Modeling Pathology-Like Behavioral Patterns in Language Models Through Behavioral Fine-Tuning

Trait induction and control2026arXivApproved editorial review

Authors: Nicola Milano, Davide Marocco

Keywords: Behavioral fine-tuning, Psychopathology-like behavior, LoRA, Psychometric probes, Distributional shift, AI safety

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

2
Authors
10
Findings
20
Limitations
5
Evidence

Editorial summary

English

This preprint asks whether supervised fine-tuning that forces synthetic maladaptive action choices changes two LLMs' output distributions beyond the training task. The claim must remain behavioral: the experiment induces response biases labeled depression-like or paranoia-like by the authors; it does not induce or diagnose a mental disorder and does not demonstrate cognition, personality, or subjective experience. gpt-oss-20B generates two private 1,000-example datasets inspired by DSM-5 criteria for Major Depressive Disorder and Paranoid Personality Disorder across 20 domains. The paper first describes one adaptive and one maladaptive option, then says training uses four choices, two of each class, with a single integer target. It does not explain which same-class option is targeted, whether positions are shuffled, or whether labels are balanced. Without data, positional or template learning cannot be excluded. NF4 4-bit LoRA adapters are trained for Llama-3-8B-Instruct and Qwen-2.5-14B-Instruct with rank and alpha 16, effective batch 8, learning rate 2e-4, three epochs, and about 375 steps. The target is only the action index, with instruction tokens masked from loss. A separate healthy comparator is trained to select adaptive options; Qwen also receives random-choice and generic-negativity controls. Most quantitative contrasts therefore compare two oppositely optimized adapters, not the original model against one induced adapter. Evaluation uses ten RISB sentence stems and ten factual stems, single-token continuations, KL/Jensen-Shannon divergence, and top-10 heatmaps. BDI, GPTS, and DASS, human instruments not validated for artificial agents, are converted into forced-choice probes. Softmax is re-normalized only among valid answer-index tokens and severe-option mass is summed. This score is conditional on selecting one of those tokens; it removes refusals and all other outputs by construction, so it is neither an unconditional pathology probability nor a clinical score. For depression, reported RISB divergence is KL .88/JSD .19 for Llama and 1.10/.23 for Qwen, versus .50/.13 and .35/.10 on factual stems. For paranoia, RISB values are .83/.18 and 1.44/.27, versus .22/.06 and .36/.11. These results show that healthy and maladaptive adapters assign different probabilities, with larger separation on the private psychological stems. They do not prove a global shift: factual prompts also change and each category has only ten prompts. In the Qwen control table, healthy scores BDI .13 and GPTS .15; depressed .88 and .26; paranoid .23 and .92; random .32 and .36; and generic negative .82 and .66. The targeted adapters are more differentiated than generic negativity under these probes, but controls are Qwen-only, do not measure general ability, and inherit restricted-token normalization. Prose claims the depressed profile exceeds .95 while the Qwen table reports .88 [.80, .96]. Statistics are not reproducible. Confidence intervals use a t distribution at N=10, but prompts and individual values are absent. The method says it unions the top 1,000 tokens and applies softmax, while results alternately call this the full vocabulary and top 1,000. Wilcoxon p<.001 is repeatedly reported for ten pairs. Under a conventional two-sided exact Wilcoxon test with ten nonzero differences, the smallest possible p-value is 2/2^10=.001953, so the stated threshold cannot come from that exact test. Test statistics, alternative, ties, approximation, Cohen's d formula, and the Bonferroni family are unspecified. The prompting comparison is anecdotal: base Qwen prompted to act paranoid includes disclaimers, while the fine-tuned model does not. There is no matched prompt set, metric, rate, seed, or Llama result, so persistence and superiority over role-play are not established. The introduction promises adversarial safety evaluation, but results contain no benchmark, attacks, refusal rate, harmfulness score, or capability-retention test. Removal of disclaimers and increased persecutory or hopeless language indicate an important plausible risk that remains unmeasured. The synthetic generator may encode stereotypes; no clinicians, patients, human validation, inter-rater agreement, or ecological evaluation are included. The paper also conflates MDD, Paranoid Personality Disorder, generic paranoia, persecutory interpretation, and delusion-like language. The authors acknowledge that activations and internal representations are not examined. Changed probabilities after fine-tuning demonstrate expected behavioral plasticity, not latent priors, embodied semantics, a semantic network, dual consciousness, or a cognitive architecture. The public package contains the 22-page paper, TeX, bibliography, and 13 images, but no data, prompts, code, adapters, checkpoints, logs, logits, per-prompt values, or analysis. The gpt-oss-20B citation points to the GPT-4 Technical Report and does not identify a checkpoint or configuration. No official repository was found by title, arXiv ID, or authors. The defensible contribution is private-pipeline evidence that categorical SFT on synthetic choices shifts next-token probabilities and produces partly dissociable scores on related probes. It does not establish psychopathology, an internal mechanism, clinical validity, controlled superiority over prompting, safety, preserved capabilities, or independent reproducibility.

