dQwen3.5: Hybrid-Attention Diffusion Language Models

TL;DR

dQwen3.5 uses hybrid attention to achieve training loss in half the tokens compared to full-attention models.

cs.CL 🔴 Advanced 2026-09-18 13 views
Anton Xue Litu Rout Aditya Akella Adam Klivans Sujay Sanghavi Sanjay Shakkottai
hybrid attention diffusion language model RNN bidirectionalization parallel decoding

Key Findings

Methodology

The study employs a hybrid attention mechanism combining attention and RNN layers, with RNN layers remaining causal and attention layers bidirectionalized. This architecture was adapted at scales from 0.8B to 9B, significantly enhancing training efficiency.

Key Results

  • dQwen3.5 achieves the same training loss with half the tokens compared to full-attention models, excelling in HumanEval tests.
  • Across scales, dQwen3.5's local AR-ness is about 0.64, and global AR-ness is approximately 0.88-0.97.
  • After 50B tokens, dQwen3.5-2B outperforms CoDA on 6/7 benchmarks.

Significance

This research demonstrates the potential of hybrid attention architectures in diffusion language models, particularly in training efficiency and decoding flexibility. It offers a new perspective for future language model research, especially in resource-constrained scenarios.

Technical Contribution

Technical contributions include proposing a hybrid architecture combining RNN and attention layers, providing an efficient AR-to-DLM adaptation method, and demonstrating adaptation performance across scales.

Novelty

This work is the first to successfully implement efficient AR-to-DLM adaptation in a hybrid architecture, particularly with causal RNN layers, overcoming the limitations of traditional full-attention models.

Limitations

  • The hybrid architecture underperforms on mathematical tasks, possibly due to training data bias.
  • Extended adaptation time may degrade performance, especially in large-scale models.

Future Work

Future research could explore optimizing hybrid architecture performance on mathematical tasks and adaptation effects on different datasets.

AI Executive Summary

dQwen3.5 addresses the inefficiency of traditional full-attention models in training by employing a hybrid attention mechanism. This method combines attention and RNN layers, with attention layers bidirectionalized and RNN layers remaining causal. The architecture was adapted at scales from 0.8B to 9B, significantly enhancing training efficiency, particularly excelling in HumanEval tests.

Experimental results show that dQwen3.5 achieves the same training loss with half the tokens compared to full-attention models. Although the causal nature of RNN layers might limit decoding order flexibility, dQwen3.5 still demonstrates strong parallel decoding capabilities.

However, the hybrid architecture underperforms on mathematical tasks, possibly due to training data bias. Future research could explore optimizing hybrid architecture performance on mathematical tasks and adaptation effects on different datasets.

Deep Analysis

Background

In recent years, diffusion language models (DLMs) have gained attention for their advantages in parallel and any-order generation. Traditional DLMs typically start from full-attention transformers, but this approach is costly. Hybrid architectures combining attention and RNN layers offer a new perspective.

Core Problem

Adapting pretrained autoregressive models to diffusion language models is challenging, especially with hybrid architectures. The causal nature of RNN layers makes bidirectionalization difficult.

Innovation

dQwen3.5 successfully implements efficient AR-to-DLM adaptation by bidirectionalizing attention layers while keeping RNN layers causal. This method was validated at different scales, showing significant improvements in training efficiency.

Methodology

  • �� Use hybrid architecture combining attention and RNN layers
  • �� Bidirectionalize attention layers, keep RNN layers causal
  • �� Adapt at scales from 0.8B to 9B
  • �� Train with a fixed data mixture ratio

Experiments

Experiments were conducted on multiple benchmarks like HumanEval and MMLU, comparing different scales of dQwen3.5 models. The model's training efficiency and decoding capabilities were evaluated.

Results

dQwen3.5 achieves the same training loss with half the tokens compared to full-attention models. Its local AR-ness is about 0.64, and global AR-ness is approximately 0.88-0.97.

Applications

dQwen3.5 is suitable for natural language processing tasks requiring efficient training and decoding, especially in resource-constrained environments.

Limitations & Outlook

The hybrid architecture underperforms on mathematical tasks, and extended adaptation time may degrade performance. Future research could explore optimizing these aspects.

Plain Language Accessible to non-experts

Imagine a factory where traditional full-attention models are like an assembly line that needs to process items sequentially from start to finish. In contrast, dQwen3.5 is like a flexible workshop where multiple stations can work simultaneously. While some stations (RNN layers) still need to operate in order, others (attention layers) can work freely at different positions. This flexibility significantly boosts dQwen3.5's production efficiency, especially when quick responses are needed.

ELI14 Explained like you're 14

Imagine you're playing a massive multiplayer online game. Traditional full-attention models are like a game where you have to complete quests in order. But dQwen3.5 is like an open-world game where you can tackle multiple quests at once! Some quests (RNN layers) still need to be done in order, but others (attention layers) can be done anywhere. This flexibility helps you level up and complete tasks faster in the game!

Glossary

Hybrid Attention

An architecture combining attention and RNN layers, with attention layers bidirectionalized and RNN layers remaining causal.

Used in dQwen3.5 to enhance training efficiency and decoding capability.

Diffusion Language Model

Generates text by iteratively decoding (unmasking) masked positions rather than strictly left-to-right.

dQwen3.5 uses this model for efficient training and decoding.

Autoregressive Model

A generative model that relies on the output of the previous time step.

dQwen3.5 adapts autoregressive models for efficient diffusion language model training.

Local AR-ness

Measures left-to-right ordering among adjacent positions.

Used to evaluate dQwen3.5's decoding flexibility.

Global AR-ness

Measures left-to-right ordering across all position pairs.

Used to evaluate dQwen3.5's decoding flexibility.

Open Questions Unanswered questions from this research

  • 1 How to optimize hybrid architecture performance on mathematical tasks? Current training data bias may lead to underperformance.
  • 2 What are the adaptation effects on different datasets? Further experimental validation is needed.

Applications

Immediate Applications

Natural Language Processing

dQwen3.5 is suitable for NLP tasks requiring efficient training and decoding, especially in resource-constrained environments.

Long-term Vision

Intelligent Assistants

By enhancing decoding efficiency and flexibility, dQwen3.5 can be applied in intelligent assistants for more natural human-computer interaction.

Abstract

Adapting a pretrained autoregressive (AR) model is a cost-efficient route to a diffusion language model (DLM). While nearly all such adaptations start from a full-attention transformer, AR modeling has shifted toward hybrid architectures that interleave attention and RNN layers. This creates an obstacle for adaptation: unlike attention, RNNs are structurally causal and nontrivial to bidirectionalize. Despite this mismatch, we investigate whether such backbones can become effective DLMs by adapting Qwen3.5 at 0.8B, 2B, 4B, and 9B scales, yielding the dQwen3.5 family. We find that hybrid backbones can be efficient starting points for adaptation: against a full-attention control, the hybrid reaches a given training loss in about half the tokens. Across scales, dQwen3.5 resembles full-attention DLMs in any-order decoding behavior and performs strongly under parallel decoding.

cs.CL cs.LG