One Token Is Enough: Improving Diffusion Language Models with a Sink Token (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing Diffusion Language Models lack a structural constraint to stabilize attention sinks. |
| Approach: | They propose a simple but effective extra sink token that is constrained to attend to itself while remaining globally visible to all other tokens. |
| Outcome: | The proposed token is able to stabilize attention sinks and improve model performance. |
Similar Papers
Attention Sinks in Diffusion Language Models (2026.findings-acl)
Copied to clipboard
Maximo Eduardo Rulli, Simone Petruzzi, Edoardo Michielon, Fabrizio Silvestri, Simone Scardapane, Alessio Devoto
| Challenge: | Masked Diffusion Language Models (DLMs) employ transformer encoders with bidirectional attention, enabling parallel token generation while maintaining competitive performance. |
| Approach: | They conduct an empirical analysis of DLM attention patterns focusing on the attention sinking phenomenon . they find that DLMs also exhibit attention sinks, but with distinct characteristics . |
| Outcome: | The proposed models employ transformer encoders with bidirectional attention, enabling parallel token generation while maintaining competitive performance. |
Beyond Hard Masks: Progressive Token Evolution for Diffusion Language Models (2026.acl-long)
Copied to clipboard
Linhao Zhong, Linyu Wu, Bozhen Fang, Tianjian Feng, Chenchen Jing, Wen Wang, Jiaheng Zhang, Hao Chen, Chunhua Shen
| Challenge: | Existing Diffusion Language Models rely on hard binary masking and discrete token assignments, which hinder the revision of early decisions. |
| Approach: | They propose a diffusion-based language modeling approach that replaces hard binary masks with evolving soft token distributions. |
| Outcome: | The proposed approach outperforms existing DLMs on multiple benchmarks. |
Focus-dLLM: Accelerating Long-Context Diffusion LLM Inference via Confidence-Guided Context Focusing (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for estimating attention importance for tokens are ineffective . dLLMs require bidirectional attention, which limits inference efficiency . |
| Approach: | They propose a training-free attention sparsification framework for efficient long-context inference . they propose 'sink-aware pruning strategy' to accurately estimate and remove redundant computation . |
| Outcome: | The proposed approach offers 29 lossless speedup under 32K context length. |
Diffusion with Truncated Blocks: Fast and High-Quality Text Generation using Truncated Block Generation (2026.findings-acl)
Copied to clipboard
| Challenge: | Diffusion-based Large Language Models (dLLMs) generate text by iteratively denoising masked sequences. |
| Approach: | They propose a method that iteratively denoises masked sequences to reduce the model's attention dilution by token-level noise while models employing sequence-level noising exhibit a reduced effect. |
| Outcome: | The proposed method improves the performance and efficiency of Diffusion-based large language models by iterating on masked sequences. |
Revealing the Attention Floating Mechanism in Masked Diffusion Models (2026.findings-acl)
Copied to clipboard
Xin Dai, Pengcheng Huang, Zhenghao Liu, Shuo Wang, Yukun Yan, Chaojun Xiao, Yu Gu, Ge Yu, Maosong Sun
| Challenge: | Masked diffusion models (MDMs) leverage bidirectional attention and a denoising process. |
| Approach: | They investigate the attention behaviors of Masked diffusion models by revealing the phenomenon of Attention Floating. |
| Outcome: | The proposed model doubles the performance of autoregressive models in knowledge-intensive tasks. |
Taking a Deep Breath: Enhancing Language Modeling of Large Language Models with Sentinel Tokens (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies have explored compression and accumulation methods to compress contexts, but these methods lose useful context information during the compression process, leading to performance degradation. |
| Approach: | They propose a method that allows LLMs to take a deep breath and insert a special token at the end of each chunk. |
| Outcome: | Experiments on language modeling and out-of-domain tasks validate the superiority of the proposed method. |
Evolving Sparsity: Leveraging Token Importance Dynamics for Efficient LLM Decoding with Sparse Attention (2026.acl-long)
Copied to clipboard
| Challenge: | Efficient long-context inference remains a major challenge for large language models (LLMs), as the cost of attention computation during auto-regressive decoding grows linearly with the context length. |
| Approach: | They propose to model token importance as a dynamic process that evolves over decoding steps and propagates through model layers. |
| Outcome: | The proposed method outperforms baseline sparse attention methods and achieves speedups of up to 5.36 for attention latency and 2.33 for end-to-end decoding. |
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. |
TokenPenalty: Alleviating Attention Sinks and Positional Decay in LVLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Multimodal large language models (MLLMs) often hallucinate due to two relevant phenomena: massive activation phenomenon and positional information decay. |
| Approach: | They propose a token-level intervention strategy that dynamically suppresses irrelevant visual tokens while preserving key contextual signals. |
| Outcome: | Experiments show that TokenTruth significantly improves factual consistency across MLLMs on standard image understanding benchmarks. |
Mind Your Special Tokens! On the Importance of Dedicated Sequence-End Tokens in Vision-Language Embedding Models (2026.eacl-short)
Copied to clipboard
| Challenge: | Large Vision-Language Models (LVLMs) are highly sensitive to end-of-input artifacts in fine-tuning and inference data, e.g., whether input sequences end with punctuation or newline characters. |
| Approach: | They propose to convert generative LVLMs into vision-language encoders via contrastive learning objectives and use supervised contrastive objectives to train them. |
| Outcome: | The proposed approach improves visual and text representations and improves retrieval and (semantic) similarity tasks. |