Atlas

Benchmarks

← All benchmarks

Concept Synth INDUCTION v1 - FullObs

General QA · 2026-05-23

The fully observed split of INDUCTION v1. A formula is valid when its extension exactly matches the target label for every object in every closed-world input structure.

Top models (higher is better)

ModelScore
Grok 450.7
GPT-5.243.7
GPT-5.443.5
Gemini 3.5 Flash34.4
DeepSeek-V4-Pro28.2
Opus 4.623.7
Grok 4.1 Fast17.9
Gemini 3.1 Pro Preview17.6
Gemini 3 Pro Preview16.0
Grok 4.313.9
DeepSeek-V3.210.4
Opus 4.58.5
GPT-4o0.0
Loading Atlas data…