Researchers have developed Cord, a new decoding policy for masked diffusion language models (dLLMs) that improves generation quality by intelligently deciding when and how to reveal masked tokens. Unlike standard methods that use a uniform reveal rule, Cord learns to adapt its strategy based on confidence, generation direction, proximity to context, and infilling needs. This approach significantly boosts task success rates across various benchmarks, including reasoning, coding, and constraint satisfaction, while adding minimal computational overhead. AI
IMPACT This research could lead to more capable and efficient language models for complex tasks like reasoning and code generation.
RANK_REASON The cluster contains an academic paper detailing a new method for language models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
