CADER: Confidence-Aware Dynamic Evidence Reasoning for Long-Video Understanding

2026-07-27

AI Summary

This work addresses the challenge that existing long-form video understanding methods employ a uniform inference pipeline, struggling to balance computational efficiency for simple queries with fine-grained temporal modeling required for complex ones. The authors propose a training-free adaptive inference framework that first performs coarse reasoning via global frame sampling and estimates prediction confidence using a logit-margin criterion. High-confidence predictions are directly output, while low-confidence cases trigger a tool-augmented refinement loop comprising temporal cropping, lightweight semantic verification, and relevance-guided resampling. Notably, this approach introduces tool invocation as a sample-level dynamic decision mechanism, enabling branched reasoning paths. Without any tool-integrated chain-of-thought training on the backbone model, the method achieves significant performance gains across multiple long-video question-answering benchmarks while reducing redundant computation.

Abstract

Long-video understanding increasingly relies on large vision-language models and tool-augmented reasoning, but most systems apply the same inference procedure to every example regardless of difficulty. This uniform strategy invokes unnecessary tool-assisted processing for easy questions and provides limited control when difficult questions require fine-grained temporal evidence. We propose CADER (Confidence-Aware Dynamic Evidence Reasoning), a training-free framework for adaptive and reliable long-video reasoning. CADER first performs global reasoning over uniformly sampled frames and estimates answer confidence with a logit-margin signal, allowing high-confidence examples to exit early. For uncertain examples, CADER activates a second-stage tool-augmented loop that combines temporal cropping, lightweight semantic verification, and Relevance-Guided Resampling to progressively localize question-relevant evidence. This design treats tool use as a sample-level decision: a single global pass handles easy cases, while additional reasoning is reserved for examples where uncertainty suggests that more evidence is needed. Experiments on multiple VideoQA benchmarks show that CADER improves long-video reasoning while bypassing Stage~2 for high-confidence samples. Moreover, when applied to a backbone trained only with tool-free chain-of-thought supervision, CADER achieves competitive performance against specialized tool-augmented frameworks, suggesting a practical inference-time route for adaptive long-video reasoning.

Problem

• long-video understanding
• adaptive reasoning
• confidence estimation
• tool-augmented reasoning
• VideoQA

Innovation

• Confidence-Aware Reasoning
• Dynamic Evidence Localization
• Tool-Augmented Inference
• Adaptive VideoQA
• Relevance-Guided Resampling