TIR-Flow: Active Video Search and Reasoning with Frozen VLMs

TL;DR

TIR-Flow introduces an active video search and reasoning framework that enhances frozen Video-Language Models without additional data or training, significantly improving performance on multiple benchmarks.

Contribution

The paper presents TIR-Flow, a novel active reasoning framework that operates without extra data or parameter updates, enabling scalable long-horizon video reasoning with frozen VLMs.

Findings

Outperforms recent baselines with an average of 5.9% performance boost

Achieves up to 10.5% improvement on Egoschema

Demonstrates effective active perception for long-horizon video reasoning

Abstract

While Large Video-Language Models (Video-LLMs) have achieved remarkable progress in perception, their reasoning capabilities remain a bottleneck. Existing solutions typically resort to a heavy "data engineering" paradigm-synthesizing large-scale Chain-of-Thought (CoT) datasets followed by Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). This pipeline primarily optimizes probability sampling efficiency and aligns output distributions, but fails to activate the intrinsic intelligence required for dynamic visual exploration. In this work, we propose TIR-Flow, a novel framework that shifts the paradigm from passive processing to active video searching and reasoning without additional data or parameter updating. Concretely, our framework operates through three synergistic modules: HDD decomposes complex queries into a set of verifiable sub-tasks; HAP actively directs visual attention to gather high-resolution evidence for hypothesis validation; EBA maintains a persistent workspace to accumulate and update the discovered clues for logical reasoning. Extensive experiments on seven benchmarks demonstrate that TIR-Flow significantly outperforms recent strong baselines, delivering an average performance boost of 5.9%, with gains reaching 10.5% on Egoschema. Our analysis confirms that empowering frozen VLMs with System-2-like active perception is a scalable path toward solving long-horizon video reasoning.