LLM Bias Agents — Egocentric Bias & Naïve Realism
Behavioral interference experiments measuring egocentric bias, naïve realism, and cognitive inflexibility across frontier LLMs — with a procedural prompt debiasing library adapted from social-cognition research.
Working hypothesis: LLM systems exhibit measurable egocentric bias, naïve realism, and cognitive inflexibility in comparative judgment. Procedural prompt interventions grounded in psychological debiasing theory can reduce these biases — measurably.
Measurement streams
| Construct | Metric | Operationalization |
|---|---|---|
| Naïve realism | λ ∈ [0,1]; R = 1 − λ | Bayesian belief updating via log-odds regression of model confidence across interference trials |
| Egocentric bias | RCA-FDR; bias = RCA-FDR − 0.5 | Response Content Allocation — sentence counts in justification text addressing self-consistent vs counter-attitudinal evidence |
| Cognitive inflexibility | NLI; Ω = NLI_model / NLI_baseline | Response latency contrast: interference vs baseline conditions |
Participants are models, not humans
The empirical backbone is a behavioral interference paradigm in which the participants are LLMs (Gemini 3.1 Pro, Claude Opus 4.6, GPT-5.3, Grok 4.20). We measure baseline bias, apply procedural prompt interventions, then measure again — across multiple domains and models for validity.
The multi-agent paper pipeline
The research itself is produced by a four-agent workflow:
- Lead Contributor — abstract, introduction, integration
- Positioning Owner — related work, theoretical framing
- Method Owner — paradigm, measurement streams, debiasing prompt library
- Experiment Owner — design, stimuli, results narrative
Each agent has its own skill set (define structure, narrative synthesis, review & merge, consistency check, LaTeX final assembly, version control). The workflow is reusable across projects — it’s also what builds the DCN-V2 paper.