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

2026-06-11 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Nous investigates the cognitive monoculture risk in LLM agents used in prediction markets, where shared foundation models lead to correlated forecast errors (r ~ 0.77). The system extracts an eight-dimension behavioral profile from real Polymarket trading activity and attempts to inject it into agents via prompts. A central finding reveals a dissociation: extraction partially works, with 8 of 14 parameters temporally stable (split-half ICC >= 0.5) and wallets identifiable (top-1 retrieval 17-22% vs. 1% chance). Two pre-specified dimensions rank-correlate with future profit, though not after controls. However, prompt-level injection fails to measurably transmit this diversity; it shows no significant advantage, nor does it reduce ensemble error correlation or improve Brier score. This null result persists because the structure-to-narrative translator emits near-uniform prompts, indicating compression before the model. This motivates exploring deeper injection methods like fine-tuning or activation steering.

Key takeaway

For LLM agent developers aiming to mitigate cognitive monoculture in prediction markets, relying solely on prompt-level injection of behavioral profiles is insufficient. Your efforts to introduce human cognitive diversity will likely be hampered by prompt compression, as demonstrated by Nous's findings. Instead, you should investigate deeper injection methods such as fine-tuning or activation steering to effectively diversify agent behavior and improve collective decision-making.

Key insights

Prompt-level injection of human cognitive diversity into LLM agents fails due to prompt compression, despite successful behavioral extraction.

Principles

• LLM agents in prediction markets risk cognitive monoculture from shared foundation models.
• Human behavioral profiles can be partially extracted and identified from trading activity.
• Prompt-level injection struggles with transmitting complex cognitive diversity effectively.

Method

Nous extracts an eight-dimension behavioral profile from real trading activity and attempts prompt-based injection into LLM agents.

In practice

• Measure cognitive monoculture in LLM agent ensembles.
• Explore fine-tuning or activation steering for diversity injection.

Topics

• LLM Agents
• Prediction Markets
• Cognitive Diversity
• Prompt Engineering
• Behavioral Profiles
• Fine-tuning

Code references

• WillChienT/nous-paper

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

Related on AIssential

• Nous: An Attempt to Extract and Inject the Cognition Behind Prediction-Market Behavior — LLM agent ensembles face cognitive monoculture; prompt-level injection of human behavioral profiles fails to induce diversity.
• 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.
• Beyond Static Responses: Multi-Agent LLM Systems as a New Paradigm for Social Science Research — A six-tier framework maps LLM agent evolution from static tools to complex adaptive systems for social science research.
• From Prompting to Cognitive Runtimes: Decoupling Cognition from Execution in LLM-based Agents (paper + code) — Decoupling LLM agent cognition from execution improves efficiency by enabling reusable reasoning steps.
• Theory of Mind in Action: The Instruction Inference Task in Dynamic Human-Agent Collaboration — Few-shot chain-of-thought prompting significantly enhances LLM Theory of Mind for collaborative task interpretation.
• Fake users generated by AI can't simulate humans — review of 182 research papers — LLMs fail to simulate human cognition and behavior due to their training on text, not diverse lived experience.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.