WAM-Diff2 Achieves 15x Speedup for Self-Driving AI via Parallel Inference
jiqizhixin · x · 2026-08-25
Fudan University and Yinwang Intelligence introduced WAM-Diff2, converting autoregressive Vision-Language-Action models into a discrete diffusion architecture.
Core Methodology
- Parallel Generation: Uses Block-Causal Attention to allow bidirectional parallel interaction of tokens within a block, maintaining causal order only across blocks.
- Knowledge Migration: Adopts a three-level hierarchical distillation strategy to progressively migrate the model from sequential to parallel generation, preserving all semantic knowledge.
Experimental Results
- Speedup: Achieves a 15.1x decoding speedup (from 44.5 to 673.4 tokens/s).
- Performance Retention: No drop in performance across multi-task benchmarks like DriveBench, LingoQA, COCO, and NAVSIM.
- Planning Improvement: Long-horizon planning error drops by 5.8%, outperforming specialized models like ReCogDrive on NAVSIM.
More from Research
- Study asks: Are LLM agents time-aware and budget-conscious? — maksym_andr · 2026-08-25
- SA-RSQ: Sparse Representation Framework for Multi-modal Recommender Systems — _reachsumit · 2026-08-25
- Revisiting N2DCG: Empirical Reformulation for Carousel Recommendation — _reachsumit · 2026-08-25
- RAG collapse: LLM answers converge when retrieving self-authored content, 79.6% simulations collapse — _reachsumit · 2026-08-25
- Semantic subword tokenization improves generative recommenders by reducing intra-item attention overload — _reachsumit · 2026-08-25
- Spotify study: better reasoning traces can hurt recommender effectiveness — _reachsumit · 2026-08-25