Research map / Vision-language and multimodal / Visual question answering
GRACE: Graph-Based Contextual Debiasing for Fair Visual Question Answering
Yifeng Zhang, Ming Jiang, Qi Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantic fairness loss LS to graph learning (ablation) | GRACE context graph [sF] · VQA-CP | accuracy: 55.38 -> 56.87 | Improved |
| add structural fairness loss LT to graph learning (ablation) | GRACE context graph [sF] · VQA-CP | accuracy: 55.38 -> 56.48 | Improved |
| add fairness losses LS and LT (ablation) | GRACE context graph [sF] · OK-VQA | accuracy: 60.36 -> 60.42 | No clear effect |
| add reasoning similarity sR to semantic similarity with both fairness losses (ablation) | GRACE · GQA-OOD tail | accuracy: 48.93 -> 50.14 | Improved |
| add reasoning similarity sR to semantic similarity with both fairness losses (ablation) | GRACE · GQA-OOD | head-tail gap: 9.31 -> 7.49 | Improved |
| add reasoning similarity sR without fairness losses (ablation) | GRACE LR baseline · VQA-CP | accuracy: 55.38 -> 54.87 | Hurt |
| combine semantic and reasoning similarity for retrieval versus semantic only (ablation) | GRACE retrieval · VQA-CP | answer hit rate: 74.82 -> 79.95 | Improved |
| swap Prophet for GRACE on different reasoners | Prophet · VQA-CP | accuracy: 54.02 -> 57.32 (Llama2-7B) | Improved |
| swap Prophet for GRACE on OK-VQA across reasoners | Prophet · OK-VQA | accuracy: 61.11 -> 60.32 (GPT-3) | Mixed |
| apply LMH and CSS debiasing to Prophet | Prophet · OK-VQA | accuracy: 61.11 -> 57.28 | Hurt |
| apply LMH and CSS debiasing to GRACE | GRACE · VQA-CP | accuracy: 57.35 -> 61.37 | Improved |
| use GRACE instead of Prophet | VQA-CP gender subsets · VQA-CP gender | male-female gap: 4.39 -> 0.90 | Improved |
Later papers that compared with this method
4 later accepted papers report GRACE as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Evaluating and Understanding Model Editing for Medical Vision Language Models (ECCV 2026 · better on 8 of 8 shared results)
- OAD-Promoter: Enhancing Zero-Shot VQA Using Large Language Models with Object Attribute Description (AAAI 2026 · better on 1 of 3 shared results)
- ReasonEdit: Editing Vision--Language Models using Human Reasoning (ICML 2026 · better on 0 of 1 shared result)
- BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing (ICML 2025 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Language‑Bias‑Resilient Visual Question Answering via Adaptive Multi‑Margin Collaborative Debiasing (NeurIPS 2025 · similarity 0.92)
- Debiased Visual Question Answering from Feature and Sample Perspectives (NeurIPS 2021 · similarity 0.92)
- Roses Are Red, Violets Are Blue... but Should VQA Expect Them To? (CVPR 2021 · similarity 0.91)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.89)
- CVQA: Culturally-diverse Multilingual Visual Question Answering Benchmark (NeurIPS 2024 · similarity 0.89)
- Multimodal Commonsense Knowledge Distillation for Visual Question Answering (Student Abstract) (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
