Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

POVQA: Preference-Optimized Video Question Answering with Rationales for Data Efficiency

Ashim Dahal, Ankit Ghimire, Saydul Akbar Murad, Nick Rahimi

TL;DR

POVQA tackles long-video VQA with data efficiency by compressing each second into a single pooled frame and interleaving it with subtitles to form a compact multimodal sequence. The approach uses four pooling operators, interleaved subtitle-text with visual tokens, and a two-stage fine-tuning pipeline (SFT via QLoRA followed by Direct Preference Optimization) on ReasonVQA, a small yet rich dataset of 239 QAs with rationales. Empirical results show substantial gains in F1, BLEU-4, and ROUGE-L over pooled baselines, with improved rationale quality and robustness across pooling strategies; zero-shot TVQA performance indicates strong transfer. The work demonstrates that careful temporal summarization plus rationale-supervised learning can achieve data-efficient, scalable VQA for long-form videos, with plans to expand datasets and incorporate external cues.

Abstract

Video Question Answering (VQA) with Large Vision Language Models (LVLMs) has gained significant traction in research ever since the Flamingo was introduced by Deepmind. Recent advancements in large context/long video question answering have allowed VQA tasks to have context window of 1500+ frames. However, this only leads to 50 seconds of video footage without losing any significant information. We introduce POVQA, a data-efficient pipeline that compresses each second of video into a single temporally pooled image (via motion blur and weighted averaging variants) and then align LVLMs with lightweight supervision. Concretely, we build 1 fps input sources using Blend Blur with Last Frame, Weighted Average, Exponential and Ramp pooling and fine-tune QWEN-2.5-VL 7B with supervised two turn target including reasoning and final answer. We apply Supervised Fine Tuning (SFT) and Direct Preference Optimization (DPO) on our novel dataset ReasonVQA consisting of 12 movies with 239 human annotated question-answer with reasoning prompts. On our ReasonVQA dataset, this method dramatically improves performance over pooled baselines: F1 score improves from 0.212 to 0.543, BLEU-4 from 0.031 to 0.291, and ROUGE-L from 0.196 to 0.528. Rationale quality also significantly increases. Cross-evaluation of SFT + DPO on various pooling functions show that the gains persist regardless of the pooling scheme used at train or test time, indicating strong robustness on summarization of temporal evidence. Similar observations were made on zero-shot in TVQA.

POVQA: Preference-Optimized Video Question Answering with Rationales for Data Efficiency

TL;DR

POVQA tackles long-video VQA with data efficiency by compressing each second into a single pooled frame and interleaving it with subtitles to form a compact multimodal sequence. The approach uses four pooling operators, interleaved subtitle-text with visual tokens, and a two-stage fine-tuning pipeline (SFT via QLoRA followed by Direct Preference Optimization) on ReasonVQA, a small yet rich dataset of 239 QAs with rationales. Empirical results show substantial gains in F1, BLEU-4, and ROUGE-L over pooled baselines, with improved rationale quality and robustness across pooling strategies; zero-shot TVQA performance indicates strong transfer. The work demonstrates that careful temporal summarization plus rationale-supervised learning can achieve data-efficient, scalable VQA for long-form videos, with plans to expand datasets and incorporate external cues.

Abstract

Video Question Answering (VQA) with Large Vision Language Models (LVLMs) has gained significant traction in research ever since the Flamingo was introduced by Deepmind. Recent advancements in large context/long video question answering have allowed VQA tasks to have context window of 1500+ frames. However, this only leads to 50 seconds of video footage without losing any significant information. We introduce POVQA, a data-efficient pipeline that compresses each second of video into a single temporally pooled image (via motion blur and weighted averaging variants) and then align LVLMs with lightweight supervision. Concretely, we build 1 fps input sources using Blend Blur with Last Frame, Weighted Average, Exponential and Ramp pooling and fine-tune QWEN-2.5-VL 7B with supervised two turn target including reasoning and final answer. We apply Supervised Fine Tuning (SFT) and Direct Preference Optimization (DPO) on our novel dataset ReasonVQA consisting of 12 movies with 239 human annotated question-answer with reasoning prompts. On our ReasonVQA dataset, this method dramatically improves performance over pooled baselines: F1 score improves from 0.212 to 0.543, BLEU-4 from 0.031 to 0.291, and ROUGE-L from 0.196 to 0.528. Rationale quality also significantly increases. Cross-evaluation of SFT + DPO on various pooling functions show that the gains persist regardless of the pooling scheme used at train or test time, indicating strong robustness on summarization of temporal evidence. Similar observations were made on zero-shot in TVQA.

Paper Structure

This paper contains 22 sections, 20 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. POVQA
  4. Temporal pooling and pooled frame construction
  5. Subtitle alignment and interleaving
  6. Tokenization and model input
  7. Supervised fine-tuning (SFT) with QLoRA
  8. Direct Preference Optimization (DPO)
  9. Token budget and attention shift
  10. Decoding
  11. Experiments
  12. Experimental Setup
  13. Data Details
  14. Model Details
  15. Training Details
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Abstract concept of inference on POVQA pipeline on evaluation set of ReasonVQA
  • Figure 2: Overview of the training process of POVQA on ReasonVQA dataset.
  • Figure 3: Qualitative analysis on a random sample of TVQA. Frames sub-sampled to fit on single page.
  • Figure 4: Qualitative analysis on a random sample on ReasonVQA. Frames sub-sampled to fit on single page.