Diffusion Language Model Knows the Answer Before It Decodes
Pengxiang Li, Yefan Zhou, Dilxat Muhtar, Lu Yin, Shilin Yan, Li Shen, Yi Liang, Soroush Vosoughi, Shiwei Liu
Abstract
Diffusion language models (DLMs) have recently emerged as an alternative to autoregressive approaches, offering parallel sequence generation and flexible token orders. However, their inference remains slower than that of autoregressive models, primarily due to the cost of bidirectional attention and the large number of refinement steps required for high-quality outputs. In this work, we highlight and leverage an overlooked property of DLMs—**early answer convergence**: in many cases, the correct answer can be internally identified by half steps before the final decoding step, under both semi-autoregressive and random remasking schedules. For example, on GSM8K and MMLU, up to 97\% and 99\% of instances, respectively, can be decoded correctly using only half of the refinement steps. Building on this observation, we introduce **Prophet**, a training-free fast decoding paradigm that enables **early commit decoding**. Specifically, Prophet dynamically decides whether to continue refinement or to go "all-in" (i.e. decode all remaining tokens in one step), using the confidence gap between the top-2 prediction candidates as the criterion. It integrates seamlessly into existing DLM implementations, incurs negligible overhead, and requires no additional training. Empirical evaluations on LLaDA-8B and Dream-7B across multiple tasks show that Prophet reduces the number of decoding steps by up to 3.4$\times$ while preserving high generation quality, and yields additional speedups when combined with existing acceleration methods. These results recast DLM decoding as a problem of *when to stop sampling*, and demonstrate that early answer convergence provides a simple yet powerful mechanism for accelerating DLMs on reasoning, code, and planning tasks with identifiable answer regions. Our code is available at \url{https://github.com/pixeli99/Prophet}.
Prophet identifies early answer convergence in diffusion language models to accelerate decoding by 3.4x on reasoning tasks.
- Identifies early answer convergence property of diffusion language models for half-step decoding
- Introduces Prophet, training-free fast decoding paradigm using confidence gap criterion for early commit decisions
- Reduces decoding steps by 3.4x while preserving generation quality without additional training
- Demonstrates compatibility with distillation and cache-based acceleration methods
- Diffusion language models
- Early stopping criteria
- Confidence scoring
- GSM8K
- MMLU
- HumanEval
- Sudoku
Designed for tasks with identifiable answer regions; less suitable for open-ended generation
from the paperMore conservative speedups on complex tasks like code generation compared to short-answer tasks
from the paper
Develop learnable judge-based termination criteria for improved robustness in tasks where confidence doesn't correlate with correctness
from the paperExplore system-level optimizations with KV Cache frameworks for immediate inference termination
from the paper
Author keywords
- diffusion language model
- discrete
Related orals
Benchmarking Empirical Privacy Protection for Adaptations of Large Language Models
Benchmarks practical privacy risks in differential privacy-adapted LLMs, revealing distribution shifts and model choice impact effectiveness.
Half-order Fine-Tuning for Diffusion Model: A Recursive Likelihood Ratio Optimizer
Proposes Recursive Likelihood Ratio optimizer for efficient fine-tuning of diffusion models with lower variance gradient estimation.
Invisible Safety Threat: Malicious Finetuning for LLM via Steganography
Demonstrates LLMs can be finetuned to generate harmful steganographically-hidden outputs while appearing benign to safety systems.
Reducing Belief Deviation in Reinforcement Learning for Active Reasoning of LLM Agents
Proposes T3 algorithm to detect belief deviation in LLM agents and truncate trajectories for improved reinforcement learning in active reasoning tasks.
RefineStat: Efficient Exploration for Probabilistic Program Synthesis
RefineStat enforces semantic constraints and applies diagnostic-aware refinement for synthesizing valid probabilistic programs from smaller language models.