arXiv:2512.02892v1 Announce Type: new 
Abstract: Diffusion large language models (dLLMs) offer a promising alternative to autoregressive models, but their practical utility is severely hampered by slow, iterative sampling. We present SchED, a training-free, model-agnostic early-exit algorithm that aggregates full-span logit margins and halts decoding once a smooth, progress-dependent confidence threshold is met. We evaluated SchED on two dLLM families (Dream and LLaDA), in base and instruction-tuned variants across ten benchmarks spanning downstream tasks including multiple-choice question answering (MCQ), math, long-form QA/summarization, and translation. SchED delivers large, stable accelerations: on instruction-tuned models, it achieves $3.8$-$4.0\times$ speedups while retaining $99.8$-$100\%$ of the baseline score on average. On base models, SchED yields consistent speedup gains with $99.1$-$100\%$ performance retention, with up to $2.34\times$ under more aggressive settings. Using a conservative speed metric that heavily penalizes quality loss (QPS, $\gamma{=}4$), we show that SchED is robust and clearly outperforms prior confidence-based early-exit methods, which break down on long-form generation. An entropy analysis of the model's token predictions reveals that instruction tuning speeds up the decay of predictive entropy. By turning genuine confidence stabilization into computational savings, SchED makes dLLM decoding substantially more efficient.

تم تقديم خوارزمية جديدة تُدعى SchED لتحسين كفاءة نماذج اللغة الانتشارية (dLLMs) من خلال السماح بالخروج المبكر أثناء عملية فك التشفير. هذه الطريقة المستقلة عن النموذج تجمع بين هوامش اللوجيت وتوقف فك التشفير عند الوصول إلى عتبة ثقة، مما يسرع الأداء بشكل كبير مع الحفاظ على دقة عالية عبر مهام متنوعة.

Se ha introducido un nuevo algoritmo llamado SchED para mejorar la eficiencia de los modelos de lenguaje de difusión (dLLMs) al permitir salidas anticipadas durante el proceso de decodificación. Este enfoque agnóstico al modelo agrega márgenes de logit y detiene la decodificación cuando se alcanza un umbral de confianza, acelerando significativamente el rendimiento mientras mantiene una alta precisión en diversas tareas.

Un nouvel algorithme nommé SchED a été introduit pour améliorer l'efficacité des modèles de langage à diffusion (dLLMs) en permettant des sorties anticipées lors du processus de décodage. Cette approche, indépendante du modèle, agrège les marges de logit et interrompt le décodage lorsque un seuil de confiance est atteint, accélérant considérablement les performances tout en maintenant une grande précision sur diverses tâches.

A new algorithm named SchED has been introduced to enhance the efficiency of diffusion large language models (dLLMs) by allowing early exits during the decoding process. This model-agnostic approach aggregates logit margins and stops decoding when a confidence threshold is reached, significantly speeding up performance while maintaining high accuracy across various tasks.

Fast-Decoding Diffusion Language Models via Progress-Aware Confidence Schedules

Was this article worth reading? Share it

LucidQuery AI

Langfuse

FastML

Langtail

CodeSpaced

Dubsmart LLC

Ready to build your own newsroom?