Turning the TIDE: Cross-Architecture Distillation for Diffusion Large Language Models

• URL: http://arxiv.org/abs/2604.26951v1
• Date: Wed, 29 Apr 2026 17:59:01 GMT
• Title: Turning the TIDE: Cross-Architecture Distillation for Diffusion Large Language Models
• Abstract summary: Distilling 8B dense and 16B MoE teachers into a 0.6B student via two heterogeneous pipelines outperforms the baseline by an average of 1.53 points across eight benchmarks.<n>HumanEval scores reach 48.78 compared to 32.3 for the AR baseline.
• Score: 25.171094765557996
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Diffusion large language models (dLLMs) offer parallel decoding and bidirectional context, but state-of-the-art dLLMs require billions of parameters for competitive performance. While existing distillation methods for dLLMs reduce inference steps within a single architecture, none address cross-architecture knowledge transfer, in which the teacher and student differ in architecture, attention mechanism, and tokenizer. We present TIDE, the first framework for cross-architecture dLLM distillation, comprising three modular components: (1) TIDAL, which jointly modulates distillation strength across training progress and diffusion timestep to account for the teacher's noise-dependent reliability; (2) CompDemo, which enriches the teacher's context via complementary mask splitting to improve predictions under heavy masking; and (3) Reverse CALM, a cross-tokenizer objective that inverts chunk-level likelihood matching, yielding bounded gradients and dual-end noise filtering. Distilling 8B dense and 16B MoE teachers into a 0.6B student via two heterogeneous pipelines outperforms the baseline by an average of 1.53 points across eight benchmarks, yielding notable gains in code generation, where HumanEval scores reach 48.78 compared to 32.3 for the AR baseline.

