TrackThenBind Study - Direct - VET Cup: leaderboard
TrackThenBind study:Direct frozen visual-language inference;Direct uses greedy2048newtokens for Instruct/Molmo and16384forThinking,TOC nativebudget;TrackThenBind uses SAM2.1-hiera-large with about11–19calls and a frozen development-chosen configuration,VET50videos/task,real PerceptionTest cups andTOC. TrackThenBind binds a visible object identity then carries it withtracker;this scaffold is separate from direct,thinking and external multi-tool agents. Cited closedmodel scores not re-ingested. Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around July 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B Instruct | 32 | #285 |
| 2 | Qwen 3 VL 8B Instruct | 28 | #418 |
| 3 | Qwen 3 VL 8B (Thinking) | 28 | |
| 4 | Qwen 3 VL 32B (Thinking) | 24 | #297 (Qwen 3 VL 32B) |
| 5 | Qwen 3 VL 235B A22B Instruct | 22 | #273 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=trackthenbind-study-direct-vet-cup · How It Works · Data refreshed daily, snapshot 2026-10-24.