Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FastV-RAG: Towards Fast and Fine-Grained Video QA with Retrieval-Augmented Generation

Gen Li, Peiyu Liu

TL;DR

This work tackles the challenge of knowledge-grounded video QA by integrating external retrieval with vision–language models through a fast, reliable framework. VideoSpeculateRAG uses speculative decoding: a lightweight drafter quickly generates multiple candidate answers from retrieved documents, which a heavier verifier calibrates to ensure accuracy, achieving about a $2\times$ latency reduction without sacrificing performance. To address fine-grained entity errors, the method introduces structured draft reasoning and a CLIP-based entity alignment in a two-stage verification process, improving grounding and robustness. Evaluations on VideoSimpleQA and Encyclopedic VQA show competitive or superior accuracy relative to standard RAG, with substantial efficiency gains, highlighting the approach’s practicality for real-time, knowledge-intensive multimodal tasks.

Abstract

Vision-Language Models (VLMs) excel at visual reasoning but still struggle with integrating external knowledge. Retrieval-Augmented Generation (RAG) is a promising solution, but current methods remain inefficient and often fail to maintain high answer quality. To address these challenges, we propose VideoSpeculateRAG, an efficient VLM-based RAG framework built on two key ideas. First, we introduce a speculative decoding pipeline: a lightweight draft model quickly generates multiple answer candidates, which are then verified and refined by a more accurate heavyweight model, substantially reducing inference latency without sacrificing correctness. Second, we identify a major source of error - incorrect entity recognition in retrieved knowledge - and mitigate it with a simple yet effective similarity-based filtering strategy that improves entity alignment and boosts overall answer accuracy. Experiments demonstrate that VideoSpeculateRAG achieves comparable or higher accuracy than standard RAG approaches while accelerating inference by approximately 2x. Our framework highlights the potential of combining speculative decoding with retrieval-augmented reasoning to enhance efficiency and reliability in complex, knowledge-intensive multimodal tasks.

FastV-RAG: Towards Fast and Fine-Grained Video QA with Retrieval-Augmented Generation

TL;DR

This work tackles the challenge of knowledge-grounded video QA by integrating external retrieval with vision–language models through a fast, reliable framework. VideoSpeculateRAG uses speculative decoding: a lightweight drafter quickly generates multiple candidate answers from retrieved documents, which a heavier verifier calibrates to ensure accuracy, achieving about a latency reduction without sacrificing performance. To address fine-grained entity errors, the method introduces structured draft reasoning and a CLIP-based entity alignment in a two-stage verification process, improving grounding and robustness. Evaluations on VideoSimpleQA and Encyclopedic VQA show competitive or superior accuracy relative to standard RAG, with substantial efficiency gains, highlighting the approach’s practicality for real-time, knowledge-intensive multimodal tasks.

Abstract

Vision-Language Models (VLMs) excel at visual reasoning but still struggle with integrating external knowledge. Retrieval-Augmented Generation (RAG) is a promising solution, but current methods remain inefficient and often fail to maintain high answer quality. To address these challenges, we propose VideoSpeculateRAG, an efficient VLM-based RAG framework built on two key ideas. First, we introduce a speculative decoding pipeline: a lightweight draft model quickly generates multiple answer candidates, which are then verified and refined by a more accurate heavyweight model, substantially reducing inference latency without sacrificing correctness. Second, we identify a major source of error - incorrect entity recognition in retrieved knowledge - and mitigate it with a simple yet effective similarity-based filtering strategy that improves entity alignment and boosts overall answer accuracy. Experiments demonstrate that VideoSpeculateRAG achieves comparable or higher accuracy than standard RAG approaches while accelerating inference by approximately 2x. Our framework highlights the potential of combining speculative decoding with retrieval-augmented reasoning to enhance efficiency and reliability in complex, knowledge-intensive multimodal tasks.
Paper Structure (31 sections, 15 equations, 5 figures, 4 tables, 1 algorithm)

This paper contains 31 sections, 15 equations, 5 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Knowledge-aware VQA.
  4. Speculative Decoding.
  5. Methods
  6. VideoSpeculateRAG
  7. Multimodal Retrieval.
  8. Speculate Generation.
  9. Fine-grained Entity Alignment
  10. Error Analysis.
  11. Structured Draft Reasoning.
  12. Two-Stage Verification.
  13. Experiments
  14. Datasets
  15. Baselines
  16. ...and 16 more sections

Figures (5)

  • Figure 1: Illustration of Video Speculate RAG.
  • Figure 2: Error Analysis. RED color stands for entities inconsistent with the video content, while GREEN color refers entities consistent with the video.
  • Figure 3: An illustrative example of how our method detects Cross-Entity Transfer.
  • Figure 4: An illustrative example of how our method detects Entity Substitution.
  • Figure 5: Accuracy variation with different $\delta$ values