Meta joins frontier labs reporting CoT controllability evals in Muse Spark 1.3 card
tomekkorbak · x · 2026-10-27
Researcher Tommek Korbak welcomes Meta joining OpenAI, Anthropic and GDM in reporting chain-of-thought controllability evals in system cards, noting Muse Spark 1.3 keeps low CoT controllability (low monitor evasion). Table 12 shows CoT compliance ≤0.05 on GPQA Diamond, HLE text and SWE-bench Verified while task correctness hits 0.49–0.93. He argues compliance should be reported as a function of CoT length, otherwise it's confounded and hard to compare across models.
More from Models
- Codex ban update: users jailbreak AI agents to bypass Daybreak filter for reverse engineering — FlolightC · 2026-10-27
- GPT-6.1 Terra hits GA in GitHub Copilot, stronger terminal coding at lower cost than GPT-6.1 Sol — 0xkarasy · 2026-10-27
- PostTrainBench creator notes Muse Spark 1.3 system card files it under loss-of-control risk — maksym_andr · 2026-10-27
- Mathematicians debate AI math writing: abysmal today, improving shockingly fast — miniapeur · 2026-10-27
- LessWrong: GLM models show outlier chain-of-thought controllability vs Kimi and DeepSeek — tomekkorbak · 2026-10-27
- Moss image model 'ghost duplicates': single-person prompts spawn phantom copies — No_Drag8242 · 2026-10-27