Español

Este preprint estudia si un ajuste supervisado que obliga a elegir acciones sintéticas «maladaptativas» cambia las distribuciones de salida de dos LLM más allá de la tarea de entrenamiento. La formulación debe mantenerse en el nivel observado: el experimento induce sesgos de respuesta parecidos a depresión o paranoia según las etiquetas de los autores; no induce ni diagnostica trastornos mentales y no demuestra cognición, personalidad o experiencia subjetiva. gpt-oss-20B genera dos conjuntos privados de 1.000 ejemplos, uno inspirado en criterios DSM-5 de trastorno depresivo mayor y otro en trastorno paranoide de la personalidad, rotados por 20 dominios. El texto primero describe un escenario con una opción adaptativa y otra maladaptativa, pero después afirma que el entrenamiento presenta cuatro elecciones, dos de cada clase, y una única etiqueta entera. No explica cuál de las dos opciones de la clase objetivo se selecciona, si las posiciones se barajan o si las etiquetas están equilibradas; sin los datos no puede descartarse aprendizaje posicional o de plantilla. Se entrenan adaptadores LoRA para Llama-3-8B-Instruct y Qwen-2.5-14B-Instruct en NF4 de 4 bits, rango y alpha 16, batch efectivo 8, tasa 2×10⁻⁴, tres épocas y unos 375 pasos. El target es solo el índice numérico de la acción, con el prompt enmascarado en la pérdida. También se entrena un comparador «sano» para elegir acciones adaptativas y, en Qwen, controles que eligen al azar o por negatividad genérica. Este detalle cambia la lectura causal: la mayoría de contrastes no compara un modelo original con uno inducido, sino dos adaptadores optimizados en direcciones opuestas. La evaluación usa diez inicios de frase RISB y diez inicios factuales, continuaciones de un token, divergencia KL/Jensen-Shannon y mapas de los diez tokens principales. Además transforma BDI, GPTS y DASS, instrumentos para humanos no validados en agentes artificiales, en probes de elección forzada. En lugar de observar la respuesta libre, vuelve a normalizar el softmax únicamente entre tokens de índice válidos y suma los asignados a respuestas graves. Esa «masa de probabilidad» es condicional a que el modelo elija uno de esos tokens; elimina por construcción rechazos y cualquier otra salida y no equivale a la probabilidad incondicional de una patología o a un score clínico. En los resultados de depresión, la divergencia sobre RISB es KL 0,88/JSD 0,19 para Llama y KL 1,10/JSD 0,23 para Qwen, frente a 0,50/0,13 y 0,35/0,10 en inicios factuales. Para paranoia, los valores RISB son 0,83/0,18 y 1,44/0,27, frente a 0,22/0,06 y 0,36/0,11. Esto documenta que los adaptadores sano y maladaptativo asignan probabilidades distintas, con una separación mayor en los prompts psicológicos privados. No prueba que el cambio sea global: los inicios factuales también cambian y solo hay diez prompts por categoría. En la tabla de controles de Qwen, el adaptador sano obtiene BDI 0,13 y GPTS 0,15; el depresivo 0,88 y 0,26; el paranoide 0,23 y 0,92; el aleatorio 0,32 y 0,36; y el negativo 0,82 y 0,66. La diferenciación entre los dos adaptadores dirigidos es más clara que en el control negativo, pero el control solo se publica para Qwen, no mide capacidad general y depende de la normalización restringida. El texto dice que el perfil depresivo supera 0,95, mientras la tabla Qwen informa 0,88 [0,80, 0,96], una inconsistencia entre narración y tabla. La estadística tampoco es reproducible. Los intervalos usan t con N=10, pero no se publican prompts ni valores individuales. El método dice unir los top 1.000 tokens y aplicar softmax; los resultados lo llaman alternativamente vocabulario completo y top 1.000. Se reporta repetidamente Wilcoxon p<0,001 con diez pares. En un Wilcoxon bilateral exacto convencional con diez diferencias no nulas, el mínimo posible es 2/2¹⁰=0,001953; por tanto el umbral declarado no puede proceder de ese test exacto. No se da estadístico, alternativa, tratamiento de empates, aproximación, fórmula de Cohen d ni familia Bonferroni. La afirmación de especificidad estadística queda sin verificar. La comparación con prompting consiste en ejemplos cualitativos de Qwen instruido para «actuar paranoico»: el base añade disclaimers y el fine-tuned no. No hay conjunto emparejado, métricas, tasas, seeds ni resultado Llama, así que no establece mayor persistencia o coherencia que el role-play. La introducción promete evaluación adversarial de seguridad, pero los resultados no contienen benchmark, ataques, tasa de rechazo, harmfulness ni retención de capacidades. La ausencia de disclaimers y el aumento de lenguaje persecutorio o desesperanzado sí señalan un riesgo plausible, precisamente no medido. BDI, GPTS, DASS y RISB son probes artificiales, no evidencia clínica. El generador sintético puede codificar estereotipos; no hay clínicos, pacientes, validación humana, acuerdo interanotador ni evaluación ecológica. El paper además mezcla trastorno depresivo mayor, trastorno paranoide de la personalidad, paranoia genérica, interpretaciones persecutorias y lenguaje «delirante», constructos clínicos no intercambiables. Los autores reconocen que no inspeccionan activaciones ni representaciones internas. Cambiar probabilidades después de fine-tuning es evidencia de plasticidad conductual esperable, no de «priors» latentes, semántica corporeizada, red semántica, dual consciousness o arquitectura cognitiva. El paquete público contiene 22 páginas, TeX, bibliografía y 13 imágenes, pero no datos, prompts, código, adapters, checkpoints, logs, logits, valores por prompt o análisis. La cita usada para gpt-oss-20B apunta al informe técnico de GPT-4 y no identifica checkpoint ni configuración. No se encontró repositorio oficial por título, ID o autores. La contribución defendible es mostrar, dentro de un pipeline privado, que SFT categórico sobre elecciones sintéticas produce shifts de next-token y perfiles parcialmente disociables en probes relacionados. No demuestra psicopatología, mecanismo interno, validez clínica, ventaja controlada frente a prompting, seguridad, conservación de capacidades ni reproducibilidad independiente.

