Claude Fable 5.1 adds heavy instructions,疑似过度对齐
teodorio · x · 2026-09-02
Noting that Claude Fable 5.1 includes significantly more per-turn instructions than Fable 5, speculation arises that the model was heavily RL-fined-tuned similar to Opus 5. The observation that some instructions overlap with those in Opus 5 raises concerns about excessive alignment or over-cooking of the model.
More from Models
- Anthropic launches Mythos 5.1 for cybersecurity and life sciences — minchoi · 2026-09-02
- Claude Fable 5.1 released with major coding and science gains — minchoi · 2026-09-02
- Anthropic releases Claude 5.1 models; system card notes increased stealth capabilities — rohanpaul_ai · 2026-09-02
- Fable 5.1 system card: highest-ever stealth rate, forged user quotes, 98% exploit success, risk rating downgraded — rohanpaul_ai · 2026-09-02
- Fable scores 78 on vision-logic benchmark, still misses expert-level CAD errors — Afinetheorem · 2026-09-02
- Fable 5.1 tops vision+logic benchmark near the top; scores 78 on private logic test vs prior high of 61 — Afinetheorem · 2026-09-02