FlatSeek Study - Held-Out Synthetic: leaderboard

The FlatSeek study tests untouched symbolic-regression baselines on 83 literature equations and 200 synthetic equations, allowing up to 64 candidate expressions. Scores are counts of oracle-accepted expressions, not percentages. The paper’s trained FlatSeek adapter is a distinct artifact. This board reports systems with an accepted candidate (count out of 200). Higher is better.

Metric: Systems with an accepted candidate (count out of 200). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1DeepSeek V4 Flash97#259
2Qwen 3.5 27B89#365
3Qwen 3.5 9B88#814
4DeepSeek R1 Distill Qwen 32B66#1822
5Qwen 3.5 2B28#5184

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=flatseek-study-held-out-synthetic · How It Works · Data refreshed daily, snapshot 2026-11-16.