Test-Time Scaling in Diffusion LLMs via Hidden Semi-Autoregressive Experts
This addresses a critical bottleneck in deploying dLLMs for reasoning tasks, offering a novel inference paradigm with broad applicability across benchmarks.
The paper tackles the problem of suboptimal inference in diffusion-based large language models (dLLMs) by showing that fixed inference schedules fail to leverage latent semi-autoregressive experts, and introduces HEX, a training-free method that ensembles diverse block schedules to boost performance, achieving accuracy improvements such as from 24.72% to 88.10% on GSM8K.
Diffusion-based large language models (dLLMs) are trained flexibly to model extreme dependence in the data distribution; however, how to best utilize this information at inference time remains an open problem. In this work, we uncover an interesting property of these models: dLLMs trained on textual data implicitly learn a mixture of semi-autoregressive experts, where different generation orders reveal different specialized behaviors. We show that committing to any single, fixed inference time schedule, a common practice, collapses performance by failing to leverage this latent ensemble. To address this, we introduce HEX (Hidden semiautoregressive EXperts for test-time scaling), a training-free inference method that ensembles across heterogeneous block schedules. By doing a majority vote over diverse block-sized generation paths, HEX robustly avoids failure modes associated with any single fixed schedule. On reasoning benchmarks such as GSM8K, it boosts accuracy by up to 3.56X (from 24.72% to 88.10%), outperforming top-K margin inference and specialized fine-tuned methods like GRPO, without additional training. HEX even yields significant gains on MATH benchmark from 16.40% to 40.00%, scientific reasoning on ARC-C from 54.18% to 87.80%, and TruthfulQA from 28.36% to 57.46%. Our results establish a new paradigm for test-time scaling in diffusion-based LLMs (dLLMs), revealing that the sequence in which masking is performed plays a critical role in determining performance during inference.