SciLitBench - Title and Abstract Screening: leaderboard

Metric: F2 score (%; recall-weighted F-score of include decisions on the frozen 1,800-record evaluation split of the manually annotated 2,000-record title and abstract seed set (52 inclusions) of a real systematic review; zero-shot prompt that states the inclusion criteria and requires explicit exclusion reasoning; Ollama, temperature 0, top-p 0.9). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 22 models tracked.

Top models

#ModelScore
1SciLitBench Llama 3.3 70B (Ollama build)78.6
2SciLitBench Llama 3.1 70B (Ollama build)75.5
3SciLitBench Qwen3 14B (Ollama build)74.9
4SciLitBench gpt-oss 120B (Ollama build)69.7
5SciLitBench Llama 3 70B (Ollama build)67.9
6SciLitBench Qwen3 32B (Ollama build)66.8
7SciLitBench Qwen3 8B (Ollama build)66.5
8SciLitBench gpt-oss 20B (Ollama build)62.8
9SciLitBench Llama 3.1 8B (Ollama build)60.9
10SciLitBench Mistral Large 123B (Ollama build)60.6

Interactive version: theaggregate.ai/benchmark?slug=scilitbench-title-and-abstract-screening · How It Works · Data refreshed daily, snapshot 2026-09-26.