ZICQ
中 Log in / Sign up
Newsroom LLMs & Foundation Models #Long-Video Question Answering #Evidence Acquisition #HKUST-KnowComp #PACE #Qwen3-VL

HKUST-KnowComp Releases PACE Framework: Revolutionizing Evidence Acquisition in Long-Video Question Answering

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 1 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:The HKUST-KnowComp team introduces PACE (Progressive Acquisition of Critical Evidence), a novel framework designed to address the challenges of evidence sparsity and discriminative insufficiency in long-video question answering. PACE operates in two stages: it first indexes clip-level descriptions guided by question-derived factors without observing candidate answers, and then uses the candidate answers to derive contrastive cues and queries the index for verification. On the MMR-V dataset with


Background and Challenges

Long-video question answering remains a significant challenge in the AI domain. Despite the continuous advancements in the visual understanding capabilities of language models, relevant evidence in long videos is often sparse and insufficient to distinguish the correct answer from plausible alternatives. Traditional methods struggle with this issue, primarily due to their limitations in acquiring discriminative evidence.

Introduction of the PACE Framework

To tackle these challenges, the HKUST-KnowComp team proposes the PACE (Progressive Acquisition of Critical Evidence) framework. PACE operates in two stages:

  1. Question-Guided Indexing Stage: In this stage, PACE indexes clip-level descriptions guided by question-derived factors without observing the candidate answers. This step aims to preliminarily filter evidence related to the question.
  2. Candidate Answer Verification Stage: PACE uses the candidate answers to generate contrastive cues and queries the index for verification. This ensures that the acquired evidence is discriminative and effectively distinguishes the correct answer from alternatives.

Experiments and Results

On the MMR-V dataset, the PACE framework, with the open-source Qwen3-VL backbone, achieves 42.6% accuracy, significantly outperforming direct inference methods and existing agentic baselines like Deep Video Discovery (DVD). Additionally, PACE recovers 66.9% of the annotated cues on the diagnostic subset, demonstrating that its performance gains are primarily due to enhanced evidence acquisition rather than stronger answer-side priors.

PACE also demonstrates consistent gains across multiple benchmarks, including LVBench, Video-MME, EgoSchema, and LongVideoBench, showcasing its versatility and potential in cross-domain long-video question answering tasks.

Technical Highlights

  • Two-Stage Processing Mechanism: By combining question-guided indexing and candidate answer verification, PACE achieves more precise evidence acquisition.
  • Contrastive Cue Generation: Utilizing candidate answers to generate contrastive cues enhances the discriminative power of the evidence.
  • Open-Source Implementation: The PACE framework's code is publicly available on GitHub (https://github.com/HKUST-KnowComp/PACE), facilitating replication and extension by researchers and developers.

Industry Impact and Developer Recommendations

The release of the PACE framework provides a new technical path for the long-video question answering field, particularly excelling in handling complex, multi-modal data. For developers, the following points are noteworthy:

  • Model Fine-Tuning: It is recommended to fine-tune the model on specific domain data to further enhance PACE's performance in particular tasks.
  • Multi-Modal Data Fusion: Combining other modalities of data (such as audio, text) for fusion processing can further enhance the comprehensiveness of evidence acquisition.
  • Real-Time Applications: The efficiency of the PACE framework makes it a potential candidate for real-time long-video question answering applications, but optimizing inference speed to meet real-time requirements is crucial.

Source: ArXiv cs.LG (2026-08-26)

— END —

Tags: #Long-Video Question Answering #Evidence Acquisition #HKUST-KnowComp #PACE #Qwen3-VL

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…