SemBlock: Semantic Boundary Dynamic Blocks for Diffusion LLMs

A new decoding framework called SemBlock aims to make diffusion language models more practical by dynamically aligning token blocks with semantic boundaries, according to a paper submitted on 3 June 2026 [1]. The approach replaces fixed-size or delimiter-based block decoding with a learned predictor that identifies where natural semantic units end [1]. Diffusion language models (DLMs) generate text through an iterative denoising process, and blockwise decoding — which commits groups of tokens at once — has been key to improving their efficiency [1]. However, existing methods typically use fixed block sizes or runtime signals based on delimiters, which often fail to align with the actual semantic structure of the generated text [1][2]. The SemBlock framework addresses this by formulating dynamic block construction as a semantic boundary prediction task [1]. It trains lightweight predictors on frozen hidden states from LLaDA, a diffusion large language model, to estimate whether a semantic unit has been completed at each candidate position [3][4]. To train these predictors, the researchers built SemBound, a supervision dataset that derives boundary labels from discourse units in natural language, reasoning steps in mathematical solutions, and implementation spans in code [3][4]. During inference, the predictor outputs boundary probabilities over a candidate window, and a scheduler selects the most confident position above a threshold as the block endpoint. If no position clears the threshold, decoding falls back to a default block size [3][4]. The work builds on earlier efforts to improve blockwise decoding. A prior system, AdaBlock-dLLM, introduced a training-free scheduler that adaptively adjusts block sizes during runtime by analyzing confidence dynamics in the denoising process [5]. That approach identified a volatility band region that encodes local semantic structure and achieved up to 5.3% accuracy improvement under the same throughput budget [5]. SemBlock goes further by learning boundary placement directly rather than relying on runtime confidence proxies [3][4]. Experiments across four benchmarks — GSM8K, IFEval, MATH, and HumanEval — showed that SemBlock consistently outperformed both fixed-block decoding and AdaBlock under the same initial block budget [1][3]. On LLaDA-1.5, SemBlock improved over AdaBlock by up to 11.60 pass@1 points on HumanEval [3][4]. Ablation studies confirmed that the gains came from placing block boundaries closer to true semantic units, not from finer segmentation or additional denoising computation [3][4]. The code has been made publicly available on GitHub [1][2].