CLASP: Class-Adaptive Layer Fusion and Dual-Stage Pruning for Multimodal Large Language Models

CLASP addresses a critical bottleneck in MLLMs: the high computational cost of processing visual tokens. Its technical novelty lies in moving away from static, single-layer feature extraction toward a dynamic, instruction-aware fusion of ViT layers combined with pruning. While technically sound, the project's defensibility is low (3) because it is primarily a research-grade algorithm. Its value is easily captured by inference engines (vLLM, TensorRT-LLM) or the model creators themselves (OpenAI, Google) who are incentivized to bake these optimizations directly into their proprietary architectures. The 7 forks within 3 days despite 0 stars suggest high interest from the research community (likely internal lab members or peer researchers), but it lacks a commercial or ecosystem moat. It competes with existing techniques like Token Merging (ToMe), DynamicViT, and the built-in pooling strategies used in LLaVA-NeXT or Qwen-VL. The 'platform domination risk' is high because as MLLMs move toward the edge or large-scale production, efficiency techniques like CLASP will be standardized into the hardware-accelerated kernels provided by NVIDIA or integrated into the core architecture of frontier models to reduce serving costs.