Chengyu Du, Xintao Wang, Aili Chen, Weiyuan Li, Rui Xu, Junteng Liu, Zishan Huang, Rong Tian, Zijun Sun, Yuhao Li, Liheng Feng, Deming Ding, Pengyu Zhao, Yanghua Xiao

Keywords: Role-playing language models, Persona simulation, Dual-layer thinking, Generative reward model, Reinforcement learning

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

14
Authors
9
Findings
12
Limitations
4
Evidence

Editorial summary

English

HER proposes a two-layer textual format for role-play. Before responding, the model generates a hidden third-person system-thinking block to plan how to portray the character; it then produces first-person role thinking, actions, and dialogue. These traces are not observations of human thought: an unidentified commercial teacher synthesizes them from CoSER literary dialogues. The pipeline adds inner monologue, diversifies the ordering of thought/action/speech, and rewrites the plan offline with access to the realized continuation. Qwen3-32B-Base receives SFT followed by GRPO/DAPO-style RL. A Qwen3-32B generative reward model compares the policy response with a frozen SFT checkpoint using case-dependent principles.

The paper reports 53.12 on CoSER and 65.73 on MiniMax Role-Play Bench for HER-RL, versus 22.86 and 50.76 for Qwen3-32B. The differences are 30.26 and 14.97 score points, not percentages. Relative to the baselines, they correspond to approximately 132.37% and 29.49%. The gain attributable to RL over HER-SFT is much smaller: 2.20 CoSER points and 7.29 MiniMax points. HER-RL ranks eighth in the main table and remains below several proprietary models. The ablation raises CoSER from 48.64 without system thinking to 50.92 with that layer and 53.12 with RL, but it also changes computation, format, and supervision; it does not identify a human cognitive mechanism.

CoSER evaluation uses 200 twenty-round conversations and a single Qwen3-235B-A22B judge to score Qwen-derived and other systems. There are no CoSER intervals, repeated training runs, or seed-level ledger. The teacher-human comparison reserves only 50 pairs for testing after 150 are used to refine the prompt: agreement with consensus is 80.5%, and agreement between two experts is 84.0%. A separate calibration reports 81.5% raw agreement without a sample size. Table 3 also mentions 4,739 expert-annotated pairs without documenting their protocol. This evidence calibrates textual role-play preferences; it does not establish human reasoning, authentic mental states, consciousness, or psychological personality. The diversity evidence is mostly tag-order and n-gram variety.

The public release is substantial but does not reproduce the study end to end. Complete BF16 HER-32B and HER-RM-32B weights, four large JSONL files, a demo, and 69 code files are available; all Python files compile. However, there are no tests, CI, lockfile, logs, or result ledger; scripts required by the documentation are missing, many paths remain /path/to placeholders, ms-swift/verl launch configurations are absent, and the CoSER data expected by the evaluator are not included. The code stopped in January, before arXiv v4 and the final ACL paper. The released dataset also differs from the paper: 76,883 versus 72,656 multi-turn samples and 342,493 versus 323,600 single-turn samples.

Two ethical contradictions matter. The paper says that no user or user-derived data are used, while Appendix F describes explicit and implicit feedback collected during normal deployment. It also says that no raw copyrighted source text is released, while the dataset card declares original literary-text fields and public samples reproduce identifiable novel passages. An Apache-2.0 label does not itself resolve rights to each underlying work. The supported contribution is a structured-generation architecture, a contextual reward model, usable weights, and benchmark evidence of stronger role-play for an adapted Qwen model. It is not evidence of human cognitive emulation or a fully reproducible release. The paper is peer reviewed in Findings of ACL 2026, DOI 10.18653/v1/2026.findings-acl.1283.

Español

HER propone un formato de role-play con dos capas textuales. Antes de responder, el modelo genera un bloque oculto de system thinking en tercera persona para planificar cómo representar al personaje; después produce role thinking en primera persona, acciones y diálogo. Esas trazas no proceden de pensamientos humanos observados: un modelo comercial no identificado las sintetiza a partir de diálogos literarios de CoSER. El pipeline añade monólogo interno, diversifica el orden de pensamiento/acción/habla y reescribe el plan con acceso offline a la continuación realizada. Qwen3-32B-Base recibe SFT y después RL tipo GRPO/DAPO. Un reward model generativo Qwen3-32B compara la respuesta de la política con la de un checkpoint SFT congelado usando principios dependientes del caso.

