CoTasks: Chain-of-Thought based Video Instruction Tuning Tasks

TL;DR

This paper introduces CoTasks, a framework that decomposes complex video questions into structured reasoning steps, significantly improving the reasoning capabilities of VideoLLMs through object-centric, step-by-step training.

Contribution

It proposes a novel decomposition of video reasoning tasks into entity-level subtasks and embeds these into training to enhance model reasoning abilities.

Findings

LLaVA-video-7B improves by +3.3 points on GPT-4 evaluation.

Qwen2.5-VL-3B gains +17.4 points, with large boosts in causal, temporal, and descriptive reasoning.

Structured CoT supervision significantly enhances compositional video reasoning.

Abstract

Despite recent progress in video large language models (VideoLLMs), a key open challenge remains: how to equip models with chain-of-thought (CoT) reasoning abilities grounded in fine-grained object-level video understanding. Existing instruction-tuned models, such as the Qwen and LLaVA series, are trained on high-level video-text pairs, often lacking structured annotations necessary for compositional, step-by-step reasoning. We propose CoTasks: Chain-of-Thought based Video Instruction Tuning Tasks, a new framework that decomposes complex video questions of existing datasets (e.g., NeXT-QA, STAR) into four entity-level foundational tasks: frame localization, entity tracking, spatial and temporal relation extraction. By embedding these intermediate CoT-style reasoning steps into the input, CoTasks enables models to explicitly perform object-centric spatiotemporal reasoning. Experiments on the NeXT-QA benchmark show that CoTasks significantly enhance inference performance: LLaVA-video-7B improves by +3.3 points in average GPT-4 evaluation score, and Qwen2.5-VL-3B gains +17.4, with large boosts in causal (+14.6), temporal (+10.9), and descriptive (+48.1) subcategories. These results demonstrate the effectiveness of CoTasks as a structured CoT-style supervision framework for improving compositional video reasoning.