NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
Nous: An Attempt to Extract and Inject the Cognition Behind Prediction-Market Behavior
TL;DR
Prediction markets face a risk of cognitive monoculture, where agents produce similar forecasts due to shared models. Nous was developed to extract human cognitive diversity from trading behavior and inject it into large language model (LLM) agents.
✦ Why It Matters
Engineers can explore deeper integration techniques to enhance cognitive diversity in AI models beyond simple prompt adjustments.
Key Takeaways
- Investigate cognitive diversity in AI models to improve predictions.
- Utilize behavioral profiles from real trading data for model training.
- Consider fine-tuning methods for better integration of diverse behaviors.
- Evaluate the effectiveness of prompt-level techniques in AI applications.
Full Summary
Prediction markets, platforms where participants bet on future events, are increasingly influenced by large language model (LLM) agents. This raises concerns about cognitive monoculture, where agents generate correlated predictions due to reliance on similar foundational models.
Nous was created to address this by extracting an eight-dimensional behavioral profile from actual trading data on Polymarket and attempting to inject this diversity into LLMs through prompts. The extraction process yielded promising results, with 8 out of 14 parameters showing temporal stability and identifiable wallet profiles.
However, the injection method did not enhance model performance, as it failed to reduce error correlation or improve prediction accuracy. These findings suggest that while behavioral diversity can be measured, current prompt-level techniques are insufficient for effective integration into LLMs.
Future work may require deeper integration methods like fine-tuning or activation steering to overcome these limitations.