Research map / Vision-language and multimodal / Visual question answering
Learning What Makes a Difference from Counterfactual Examples and Gradient Supervision
Damien Teney, Ehsan Abbasnedjad, Anton van den Hengel
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 GS loss on top of counterfactual data | strong VQA baseline · VQA-CP v2 focused test (All) | accuracy: 44.2 -> 46.2 | Improved |
| add GS loss on top of counterfactual data | strong VQA baseline · VQA-CP v2 test (All) | accuracy: 46.0 -> 46.8 | Improved |
| add GS loss on top of counterfactual data | strong VQA baseline · VQA-CP v2 val (All) | accuracy: 63.3 -> 62.4 | Hurt |
| add counterfactual data | weak BUTD baseline (Other only) · VQA-CP v2 test (Other) | accuracy: 43.3 -> 45.0 | Improved |
| add GS loss on top of counterfactual data | weak BUTD baseline (Other only) · VQA-CP v2 focused test (Other) | accuracy: 40.6 -> 38.3 | Hurt |
| add edited training images | COCO multi-label classifier · COCO edited images | mAP: 58.1 -> 64.0 | Improved |
| add GS loss with counterfactual relations | COCO classifier with edited data · COCO hard edited images | mAP: 56.0 -> 57.7 | Improved |
| use random relations in GS instead of counterfactual relations (ablation) | COCO classifier with GS · COCO original images | mAP: 71.8 -> 72.9 | Improved |
| add edited training data | sentiment model · IMDb test edited | accuracy: 55.3 -> 88.7 | Improved |
| add GS loss with counterfactual relations | sentiment model with edited data · IMDb test edited | accuracy: 88.7 -> 91.2 | Improved |
| use counterfactual relations instead of random relations in GS (ablation) | sentiment model · IMDb val | accuracy: 50.8 -> 89.8 | Improved |
| add edited training data | NLI model · MultiNLI dev (zero-shot) | accuracy: 46.0 -> 42.4 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- RUBi — RUBi: Reducing Unimodal Biases for Visual Question Answering (NeurIPS 2019) · better on 0 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
- Counterfactual Vision and Language Learning (CVPR 2020 · similarity 0.89)
- V-CECE: Visual Counterfactual Explanations via Conceptual Edits (NeurIPS 2025 · similarity 0.86)
- A General Search-Based Framework for Generating Textual Counterfactual Explanations (AAAI 2024 · similarity 0.86)
- Counterfactual reasoning: an analysis of in-context emergence (NeurIPS 2025 · similarity 0.86)
- From Black-box to Causal-box: Towards Building More Interpretable Models (NeurIPS 2025 · similarity 0.86)
- CF-OPT: Counterfactual Explanations for Structured Prediction (ICML 2024 · similarity 0.85)
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.
