Research map / Vision-language and multimodal / Multimodal LLMs
Wake up for Touch! Mask-isolated Tactile Alignment Learning in MLLMs
Yoonhyung Park, Minji Kim, Sungwon Moon, Jiyoung Lee
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 |
|---|---|---|---|
| use mask-guided dormant-subspace tuning instead of LoRA (TVL) on the same backbone | Qwen2.5-VL-3B · VTL average | LLM-judge score: 4.50 -> 4.91 | Improved |
| use Splash instead of LoRA on TacQuad | Qwen2.5-VL-3B · TacQuad | LLM-judge score: 4.22 -> 4.86 | Improved |
| use Splash instead of LoRA, VL retention on MathVista | Qwen2.5-VL-3B · MathVista | accuracy: 52.8 -> 65.3 | Improved |
| use Splash instead of LoRA, VL retention on MMMU | Qwen2.5-VL-3B · MMMU val | accuracy: 50.0 -> 55.3 | Improved |
| tune with sparsity 100% (full LLM tuning) instead of 60% (ablation) | Splash-3B · VL benchmarks | MMMU val: 55.3 -> 22.0 | Hurt |
| tune with sparsity 80% instead of 60% (ablation) | Splash-3B · VL benchmarks | MMMU val: 55.3 -> 52.6 | Hurt |
| lower sparsity 30% instead of 60% (ablation) | Splash-3B · VTL average | LLM-judge score: 4.91 -> 4.79 | Hurt |
| calibrate mask with unpaired noise images instead of CC3M paired (ablation) | Splash-3B · VTL average | LLM-judge score: 4.91 -> 4.88 | No clear effect |
| calibrate with TVL paired data instead of CC3M (ablation) | Splash-3B · VTL average | LLM-judge score: 4.91 -> 4.75 | No clear effect |
| use uniform sparsity instead of OWL adaptive layer-wise sparsity (ablation) | Splash-3B · VTL average | LLM-judge score: 4.83 -> 4.91 | Improved |
| freeze first and last transformer blocks (ablation) | Splash-3B · VTL average | LLM-judge score: 0.18 point drop when unfrozen | Improved |
| use Splash instead of (IA)3 on TacQuad | Qwen2.5-VL-3B · TacQuad | LLM-judge score: 3.84 -> 4.86 | Improved |
Closest papers by meaning
- Touch-R1: Reinforcing Touch Reasoning in MLLMs (NeurIPS 2026 · similarity 0.91)
- Universal Visuo-Tactile Video Understanding for Embodied Interaction (NeurIPS 2025 · similarity 0.88)
- Binding Touch to Everything: Learning Unified Multimodal Tactile Representations (CVPR 2024 · similarity 0.88)
- PhysTacGen: Physics-Aware Visual-Tactile Sensor Image Generation (NeurIPS 2026 · similarity 0.87)
- AnyTouch: Learning Unified Static-Dynamic Representation across Multiple Visuo-tactile Sensors (ICLR 2025 · similarity 0.87)
- A Touch, Vision, and Language Dataset for Multimodal Alignment (ICML 2024 · 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.
