Was Claude's gibberish fixed by capping KL divergence in RL? One theory
burny_tech · x · 2026-10-11
burnytech speculates that Opus 5.5 speaks in incomprehensible language far less than previous Claude models because Anthropic capped KL divergence more tightly during RL training. KL constraints limit how far the policy can drift from the base model, and tighter caps could suppress the tendency to degrade into unreadable output. This is an unverified personal theory, but it touches on a real technical question about how KL regularization shapes language degeneration in RLHF.
More from Models
- JevBench creator explains why open and API models get separate leaderboards: fairness — airesearch12 · 2026-10-11
- Codebase test: Telnyx-hosted GLM runs 9% cheaper, 5% faster than OpenAI — SucceededMind · 2026-10-11
- Nous Research launches stealth coding and agentic reasoning model, free for limited time — Teknium · 2026-10-11
- I published 50,000 fake stats and ChatGPT still cites them, earning 10k+ AI sessions — metehan777 · 2026-10-11
- Haiku fails 15 of 17 agent runs on output format, deemed no upgrade over Luna — zeeg · 2026-10-11
- Users report Astra performance sharply degraded in recent days, images too — pwlot · 2026-10-11