A collaborative paper from Alibaba and Tsinghua University, titled "The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models," has been recognized with an Outstanding Paper award at ICML 2026. The research challenges the prevailing notion that diffusion large language models (dLLMs) benefit from arbitrary token generation order. Instead, the study demonstrates that this flexibility can hinder reasoning capabilities, particularly in tasks requiring logical deduction, by causing "entropy degradation." The researchers propose a simplified approach called "JustGRPO," which enforces a left-to-right generation order during reinforcement learning, leading to improved reasoning performance and maintaining inference speed. AI
IMPACT Challenges a core assumption in dLLM architecture, potentially simplifying future model development and improving reasoning performance.
RANK_REASON Academic paper recognized with a top award at a major AI conference.
- Alibaba
- Gemini Diffusion
- GPT
- GRPO
- GSM8K
- HumanEval
- JustGRPO
- Qwen
- Renmin University of China
- The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models
- Tsinghua University
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