Related papers

• Trajectory-Level Speculative Decoding for Diffusion Language Models [16.97093186372498]
  Diffusion-based language models (dLLMs) enable parallel token generation through iterative denoising, but existing decoding strategies collapse to single-token generation under low confidence.<n>We develop a trajectory-level speculative framework that constructs draft denoising trajectories via confidence-stratified tree exploration.<n>Our framework reduces denoising by 30-40% and increases tokens-per-step from 2.6 to 4.3, achieving 7-14x speedup over vanilla dLLMs and 1.3x over Fast-dLLM with less than 1% accuracy change across reasoning and code benchmarks.
  arXiv  Detail & Related papers  (2026-08-27T09:42:20Z)
• LLaDA MoE v2: Scaling Mixture-of-Experts Diffusion Language Models [129.64558694024126]
  Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling.<n>We characterize how optimization hyper parameters, compute allocation, and architecture scale for MoE dLLMs.<n>We train LLaDA MoE v2, a 30B-A3B dLLM, from scratch on 23.5T tokens.
  arXiv  Detail & Related papers  (2026-08-04T10:53:02Z)
• DiffusionGemma Technical Report [72.10570774996219]
  We introduce DiffusionGemma, a language model that uses discrete diffusion to generate text at exceptionally high speed.<n>Rather than decoding one token at a time, DiffusionGemma iteratively refines blocks of 256 tokens in parallel.<n>Our compute-efficient two-stage training pipeline uses fewer than 10% of the starting AR model's total training token budget.
  arXiv  Detail & Related papers  (2026-07-31T16:11:46Z)
• FLARE: Diffusion for Hybrid Language Model [72.60770374799634]
  FLARE is a systematic conversion framework for hybrid-attention large language models.<n>It combines a token-equal AR-and-diffusion objective, hardware-aware kernels, and unified inference, enabling one checkpoint to support both AR-style verified decoding and diffusion-style parallel denoising.<n>Our results suggest practical dLLMs are limited not only by decoding algorithms, but also by transfer data quality and the training inefficiency of current block-diffusion objectives.
  arXiv  Detail & Related papers  (2026-06-01T06:58:15Z)
• Pruning and Distilling Mixture-of-Experts into Dense Language Models [13.03679207895992]
  Mixture-of-Experts (MoE) is now the dominant architecture for frontier language models.<n>We present the first systematic framework for converting a trained MoE into a standard fully dense architecture.<n>We evaluate 7 scoring, 5 grouping, and 2 magnitude scaling methods across a range of selected expert counts on Qwen3-30B-A3B, yielding 350 configurations.
  arXiv  Detail & Related papers  (2026-05-27T09:27:36Z)
• ES-dLLM: Efficient Inference for Diffusion Large Language Models by Early-Skipping [26.560813832545563]
  Diffusion large language models (dLLMs) are emerging as a promising alternative to autoregressive models (ARMs)<n>We analyze the generation dynamics of dLLMs and find that intermediate representations, including key, value, and hidden states, change only subtly across successive iterations.<n>We propose textbfES-dLLM, a training-free inference acceleration framework for dLLM that reduces computation by skipping tokens in early layers based on the estimated importance.
  arXiv  Detail & Related papers  (2026-03-10T14:31:19Z)
• Prism: Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models [96.0074341403456]
  Inference-time compute has re-emerged as a practical way to improve LLM reasoning.<n>Most test-time scaling (TTS) algorithms rely on autoregressive decoding.<n>We propose Prism, an efficient TTS framework for dLLMs.
  arXiv  Detail & Related papers  (2026-02-02T09:14:51Z)
• MoDES: Accelerating Mixture-of-Experts Multimodal Large Language Models via Dynamic Expert Skipping [52.02659589971978]
  We propose MoDES, the first training-free framework that adaptively skips experts to enable efficient and accurate MoE MLLM inference.<n>MoDES significantly enhances inference speed, improving the prefilling time by 2.16$times$ and the decoding time by 1.26$times$.
  arXiv  Detail & Related papers  (2025-11-19T18:48:27Z)
• Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models [49.911784762244814]
  TraceRL is a trajectory-aware reinforcement learning framework for diffusion language models (DLMs)<n>We derive a series of state-of-the-art diffusion language models, namely TraDo.<n>TraDo-8B-Instruct achieves relative accuracy improvements of 6.1% over Qwen2.5-7B-Instruct and 51.3% over Llama3.1-8B-Instruct on mathematical reasoning benchmarks.
  arXiv  Detail & Related papers  (2025-09-08T17:58:06Z)
• ASMA-Tune: Unlocking LLMs' Assembly Code Comprehension via Structural-Semantic Instruction Tuning [33.53059396922164]
  Assembly code analysis and comprehension play critical roles in applications like reverse engineering.<n>Traditional masked language modeling approaches do not explicitly focus on natural language interaction.<n>We present Assembly Augmented Tuning, an end-to-end structural-semantic instruction tuning framework.
  arXiv  Detail & Related papers  (2025-03-14T17:36:08Z)
• TinyR1-32B-Preview: Boosting Accuracy with Branch-Merge Distillation [19.938309176933902]
  We introduce the Branch-Merge distillation approach, which enhances model compression through two phases.<n>We validate our distillation approach using DeepSeek-R1 as the teacher and DeepSeek-R1-Distill-Qwen-32B as the student.<n>The resulting merged model, TinyR1-32B-Preview, outperforms its counterpart DeepSeek-R1-Distill-Qwen-32B across multiple benchmarks.
  arXiv  Detail & Related papers  (2025-03-06T16:25:53Z)
• Adapted Multimodal BERT with Layer-wise Fusion for Sentiment Analysis [84.12658971655253]
  We propose Adapted Multimodal BERT, a BERT-based architecture for multimodal tasks. adapter adjusts the pretrained language model for the task at hand, while the fusion layers perform task-specific, layer-wise fusion of audio-visual information with textual BERT representations. In our ablations we see that this approach leads to efficient models, that can outperform their fine-tuned counterparts and are robust to input noise.
  arXiv  Detail & Related papers  (2022-12-01T17:31:42Z)
• Squeezeformer: An Efficient Transformer for Automatic Speech Recognition [99.349598600887]
  Conformer is the de facto backbone model for various downstream speech tasks based on its hybrid attention-convolution architecture. We propose the Squeezeformer model, which consistently outperforms the state-of-the-art ASR models under the same training schemes.
  arXiv  Detail & Related papers  (2022-06-02T06:06:29Z)
• A Mixture of Expert Based Deep Neural Network for Improved ASR [4.993304210475779]
  MixNet is a novel deep learning architecture for acoustic model in the context of Automatic Speech Recognition (ASR) In natural speech, overlap in distribution across different acoustic classes is inevitable, which leads to inter-class mis-classification. Experiments are conducted on a large vocabulary ASR task which show that the proposed architecture provides 13.6% and 10.0% relative reduction in word error rates.
  arXiv  Detail & Related papers  (2021-12-02T07:26:34Z)
• G-DetKD: Towards General Distillation Framework for Object Detectors via Contrastive and Semantic-guided Feature Imitation [49.421099172544196]
  We propose a novel semantic-guided feature imitation technique, which automatically performs soft matching between feature pairs across all pyramid levels. We also introduce contrastive distillation to effectively capture the information encoded in the relationship between different feature regions. Our method consistently outperforms the existing detection KD techniques, and works when (1) components in the framework are used separately and in conjunction.
  arXiv  Detail & Related papers  (2021-08-17T07:44:27Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.