Atlas

Benchmarks

← All benchmarks

Concept Synth ABD v1.2 - Partial

General QA · 2026-02-21

The partially observed, optimistic-completion split of ABD v1.2. A rule is valid when each prompt world has at least one completion of its unknown atoms that satisfies the repaired theory.

Top models (higher is better)

ModelScore
Opus 4.697.9
Gemini 3.1 Pro Preview97.1
DeepSeek-V3.296.0
Grok 4.1 Fast94.7
GPT-5.292.4
Grok 491.5
GPT-5.486.4
Kimi K2 Thinking73.3
Gemini 3 Pro Preview65.3
GPT-4o22.6
Loading Atlas data…