5arXiv cs.CL (Computation and Language)·
AngelSpec: Unified training framework for speculative decoding combining MTP and block-parallel diffusion
AngelSpec is a new training framework that unifies autoregressive multi-token prediction (MTP) and block-parallel speculative decoding, co-specializing each drafter to different data domains (conversational vs. code/math). The paper introduces DFly, a block-diffusion architecture with hybrid target-conditioning and adaptive verification depth that treats verification as a shared batch-level resource. On the Hy3-A21B model, DFly achieves 1.98–2.40x speedup over autoregressive decoding and 10.5–11.8% higher throughput than DFlash across concurrency levels from 4 to 64. The framework and code are released publicly.
Related guides (3)
Related events (8)
AdaFlash: Adaptive speculative decoding with on-policy distilled diffusion drafters
AdaFlash is a new framework for accelerating LLM inference via speculative decoding that addresses variance problems in diffusion-based draft models. It combines an on-policy distillation algorithm using reverse-KL divergence with an adaptive length head that dynamically adjusts candidate sequence length. Experiments show up to ~66% higher throughput than prior state-of-the-art methods, with especially large gains in high-concurrency deployment scenarios.
SimSD: Speculative Decoding Adapted for Diffusion Language Models
SimSD introduces a training-free speculative decoding algorithm for diffusion large language models (dLLMs), which previously could not use standard token-level speculative decoding due to their bidirectional attention and masked language modeling formulation. The method uses a plug-and-play masking strategy that introduces reference tokens from a draft model and a custom attention mask, enabling valid logit computation for drafted tokens in a single forward pass. Evaluated on SDAR-family dLLMs across four benchmarks, SimSD achieves up to 7.46x decoding throughput improvement while maintaining or improving generation quality. The approach is compatible with other acceleration techniques such as KV cache and blockwise decoding.
DeLS-Spec: Decoupled long-short context speculative decoding improves drafting efficiency
DeLS-Spec is a new speculative decoding method that combines a fixed block-parallel draft model (DFlash) as a long-context expert with a lightweight locally-trained short-context head, avoiding joint training with the target model. The approach introduces intra-block causal conditioning at low training cost and is modular across DFlash checkpoints. Experiments on Qwen3 models show consistent speedup and acceptance-length improvements over DFlash on math, code, and dialogue benchmarks.
DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation for LLM Inference
DSpark is a new speculative decoding framework that combines a semi-autoregressive draft architecture with confidence-scheduled verification to improve LLM inference throughput. The semi-autoregressive design introduces intra-block token dependencies to reduce acceptance decay common in parallel drafters, while dynamic verification length tuning reduces wasted batch capacity in high-concurrency settings. Deployed within the DeepSeek-V4 serving system under live traffic, DSpark achieves 60–85% faster per-user generation speeds compared to the MTP-1 production baseline at matched throughput, and expands the Pareto frontier of latency-throughput tradeoffs.
ADAS: Attention-Discounted Adaptive Sampler improves parallel decoding for masked diffusion language models
Researchers propose ADAS, a training-free reranking rule for masked diffusion language model decoding that addresses token interaction failures in parallel token commitment. The method greedily penalizes candidates that attend strongly to already-selected uncertain positions, using attention weights as soft marginal penalties rather than hard constraints. Evaluated on LLaDA-8B-Base and Dream-7B-Base across GSM8K, MATH500, HumanEval, and MBPP, ADAS improves low-NFE performance by 9–10 percentage points on average when plugged into existing samplers with only 3.1% runtime overhead.
Practical investigation of training-free relaxed speculative decoding techniques
A new arXiv preprint systematically investigates training-free relaxed speculative decoding methods, which trade the strict distribution-preservation guarantee of standard speculative decoding for potential speed or capability gains. The authors unify existing approaches in a shared framework, benchmark them on contemporary settings, and surface practical findings for practitioners. Key takeaways include that relaxed approaches require careful capability evaluation and tend to depend on high-quality drafter models, making them poorly suited for lightweight multi-token-prediction drafters.
VIA-SD: Multi-tier speculative decoding via intra-model routing cuts rejection rates and boosts inference speed
VIA-SD introduces a three-tier verification framework for speculative decoding that routes draft tokens to a lightweight 'slim verifier' submodel for medium-confidence cases, reserving full-model verification only for uncertain tokens. Across four tasks and multiple model families, the method reduces rejection rates by 0.10–0.22 and achieves 10–20% speedups over strong speculative decoding baselines, with 2.5–3x acceleration over standard decoding. The approach is compatible with existing speculative decoding frameworks without retraining. The work proposes multi-tier speculative decoding as a general paradigm for scalable LLM inference.
DeepSeek open-sources DSpark speculative decoding module, achieving 50%+ throughput gains in production
Researchers from Peking University and DeepSeek introduced DSpark, a speculative decoding module that dynamically adjusts verification depth based on server load, achieving 57–85% faster per-user token generation and 51–52% higher total throughput compared to DeepSeek's previous production drafter. The team released checkpoints DeepSeek-V4-Pro-DSpark and DeepSeek-V4-Flash-DSpark on Hugging Face under an MIT license, with the draft module attaching to frozen target model weights. Key innovations include a parallel drafting backbone (adapted from DFlash), a Markov head for sequential token coherence correction, a calibrated confidence head, and a load-aware scheduler that trades draft length against server capacity. Results generalize across model families including Qwen3 and Gemma4.