Nous: An Attempt to Extract and Inject the Cognition Behind Prediction-Market Behavior

Recent advances in large language models (LLMs) have significantly improved individual prediction accuracy, exemplified by systems like Halawi et al.'s retrieval-augmented reasoning pipeline and Mantic's reinforcement learning approach. These methods have approached or surpassed human-level performance on various benchmarks. However, the focus has largely been on single-agent accuracy, neglecting the importance of diversity within populations. Behavioral economics and cognitive science emphasize that individual differences in reasoning and perception are crucial for collective intelligence. Yet, current models tend to produce highly correlated errors due to shared training data and architectures, leading to a phenomenon known as epistemic monoculture, which hampers ensemble robustness and decision quality.

The core challenge is that homogeneous large models, when deployed as multiple agents, tend to produce correlated errors, undermining the benefits of ensemble methods. Existing approaches relying on architectural diversity or manual persona design are limited in scalability and authenticity. The key question is whether behavioral heterogeneity, grounded in real human data, can be extracted and effectively injected into models to promote independent reasoning pathways, thereby reducing error correlation and enhancing collective decision-making.

This work introduces a structured behavioral schema capturing eight key cognitive dimensions derived from real trading behavior. It employs a data-driven extraction process to quantify trader profiles, demonstrating their stability and predictive relevance. The core innovation lies in designing prompt-based mechanisms to inject these behavioral features into LLMs, aiming to emulate human cognitive diversity. The study systematically evaluates the transfer effectiveness using multiple metrics, revealing the limitations of prompt-level injection and highlighting the need for deeper model interventions. This approach bridges behavioral economics, AI prompting, and ensemble theory, offering a novel pathway to foster cognitive heterogeneity.

• �� Collect Polymarket trading data, define 14 behavioral parameters across eight dimensions such as risk perception, contrarianism, and belief update inertia.
• �� Assess parameter stability via split-half ICC and bootstrap confidence intervals.
• �� Use a structure-to-narrative translator to generate prompts embedding behavioral profiles.
• �� Inject prompts into multiple LLMs, compare output diversity using Jensen-Shannon divergence and Brier scores.
• �� Conduct ablation studies varying prompt length, profile diversity, and question difficulty.
• �� Analyze whether behavioral parameters are transmitted into model outputs and whether they influence ensemble error correlation.

The experimental setup involves extracting behavioral parameters from 100 wallets, evaluating their stability and identification accuracy. Prompts are constructed with embedded profiles and tested across different models and settings. The evaluation metrics include ICC for stability, Top-1 retrieval accuracy, Jensen-Shannon divergence for output diversity, and Brier score for forecast accuracy. Control prompts with matched length but no behavioral content are used to isolate the effect of structured injection. Multiple exploratory tests adjust sampling temperature, profile diversity, and question difficulty to validate robustness.