Papers with Diffusion large language models
Mask Tokens as Prophet: Fine-Grained Cache Eviction for Efficient dLLM Inference (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing cache eviction strategies for autoregressive language models fail to account for the role of mask tokens and specific characteristics in dLLMs. |
| Approach: | They propose a training-free cache eviction framework tailored to dLLMs that denies a fully masked sequence and allows parallel decoding at the expense of memory and computation. |
| Outcome: | The proposed framework reduces the cost of memory and cache eviction and improves efficiency by reducing allocation in intermediate layers and concentrating resources on prompt-preferring heads. |
DIFFA-2: A Practical Diffusion Large Language Model for General Audio Understanding (2026.findings-acl)
Copied to clipboard
| Challenge: | Autoregressive (AR) large audio language models are expensive in data and computation . prior work shows diffusion-based LALMs can improve audio understanding under matched settings . |
| Approach: | They propose a diffusion-based LALM that upgrades the speech encoder and employs dual semantic and acoustic adapters. |
| Outcome: | a new model improves over existing autoregressive large language models and is competitive to strong AR models . the proposed model can make use of limited training data and improve inference efficiency . a recent study shows that diffusion-based models can improve audio understanding . |
Efficient Self-Evaluation for Diffusion Language Models via Sequence Regeneration (2026.acl-long)
Copied to clipboard
| Challenge: | Non-sequential and bidirectional nature of diffusion large language models makes direct likelihood-based self-evaluation challenging. |
| Approach: | They propose a self-evaluation confidence quantification method for diffusion large language models that quantifies confidence by computing the probability of regenerating tokens in the entire generated sequence, given the full context. |
| Outcome: | The proposed method is correlated with semantic coherence and answer accuracy. |
CreditDecoding: Accelerating Parallel Decoding in Diffusion Large Language Models with Trace Credit (2026.acl-long)
Copied to clipboard
| Challenge: | Diffusion large language models generate text through iterative denoising with bidirectional attention, enabling richer contextual dependencies. |
| Approach: | They propose a training-free parallel decoding method that fuses Trace Credit with current logits to boost the confidence of correct but underconfident tokens. |
| Outcome: | The proposed method achieves 5.48 times speedup with +0.48 accuracy on LLaDA-8B and is orthogonal to mainstream inference optimizations. |
Dynamic Infilling Anchors for Format-Constrained Generation in Diffusion Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Diffusion large language models (dLLMs) offer bidirectional attention and parallel generation . fixed anchors can enforce constraints, but they often impose rigid spans, leading to truncated reasoning . |
| Approach: | They propose a method that dynamically estimates end-anchor positions to adjust generation length before iterative infilling. |
| Outcome: | The proposed method improves format compliance and answer accuracy on GSM8K and MATH. |
TA-GRPO-d: Trajectory-Aware GRPO for Optimizing Denoising Trajectories in Diffusion LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Existing dLLMs rely on fixed denoising schedules and cannot learn efficient unmasking orders. |
| Approach: | They propose a framework that transforms dLLM decoding into a trajectory-aware policy . it uses a confidence-gated denoising strategy that decides which tokens to unmask . |
| Outcome: | The proposed model can learn which tokens to unmask and how many to unmak per step . it can learn the output quality and efficiency of the decoding path itself . |
Fast-Decoding Diffusion Language Models via Progress-Aware Confidence Schedules (2026.findings-acl)
Copied to clipboard
| Challenge: | *SchED* is a training-free, model-agnostic early-exit algorithm that terminates diffusion decoding using a progress-aware confidence threshold. |
| Approach: | They propose a training-free, model-agnostic early-exit algorithm that terminates diffusion decoding using a progress-aware confidence threshold. |
| Outcome: | The proposed algorithm achieves 4 speedups on instruction-tuned models while maintaining baseline performance on average. |
Beyond Fully Random Masking: Attention-Guided Denoising and Optimization for Diffusion Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for full-attention dLLMs rely on random masking strategies that overlook intrinsic token dependencies. |
| Approach: | They propose an attention-guided denoising and optimization framework that aligns training and optimization with attention-derived dependencies. |
| Outcome: | The proposed framework outperforms state-of-the-art methods on mathematical and coding benchmarks. |
The Confidence Paradox: Unveiling the Latent Discriminative Power of Diffusion Large Language Models in Mathematical Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Diffusion large language models (DLLMs) are a promising alternative to autoregressive (AR) generation, offering token-level probabilities under bidirectional context. |
| Approach: | They propose to use diffusion large language models to generate token-level probabilities under bidirectional context and to examine the calibration paradox inherent to their native uncertainty estimates. |
| Outcome: | The proposed model outperforms AR baselines on mathematical reasoning benchmarks and is highly miscalibrated on reasoning benchmark. |