Reflexive Intelligence: Decision-Making in Observer-Participant Environments

Author: Mian Zhang | Independent Researcher
Date: April 12, 2026
DOI: 10.5281/zenodo.19557261
Status: Published (Zenodo Open Access)

Abstract

We introduce Reflexive Intelligence — the capacity for AI systems to make decisions while accounting for their own causal impact on the environment. Current AI excels in observer-invariant settings (games, math, protein folding) but fails in observer-participant environments where the agent's actions alter the system being modeled (financial markets, policy, recommendation). We formalize this distinction, propose a neuroscience-motivated cognitive architecture using multi-reward GRPO, and provide preliminary evidence that a 3B-parameter model trained with this framework exhibits reflexive reasoning absent from generic models 600× larger.

Files

• paper1_reflexive_intelligence.pdf — Full paper (PDF)
• paper1_reflexive_intelligence.tex — LaTeX source
• paper1_reflexive_intelligence.md — Markdown source

Citation

@article{zhang2026reflexive,
  title={Reflexive Intelligence: Decision-Making in Observer-Participant Environments},
  author={Zhang, Mian},
  journal={Zenodo},
  doi={10.5281/zenodo.19557261},
  year={2026}
}

License

CC BY-NC-SA 4.0