SVFSearch - RAG Workflow: leaderboard

A fixed retrieval workflow (image retrieval, element voting, then text retrieval from a frozen offline knowledge base) on SVFSearch (2026), four-choice questions about paused frames of Chinese gaming short videos that need game-specific knowledge, over its 5,000-question test split. Accuracy in percent; higher is better.

Metric: Accuracy (%) on the 5,000-item SVFSearch test split (four-choice questions about a paused Chinese game video frame), given only the frame, question and options, with the fixed RAG workflow (image retrieval over the frozen game image index, element voting, then top-5 text retrieval) whose evidence is given to the model; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3.5 27BNon-reasoning69.4#375
2Qwen 3.5 9BNon-reasoning66.5#828
3Qwen 3 VL 8B63.5#917
4Qwen 3 VL 32B62.2#622
5Qwen 2.5 VL 7B57.3#1862

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

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