Research map / Vision-language and multimodal / Human and embodied modeling
Learning Taxonomic Trees with Hierarchical Representation Regularization for Large Multimodal Models
Hulingxiao He, Zhi Tan, Yuxin Peng
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 HiR2 regularizer to SFT | Qwen2.5-VL-3B SFT · CUB-200 base classes | HCA: 9.36 -> 35.38 | Improved |
| add HiR2 regularizer to SFT | Qwen2.5-VL-3B SFT · iNat-Animal base classes | HCA: 14.06 -> 18.07 | Improved |
| add HiR2 regularizer to DFT | Qwen2.5-VL-3B DFT · iNat-Plant base classes | HCA: 8.15 -> 11.66 | Improved |
| add HiR2 regularizer to SFT on other LMMs | Intern3.5-VL-1B SFT · iNat-Animal base classes | HCA: 2.04 -> 3.04 | Improved |
| add HiR2 regularizer to SFT on LLaVA-OV-1.5-4B | LLaVA-OV-1.5-4B SFT · iNat-Animal novel classes | HCA: 6.05 -> 6.09 | No clear effect |
| add entailment loss to SFT baseline (ablation) | HiR2 · iNat-Animal base classes | HCA: 14.06 -> 17.89 | Improved |
| add dispersive loss on top of entailment loss (ablation) | HiR2 · iNat-Animal base classes | HCA: 17.89 -> 18.07 | Improved |
| add dispersive loss on top of entailment loss (ablation) | HiR2 · iNat-Animal base classes | leaf accuracy: 40.78 -> 41.37 | Improved |
| use spherical angular dispersive loss instead of hyperbolic geodesic variant (ablation) | HiR2 dispersive loss · iNat-Animal base classes | HCA: 15.33 -> 18.07 | Improved |
| use spherical angular dispersive loss instead of tangent-space variants (ablation) | HiR2 dispersive loss · iNat-Animal base classes | leaf accuracy: 40.67 -> 41.37 (vs variant III) | Improved |
| regularize last layer (28) instead of layer 21 (ablation) | HiR2 · iNat-Animal base classes | HCA: 11.32 -> 18.07 | Improved |
| regularize all layers instead of only the last layer (ablation) | HiR2 · iNat-Animal base classes | HCA: 18.07 -> 12.06 | Hurt |
Closest papers by meaning
- Deep Taxonomic Networks for Unsupervised Hierarchical Prototype Discovery (NeurIPS 2025 · similarity 0.88)
- Predictive Regularization Against Visual Representation Degradation in Multimodal Large Language Models (CVPR 2026 · similarity 0.87)
- Vision as a Dialect: Unifying Visual Understanding and Generation via Text-Aligned Representations (NeurIPS 2025 · similarity 0.87)
- Analyzing and Boosting the Power of Fine-Grained Visual Recognition for Multi-modal Large Language Models (ICLR 2025 · similarity 0.87)
- Bridging Vision and Language Spaces with Assignment Prediction (ICLR 2024 · similarity 0.87)
- Taxonomy-Aware Evaluation of Vision-Language Models (CVPR 2025 · similarity 0.87)
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.