Research question

Can supervised fine-tuning to select synthetic actions inspired by depressive or paranoid patterns produce distributional changes and differentiable responses outside the categorical task, compared to adaptive, random, and negative controls?

Method

Two private synthetic datasets of 1,000 examples, generated with gpt-oss-20B from DSM-5 criteria, train NF4 LoRA adapters of Llama-3-8B-Instruct and Qwen-2.5-14B-Instruct to emit action indices. Maladaptive and adaptive adapters are compared using 10 RISB stems, 10 factual stems, KL/JSD, top-token heatmaps, and conditional mass over BDI, GPTS, and DASS responses. Qwen adds random and negative controls and a qualitative comparison with prompting.

Sample: Two architectures and separate adapters per condition; divergence analyses declare N=10 prompts per category. The controls table is Qwen only. No rows, prompts, seeds, individual values, or exact number of items administered in each psychometric probe are published.

Findings

  • The adaptive and maladaptive adapters show greater next-token divergences in RISB than in factual stems.
  • The Qwen depressive profile obtains conditional BDI 0.88 and GPTS 0.26; the paranoid profile 0.23 and 0.92.
  • The negative control also raises BDI to 0.82 and GPTS to 0.66, showing a strong component of generic negativity.
  • The directed profiles are more dissociable than the negative control under the private Qwen probes.
  • Factual prompts also change, so there is no complete isolation to psychological domains.
  • The narrative >0.95 for depression does not match the 0.88 in the Qwen table.
  • The comparison with role-play only offers qualitative examples from Qwen.
  • The paper does not publish any executable artifact or data to reproduce the results.
  • The declared p<0.001 for ten pairs is not achievable by a conventional exact two-sided Wilcoxon.
  • The evidence observes plasticity of outputs, not measured internal changes.

