CLBench-V - Knowledge Learning (L2): leaderboard
CLBench-V measures multimodal context learning: whether a model can ground an answer in the visual context it is given, apply new information supplied in that context, and learn new knowledge from it. This board covers the datasets where the model must learn context-defined rules or infer the findings of a paper from the evidence that precedes them. Dataset scores are combined weighted by their instance counts, and each dataset keeps its own evaluator, from option extraction and numeric matching to route-topology checks and semantic judging.
Metric: Score (%; instance-weighted mean of the dataset scores, each 0-1 and scaled to percent, over the L2 new-knowledge datasets (CL-Bench-Table, Paper Conclusion); task-specific evaluators: option extraction, normalized or tolerance numeric match, a route-topology judge, and Qwen3.6-27B semantic or rubric judging for open-ended answers). Source: arxiv.org. Saturation forecast: Around June 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.6 | 21.86 | #178 |
| 2 | Qwen 3.5 Plus | 19.64 | #232 |
| 3 | Seed 2.0 Lite | 16.55 | #289 |
| 4 | Qwen 3.6 27B | 13.98 | #341 |
| 5 | GPT-5.4 | 6.94 | #123 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=clbench-v-knowledge-learning-l2 · How It Works · Data refreshed daily, snapshot 2026-11-13.