Research map / Image generation / Fast diffusion sampling

Latent Refinement Decoding: Enhancing Diffusion Language Models by Refining Belief States

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Masked diffusion language model decoding

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 Latent Refinement Decoding (LRD), a two-stage training-free decoding scheme for diffusion LLMs that mixes [MASK] embeddings with entropy-weighted top-p token embeddings, then uses KL-based soft-to-hard scheduling and early stopping.

Problem

Parallel decoding in diffusion language models discards predictive distributions of non-finalized tokens (hard masking), causing information loss, premature commitment, and fixed iteration counts that waste compute.

Approach

Keep uncertain masked positions as soft mixtures of the [MASK] embedding and the top-p expected token embedding (weight set by normalized entropy), first refine these soft beliefs until KL-stable, then progressively commit confident tokens while keeping the rest soft, with KL-based early stopping.

Key results

• HumanEval +6.3 pass@1 for Dream-Base-7B at length 256 over baseline (50.6 to 56.9), with 1.2x speed.
• Dream-Ins-7B at length 1024 reaches 10.6x speedup on MBPP and 9.3x on HumanEval, with HumanEval +5.0 accuracy.
• GSM8K +2.9 (Dream-Base-7B, 256) and MATH500 +3.8 (Dream-Ins-7B, 1024) gains reported.

Surprising findings

• Removing early stopping drops speed far below baseline (e.g., 0.8x) while token counts barely change, so early stopping is the main speed driver.
• Mixing a very small fraction of tokens (top-p up to about 0.2) already yields most of the accuracy gain, and full mixing (r_f=1) collapses the model.

Limitations

• Stability analysis is only local (single self-attention layer) and no global convergence guarantee is provided.
• Speedups vary widely across settings; some configurations (e.g., Dream-Ins MATH500 at 1.1-1.2x) show modest gains, and speed is measured as relative runtime.
• Results rely on hyperparameters (r_f, top-p, thresholds) tuned per setting; evaluation limited to LLaDA and Dream 7-8B models and four benchmarks.

Experimental setup

Datasets: GSM8K MATH500 MBPP HumanEval

Benchmarks: HumanEval MBPP GSM8K MATH500

Metrics: accuracy pass@1 relative speed effective token number

Baselines: LLaDA hard-masking decoding Dream hard-masking decoding

Builds on: LLaDA Dream Fast-dLLM

Closest papers by meaning

• Beyond Confidence: Adaptive and Coherent Decoding for Diffusion Language Models (ICML 2026 · similarity 0.91)
• Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.91)
• Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
• From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
• Confidence-Based Decoding is Provably Efficient for Diffusion Language Models (NeurIPS 2026 · similarity 0.89)
• Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.89)