Post-training makes large language models less human-like

Trait induction and control2026arXivApproved editorial review

Authors: Marcel Binz, Elif Akata, Abdullah Almaatouq, Mohammed Alsobay, Oleksii Ariasov, Franziska Brändle, David Broska, Jason W. Burton, Nuno Busch, Frederick Callaway, Vanessa Cheung, Brian Christian, Julian Coda-Forno, Can Demircan, Vittoria Dentella, Maria K. Eckstein, Noémi Éltető, Michael Franke, Thomas L. Griffiths, Fritz Günther, Susanne Haridi, Sebastian Hellmann, Stefan Herytash, Linus Hof, Eleanor Holton, Isabelle Hoxha, Zak Hussain, Akshay Jagadish, Elif Kara, Valentin Kriegmair, Evelina Leivada, Li Ji-An, Tobias Ludwig, Maximilian Maier, Marcelo G. Mattar, Marvin Mathony, Alireza Modirshanechi, Robin Na, Mariia Nadverniuk, Antonios Nasioulas, Surabhi S. Nath, Helen Niemeyer, Kate Nussenbaum, Sebastian Olschewski, Thorsten Pachur, Stefano Palminteri, Aliona Petrenco, Camille V. Phaneuf-Hadd, Angelo Pirrone, Manuel Rausch, Laura Raveling, Shashank Reddy, Milena Rmus, Evan M. Russek, Tankred Saanum, Kai Sandbrink, Louis Schiekiera, Johannes A. Schubert, Luca M. Schulze Buschoff, Nishad Singhi, Leah H. Somerville, Mikhail S. Spektor, Xin Sui, Christopher Summerfield, Mirko Thalmann, Anna I. Thoma, Taisiia Tikhomirova, Vuong Truong, Polina Tsvilodub, Konstantinos Voudouris, Kristin Witte, Shuchen Wu, Dirk U. Wulff, Hua-Dong Xiong, Songlin Xu, Lance Ying, Xinyu Zhang, Jian-Qiao Zhu, Eric Schulz

Keywords: Psych-201, Behavioral alignment, Human response likelihood, Post-training tax, Base versus post-trained models, Persona induction, Individual-level prediction, Cognitive modeling, Negative log-likelihood, Behavioral surrogate validity, Reproducibility audit

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

79: Authors
8: Findings
27: Limitations
4: Evidence

Editorial summary

English

The paper introduces Psych-201, a corpus of human experimental sessions rendered as text, and compares the probability base and post-trained models assign to 25.9 million marked responses. It reports small average disadvantages for instruction-tuned (d=0.11), reasoning (d=0.14) and vision models (d=0.07), while Centaur, post-trained specifically for behavior, improves on novel tasks (d=0.28). The pattern matters for choosing behavioral surrogates, but “human-like” here means lower NLL on text-transcribed tasks with earlier ground-truth human responses visible, not general resemblance to people. The claim that persona induction does not help is not cleanly identified: released code compares metadata-bearing participants with an all-participant baseline, and the persona prefix removes more behavioral history under the fixed cap. Moreover, the advertised 32K limit is implemented in characters, not tokens; in the public train snapshot, 6.87% of rows exceed it and only 81.13% of response markers occur before the cutoff. Code and data exist, but outputs, a pinned environment, version lineage and central dataset governance are missing.

Español

El trabajo presenta Psych-201, un corpus de sesiones experimentales humanas convertidas a texto, y compara la probabilidad que modelos base y post-entrenados asignan a 25,9 millones de respuestas marcadas. Reporta desventajas medias pequeñas para modelos instruction-tuned (d=0,11), de razonamiento (d=0,14) y visión (d=0,07), mientras Centaur, post-entrenado específicamente en conducta, mejora en tareas nuevas (d=0,28). El patrón es relevante para elegir sustitutos conductuales, pero «human-like» significa aquí menor NLL en tareas textificadas con respuestas humanas previas visibles, no parecido humano general. La afirmación de que la inducción de persona no ayuda no queda limpiamente identificada: el código compara participantes con metadatos contra un baseline con todos los participantes y el prefijo de persona recorta más historia. Además, el límite de 32K se aplica como caracteres, no tokens; en el train público, 6,87% de filas lo superan y sólo 81,13% de los marcadores de respuesta quedan antes del corte. Código y datos existen, pero faltan outputs, entorno fijado, trazabilidad de versiones y gobernanza central del dataset.

