A Reddit Proposal: Iteratively Self-Debiasing LLMs via Adversarial Model-to-Model Training

jcastroarnaud · reddit · 2026-11-18

A Reddit user proposes an iterative bias-mitigation scheme: a commercial model M1, explicitly told its known biases and given mitigation literature, labels a corpus and designs training for M2; M1 and M2 then role-play as humans chatting with each other, and those adversarial interaction logs guide training of a less-biased M3. Repeat with Mn built from M(n-1) and M(n-2) until convergence, then hand off to human experts for bias measurement. The post cites three papers on LLM bias sources and the author openly asks whether the idea would work.

Original post →

More from Research

Research channel →