Research map / Vision-language and multimodal / Visual question answering
Large-scale Pretraining for Visual Dialog: A Simple State-of-the-Art Baseline
Vishvak Murahari, Dhruv Batra, Devi Parikh, Abhishek Das
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 MLM loss to NSP pretraining (ablation) | language-only BERT · VisDial v1.0 val | NDCG: 56.17 -> 57.26 | Improved |
| increase dialog history from 0 to 4 rounds (ablation) | language-only NSP+MLM · VisDial v1.0 val (language-only) | NDCG: +4.56 on NDCG, +8.54 on MRR | Improved |
| increase dialog history from 4 to 10 rounds (ablation) | language-only NSP+MLM · VisDial v1.0 val (language-only) | NDCG: +2.16 on NDCG, +1.57 on MRR | Improved |
| add visual stream initialized from language-only weights | ViLBERT · VisDial v1.0 val | NDCG: 57.26 -> 62.64 | Improved |
| pretrain on CC before VisDial finetuning (vs L-only init) (ablation) | ViLBERT · VisDial v1.0 val | NDCG: 62.64 -> 60.80 | Hurt |
| add VQA pretraining after CC (ablation) | ViLBERT · VisDial v1.0 val | NDCG: 60.80 -> 64.94 | Improved |
| finetune on dense annotations with CE loss | ViLBERT w/ CC+VQA · VisDial v1.0 val | NDCG: 64.94 -> 75.24 NDCG; 69.10 -> 52.22 MRR | Mixed |
| add NSP loss to dense CE finetuning (ablation) | ViLBERT dense CE · VisDial v1.0 val | MRR: 52.22 -> 65.88 MRR; 75.24 -> 69.24 NDCG | Improved |
| zero-shot BERT (Wikipedia+BooksCorpus) with no VisDial finetuning (ablation) | language-only NSP+MLM · VisDial v1.0 val | NDCG: 57.26 -> 11.63 | Hurt |
| finetune only loss heads with other layers frozen (ablation) | language-only NSP+MLM · VisDial v1.0 val | NDCG: 11.63 -> 19.69 | Improved |
| add inconsistency loss in same batches (ablation) | language-only NSP+MLM · VisDial v1.0 val | NDCG: 57.26 -> 56.22 | Hurt |
| use jumbled inconsistency loss with NSP (ablation) | language-only NSP · VisDial v1.0 val | NDCG: 55.66 -> 55.83 | No clear effect |
Closest papers by meaning
- Structure-Aware Multimodal Sequential Learning for Visual Dialog (AAAI 2024 · similarity 0.90)
- V^2Dial: Unification of Video and Visual Dialog via Multimodal Experts (CVPR 2025 · similarity 0.90)
- Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models (ICML 2024 · similarity 0.89)
- MMDU: A Multi-Turn Multi-Image Dialog Understanding Benchmark and Instruction-Tuning Dataset for LVLMs (NeurIPS 2024 · similarity 0.89)
- Tackling Vision Language Tasks through Learning Inner Monologues (AAAI 2024 · similarity 0.89)
- EfficientLLaVA: Generalizable Auto-Pruning for Large Vision-language Models (CVPR 2025 · similarity 0.88)
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.
