Test-Time Scaling in Diffusion LLMs via Hidden Semi-Autoregressive Experts

• URL: http://arxiv.org/abs/2510.05040v1
• Date: Mon, 06 Oct 2025 17:16:41 GMT
• Title: Test-Time Scaling in Diffusion LLMs via Hidden Semi-Autoregressive Experts
• Abstract summary: Diffusion-based large language models implicitly learn a mixture of semi-autoregressive experts.<n>We show that committing to any single, fixed inference time schedule, collapses performance by failing to leverage this latent ensemble.<n>We introduce HEX, a training-free inference method that ensembles across heterogeneous block schedules.
• Score: 32.164728234520915
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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.

Related papers

• TimeSRL: Generalizable Time-Series Behavioral Modeling via Semantic RL-Tuned LLMs -- A Case Study in Mental Health [10.41600839007471]
  TimeSRL is a framework that routes predictions through an explicit semantic bottleneck.<n>It achieves state-of-the-art performance on a benchmark designed to stress-test cross-cohort generalization.<n>TimeSRL significantly outperforms prior methods in cross-benchmark transfer across different sensing pipelines.
  arXiv  Detail & Related papers  (2026-05-20T15:25:46Z)
• Learning-Zone Energy: Online Data Selection for Efficient RL Post-Training [28.00739954235118]
  We propose a theoretically grounded, fully online data selection framework that concentrates on the model's active learning frontier.<n>A forward pruner with replay further reduces wall-clock time cost by skipping rollout generation for persistently solved prompts.<n>Our method retains only 40% of the training data per step yet matches or surpasses full-data baselines.
  arXiv  Detail & Related papers  (2026-05-16T14:01:12Z)
• eXplaining to Learn (eX2L): Regularization Using Contrastive Visual Explanation Pairs for Distribution Shifts [0.0]
  We propose a framework that decorrelates confounding features from latent representations during training.<n>eX2L achieves an average accuracy (AA) of 82.24% +/- 3.87% and a worst-group accuracy (WGA) of 66.31% +/- 8.73% on the Spawrious Many-to-Many Hard Challenge benchmark.
  arXiv  Detail & Related papers  (2026-05-07T14:46:08Z)
• Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis [68.28714988482703]
  Process Reward Models (PRMs) have achieved remarkable success in augmenting the reasoning capabilities of Large Language Models (LLMs)<n>We present a study revealing that general-domain PRMs struggle to supervise data analysis agents.<n>We introduce DataPRM, a novel environment-aware generative process reward model.
  arXiv  Detail & Related papers  (2026-04-27T09:00:30Z)
• Online Reasoning Calibration: Test-Time Training Enables Generalizable Conformal LLM Reasoning [18.69627681731888]
  Online Reasoning calibration (ORCA) is a framework for calibrating the sampling process that draws upon conformal prediction and test-time training.<n>At risk level $=0.1$, ORCA improves Qwen2.5-32B efficiency on in-distribution tasks with savings up to 47.5% with supervised labels and 40.7% with self-consistency labels.
  arXiv  Detail & Related papers  (2026-04-01T17:21:50Z)
• Learning Hierarchical Procedural Memory for LLM Agents through Bayesian Selection and Contrastive Refinement [23.31711942240935]
  We present MACLA, a framework that decouples reasoning from learning by maintaining a frozen large language model while performing all adaptation in an external hierarchical procedural memory.<n> MACLA extracts reusable procedures from trajectories, tracks reliability via Bayesian posteriors, selects actions through expected-utility scoring and refines procedures by contrasting successes and failures.<n>Across four benchmarks (ALFWorld, WebShop, TravelPlanner, InterCode), MACLA achieves 78.1 percent average performance, outperforming all baselines.
  arXiv  Detail & Related papers  (2025-12-22T01:56:28Z)
• Scaling Behavior of Discrete Diffusion Language Models [74.72926629897636]
  We study the scaling behavior of discrete diffusion language models (DLMs) on different noise types.<n>Our experiments reveal that the scaling behavior of DLMs strongly depends on the noise type and is considerably different from ALMs.<n>We scale our uniform diffusion model up to 10B parameters trained for $1022$ FLOPs, confirming the predicted scaling behavior and making it the largest publicly known uniform diffusion model to date.
  arXiv  Detail & Related papers  (2025-12-11T17:54:10Z)
• EDIT: Early Diffusion Inference Termination for dLLMs Based on Dynamics of Training Gradients [6.736735746633275]
  Diffusion-based large language models (dLLMs) refine token generations through iterative denoising, but answers often stabilize before all steps complete.<n>We propose EDIT, an inference-time criterion that adaptively stops denoising once sufficient reasoning stability relative to training-time reasoning is detected.
  arXiv  Detail & Related papers  (2025-11-29T23:47:47Z)
• Bias Testing and Mitigation in Black Box LLMs using Metamorphic Relations [12.095552079739933]
  We propose a unified framework for both systematic bias evaluation and targeted mitigation.<n>Our approach introduces six novel Metamorphic Relations (MRs)<n>MRs transform direct bias-inducing inputs into semantically equivalent yet adversarially challenging variants.
