Video-language models lose track of an object once the evidence that identified it is gone. Track, Then Bind asks the model about identity once, at the moment the evidence is visible, lets a tracker carry that identity, and answers from the tracked state. Every example below can be inspected: the moment of the identity decision, the numbered tracks the model saw, and the facts behind the answer.
| Accuracy on the full test sets (%) | VET cup game | VET card monte | Real cup games | TOC-Bench |
|---|---|---|---|---|
| Track, Then Bind | 100 | 100 | 51.4 | 43.0 |
| Same VLM, direct answer | 28.0 | 28.0 | 33.6 | 25.1 |
| Same tracks drawn, question asked directly | 44.0 | 40.0 | 32.7 | — |
| Random guess | 33.3 | 33.3 | 33.3 | 30.3 |
All test questions: 50 per VET task, 107 real cup games from the Perception Test, 1,845 TOC-Bench questions. The examples below are selected: questions where Track, Then Bind is right and the same model is wrong both ways.