Temporal Backtracking Search for Test-time Generative Video Reasoning
About
While test-time scaling has revolutionized reasoning in large language models, generative video reasoning remains bottlenecked by a single-shot paradigm. We demonstrate that searching over denoising steps cannot rescue logically flawed rollouts because spatial trajectories commit early in the diffusion process. Root-level Best-of-N (BoN) sampling is similarly inefficient: reasoning errors cluster early in the temporal axis, and resampling blindly discards verified upstream progress. To unlock effective test-time scaling for video models, we introduce Temporal Backtracking Search (TBS), which shifts the search space to the temporal axis. TBS transforms video generation into an iterative generate-verify-restart loop via three core mechanisms: (1) variable-K conditioning to resume generation from arbitrary clean prefixes; (2) temporal process verification to localize failures and extract valid restart anchors; and (3) prefix-based search to reallocate compute toward extending correct trajectories rather than root resampling. Across algorithmic, navigation, and robotics domains, TBS Pareto-dominates matched-budget BoN. In a strict out-of-distribution setting where one-shot generation collapses (0.7% for BoN), TBS achieves 22.7%, with every solved episode stemming from a restarted branch. Ultimately, TBS reveals that the local reasoning competence of video models far exceeds what single-shot rollouts indicate, providing a scalable test-time framework to unlock it.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Structured video reasoning | Maze Overall | Exact Match Percentage90.3 | 6 | |
| Structured video reasoning | Maze Med | Exact Match Rate99.5 | 6 | |
| Structured video reasoning | Maze Hard | Exact Match Rate81 | 6 | |
| Structured video reasoning | Maze3D Overall | Exact Match (%)97.7 | 6 | |
| Structured video reasoning | Maze3D Med. | Exact Match Rate100 | 6 | |
| Structured video reasoning | Maze3D Hard | Exact Match95.3 | 6 | |
| Structured video reasoning | Sokoban Overall | Exact Match Rate82.7 | 6 | |
| Structured video reasoning | Sokoban Easy | Exact Match (%)99 | 6 | |
| Structured video reasoning | Sokoban Hard | Exact Match79.7 | 6 | |
| Structured video reasoning | Sliding Puzzle Overall | Exact Match (%)86.5 | 6 |