Skip to content

About

Code for CVPR 2026 "Distilling Counterfactual Reasoning from Language to Vision: Causal Graph Guided Post-Training for Video Understanding"

Resources

Stars

1 star

Watchers

0 watching

Forks

 
 

Latest commit

 

History

6 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 

Repository files navigation

CounterVQA: Evaluating and Improving Counterfactual Reasoning in Vision-Language Models for Video Understanding

Yuefei Chen1, Jiang Liu2, Xiaodong Lin1, Ruixiang Tang1†
1Rutgers University   2Advanced Micro Devices
†Corresponding author


🔥 News

  • [2026/03] CounterVQA is accepted to CVPR 2026 Findings!
  • Dataset & Code will be released.

📖 Overview

CounterVQA is the first video benchmark that systematically evaluates counterfactual reasoning in Vision-Language Models (VLMs) using explicit causal graphs.

CFGPT (Counterfactual Graph-based Post-Training) is a two-stage framework that distills counterfactual reasoning capability from language to vision, achieving +12.5% improvement over the leading 8B-scale VLM.

Key Contributions

  • CounterVQA Benchmark — 712 videos, 3,987 QA pairs across three progressive difficulty levels:
    • Level 1: Adjacent counterfactual inference (single-hop causal reasoning)
    • Level 2: Long-chain counterfactual inference (multi-hop causal chains)
    • Level 3: Counterfactual inference with non-existent events (hallucination detection)
  • CFGPT Framework — Two-stage post-training combining cross-modal causal transfer (SFT) and visual-causal alignment optimization (GRPO with causal graph rewards)
  • Comprehensive Evaluation — Benchmarking 5 state-of-the-art open- and closed-source VLMs

📊 Main Results

Model Level 1 Level 2 Level 3 Avg.
ChatGPT-4o 9.6 8.0 1.3 6.2
Gemini-2.5-Pro 41.1 40.5 42.4 41.5
Qwen-2.5-VL-7B 12.3 10.3 2.7 8.3
Qwen-2.5-VL-32B 33.0 46.2 44.0 39.4
Qwen-3-VL-8B 54.6 62.1 65.3 60.1
Qwen-3-VL-8B + CFGPT (Ours) 70.1 71.6 76.0 72.6

If you have any questions, please open an issue or contact chen.yuefei@rutgers.edu

About

Code for CVPR 2026 "Distilling Counterfactual Reasoning from Language to Vision: Causal Graph Guided Post-Training for Video Understanding"

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages