Improving the Throughput of Diffusion-based Large Language Models via a Training-Free Confidence-Aware Calibration
Not on the map. Why?
Abstract (arXiv)
We present CadLLM, a training-free method to accelerate the inference throughput of diffusion-based LLMs (dLLMs). We first investigate the dynamic nature of token unmasking confidence across blocks and steps. Based on this observation, we present a lightweight adaptive approach that controls the generation block size, step size, and threshold based on the average confidence of unmasked tokens. We further reduce softmax overhead by dynamically leveraging a subset of the vocabulary to regulate sampling breadth. CadLLM is a plug-and-play, model-agnostic method compatible with KV-cache-based dLLMs. Extensive experiments on four popular tasks demonstrate that CadLLM yields up to 1.1-2.28x throughput improvement over the state-of-the-art baseline with competitive accuracy.
Key facts. Cited by 7 papers in the map (data through 7 Oct 2026). Builds on 7 of its 20 references that are in the map. First posted on arXiv 8 Dec 2025.
Builds on
7 of 20 references.
References · 7
- 2021Structured Denoising Diffusion Models in Discrete State-Spacescited by 1,665
- 2018Hierarchical Neural Story Generationcited by 1,182
- 2022A Continuous Time Framework for Discrete Denoising Modelscited by 437
- 2025Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decodingcited by 404
- 2018Learning to Write with Cooperative Discriminatorscited by 119
- 2025Diffusion Language Models Know the Answer Before Decodingcited by 68
- 2025AdaBlock-dLLM: Semantic-Aware Diffusion LLM Inference via Adaptive Block Sizecited by 30
Built on by
Citing papers per year: 2025: 1; 2026: 6.
7 papers.
- 2025Learning Unmasking Policies for Diffusion Language Modelscited by 29
- 2026Beyond Hard Masks: Progressive Token Evolution for Diffusion Language Modelscited by 6
- 2026R2-dLLM: Accelerating Diffusion Large Language Models via Spatio-Temporal Redundancy Reductioncited by 6
- 2026TAD: Temporal-Aware Trajectory Self-Distillation for Fast and Accurate Diffusion LLMcited by 2
- 2026GeoBlock: Inferring Block Granularity from Dependency Geometry in Diffusion Language Modelscited by 1
- 2026Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMscited by 1
- 2026OLED-MoE: Accelerating MoE-Based dLLM Inference via Inter-Iteration Locality-Aware Expert Offloadingcited by 0
- 2026OLED-MoE: Accelerating MoE-Based dLLM Inference via Inter-Iteration Locality-Aware Expert Offloading27 Sep 2026
- 2026TAD: Temporal-Aware Trajectory Self-Distillation for Fast and Accurate Diffusion LLM10 May 2026
- 2026Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs7 May 2026
- 2026R2-dLLM: Accelerating Diffusion Large Language Models via Spatio-Temporal Redundancy Reduction21 Apr 2026
- 2026GeoBlock: Inferring Block Granularity from Dependency Geometry in Diffusion Language Models4 Mar 2026
- 2026Beyond Hard Masks: Progressive Token Evolution for Diffusion Language Models12 Jan 2026
- 2025Learning Unmasking Policies for Diffusion Language Models9 Dec 2025
Similar, not cited
Close in wording, no citation either way
- 2026CAI-DLLM: Convergence Aware Inference for Diffusion Language Models
- 2026Fast-dLLM++: Fr'echet Profile Decoding for Faster Diffusion LLM Inference
- 2026Beyond Scattered Acceptance: Fast and Coherent Inference for DLMs via Longest Stable Prefixes
- 2026CD4LM: Consistency Distillation and aDaptive Decoding for Diffusion Language Models
- 2025Finish First, Perfect Later: Test-Time Token-Level Cross-Validation for Diffusion Large Language Models
- 2026Roll Out and Roll Back: Diffusion LLMs are Their Own Efficiency Teachers
- 2026Window-Diffusion: Accelerating Diffusion Language Model Inference with Windowed Token Pruning and Caching
- 2025d2Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching