Claude Sonnet 4.6 · Humanity's Last Exam(2,500-question benchmark) · 附件证据(ev46-hle-method)

Claude Sonnet 4.6 的 Humanity's Last Exam · 2,500-question benchmark 官方出处。见《Claude Sonnet 4.6 System Card》§2.20.2, PDF p.46 and Figure 2.20.2.A, PDF p.47。证据 ev46-hle-method。

原文与定位

Claude Sonnet 4.6 · Humanity's Last Exam(2,500-question benchmark),Adaptive max, without tools配置,Accuracy 33.2%(具体成绩公布时间未确认)。见《Claude Sonnet 4.6 System Card》§2.20.2, PDF p.46 and Figure 2.20.2.A, PDF p.47。 查看成绩条件

Claude Sonnet 4.6 · Humanity's Last Exam(2,500-question benchmark),Adaptive thinking, max effort配置,Accuracy 49%(具体成绩公布时间未确认)。见《Claude Sonnet 4.6 System Card》§2.20.2, PDF p.46 and Figure 2.20.2.A, PDF p.47。 查看成绩条件

Humanity’s Last Exam is described by its developers as “a multi-modal benchmark at the frontier of human knowledge.” It includes 2,500 questions. For this evaluation, we tested Claude Sonnet 4.6 in two different configurations: 1. Reasoning-only without tools, and 2. Reasoning, web search, and web fetch with programmatic tool calling, code execution, context compaction that triggers every 50k tokens up to 3M tokens and adaptive thinking enabled. We used Claude Sonnet 4.5 as our model grader.
官方材料
Claude Sonnet 4.6 System Card · PDF 第 46 页
定位
附件 · §2.20.2, PDF p.46 and Figure 2.20.2.A, PDF p.47
材料版本
PDF with changelog dated 2026-03-06; 135 pages; collected 2026-11-10
版本公布
未确认
披露时间
未确认 · 首发/后补未确认
采集时间
历史依据
暂无可靠历史版本依据
引用文本

Claude Sonnet 4.6 · Humanity's Last Exam(2,500-question benchmark),Adaptive max, without tools配置,Accuracy 33.2%(具体成绩公布时间未确认);Claude Sonnet 4.6 · Humanity's Last Exam(2,500-question benchmark),Adaptive thinking, max effort配置,Accuracy 49%(具体成绩公布时间未确认)。官方出处 Claude Sonnet 4.6 System Card(附件,PDF with changelog dated 2026-03-06; 135 pages; collected 2026-11-10)§2.20.2, PDF p.46 and Figure 2.20.2.A, PDF p.47,https://www.anthropic.com/claude-sonnet-4-6-system-card#page=46。Benchlog https://benchlog.lekshome.cc/evidence/claude-sonnet-4-6/ev46-hle-method/

关联成绩与条件

Claude Sonnet 4.6 · Humanity's Last Exam 成绩与配置
指标与成绩配置与评测条件证据关系
Accuracy33.2%原文 33.2% · 单位 %
公布 未确认
Adaptive max, without tools
设置
Adaptive thinking enabled; max effort; no tools.
未知项
Default sampling settings apply unless otherwise stated.
工具
禁止工具
推理预算
max
框架
未记录
协议
Humanity's Last Exam, 2,500 questions; reasoning-only configuration.
条件
No tools; adaptive thinking and max effort per Table 2.1.A. Contamination handling and the tool-assisted configuration are described in §2.20.2.

直接支持该成绩

在排名中查看此成绩
Accuracy49%原文 49.0% · 单位 %
公布 未确认
Adaptive thinking, max effort
设置
Adaptive thinking enabled; max effort; default temperature and top_p unless an evaluation specifies otherwise.
未知项
Context window is evaluation-dependent; Table 2.1.A states it does not exceed 1M tokens.
工具
允许工具
推理预算
max
框架
未记录
协议
Humanity's Last Exam, 2,500 questions; web search, web fetch, code execution and programmatic tool calling.
条件
Context compaction at 50K tokens up to 3M total tokens, adaptive thinking, max effort; domain blocklist and transcript review used for contamination checks.

直接支持该成绩

在排名中查看此成绩