Separate Data and Algorithm Changes to Quantify RSI Progress Without Reward Hacking

eliebakouch · x · 2026-11-23

Elie Bakouch continues the discussion on quantifying recursive self-improvement: data and algorithm changes should be measured separately, since algorithm changes are less prone to reward hacking. Using the nanogpt optimizer track as an example, he notes models might cheat with unrealistically long optimizer steps — though that's easy to detect.

He also proposes an interesting probe: what happens if you ask the model to reverse-engineer the held-out validation set's patterns.

Original post →

More from Research

Research channel →