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.
More from Research
- A fill-in-the-middle scheme to distill hidden reasoning from proprietary models — OwariDa · 2026-11-23
- Prune, train, upcycle: one model that scales with your RAM — minhnguyen · 2026-11-23
- Claude formalizes GRH proof approach in 17M lines of Lean, code checks out — elonmusk · 2026-11-23
- Riemann proof built with Claude spans 17M lines of Lean — no human will ever read it — r0ck3t23 · 2026-11-23
- CLM author on extending editable context to video: RL may teach models where to look, prefill cut 30% — RulinShao · 2026-11-23
- Anthropic researcher alpoge reportedly solved one of the biggest open math problems — and went silent — kimmonismus · 2026-11-23