Builder Uses GPT-6 Astra to Craft Deterministic Readability Scorer for RL Reward, Avoiding N^2 LLM Judge Comparisons

ivan_bezdomny · x · 2026-09-06

ivanbezdomny shares an RL training practice from his news project: with thousands of headlines and articles generated by GPT, Claude, and his own finetuned models, pairwise LLM judgments are easy to get but too slow (N^2 comparisons) and noisy to serve as RL reward.

His fix: ask GPT-6 Astra to build a deterministic function scoring how well or awkwardly a headline reads. His existing reward function, evolved over weeks of RL training and analysis, handles quality scoring well but fails on readability — pre-Astra models got stuck on headline length and other superficial features. This is a targeted attempt to close that gap.

Related event: Using Google Astra to build deterministic RL reward functions for headlines(2 posts)→

Original post →

More from coding & agent

coding & agent channel →