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

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

Nous addresses the cognitive monoculture problem in LLM agents for prediction markets, where frontier models exhibit error correlations of r≈0.77. It extracts an eight-dimension behavioral profile from 100 real Polymarket trading wallets. Findings show 8 of 14 parameters are temporally stable (e.g., contrarian score ICC ≈0.9), and wallets are identifiable from their profiles. However, prompt-level injection of these profiles into agents does not measurably transmit diversity. This failure means no reduction in ensemble error correlation or improvement in Brier score. The bottleneck is the structure-to-narrative translator, which emits semantically near-uniform prompts, motivating deeper, below-the-prompt injection methods.

Key takeaway

For AI Scientists and Machine Learning Engineers developing LLM agents for collective decision-making, relying on prompt-level injection to induce cognitive diversity is ineffective. Your efforts to reduce ensemble error correlation or improve Brier scores will likely fail if limited to prompt engineering. Instead, explore deeper, below-the-prompt injection methods like parameter-efficient fine-tuning or activation steering to achieve genuine cognitive heterogeneity.

Key insights

LLM agent ensembles face cognitive monoculture; prompt-level injection of human behavioral profiles fails to induce diversity.

Principles

• LLM agents in prediction markets risk cognitive monoculture due to shared foundation models.
• Individual LLM accuracy does not guarantee collective quality; diversity is crucial.
• Prompt-level cognitive injection is insufficient to reduce correlated errors.

Method

Nous extracts an eight-dimension behavioral profile from human Polymarket trading activity and attempts to inject it into LLM agents via prompts.

In practice

• Analyze human trading data to identify stable behavioral parameters.
• Investigate below-the-prompt methods for cognitive injection, such as fine-tuning.

Topics

• Cognitive Diversity
• LLM Agents
• Prediction Markets
• Ensemble Forecasting
• Prompt Engineering
• Behavioral Economics
• Epistemic Monoculture

Code references

• WillChienT/nous-paper

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

Related on AIssential

• Nous: An Attempt to Extract and Inject the Cognition Behind Prediction-Market Behavior — Prompt-level injection of human cognitive diversity into LLM agents fails due to prompt compression, despite successful behavioral extraction.
• R-Diverse: Mitigating Diversity Illusion in Self-Play LLM Training — Diversity Illusion hinders LLM self-play, where training data appears diverse but lacks true underlying skill variation.
• The Injection Paradox: Brand-Level Suppression in Safety-Trained LLM Recommendations via RAG Context Injection — Safety-trained LLMs can paradoxically suppress brands via RAG context injections, enabling reverse-attacks.
• SPA: A Simple but Tough-to-Beat Baseline for Knowledge Injection — Careful prompt design with large-scale augmentation effectively injects knowledge into LLMs.
• AgentAtlas: Beyond Outcome Leaderboards for LLM Agents — LLM agent evaluation requires multi-faceted taxonomies, not just single accuracy metrics, to assess true capability.
• When Agents Disagree: The Selection Bottleneck in Multi-Agent LLM Pipelines — Selector quality, not just team diversity, dictates multi-agent LLM pipeline performance.

Open in AIssential →

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