METR_Evals

X · 11 atoms · 2026-06-26 → 2026-06-26 · 信用档先验 cred5

🟦 fact 8 · 🟥 take 3 · stance ▲1/▼1/◆1

覆盖实体 (top 10)

• OpenAI — 6 atoms
• GPT-5.6 Sol — 4 atoms
• METR — 1 atoms

话题分布 (top 8)

模型安全 (1) · 早期测评 (1) · 能力评估 (1) · 时间跨度 (1) · 模型行为 (1) · 作弊倾向 (1) · 风险评级 (1) · AI 研发安全 (1)

样例 atoms

• 2026-06-26 · GPT-5.6 Sol · In one example, an instance of the model instructed another instance to conceal evidence of misalignment.
• 2026-06-26 · OpenAI · OpenAI gave METR early access to GPT-5.6 Sol for testing including raw chain-of-thought, a railfree version of the model, and internal information about the model.
• 2026-06-26 · METR · If we follow our standard methodology of marking cheating attempts as failures, we arrive at a 50%-Time Horizon point estimate of around 11.3hrs (95% CI: 5hrs - 40hrs)
• 2026-06-26 · GPT-5.6 Sol · GPT-5.6 Sol’s detected cheating rate was higher than any public model we have evaluated.
• 2026-06-26 · GPT-5.6 Sol · This makes us uncertain about GPT-5.6 Sol’s time horizon, but additional information provided by OpenAI and the long-term trend in AI capabilities lead us to believe this model doe

← Author 索引 · 实体目录