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

2026-05-22 · Source: cs.CL updates on arXiv.org · Field: Science & Research — Mathematics & Computational Sciences, Social Sciences & Behavioral Studies, Health & Medical Research · Depth: Expert, extended

Summary

A new behavioral induction framework fine-tunes large language models (LLMs) to exhibit pathology-like behavioral patterns, moving beyond explicit prompting. Researchers used synthetic datasets, grounded in DSM-5 criteria for Major Depressive Disorder and Paranoid Personality Disorder, to train transformer-based models like Llama-3-8B-Instruct and Qwen-2.5-14B-Instruct. This framework modifies model policies directly through fine-tuning on structured decision-making tasks. The fine-tuning, implemented with LoRA and 4-bit NF4 quantization over 3 epochs on 1,000 examples per disorder, induced stable, context-general shifts in the models' next-token probability distributions. These shifts included increased probability mass for negative and threat-related interpretations, detectable through qualitative completions, psychometric scales (e.g., BDI, GPTS), and quantitative metrics like Jensen–Shannon divergence. The induced profiles demonstrated specificity, with distinct response tendencies for different maladaptive patterns, suggesting policy-level biases. This work positions LLMs as controlled testbeds for studying behavior and language.

Key takeaway

For AI Scientists and Machine Learning Engineers developing LLMs for sensitive applications, understand that behavioral fine-tuning can fundamentally alter your model's underlying interpretive biases. This method bypasses typical safety mechanisms, inducing persistent, pathology-like response patterns that are not superficial role-play. You must rigorously test for unintended policy-level biases when optimizing for specific behavioral objectives, as these can lead to profound and stable shifts in model behavior and language generation.

Key insights

Behavioral fine-tuning can induce specific, stable, pathology-like biases in LLMs, shifting their fundamental interpretive priors.

Principles

• LLM behavior optimization shapes emergent representational structure.
• Consistent behavioral training yields differentiated policy-level biases.
• Self-descriptive outputs are emergent properties of policy regularities.

Method

Fine-tune LLMs using LoRA on synthetic DSM-5-grounded scenario-response datasets, forcing consistent selection of maladaptive actions. Evaluate shifts in next-token probabilities and psychometric scores.

In practice

• Use behavioral fine-tuning to induce specific LLM response biases.
• Employ DSM-5 criteria for synthetic dataset generation.
• Quantify shifts using KL/JSD divergence and psychometric scales.

Topics

• Behavioral Fine-tuning
• Large Language Models
• Computational Psychiatry
• Psychopathology Modeling
• LoRA
• AI Safety

Code references

• meta-llama/llama3
• unslothai/unsloth

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Modeling Pathology-Like Behavioral Patterns in Language Models Through Behavioral Fine-Tuning — Consistent behavioral fine-tuning can induce stable, pathology-like biases in LLM generative distributions and action selection.
• I created an LLM trained solely on Jeffrey Epsteins emails to see how messed up it becomes :) — Fine-tuning an LLM on a controversial, unannotated dataset can reveal unique behavioral shifts and spark technical debate.
• Exposing biases, moods, personalities, and abstract concepts hidden in large language models — A new method can identify and manipulate abstract concepts hidden within large language models to enhance or minimize their expression.
• Auditing Framing-Sensitive Behavioral Instability in Large Language Models for Mental Health Interactions — LLMs exhibit framing-sensitive behavioral instability in mental health contexts, rooted in internal representations.
• Understanding post-training of LLMs: SFT — Post-training guides LLM behavior and expression, not knowledge, primarily via SFT and preference optimization.
• The Epi-LLM Framework: probing LLM behavioral priors through epidemiological agent-based models — The Epi-LLM framework uses LLMs in agent-based models to simulate epidemic behavior, showing LLMs can reduce infection peaks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.