BAGEL - GloBI Interactions: leaderboard
Metric: Accuracy (%, times 100) on the 3,500 GloBI questions (ecological interaction reasoning: masked interaction type or masked participant) of BAGEL, closed-book four-option single-answer multiple choice (the source passage is withheld), one unified prompt, greedy decoding with seed 0, a single run per model; questions written by GPT-4o-mini from Wikipedia species articles, GloBI interaction records, bioRxiv preprints and Xeno-canto recordings, options shuffled; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 81 |
| 2 | GPT-5.4 | 77.2 |
| 3 | Gemma 3 27B (IT) | 74.71 |
| 4 | Phi-4 | 72.51 |
| 5 | Qwen 3 32B (Non-reasoning) | 70.74 |
| 6 | Qwen 3 14B (Non-reasoning) | 69.54 |
| 7 | Qwen 3 8B (Non-reasoning) | 65.86 |
| 8 | Llama 3.1 8B Instruct | 63.86 |
| 9 | Mistral 7B Instruct (v0.3) | 60.6 |
| 10 | Qwen 3 4B (Non-reasoning) | 58.66 |
| 11 | gemma-7B | 48.06 |
| 12 | Qwen 3 0.6B (Non-reasoning) | 25.23 |
Interactive version: theaggregate.ai/benchmark?slug=bagel-globi-interactions · How It Works · Data refreshed daily, snapshot 2026-10-07.