Deep Thought Alignment: Trajectory-Level Latent Distillation for Video Reasoning

2026-08-17

AI Summary

This study addresses the computational inefficiency of multimodal large language models in video reasoning and the limitations of traditional distillation in constraining latent representations. We propose Latent-OPD, a novel framework incorporating trajectory-level latent distillation and a progressive teacher lookahead strategy to transcend output-only supervision. By precisely aligning hidden states, this approach facilitates efficient visual evidence aggregation and effective reasoning transfer. Extensive experiments across six benchmarks demonstrate that Latent-OPD consistently outperforms conventional methods, yielding substantial improvements particularly in few-frame sampling, long-video understanding, and complex evidence aggregation scenarios. These results validate the framework’s efficacy in achieving efficient video reasoning while maintaining robust performance through enhanced latent space alignment.

Abstract

Large Multimodal Models (LMMs) for video reasoning have long been hindered by the high computational cost of processing vast amounts of visual information. This dilemma motivates the transfer of the reasoning capabilities of large models to smaller, more efficient ones. On-Policy Distillation (OPD) offers a promising solution by matching output-token distributions along student-generated trajectories. However, video reasoning often depends on evidence accumulated across multiple frames. In this context, output-level supervision only captures information expressed through token predictions and does not directly constrain the latent representations formed during reasoning. To address this limitation, we propose Latent-OPD, which augments OPD with trajectory-level latent distillation. Specifically, our method focuses on the position at the end of each trajectory, where hidden states effectively summarize the accumulated visual evidence and reasoning context. Furthermore, we introduce a progressive teacher-lookahead strategy, which aligns middle-to-late student layers with increasingly deeper teacher layers. Experiments on six video reasoning benchmarks show that Latent-OPD consistently outperforms output-only OPD. Notably, the improvements are particularly pronounced in scenarios with limited frames, long videos, or tasks requiring complex evidence aggregation. These results establish Latent-OPD as a highly effective approach to frame-efficient video reasoning.

Problem

• Video Reasoning
• On-Policy Distillation
• Latent Representation
• Knowledge Distillation
• Large Multimodal Models

Innovation

• Latent-OPD
• Trajectory-Level Latent Distillation
• Progressive Teacher-Lookahead
• Video Reasoning
• On-Policy Distillation