Research map / Vision-language and multimodal / Visual question answering
Introspective Distillation for Robust Question Answering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 introspective distillation to student | CSS · VQA v2 val (ID) | accuracy (%): 56.98 -> 62.57 | Helped |
| add introspective distillation to student | CSS · VQA-CP v2 test (OOD) | accuracy (%): 58.95 -> 60.17 | Helped |
| add introspective distillation to student | LMH · VQA-CP v2 test (OOD) | accuracy (%): 52.01 -> 51.31 | Hurt |
| add introspective distillation to student | CF-VQA · VQA v2 val (ID) | accuracy (%): 60.94 -> 63.40 | Helped |
| add introspective distillation to student | LM teacher on BERT · SQuAD k=1 dev OOD | EM: 71.31 -> 72.14 | Helped |
| add introspective distillation to student | LM teacher on XLNet · SQuAD k=1 dev OOD | F1: 69.49 -> 68.84 | Hurt |
| use cross-entropy matching score instead of predicted probability (ablation) | IntroD (LMH) · VQA-CP v2 / VQA v2 | harmonic mean (%): 52.79 -> 56.17 | Helped |
| use weights inversely proportional to score instead of proportional (ablation) | IntroD (LMH) · VQA-CP v2 / VQA v2 | harmonic mean (%): 51.50 -> 56.17 | Helped |
| use introspective weights instead of fixed 0.5 average (ablation) | IntroD (CSS) · VQA-CP v2 / VQA v2 | harmonic mean (%): 56.51 -> 61.35 | Helped |
| add ID-knowledge blending instead of learning only from OOD teacher (ablation) | IntroD (CSS) · VQA-CP v2 / VQA v2 | harmonic mean (%): 59.17 -> 61.35 | Helped |
| add ID-knowledge blending instead of learning only from OOD teacher (ablation) | IntroD (LMH) · VQA-CP v2 / VQA v2 | harmonic mean (%): 55.81 -> 56.17 | Helped |
| use ground-truth labels as ID-knowledge instead of ID-teacher predictions (ablation) | IntroD (LMH) · VQA-CP v2 / VQA v2 | harmonic mean (%): 54.14 -> 56.17 | Helped |
Closest papers by meaning
- Debiased Visual Question Answering from Feature and Sample Perspectives (NeurIPS 2021 · similarity 0.89)
- Look, Listen, and Answer: Overcoming Biases for Audio-Visual Question Answering (NeurIPS 2024 · similarity 0.86)
- Discovering Bias in Latent Space: An Unsupervised Debiasing Approach (ICML 2024 · similarity 0.86)
- FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.86)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.86)
- A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering (NeurIPS 2026 · 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.