返回资讯中心
外部精选
软件工程
#开发者实时流

Singularity Gate: Benchmark for AI-driven paradigm-shifting scientific discovery

Benchmark tasks are built around real scientific findings first made public after the newest model's training cutoff, so no model can have seen them in training. Passing the Singularity Gate is necessary, though not suf…

Hacker News 最新leumon2026/11/10 · 7 天前5 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Benchmark tasks are built around real scientific findings first made public after the newest model's training cutoff, so no model can have seen them in training. Passing the Singularity Gate is necessary, though not sufficient, for autonomous AI-driven discovery. A model that cannot predict paradigm-breaking discoveries is not Einstein-level, and one that can may still fall short.

Score vs subscription usage for all tasks

Each model's score is its mean over the tasks it answered.

The Hard set is the 12 tasks rated most paradigm-breaking.

Cost covers the whole task set (80 tasks, or 12 for Hard) and is plotted on a log scale.

Every model ran at xhigh reasoning effort in its native harness (Claude Code or Codex), with tools on and web search off.

The v2.1 corpus is 80 new findings whose first public trace (paper, preprint, talk or any other disclosure) came in July 2026 or later.

GitHub Methodology

Leaderboard

Each bar shows the score over all 80 tasks. Its brighter part is the Hard score.

Costs are for all 80 tasks, with subscription usage given as a share of one week's limit on a $200 plan and API price in USD. A model that answered fewer tasks is scaled to 80 from its average cost per answered task.

Fable 5.2 and Opus 5.5 answered 72 of 80 tasks, Fable 5.1 67, Sonnet 5.5 and Haiku 5.5 79, and the GPT models all 80. Tasks a model's safety filters stopped are left out of its score.

Per-field breakdown

Fable 5.2 leads Physics and Astronomy, Chemistry and Materials, Mathematics and TCS, and Life Sciences. Opus 5.5 leads Earth and Planetary. The leader in each field is highlighted. A dash means the model answered fewer than 7 tasks in that field (Fable 5.1 in Life Sciences).

The Singularity Gate · benchmark and methodology · see Contact for outreach.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →