Papers with bidirectional attention
MoCa: Modality-aware Continual Pre-training Makes Better Bidirectional Multimodal Embeddings (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to embed multimodal models face limitations such as suboptimal causal attention in VLMs and limited diversity in training objectives and data. |
| Approach: | They propose a framework for transforming pre-trained VLMs into bidirectional multimodal embedding models. |
| Outcome: | The proposed model improves performance across MMEB and ViDoRe-v2 benchmarks and exhibits strong scalability with both model size and training data on MMEF. |
Diffusion-Pretrained Dense and Contextual Embeddings (2026.acl-industry)
Copied to clipboard
| Challenge: | pplx-embed uses diffusion-based pretraining to capture bidirectional context within passages. |
| Approach: | They propose a family of multilingual embedding models that leverage bidirectional attention through diffusion-based pretraining to capture bidirectional context within passages. |
| Outcome: | The proposed models achieve competitive performance on the MTEB(Multilingual, v2), MTEF(Code), BERGEN, and ToolRet retrieval benchmarks while pplx-embed-context-v1 sets new records on the ConTEB benchmark. |
Microblog Hashtag Generation via Encoding Conversation Contexts (N19-1)
Copied to clipboard
| Challenge: | Automated hashtag annotation plays an important role in content understanding for microblog posts. |
| Approach: | They propose to annotate hashtags with a novel sequence generation framework via viewing the hashtag as a short sequence of words. |
| Outcome: | The proposed model outperforms existing models on two large-scale datasets . it can generate rare and even unseen hashtags, which is not possible with existing 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. |
On the Role of Bidirectionality in Language Model Pre-Training (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Prior work on language model pre-training explored different architectures and learning objectives, but differences in data, hyperparameters and evaluation make a principled comparison difficult. |
| Approach: | They propose a framework that generalizes prior approaches to pre-training language models by focusing on bidirectionality and controlling each of them separately. |
| Outcome: | The proposed framework generalizes prior approaches including fully unidirectional models like GPT, fully bidirectional models and hybrid models like CM3 and prefix LM. |
ControlSpeech: Towards Simultaneous and Independent Zero-shot Speaker Cloning and Zero-shot Language Style Control (2025.acl-long)
Copied to clipboard
Shengpeng Ji, Qian Chen, Wen Wang, Jialong Zuo, Minghui Fang, Ziyue Jiang, Hai Huang, Zehan Wang, Xize Cheng, Siqi Zheng, Zhou Zhao
| Challenge: | Prior zero-shot TTS models only mimic the speaker’s voice without further control and adjustment capabilities while prior controllable TTS systems cannot perform speaker-specific voice generation. |
| Approach: | They propose a style control module that captures codec representations corresponding to timbre, content, and style in a discrete decoupling codec space. |
| Outcome: | The proposed system can fully clone the speaker's voice and perform speech-specific adjustment and control functions. |
Bitune: Leveraging Bidirectional Attention to Improve Decoder-Only LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Decoder-only large language models rely on masked causal attention, which limits expressiveness by restricting information flow to one direction. |
| Approach: | They propose a method that incorporates bidirectional attention into prompt processing to enhance pretrained decoder-only LLMs. |
| Outcome: | The proposed method shows significant improvements in commonsense reasoning, arithmetic, and language understanding tasks. |
From AR to Diffusion: Efficiently Adapting Large Language Models with Strictly Causal and Elastic Horizons (2026.acl-long)
Copied to clipboard
| Challenge: | Autoregressive (AR) models rely on bidirectional attention, creating a structural mismatch with pre-trained Autoregression models. |
| Approach: | They propose a framework that efficiently adapts autoregressive (AR) models to the diffusion paradigm. |
| Outcome: | The proposed framework reduces training costs by orders of magnitude while maintaining state-of-the-art performance. |
Look Both Ways and No Sink: Converting LLMs into Text Encoders without Training (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for converting large language models into powerful text encoders require extensive training on large datasets. |
| Approach: | They propose a training-free approach that enables bidirectional attention and suppresses the attention sink phenomenon, resulting in superior performance. |
| Outcome: | The proposed approach enables bidirectional attention and suppresses the attention sink phenomenon, resulting in superior performance. |
Learning to Look at the Other Side: A Semantic Probing Study of Word Embeddings in LLMs with Enabled Bidirectional Attention (2025.acl-long)
Copied to clipboard
| Challenge: | Autoregressive Large Language Models (LLMs) demonstrate exceptional performance in language understanding and generation tasks, but their application in text embedding tasks has been relatively slow due to the constraints of the unidirectional attention mechanism. |
| Approach: | They propose to enable bidirectional attention in LLMs to improve their ability to represent subsequent context but weaken their utilization of preceding context. |
| Outcome: | The proposed model improves the LLMs’ ability to represent subsequent context but weakens their utilization of preceding context, while contrastive learning training can help to maintain both abilities. |
Frozen LLMs are Native Decoders for High-Norm Semantic Vectors (2026.acl-long)
Copied to clipboard
| Challenge: | Existing compression methods selectively prune tokens based on information-theoretic metrics, resulting in interpretability but risking the loss of fine-grained information. |
| Approach: | They propose a landmark-based compression framework for long contexts that captures global dependencies over landmark tokens. |
| Outcome: | The proposed framework outperforms soft compression baselines on four QA benchmarks. |
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. |