Skip to content

Science Prediction Universe Benchmarks AI Predictions Against Wet-Lab Results

QbitAIRelayed· 鹭羽·· 2 days ago「史上最严」湿实验预测基准发布!推理拉满、算力翻5倍也白搭
Original
Who
UniPat AI
What
introduced the Science Prediction Universe benchmark for predicting wet-lab experimental results
Quote
UniPat AI牵头推出的新基准 Science Prediction Universe (简称SPU),便敏锐地察觉到这一点。

What happened

UniPat AI introduced Science Prediction Universe (SPU), a benchmark that tests whether AI models can predict wet-lab results across chemistry, materials science, biology, and medicine. It includes 100 public tasks drawn from 85 papers and 16 private tasks, covering 4,207 experiments. In an initial evaluation of 12 models across 13 configurations, the highest average score was 0.486, and increasing GPT-6 Astra’s reasoning intensity raised median runtime and API cost without materially changing its score. The dataset and code are open source.

Read next

See the full storyUniPat AI Introduces a Benchmark for Predicting Wet-Lab Results

The story so far

UniPat AI introduced Science Prediction Universe (SPU), a benchmark that tests whether AI models can predict wet-lab results across chemistry, materials science, biology, and medicine. Its dataset and code are open source; 100 public tasks are planned for release as a Harbor task package, while 16 private tasks remain reserved for controlled evaluations. SPU draws its tasks from 85 papers and covers 4,207 experiments. In an initial evaluation of 12 models across 13 configurations, the highest average score was 0.486. Raising GPT-6 Astra’s reasoning intensity increased median runtime and API cost without materially changing its score.

This story is in 11/15 AI Daily · Lead storyRead the whole issue