MedSTP: Preserving Subspace and Transition Evidence for Medical VQA Token Pruning
Medical vision-language models (VLMs) for visual question answering (VQA) often pass long visual-token sequences to the language model, making visual-token pruning attractive but risky for medical images, where answers may depend on localized boundary and texture evidence.
ProofPaper ↗
Key points
- We study training free pre-decoding pruning and treat pruning as a subspace stable evidence preservation problem.
- To this end, we propose Subspace-Transition Token Pruning (\method{}), which estimates each token's spectral support from principal component space and constructs a compact representation subspace.
- To better reflect medical image structure, \method{} further uses a bounded token-grid transition prior to promote boundary and texture sensitive tokens.
- These results show that \method{} is an effective training free pruning framework for medical domain VQA.
Sources (1)
- [1]MedSTP: Preserving Subspace and Transition Evidence for Medical VQA Token PruningarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 11:38 PM
Medical vision-language models (VLMs) for visual question answering (VQA) often pass long visual-token sequences to the language model, making visual-token pruning attractive but risky for medical images, where answers may depend on localized boundary and texture evidence.
We study training free pre-decoding pruning and treat pruning as a subspace stable evidence preservation problem.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026EgoP$^{2}$C: Internalizing Plan-Perceive-Compose for Spatial Reasoning in Egocentric Video
- Oct 15, 2026Twice Upon a Time: Complementary Representations for Time-Series Reasoning
- Oct 15, 2026Each Reward to Its Relevant Tokens: Reward-Specific Token-Level Credit Assignment for Multi-Reward RL
- Oct 15, 2026Learning to Think and Explore in 3D World
- Oct 15, 2026Exploring The Empirical Privacy Landscape of Synthetic Data
- Oct 15, 2026CoWorld Reasoner: Agentic Multimodal Reasoning with the Evolving Video World Model-Driven Harness