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

A new study introduces Nous, a system designed to extract human cognitive diversity from prediction-market behavior and inject it into large language model agents, aiming to counter the risk of a cognitive monoculture where agents produce correlated forecasts [1]. The research, posted to arXiv, addresses a growing concern as LLM agents become more common in prediction markets and collective decision-making. The authors note that agents built on shared foundation models can create a cognitive monoculture, with recent measurements showing frontier-model errors correlated at r ~ 0.77 [1]. The Nous system attempts to solve this by first extracting a structured eight-dimension behavioral profile from real Polymarket trading activity, then injecting it into agents through prompts [1]. The central finding is a dissociation between the two halves of the pipeline [1]. Extraction works partially: across 100 wallets, 8 of 14 parameters are temporally stable, with a split-half ICC of at least 0.5 and a bootstrap CI lower bound above 0.3 [1]. The contrarian score reaches an ICC of approximately 0.9 [1]. Wallets are identifiable from their profiles well above chance, with top-1 retrieval at 17-22% versus a 1% chance level [1]. Two of four pre-specified dimensions rank-correlate with future realized profit out-of-sample, though the correlations do not survive behavioral-confound controls [1]. Prompt-level injection, however, does not measurably transmit cognitive diversity. On a semantic embedding metric, structured injection shows no significant advantage over a length-matched control on any model [1]. The diversity it induces neither reduces ensemble error correlation nor improves Brier score, a null result that persists across exploratory checks on sampling temperature, profile diversity, and question difficulty [1]. Measuring the prompts themselves locates the compression before the model: the structure-to-narrative translator emits near-uniform prompts whose spread does not track profile spread [1]. The authors position Nous as measuring the cognitive-monoculture problem and the limits of a prompt-level remedy, motivating deeper, below-the-prompt injection methods such as fine-tuning and activation steering [1]. Code, frozen profiles, prompts, and model outputs are available on GitHub [1].