ShellfishNet Captioning: leaderboard

ShellfishNet Captioning tests how well multimodal models describe photographs of shellfish, covering shell shape, color and texture, visible anatomy and surroundings, on a 500-image subset of ShellfishNet (2026) with expert-reviewed reference captions. The score is BERTScore F1 against the reference on a 0 to 100 scale, higher is better.

Metric: BERTScore F1 (0-100; the paper's 0-1 score times 100) of the model caption against expert-reviewed reference captions covering morphology, color and texture, anatomical features and environment, on the 500-image ShellfishNet captioning subset (32 shellfish taxa in field, aquaculture and market scenes); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)90.06#99
2GPT-5.488.8#135
3Qwen 2.5 VL 7B Instruct87.5#1902
4InternVL2.5-4B87.12#2549
5InternVL2-4B86.63#4055
6InternVL2.5-8B86.27#2370
7InternVL2-8B86.15#3227
8Qwen 2 VL 7B85.93#2275
9LLaVA 1.6 Mistral 7B85.83#8226
10Qwen 2.5 VL 3B Instruct85.64#2713
11Qwen 2 VL 2B85.15#6157
12LLaVA-v1.5-7B84.42#7630
13IDEFICS2-8B83.67#5013

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=shellfishnet-captioning · How It Works · Data refreshed daily, snapshot 2026-11-21.