BioMedHop (Direct Prompting): leaderboard
Multi-hop biomedical question answering over knowledge-graph motifs from Monarch and SPOKE, with multiple-choice, open-answer and counting questions in four task families; BioMedHop (Tan et al., 2026). The score is the mean of the four family scores in percent, where each entity family averages multiple-choice and open-answer accuracy; higher is better. The model answers from the question alone with direct prompting, without the knowledge-graph, literature or web evidence the benchmark can supply.
Metric: Overall accuracy (%; mean of the four task-family scores, which are MCQ/open-answer accuracy averages and exact-count accuracy; stratified evaluation subsets of the 10,045 instances; direct input-output prompting with no retrieved evidence (the paper's IO baseline), temperature 0, at most 1,024 output tokens). Source: arxiv.org. Saturation forecast: Around 2030. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 18.6 | #262 |
| 2 | GPT-4 Turbo | 18.2 | #1002 |
| 3 | Claude 3.7 Sonnet | 17.8 | #455 |
| 4 | Qwen 3 Next 80B A3B Instruct | 16.1 | #763 |
| 5 | Claude 3.5 Haiku | 14.5 | #1563 |
| 6 | MiMo-V2.5-Pro | 13.7 | #274 |
| 7 | GPT-4o Mini | 13.2 | #1627 |
| 8 | Qwen 3 4B | 9.3 | #2297 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=biomedhop-direct-prompting · How It Works · Data refreshed daily, snapshot 2026-11-12.