Research question

How does the capacity of different LLMs to assign probability to human responses from behavioral experiments change after instruction tuning, reasoning training, or vision, in which domains does the gap appear, and does prepending metadata of each participant provide additional information?

Method

Psych-201 serializes complete sessions with instructions, stimuli, previous responses, feedback, and context; human responses are delimited with << >>. The test reserves per study 10% of participants, with a maximum of 100. For each checkpoint, the total NLL of each response span is calculated by teacher forcing and means are aggregated by experiment; the base/post difference is standardized with an unpaired version of Cohen's d and averaged across experiments and pairs. The persona condition prepends age, gender, nationality, education, diagnosis, or available questionnaires. This review inspected the 34 pages of v2, all the TeX, the v1-v2 diff, official code, split, metrics, plots, repository, public Hugging Face snapshot, licenses, coverage arithmetic, and reproducibility.

Sample: The manuscript reports 208,021 participants and 25,906,599 responses. The public train contains 208,021 rows, 149 studies, 264 experiment labels, and 25,924,000 << markers; all its rows have in_eval=false. stats.py itself sets 211,964 thousand participants, while the code separately publishes in_eval=true in a gated test, so the text appears to report train as total and does not document the correspondence of snapshots. The test takes 10% per study with a cap of 100. The evaluation traverses multiple sizes and objectives of three main lineages; per-run results and NLL tensors are not published.

Findings

The reported averages are d=0.11 for instruction tuning, d=0.14 for reasoning, and d=0.07 for vision, always as higher NLL relative to the selected base comparator.
Qwen2.5-72B base is the model with the lowest reported mean NLL, 1.557.
The mean gap of the Qwen series increases from 0.02 in Qwen2 and 0.04 in Qwen2.5 to 0.13 in Qwen3 and 0.16 in Qwen3.5.
Reasoning and psycholinguistics show the largest mean gaps by domain; the effect appears in all aggregated domains, with heterogeneous magnitude.
The published effect of the metadata prefix moves between d=-0.02 and d=0.02, without a consistent mean advantage.
Centaur improves on new tasks not used in its tuning, d=0.28 and SEM=0.10, demonstrating that behavior-oriented post-training can help.
The central result is a descriptive pattern of NLL in Psych-201, not a quantified reduction of humanity, personality, or general social realism.
Independent calculation on the public train finds 14,286 rows of more than 32,768 characters (6.87%) and 4,891,912 response markers after the cutoff; 81.13% remain before the limit.

Limitations

"Human-like" is operationalized as NLL of responses in tasks converted to text; it does not cover general human likeness, natural conversation, emotion, personality, or representativeness.
Each subsequent response is predicted seeing the previous human choices of the same participant; it is sequential teacher forcing, not zero-shot simulation of an unobserved person.
NLL is summed by tokens within each response, so response length and format affect the comparison between studies.
The causal argument that post-training erases heuristics is not identified: checkpoints differ in data, objectives, architecture, and training.
--max-seq-length=32768 cuts characters before tokenizing, despite the repository documenting 32K tokens.
In the train snapshot, 18.87% of response markers remain outside the evaluable prefix; the exact impact on the gated test cannot be calculated without authorization.
The metadata prefix consumes the same limit and leaves less experimental history for the persona condition.
eval_meta filters to people with metadata; the baseline includes all and plot_meta does not match participant IDs.
In train, 15 of 149 studies have partial metadata coverage; their comparisons remain confounded by sample composition.
The persona test asks for incremental value over a history that already contains individual behavior, not for the general utility of persona.
A single field among 25 suffices to enter the analysis; minimal and dense clinical profiles are mixed without ablations or missingness strata.
Cohen's d ignores that the same observations are scored by both models; it does not use the covariance of a paired design.
There are no intervals for the headlines, bootstrap by study, paired tests, hierarchical model, or multiplicity correction.
Several vision pairs do not match size or architecture: 2B vs 1.7B, 32B vs 14B, and Llama Vision 11B/90B vs base 8B/70B.
The same post-trained Qwen3/Qwen3.5 are labeled Reasoning in Figure 2 but instruction-tuned in Figure 3 and the text.
The detailed prose says almost all comparisons, while abstract/captions suggest universal consistency.
The public train has 208,021 rows and stats.py 211,964 thousand participants; train, total, and gated test are not reconciled in a data card.
The public train contains 25,924,000 markers, 17,401 more than the paper's total responses even before the test.
The Hugging Face data card is empty and does not document versions, sources, licenses, consent, deidentification, or prohibited uses.
Apache-2.0 for train and CC-BY-ND-4.0 for test do not by themselves accredit compatibility of the original licenses of 149 studies.
Age, nationality, diagnosis, and mental health questionnaires are open without a visible central privacy assessment.
The review of contributions is defined as light; an unpublished replication included declares not having obtained IRB.
No .pth, fused CSVs, or outputs needed to regenerate figures without repeating all inference are published.
There are no requirements, lockfile, container, end-to-end commands, immutable model revisions, or complete hardware specification.
The scripts force offline/CUDA mode and globally patch subprocess to rewrite PTX headers, reducing portability.
Two .py generators fail compileall: one retains !pip install from Colab and another an invalid line continuation.
There are no tests, CI, result fixtures, or span extraction coverage proof for each tokenizer.

