EFlow: Learning Evidence Flow for Long-Video Reasoning with Adaptive Reflection

2026-07-01

AI Summary

Existing video reasoning methods often suffer from incomplete evidence retrieval due to premature semantic commitment during the coupled process of temporal grounding and answer reasoning, leading to erroneous conclusions. To address this, this work proposes EFlow, a novel framework built upon Qwen3-VL that introduces an “evidence-first” reasoning paradigm. EFlow explicitly decouples temporal grounding from logical reasoning, employing dual chain-of-thought (CoT) processes—CoT for Temporal Grounding and CoT for Reasoning—alongside a confidence-aware adaptive reflection mechanism that re-examines the entire video when evidence is insufficient. Trained via a hybrid strategy combining supervised fine-tuning, reinforcement learning, and reinforced fine-tuning on a dedicated trajectory dataset, EFlow achieves significant performance gains across five long-form video understanding benchmarks, demonstrating its effectiveness and strong generalization capability.

Abstract

Long-video reasoning is fundamentally constrained by how models acquire and utilize visual evidence. Existing tool-augmented video frameworks often interleave temporal grounding and answer reasoning within a single trajectory, causing early semantic hypotheses to bias evidence localization. We term this failure mode premature semantic commitment, where biased grounding retrieves incomplete evidence and incomplete evidence further reinforces incorrect reasoning. To address this issue, we propose EFlow, an evidence-first video reasoning framework built upon Qwen3-VL. EFlow explicitly separates temporal grounding and logical reasoning through CoT for Temporal Grounding and CoT for Reasoning, enabling the model to retrieve relevant evidence before answer inference. In addition, EFlow introduces a confidence-aware reflection mechanism that re-evaluates the full video when retrieved evidence is potentially insufficient. We further construct dedicated trajectory datasets and train EFlow through supervised fine-tuning, reinforcement learning, and reinforcement fine-tuning. Extensive experiments across five video understanding benchmarks demonstrate that EFlow consistently improves long-video reasoning performance.

Problem

• long-video reasoning
• temporal grounding
• premature semantic commitment
• evidence localization
• visual evidence

Innovation

• evidence-first reasoning
• temporal grounding
• chain-of-thought
• confidence-aware reflection
• long-video understanding