Research map / Large language models / Language, safety and interpretability
Enhancing Numerical Prediction of MLLMs with Soft Labeling
Pei Wang, Zhaowei Cai, Hao Yang, Davide Modolo, Ashwin Swaminathan
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 |
|---|---|---|---|
| replace one-hot hard labels with soft labels on digit tokens | LLaVA-7B specialist · RefCOCOg val | Acc@0.5: 79.0 -> 82.0 | Improved |
| replace one-hot hard labels with soft labels on digit tokens | LLaVA-7B generalist · RefCOCOg val | Acc@0.5: 72.1 -> 78.5 | Improved |
| soft labeling on stronger 13B model | LLaVA-13B specialist · RefCOCOg val | Acc@0.5: 86.0 -> 86.7 | Improved |
| soft labeling for chart numeric answers | LLaVA-13B generalist · ChartQA numerical subset | accuracy: 69.7 -> 72.6 | Improved |
| soft labeling gain larger on numerical subset than full ChartQA set | LLaVA-7B specialist · ChartQA | accuracy: +1.7 (all) vs +2.3 (numerical) | Improved |
| soft labeling for counting on balanced test | LLaVA-7B specialist · TallyQA simple balanced | RMSE: 3.84 -> 3.20 | Improved |
| soft labeling on standard unbalanced TallyQA simple | LLaVA-7B specialist · TallyQA simple | Acc.: 78.5 -> 79.0 | Improved |
| use triangular/binomial instead of uniform soft distribution (ablation) | soft labeling · ChartQA numerical | accuracy: 66.8 -> 68.6 | Improved |
| uniform distribution (label smoothing) vs hard labeling (ablation) | LLaVA-7B specialist · ChartQA numerical | accuracy: 66.3 -> 66.8 | Improved |
| set eta too small or too large (ablation) | soft labeling · RefCOCOg val, ChartQA, TallyQA | accuracy/RMSE | Hurt |
| increase lambda beyond 5 (ablation) | soft labeling · RefCOCOg val, ChartQA numerical, TallyQA | accuracy/RMSE | Hurt |
| soft labeling effect on generic tasks | soft labeling · generic tasks (Supp.) | performance | No clear effect |
Closest papers by meaning
- Dirichlet-Based Prediction Calibration for Learning with Noisy Labels (AAAI 2024 · similarity 0.87)
- DenseMLLM: Standard Multimodal LLMs for Dense Prediction (ICML 2026 · similarity 0.87)
- Accelerating Multimodal Large Language Models by Searching Optimal Vision Token Reduction (CVPR 2025 · similarity 0.87)
- Fit and Prune: Fast and Training-free Visual Token Pruning for Multi-modal Large Language Models (AAAI 2025 · similarity 0.86)
- The Closeness of In-Context Learning and Weight Shifting for Softmax Regression (NeurIPS 2024 · similarity 0.86)
- Attention with Trained Embeddings Provably Selects Important Tokens (NeurIPS 2025 · similarity 0.86)
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.
