DSV-Mem: leaderboard
DSV-Mem tests dense stateful visual memory over 1,000 compositional questions in ten professional workflows. Models must reconcile changing visual and textual artifacts, recover past state, derive state and reject stale premises. Total accuracy is item-weighted, averaged over three runs; memory-management and prompt-variant experiments are separate from this baseline.
Metric: Accuracy (%), 1,000 stateful visual-memory questions, mean of three runs. Source: arxiv.org. Saturation forecast: Around May 2027. 14 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Gemini 3.7 Flash | Medium | 39.6 | #42 |
| 2 | Gemini 3.1 Pro (Preview) | Medium | 27.5 | #102 |
| 3 | Kimi K3 | High | 24.3 | #62 |
| 4 | Gemini 3.6 Flash | Medium | 23 | #84 |
| 5 | MiniMax-M3 | Thinking | 13.1 | #290 |
| 6 | Qwen 3.5 397B A17B | Thinking | 11.3 | #270 |
| 7 | Qwen 3.5 27B | Thinking | 9.6 | #380 |
| 8 | Gemma 4 31B (IT) | 8.5 | #354 | |
| 9 | Gemini 3.5 Flash Lite | Medium | 8.2 | #308 |
| 10 | Qwen 3.5 9B | Thinking | 7.2 | #838 |
| 11 | Kimi K2.5 | Thinking | 5.8 | #272 |
| 12 | Gemma 4 26B A4B (IT) | 3.9 | #563 | |
| 13 | GLM-4.6V | Thinking | 3.4 | #721 |
| 14 | Qwen 3.5 4B | Thinking | 3.3 | #1270 |
No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=dsv-mem · How It Works · Data refreshed daily, snapshot 2026-11-22.