  arXiv  Detail & Related papers  (2025-11-29T16:56:38Z)
• MoDES: Accelerating Mixture-of-Experts Multimodal Large Language Models via Dynamic Expert Skipping [52.02659589971978]
  We propose MoDES, the first training-free framework that adaptively skips experts to enable efficient and accurate MoE MLLM inference.<n>MoDES significantly enhances inference speed, improving the prefilling time by 2.16$times$ and the decoding time by 1.26$times$.
  arXiv  Detail & Related papers  (2025-11-19T18:48:27Z)
• Relative Scaling Laws for LLMs [91.73497548097775]
  Scaling laws describe how language models improve with additional data, parameters, and compute.<n>We introduce relative scaling laws, which track how performance gaps between test distributions evolve with scale.<n>These results show that although scaling improves overall performance, it is not a universal equalizer.
  arXiv  Detail & Related papers  (2025-10-28T16:55:22Z)
• MDPO: Overcoming the Training-Inference Divide of Masked Diffusion Language Models [28.79185891706149]
  Diffusion language models suffer from a key discrepancy between training and inference.<n>We propose a novel Masked Diffusion Policy Optimization (MDPO) to exploit the Markov property diffusion.<n>Our findings establish great potential for investigating the discrepancy between pre-training and inference of MDLMs.
  arXiv  Detail & Related papers  (2025-08-18T17:58:13Z)
• LookAlike: Consistent Distractor Generation in Math MCQs [42.19039301965107]
  We propose LookAlike, a method that improves error-distractor consistency via preference optimization.<n>Our two main innovations are: (a) mining synthetic preference pairs from model inconsistencies, and (b) alternating supervised fine-tuning.<n>LookAlike achieves 51.6% accuracy in distractor generation and 57.2% in error generation under LLM-as-a-judge evaluation.
  arXiv  Detail & Related papers  (2025-05-03T19:18:06Z)
• Adaptive Distraction: Probing LLM Contextual Robustness with Automated Tree Search [76.54475437069395]
  Large Language Models (LLMs) often struggle to maintain their original performance when faced with semantically coherent but task-irrelevant contextual information.<n>We propose a dynamic distraction generation framework based on tree search, where the generation process is guided by model behavior.
  arXiv  Detail & Related papers  (2025-02-03T18:43:36Z)
• MMLU-CF: A Contamination-free Multi-task Language Understanding Benchmark [57.999567012489706]
  We propose a contamination-free and more challenging benchmark called MMLU-CF.<n>This benchmark reassesses LLMs' understanding of world knowledge by averting both unintentional and malicious data leakage.<n>Our evaluation of mainstream LLMs reveals that the powerful GPT-4o achieves merely a 5-shot score of 73.4% and a 0-shot score of 71.9% on the test set.
  arXiv  Detail & Related papers  (2024-12-19T18:58:04Z)
• FactorLLM: Factorizing Knowledge via Mixture of Experts for Large Language Models [50.331708897857574]
  We introduce FactorLLM, a novel approach that decomposes well-trained dense FFNs into sparse sub-networks without requiring any further modifications. FactorLLM achieves comparable performance to the source model securing up to 85% model performance while obtaining over a 30% increase in inference speed.
  arXiv  Detail & Related papers  (2024-08-15T16:45:16Z)
• Uncertainty Aware Learning for Language Model Alignment [97.36361196793929]
  We propose uncertainty-aware learning (UAL) to improve the model alignment of different task scenarios. We implement UAL in a simple fashion -- adaptively setting the label smoothing value of training according to the uncertainty of individual samples. Experiments on widely used benchmarks demonstrate that our UAL significantly and consistently outperforms standard supervised fine-tuning.
  arXiv  Detail & Related papers  (2024-06-07T11:37:45Z)
• Cycles of Thought: Measuring LLM Confidence through Stable Explanations [53.15438489398938]
  Large language models (LLMs) can reach and even surpass human-level accuracy on a variety of benchmarks, but their overconfidence in incorrect responses is still a well-documented failure mode. We propose a framework for measuring an LLM's uncertainty with respect to the distribution of generated explanations for an answer.
  arXiv  Detail & Related papers  (2024-06-05T16:35:30Z)
• Distributionally Robust Multiclass Classification and Applications in Deep Image Classifiers [3.179831861897336]
  We develop a Distributionally Robust Optimization (DRO) formulation for Multiclass Logistic Regression (MLR) We demonstrate reductions in test error rate by up to 83.5% and loss by up to 91.3% compared with baseline methods, by adopting a novel random training method.
  arXiv  Detail & Related papers  (2021-09-27T02:58:19Z)
• Examining and Combating Spurious Features under Distribution Shift [94.31956965507085]
  We define and analyze robust and spurious representations using the information-theoretic concept of minimal sufficient statistics. We prove that even when there is only bias of the input distribution, models can still pick up spurious features from their training data. Inspired by our analysis, we demonstrate that group DRO can fail when groups do not directly account for various spurious correlations.
  arXiv  Detail & Related papers  (2021-06-14T05:39:09Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.