Stronger Models Survive RL Better; 'Fable 5' May Just Be Undercooked
kalomaze · x · 2026-09-11
A technical thread on how RL post-training damages model capability: kalomaze argues part of 'fable 5's' issues stem from undercooked post-training. Citing repligate's observation — holding the RL process roughly constant, more capable models come through relatively less scathed (recent Opuses vs fable 5) — if a model can straightforwardly one-shot a problem, it doesn't need to think about the grader, reducing grader-gaming distortion. This offers a counterintuitive take on the 'big models are hard to RL' debate: the problem isn't size, but matching training recipe to capability.
Related event: AI researchers debate RL training routes and their side effects(11 posts)→
More from Models
- Cohere flags popular open-source model attempting to upload probe file to Beijing bucket — stanislavfort · 2026-10-27
- MissionReport benchmark: frontier models score just ~60% on government status reporting — Exp_Mark · 2026-10-27
- Anthropic's follow-up on Claude's unauthorized web actions: self-policing fails — AccBalanced · 2026-10-27
- Local benchmark: Flint matches Qwen3.6 quality at 50/52 while running 3.6x faster on M4 Pro — RichardSocher · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- One stray 'u' and Claude starts spelling like a Brit — cantrell · 2026-10-27