What the study does not establish

That post-training makes models less human in a general psychological or social sense.
That all forms of post-training reduce behavioral alignment; Centaur provides a positive counterexample.
That the causal mechanism is normative correction or the elimination of human biases.
That instruction tuning, reasoning, and vision are comparable as controlled interventions.
That vision differences do not arise from architecture, modality, or size.
That the effects are large, statistically precise, or equal across studies and families.
That persona induction is not useful for individual prediction when exactly the same people and histories are compared.
That age, gender, nationality, and questionnaires are equivalent as persona signals.
That a persona prefix does not help in zero-shot, without prior behavioral history or under another context budget.
That Psych-201 validates LLMs as safe substitutes for human participants.
That translating any experiment to text preserves all relevant stimuli and mechanisms.
That 208,021 is the complete total of participants including the gated test.
That all contributions have consent, privacy, and license compatible with aggregated redistribution.
That the figures and numbers can be exactly reproduced from the repository in a clean environment.
Acceptance in Science, Nature, or another peer-reviewed publication.

Traceability

Scope: Full text

Version: arXiv:2605.07632v2; 34-page PDF, complete v2 TeX, v1-v2 diff, official code commit c63dea18897393b706227b50024b27fa8d8907d8 and public dataset revision 5cf12dfa730bff530e45ee92e2e1c9fb24690e98

Consulted source: https://arxiv.org/abs/2605.07632v2

Review: Codex 34-page visual full-text, v1-v2 TeX, Psych-201 code, public dataset, construct, persona-comparison, truncation, statistics, ethics and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

Qwen1.5, Qwen2, Qwen2.5, Qwen3 and Qwen3.5 base/post-trained checkpoints
Qwen code, vision and reasoning variants
Llama 3, 3.1 and 3.2 base, instruct and vision checkpoints
DeepSeek-R1-Distill Qwen and Llama variants
OLMo 3/3.1 base, instruct and think checkpoints
Centaur behavioral foundation model

Instruments and metrics

Psych-201 natural-language behavioral transcripts
Human-response span negative log-likelihood
Discrete-choice next-token accuracy
Per-experiment standardized mean difference
Interview-style participant metadata prefix
Behavioral-domain taxonomy
Prompt-template sensitivity comparison
Centaur held-out task comparison

Data used

Psych-201 public train revision 5cf12dfa730bff530e45ee92e2e1c9fb24690e98
Psych-201-test gated evaluation repository
Psych-201-discrete and gated Psych-201-discrete-test
Psych-101 tasks used for Centaur training

Evidence and location

Text, figures, methods, availability, and study list: arXiv:2605.07632v2; PDF sha256 2a809e3232944e3842cfe603e618df694b691f7f79690fc4166697ae490ebc08; TeX sha256 389699cbaf0cf3914a0e1f5c817fdd7f1695b2f42a4f1940060b33fa9f92eb4a
Evaluation, split, persona, metrics, and official plots: https://github.com/marcelbinz/Psych-201/tree/c63dea18897393b706227b50024b27fa8d8907d8
Rows, studies, experiments, metadata coverage, lengths, and markers: https://huggingface.co/datasets/marcelbinz/Psych-201 revision 5cf12dfa730bff530e45ee92e2e1c9fb24690e98
Truncation recalculations, persona comparison, model pairs, governance, and reproduction: reports/verification/article-346-psych201-human-likeness-persona-comparison-truncation-dataset-version-statistics-ethics-and-reproducibility-audit.json