El paper informa 53,12 en CoSER y 65,73 en MiniMax Role-Play Bench para HER-RL, frente a 22,86 y 50,76 para Qwen3-32B. Las diferencias son 30,26 y 14,97 puntos, no porcentajes. Respecto al baseline, equivalen aproximadamente a 132,37% y 29,49% relativos. El salto atribuible a RL sobre HER-SFT es mucho menor: 2,20 puntos CoSER y 7,29 MiniMax. En la tabla principal HER-RL ocupa el octavo lugar y queda por debajo de varios modelos propietarios. La ablación sube CoSER de 48,64 sin system thinking a 50,92 con esa capa y a 53,12 con RL, pero también cambia cómputo, formato y supervisión; no identifica un mecanismo cognitivo humano.

La evaluación CoSER usa 200 conversaciones de 20 rondas y un único juez Qwen3-235B-A22B para puntuar modelos derivados de Qwen y otros sistemas. No hay intervalos de CoSER, repeticiones de entrenamiento ni ledger por semilla. El contraste teacher-humano reserva solo 50 pares para test después de usar 150 para ajustar el prompt: 80,5% de acuerdo con consenso y 84,0% entre dos expertos. Otra calibración declara 81,5% de acuerdo bruto sin tamaño muestral. Tabla 3 menciona además 4.739 pares anotados por expertos sin documentar su protocolo. Esta evidencia calibra preferencias textuales de role-play; no demuestra razonamiento humano, estados mentales auténticos, conciencia ni personalidad psicológica. La diversidad medida es principalmente variedad de órdenes de etiquetas y n-gramas.

La liberación pública es sustancial pero no reproduce el trabajo de extremo a extremo. Existen los pesos completos BF16 de HER-32B y HER-RM-32B, cuatro JSONL grandes, una demo y 69 archivos de código; todos los Python compilan. Sin embargo, no hay tests, CI, lockfile, logs ni resultados, faltan scripts que exige la propia guía, muchas rutas siguen como /path/to, no se publican las configuraciones ms-swift/verl ni los datos CoSER que espera el evaluador. El código quedó en enero, antes de arXiv v4 y ACL final. El dataset público tampoco coincide con el paper: 76.883 frente a 72.656 muestras multi-turn y 342.493 frente a 323.600 single-turn.

Hay dos contradicciones éticas importantes. El texto afirma que no usa datos de usuarios, pero el Apéndice F describe feedback explícito e implícito recogido durante despliegue normal. También afirma que no libera texto fuente protegido, mientras la ficha del dataset declara campos con texto literario original y las muestras públicas reproducen pasajes identificables de novelas. La etiqueta Apache-2.0 no resuelve por sí sola los derechos de cada obra. La contribución válida es una arquitectura de generación estructurada, un reward model contextual, pesos utilizables y evidencia benchmark de mejor role-play para un Qwen adaptado. No es prueba de emulación cognitiva humana ni un release plenamente reproducible. El trabajo sí está revisado por pares en Findings of ACL 2026, DOI 10.18653/v1/2026.findings-acl.1283.

Research question

Do synthetic traces that separate model planning from character monologue, together with a contextual reward model and RL, improve literary character role-play compared to SFT or the base Qwen3-32B?

Method

HER converts literary dialogues from CoSER via a commercial teacher into trajectories with system thinking, role thinking, action, and speech; diversifies their patterns and builds plans through forward generation and retrospective rewriting. It trains Qwen3-32B-Base with SFT and then with GRPO/DAPO. A GenRM Qwen3-32B generates principles per case, compares the policy with frozen HER-SFT, and returns a reward of -1/0/+1. It evaluates 200 CoSER conversations of 20 rounds with Qwen3-235B-A22B and the MiniMax protocol of 100 turns, plus ablations, diversity analysis, and small human calibrations.

Sample: After cleaning CoSER, the paper declares 760 books, 30,069 plots, 29,081 conversations, 17,966 characters, and 383,654 utterances. Table 12 assigns 107,800 samples to role-play SFT, 26,800 to role-play RL, 108,800 to GRM SFT, 80,000 to GRM RL, and 200 to GRM test: 323,600 single-turn derived from 72,656 multi-turn. The release contains 76,883 and 342,493 rows, respectively. CoSER test uses 200 conversations x 20 rounds. The teacher-human validation uses 150 pairs for development and 50 for test; another calibration does not publish N. Table 3 declares 4,739 expert preferences and the balanced GRM is tested on 800 comparisons.

Findings

  • HER-RL obtains 53.12 on CoSER compared to 50.92 for HER-SFT and 22.86 for the base Qwen3-32B.
  • On MiniMax it obtains 65.73 compared to 58.44 for HER-SFT and 50.76 for the base.
  • 30.26 and 14.97 are differences in points; the correct relative percentages over the base are approximately 132.37% and 29.49%.
  • The incremental gain of RL over SFT is 2.20 points on CoSER and 7.29 points on MiniMax.
  • System thinking raises the CoSER ablation from 48.64 to 50.92; RL brings it to 53.12.
  • The balanced GenRM reaches 73.99% compared to 69.91% for the unbalanced one on 800 comparisons and reduces pattern collapse.
  • The teacher-human calibration reports 80.5% agreement on 50 test pairs and 84.0% between two experts.
  • The full weights of the policy and reward model are published and the released Python scripts compile.
  • The counts of the public dataset do not match those of the paper and the full pipeline does not reproduce from the checkout.

Limitations

  • The thought traces are syntheses from the teacher and retrospective rationalizations, not observed mental states.
  • A single Qwen judge evaluates Qwen systems without comparison across judge families.
  • CoSER has no intervals or per-seed results; no training repetitions are reported.
  • The main figures are incorrectly labeled as percentages.
  • The human studies are small or have unpublished N/protocol, and do not directly evaluate the full output of HER-RL.
  • Label and n-gram diversity does not demonstrate semantic or psychological diversity.
  • The commercial teacher, its operational prompts, and its snapshots are not fully fixed.
  • Dataset, code, weights, and manuscript do not share an immutable release and their counts diverge.
  • Documented scripts, training configurations, evaluation data, results, tests, and CI are missing.
  • The claim of not using user-derived data contradicts the deployment signals section.
  • The claim of not releasing protected text contradicts the schema and dataset samples.
  • There is no safety evaluation of the released model for impersonation, manipulation, or biases.

What the study does not establish

  • That HER reasons like a human or possesses cognition, consciousness, or authentic internal states.
  • That role thinking is the actual thought of a character or of a person.
  • That stable psychological personality has been measured.
  • That 30.26 and 14.97 are relative percentage improvements.
  • That RL alone explains the entire difference over the base model.
  • That a judge from the same family is independent or impartial.
  • That human agreements generalize to all conversations, dimensions, and models.
  • That structural output variety equates to narrative or psychological depth.
  • That no signals derived from users were used.
  • That no protected literary text was released or that Apache-2.0 resolves all its rights.
  • That the public dataset corresponds exactly to the splits and counts of the paper.
  • That the repository allows reproducing training and tables end to end.
  • That the results generalize to real people, other languages, domains, or deployment risks.
  • That publication in Findings validates broader claims than the reported experimental evidence.

Traceability

Scope: Full text

Version: Findings of ACL 2026 final paper / arXiv:2601.21459v4, 38 pages; ACL record, GitHub commit 1e5fd85, Hugging Face dataset and HER-32B/HER-RM-32B releases also audited

Consulted source: https://aclanthology.org/2026.findings-acl.1283/

Review: Codex 38-page Findings/arXiv-v4 visual, ACL-publication, construct, score-arithmetic, judge/human-study, dataset-count, privacy/copyright, GitHub and Hugging Face artifact audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Qwen3-32B-Base (policy backbone)
  • HER-SFT
  • HER-RL / ChengyuDu0123/HER-32B
  • HER-RM-32B generative reward model
  • Qwen3-235B-A22B evaluation judge
  • Unidentified commercial teacher model
  • Commercial and open comparison models listed in Table 2

Instruments and metrics

  • Dual-layer system-thinking and role-thinking format
  • Three-stage reverse-synthesis pipeline
  • Case-dependent principle generation
  • Pairwise generative reward model
  • GRPO with DAPO techniques
  • CoSER SC, AN, CF and SQ scores
  • MiniMax Worlds, Stories and Preferences scores
  • Human win/lose/tie preference labels
  • Pattern concentration, Shannon entropy, Distinct-n and Self-BLEU

Data used

  • CoSER literary role-play dialogues
  • HER-Dataset full_info and clean releases
  • HER-Dataset sft_multi_turn and sft_single_turn releases
  • CoSER held-out evaluation conversations
  • MiniMaxAI Role-Play Bench
  • Principle and preference data derived from simulated, expert and reported production interaction signals

Evidence and location

  • Publication, architecture, training, results, appendices, limitations, and ethics: Findings of ACL 2026, pages 25725-25762; ACL Anthology 2026.findings-acl.1283; arXiv:2601.21459v4
  • Weights, dataset, public counts, schema, samples, and licenses: Hugging Face ChengyuDu0123/HER-32B, HER-RM-32B and HER-Dataset revisions checked 2026-07-16
  • Code scope, documentary drift, absent files, placeholders, tests, and CI: cydu24/HER commit 1e5fd856b9e23d3af41445e3c4ca734e24e012ef
  • Result arithmetic, construct validity, human evaluation, privacy, copyright, and reproducibility: reports/verification/article-266-her-publication-construct-metric-human-data-privacy-and-artifact-audit.json