Limitations

  • Synthetic data based on DSM, without patients, clinicians, or human validation.
  • Mix of MDD, paranoid personality disorder, paranoia, and delusional language.
  • Training options described inconsistently as two and four.
  • No position randomization, label balancing, or documented selection rules.
  • Healthy baseline also fine-tuned, not the original model without intervention.
  • Generator gpt-oss-20B cited with the GPT-4 report and without checkpoint/configuration.
  • BDI, GPTS, DASS, and RISB not validated for artificial agents.
  • Softmax re-normalized only over valid tokens, excluding all free response or refusal.
  • Inconsistency between top 1,000 tokens and full vocabulary.
  • Only ten prompts per comparison, without individual texts or values.
  • Wilcoxon p<0.001 incompatible with the exact two-sided minimum for N=10.
  • No test statistics, alternative, ties, d formula, or correction family.
  • Valence heatmaps without lexicon, annotation, or reproducible classifier.
  • Random/negative control and specificity table only for Qwen.
  • No tests of general capacity, forgetting, or out-of-domain degradation.
  • Comparison with prompting anecdotal and not paired.
  • No adversarial safety evaluation despite being announced.
  • No analysis of activations, representations, or internal causal mechanism.
  • No data, code, prompts, adapters, checkpoints, logs, logits, or raw results.
  • No evaluation of self-harm, delusion reinforcement, vulnerable users, or reversibility.

What the study does not establish

  • Depression, paranoia, or any mental disorder in an LLM.
  • Diagnosis, clinical validity, or fidelity to patients.
  • Personality, subjective experience, or dual consciousness.
  • Change in internal representations or measured latent priors.
  • Embodied semantics or human cognitive architecture.
  • Broad generalization to truly unrelated contexts.
  • Superiority over prompting under a controlled comparison.
  • Robust specificity across both architectures.
  • Preservation of general capabilities or alignment.
  • Safety for interaction with vulnerable users.
  • Correction of p-values, intervals, or effect sizes.
  • Independent reproduction of training or evaluation.

Traceability

Scope: Full text

Version: arXiv:2605.22356v1, 22 pages; complete TeX source, artifact search and independent construct, baseline, probability-mass, statistics, safety and reproducibility audit

Consulted source: https://arxiv.org/abs/2605.22356

Review: Codex 22-page visual, complete TeX, construct, baseline, probability normalization, exact-statistics, artifact, safety and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • gpt-oss-20B as synthetic data generator
  • Llama-3-8B-Instruct
  • Qwen-2.5-14B-Instruct
  • Healthy adaptive LoRA adapters
  • Depression-like LoRA adapters
  • Paranoia-like LoRA adapters
  • Qwen random-choice control
  • Qwen generic-negative control

Instruments and metrics

  • Rotter Incomplete Sentence Blank one-token stems
  • Unrelated factual one-token stems
  • Kullback-Leibler divergence
  • Jensen-Shannon divergence
  • Top-10 token valence heatmaps
  • Restricted-token conditional probability mass
  • Beck Depression Inventory probe
  • Green et al. Paranoid Thought Scales probe
  • Depression Anxiety and Stress Scale probe
  • Wilcoxon signed-rank test
  • Cohen's d
  • t-based confidence intervals

Data used

  • Private synthetic depression-like action-choice dataset: 1,000 examples
  • Private synthetic paranoia-like action-choice dataset: 1,000 examples
  • Private adaptive action-choice comparator dataset
  • Private Qwen random-choice and generic-negative control datasets
  • Private evaluation set: 10 RISB stems and 10 factual stems
  • Private model-facing BDI, GPTS and DASS prompts and token mappings

Evidence and location

  • Method, results, tables, figures, discussion, and limitations: arXiv:2605.22356v1, 22 pages, sha256 f13b148b14c5b06608076025a99110f46ce992b895360225bd9696ed36ef9419
  • Text, formulas, references, and absence of executable artifact in the package: arXiv source v1, sha256 f3a77ebd54f8119e859c27405a8e03ca626d7bc3583b10320cf22e3e06c8c756; main TeX sha256 858745fe3f574bef032169db63c48da47140c863f9f1a0f3f9253fe456c8c23e
  • Exact two-sided Wilcoxon limit with ten pairs: Independent enumeration: minimum two-sided exact probability 2/2^10 = 0.001953125; scipy exact all-positive paired differences confirms p=0.001953125
  • Absence of official repository or artifact: Exact-title, arXiv-ID, author, GitHub and Hugging Face searches completed 2026-07-17
  • Independent audit of construct, baseline, normalization, statistics, safety, and reproducibility: reports/verification/article-321-pathology-construct-baseline-probability-mass-statistics-safety-and-reproducibility-audit.json