AI Paper Atlas

Research map / Image generation / Fast diffusion sampling

Diffusion Language Model Knows the Answer Before It Decodes

Pengxiang Li, Yefan Zhou, Dilxat Muhtar, Lu Yin, Shilin Yan, Li Shen, Yi Liang, Soroush Vosoughi, Shiwei Liu

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

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 changedWhereMeasuredEffect
apply Prophet early-commit decodingLLaDA-8B · TruthfulQAaccuracy (%): 34.4 -> 46.1Improved
apply Prophet early-commit decodingLLaDA-8B · WinoGrandeaccuracy (%): 73.8 -> 70.5Hurt
apply Prophet early-commit decodingDream-7B · MMLUaccuracy (%): 67.6 -> 66.1Hurt
apply Prophet early-commit decodingLLaDA-8B · GSM8Kaccuracy (%): 77.1 -> 77.9No clear effect
apply Prophet early-commit decodingLLaDA-8B · HumanEvalaccuracy (%): 30.5 -> 30.5No clear effect
add suffix prompt 'Answer:' (ablation)LLaDA-8B low-confidence remasking · GSM8Kfraction correct by 25% step: 7.9% -> 59.7%Improved
switch from low-confidence to random remasking (ablation)LLaDA-8B · GSM8Kfraction correct by 50% step: 24.2% -> 97.2%Improved
use adaptive Prophet instead of static 128-step cap (ablation)LLaDA-8B · GSM8K L=256accuracy (%): 76.2 -> 77.9Improved
add Prophet to Fast-dLLMFast-dLLM · GSM8Kspeedup: 6.82x -> 7.66xImproved
add Prophet on top of SDTT-distilled studentSDTT · GSM8Kspeedup: 2.00x -> 3.21xImproved
add Prophet with different remasking strategies (ablation)LLaDA-8B · GSM8Kaccuracy (%): 63.8 -> 66.6 (random); 71.3 -> 72.7 (low-confidence); 72.4 -> 73.1 (top-k margin)Improved
add Prophet at large block lengths (ablation)LLaDA-8B · GSM8K block length 128accuracy (%): 33.1 -> 52.2Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy