Research map / Image generation / Fast diffusion sampling

Self-Rewarding Sequential Monte Carlo for Masked Diffusion Language Models

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Masked diffusion language model decoding · 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 self-rewarding SMC, an inference-time particle-resampling method for masked diffusion language models that uses trajectory-level token confidence as the importance weight, improving MDLM, BD3-LM, LLaDA-1.5 and Dream-7B without extra training or reward models.

Problem

Confidence-based low-confidence remasking in masked diffusion language models makes greedy, myopic decoding that collapses path diversity and yields noise-sensitive outputs; existing inference-time scaling relies on task-specific external rewards.

Approach

Run N interacting diffusion particles, weight each by its accumulated token confidence (shown to equal the SMC incremental weight under a Feynman-Kac model with the diffusion kernel as proposal), and resample adaptively by ESS, so no external reward or training is needed.

Key results

• On GSM8K/MATH/HumanEval/MBPP averages at length 256, SR-SMC raises LLaDA-1.5 from 49.3 to 52.1 and Dream-7B from 51.9 to 56.4 (N=4).
• For MDLM at L=1024, generative perplexity drops from 46.8 to 25.8 with SR-SMC; BD3-LMs with block size 4 reaches 16.1 (vs 25.7).
• Dream-7B baseline collapses at low Gumbel temperatures (e.g., 1.80 on MBPP at tau=0.1) while SR-SMC stays around 43-49.

Surprising findings

• Dream-7B greedy decoding collapses to near-zero MBPP accuracy at tau=0.1 due to repetition, while SR-SMC remains stable.
• Overtake events, where the non-dominant particle at block start becomes dominant at block end, occur in about 24-31% of blocks.

Limitations

• Increases inference compute roughly linearly with the number of particles.
• Trajectory confidence relies only on model likelihood and does not optimize for correctness or human preference.
• Gains on zero-shot GSM8K are not uniform (e.g., Dream-7B at L=512 unchanged).

Experimental setup

Datasets: OpenWebText GSM8K MATH HumanEval MBPP

Benchmarks: GSM8K MATH HumanEval MBPP OpenWebText generative perplexity

Metrics: accuracy generative perplexity entropy NFEs

Baselines: MDLM BD3-LMs SEDD SSD-LM LLaDA-1.5 Dream-7B ReMDM

Builds on: sequential Monte Carlo masked diffusion models low-confidence remasking Feynman-Kac model particle Gibbs sampling

Closest papers by meaning

• Lookahead Path Likelihood Optimization for Diffusion LLMs (ICML 2026 · similarity 0.90)
• KLASS: KL-Guided Fast Inference in Masked Diffusion Models (NeurIPS 2025 · similarity 0.90)
• UnMaskFork: Test-Time Scaling for Masked Diffusion via Deterministic Action Branching (ICML 2026 · similarity 0.89)
• Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (NeurIPS 2025 · similarity 0.89)
• Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · similarity 0.89)
• Accelerated Diffusion Models via Speculative Sampling (ICML 2025 · similarity 0.89)