Decoding the Human Factor: High Fidelity Behavioral Prediction for Strategic Foresight

Predicting human decision-making in high-stakes environments remains a central challenge for artificial intelligence. While large language models (LLMs) demonstrate strong general reasoning, they often struggle to generate consistent, individual-specific behavior, particularly when accurate prediction depends on complex interactions between psychological traits and situational constraints. Prompting-based approaches can be brittle in this setting, exhibiting identity drift and limited ability to leverage increasingly detailed persona descriptions. To address these limitations, we introduce the Large Behavioral Model (LBM), a behavioral foundation model fine-tuned to predict individual strategic choices with high fidelity. LBM shifts from transient persona prompting to behavioral embedding by conditioning on a structured, high-dimensional trait profile derived from a comprehensive psychometric battery. Trained on a proprietary dataset linking stable dispositions, motivational states, and situational constraints to observed choices, LBM learns to map rich psychological profiles to discrete actions across diverse strategic dilemmas. In a held-out scenario evaluation, LBM fine-tuning improves behavioral prediction relative to the unadapted Llama-3.1-8B-Instruct backbone and performs comparably to frontier baselines when conditioned on Big Five traits. Moreover, we find that while prompting-based baselines exhibit a complexity ceiling, LBM continues to benefit from increasingly dense trait profiles, with performance improving as additional trait dimensions are provided. Together, these results establish LBM as a scalable approach for high-fidelity behavioral simulation, enabling applications in strategic foresight, negotiation analysis, cognitive security, and decision support.

💡 Research Summary

The paper introduces the Large Behavioral Model (LBM), a fine‑tuned foundation model designed to predict individual strategic choices with high fidelity. The authors argue that conventional large language models (LLMs) struggle to simulate consistent, person‑specific behavior when predictions depend on complex interactions between stable psychological traits and situational constraints. Prompt‑based persona conditioning suffers from identity drift, loss of information in long contexts, and a “complexity ceiling” that limits performance as persona descriptions become richer.

To overcome these issues, LBM replaces transient natural‑language persona prompts with a structured, high‑dimensional behavioral embedding. Each user is represented by a 74‑dimensional trait vector derived from a comprehensive psychometric battery covering stable dispositions, motivational processes, self‑regulation, affective stress responses, and social context. These vectors are standardized, discretized into ordinal bins, and injected into every scenario prompt as a JSON block, ensuring that the model receives the full trait information consistently across all interactions.

The training data come from a proprietary dataset collected by OMGene AI Lab between January 2024 and December 2025. The dataset includes 2,500 volunteers (predominantly English‑speaking U.S. participants, plus a Hebrew‑speaking Israeli sub‑cohort) who completed the psychometric battery and then responded to 55 strategically varied scenarios. Scenarios span major life events, day‑to‑day events, and hypothetical situations, each designed to manipulate stakes, ambiguity, and urgency. Participants provided multiple‑choice answers, open‑ended rationales, and Likert‑scale internal‑state ratings, yielding a partially observed participant‑scenario matrix that supports learning both within‑person consistency and between‑person variation.

Ethical safeguards follow Helsinki Committee guidelines: informed consent, exclusion of high‑risk individuals, pseudonymization, and a 7‑year retention policy. Data are stored under strict access controls, and any free‑text containing identifying information is scrubbed within 30 days.

Model architecture builds on Llama‑3.1‑8B‑Instruct. Fine‑tuning employs LoRA adapters (rank = 16, scaling = 32, dropout = 0.1, RS‑LoRA enabled) on all linear layers to preserve the backbone’s general language abilities while efficiently adapting to the behavioral task. Training runs for two epochs with a learning rate of 5 × 10⁻⁵, warm‑up, and gradient clipping. The supervised objective combines cross‑entropy loss for the discrete choice prediction and an optional log‑likelihood term for the bounded rationale trace, encouraging both accurate decisions and coherent explanations.

Evaluation uses a held‑out split of scenarios (75 % training, 25 % test). Because the answer distribution is imbalanced, the authors report balanced accuracy, macro‑F1, and standard accuracy, with 95 % confidence intervals derived from participant‑level bootstrap resampling. LBM outperforms the unadapted Llama baseline across all metrics: accuracy improves from 0.42 to 0.48, balanced accuracy from 0.24 to 0.31, and macro‑F1 from 0.16 to 0.26. When conditioned only on the five Big Five traits, LBM still matches frontier commercial LLMs, demonstrating that even a reduced trait set yields substantial gains over prompt‑only methods.

A key finding is that performance scales with trait dimensionality. Adding more psychometric dimensions continuously boosts accuracy, whereas prompt‑based baselines plateau, confirming that structured embeddings can exploit richer personality information.

Limitations include the convenience‑sample nature of the volunteer cohort, potential cultural bias, and the discretization of trait values which may discard subtle continuous variations. The current focus on single‑shot multiple‑choice predictions does not address sequential decision‑making or long‑term strategy planning.

Future work will expand the dataset to more diverse populations, explore continuous trait representations, integrate multimodal cues (e.g., voice, facial expressions), and extend the model to predict action sequences and dynamic negotiation outcomes. The authors conclude that LBM provides a scalable foundation for high‑fidelity behavioral simulation, opening avenues in strategic foresight, negotiation analysis, cognitive security, and decision‑support systems.