Research map / Image generation / Autoregressive and latent generation
LUGS: Latent-aware Guidance for Efficient Unmasking in Diffusion Large Language Models
Nuanqiao Shan, Kairong Han, Xinpeng Dong, Kun Kuang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add latent-aware guidance to beam search | Beam Search, LLaDA-8B-Instruct · HumanEval | Pass@1: 35.37 -> 42.68 | Improved |
| add latent-aware guidance to beam search | Beam Search, LLaDA-8B-Instruct · Countdown | accuracy: 23.14 -> 26.07 | Improved |
| add latent-aware guidance to beam search | Beam Search, LLaDA-8B-Instruct · GSM8K | accuracy: 79.38 -> 79.38 | No clear effect |
| add latent-aware guidance to beam search | Beam Search, LLaDA-1.5 · code tasks (HumanEval, MBPP) | Pass@1: 40.02 -> 42.04 | Improved |
| add latent-aware guidance to beam search | Beam Search, LLaDA-1.5 · Math500 | accuracy: 35.40 -> 37.60 | Improved |
| add latent-aware guidance to beam search | Beam Search, LLaDA-1.5 · ARC-Challenge | accuracy: 84.04 -> 84.56 | No clear effect |
| use beam search instead of probability-based greedy unmasking | LLaDA-8B-Instruct · Countdown | accuracy: 19.34 -> 23.14 | Improved |
| pure latent score (beta=0.0) in selection score (ablation) | LUGS · Countdown | accuracy: 23.14 -> 23.83 | No clear effect |
| vary guidance weight beta within 0.3 to 0.8 (ablation) | LUGS · Countdown | accuracy | No clear effect |
| increase beam width from 2 to 9 (ablation) | LUGS · Countdown, HumanEval, MBPP | accuracy | No clear effect |
| increase beam width (ablation) | LUGS · inference time | time per sample | Hurt |
| add LUGS guidance across beam widths (ablation) | Beam Search · HumanEval | Pass@1: average improvement 3.51% | Improved |
Closest papers by meaning
- UnMaskFork: Test-Time Scaling for Masked Diffusion via Deterministic Action Branching (ICML 2026 · similarity 0.89)
- Latent Refinement Decoding: Enhancing Diffusion Language Models by Refining Belief States (NeurIPS 2026 · similarity 0.89)
- UltraLLaDA: Scaling the Context Length to 128K for Diffusion Large Language Models (ICLR 2026 · similarity 0.88)
- Uncertainty Quantification for Large Language Diffusion Models (NeurIPS 2026 · similarity 0.88)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.88)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
