Research map / Large language models / Reasoning and chain of thought

TEST-TIME SCALING IN DIFFUSION LLMS VIA HIDDEN SEMI-AUTOREGRESSIVE EXPERTS

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HEX, a training-free test-time method that ensembles diffusion LLM decoding over multiple semi-autoregressive block schedules and aggregates answers by majority vote, matching or beating GRPO-tuned d1 on reasoning benchmarks.

Problem

Confidence-based (top-k margin) unmasking order in diffusion LLMs can collapse into degenerate outputs, and the best decoding order is unclear; the authors ask how to use the learned masking flexibility at inference time.

Approach

View a masked diffusion LLM as an implicit mixture of semi-autoregressive experts indexed by block size; marginalize over block schedules (block sizes 8-128, five samples each) and take a majority vote over final answers, with no retraining.

Key results

• HEX reaches 88.10% on GSM8K vs 79.80% for d1 (GRPO) and 24.72% for top-k margin decoding, with 25 samples at output length 256.
• On MATH HEX gets 40.00% vs 37.20% for d1 (GRPO); on ARC-C 87.80% vs 82.68%.
• Accuracy rises monotonically with more voting samples (1 to 6 seeds) while tie rate falls; HEX decoding costs about 5.26x the latency of random decoding.

Surprising findings

• Top-k margin (confidence-based) unmasking underperforms random unmasking on GSM8K (24.72% vs 50.87%) because it collapses into generating end-of-text tokens from the back of the sequence.
• Best-of-N with majority vote over non-semi-AR samples (44.35%) is far worse than semi-AR HEX (82.18%).

Limitations

• Requires substantially more test-time compute (about 5x latency) for majority voting.
• Evaluated mainly on reasoning tasks; not on creative or open-ended generation.
• No theoretical analysis of why the mixture-of-experts view yields gains.

Experimental setup

Datasets: GSM8K MATH ARC-C TruthfulQA

Benchmarks: GSM8K MATH ARC-C TruthfulQA

Metrics: accuracy tie rate wall-clock latency

Baselines: Random unmasking top-k margin decoding top-k confidence decoding d1 (GRPO) NLL candidate selection

Builds on: LLaDA d1 masked diffusion language models self-consistency

Closest papers by meaning

• Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · similarity 0.92)
• Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.90)
• Large Language Models to Diffusion Finetuning (ICML 2025 · similarity 0.90)
• From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
• UnMaskFork: Test-Time Scaling for Masked Diffusion via Deterministic Action Branching (ICML 2026 · similarity 0.90)
• Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.89)