Breaking Block Boundaries: Anchor-based History-stable Decoding for Diffusion Large Language Models

This project provides a reference implementation for a specific decoding optimization for Diffusion LLMs (dLLMs). While dLLMs are a high-growth research area (e.g., following the success of models like SEDD), this specific contribution is a technical refinement of the Semi-AR decoding process. The defensibility is low (3) because the 'moat' consists entirely of the mathematical insight described in the paper; once the technique is validated, it can be trivially reimplemented by any inference engine provider (vLLM, TGI, NVIDIA TensorRT-LLM). The high fork-to-star ratio (7 forks, 0 stars) suggests immediate interest from research peers looking to replicate or build upon the results, but it lacks the community gravity or infrastructure complexity required for a higher score. Frontier labs like OpenAI or Google, who are heavily incentivized to reduce inference costs, are the primary 'threats' as they would likely integrate such optimizations directly into their proprietary stacks if they move toward diffusion-based architectures. The displacement horizon is short (6 months) because inference optimization is a fast-moving field where paper-to-production pipelines are becoming increasingly streamlined.