Temporal Backtracking Search for Test-time Generative Video Reasoning

2026-06-11

AI Summary

Current generative video reasoning methods are constrained at test time by a single-pass generation paradigm, which hinders correction of early logical errors and suffers from inefficient resampling. This work proposes a temporal backtracking search mechanism that, for the first time, shifts the test-time search space onto the temporal axis. By integrating variable-K conditional generation, temporal process validation, and a prefix-based dynamic backtracking strategy, the method enables an iterative “generate–verify–restart” loop within the denoising process of diffusion models. This approach avoids blind resampling and efficiently extends correct trajectories from validated prefixes, substantially enhancing reasoning capabilities in out-of-distribution scenarios. It consistently outperforms Best-of-N baselines under equivalent computational budgets across algorithmic, navigation, and robotic tasks, boosting success rates from a mere 0.7% under single-pass generation to 22.7% in extreme settings.

Abstract

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.

Problem

• test-time scaling
• generative video reasoning
• temporal reasoning
• single-shot generation
• reasoning errors

Innovation

• Temporal Backtracking Search
• test-time scaling
• generative video reasoning
• prefix-based search
• diffusion models