README.md

Overview

Diffusion Language Models (DLMs) are often described as naturally parallel decoders. In practice, many decoding trajectories still collapse into autoregressive (AR-like) left-to-right behavior.

This project presents NAP (Non-Autoregressive Parallel DLMs), a data-decoding co-design framework:

1. Parallel data curation: each training sample contains multiple independent reasoning trajectories, not a single privileged chain.
2. Parallel-forced decoding: generation updates are enforced across multiple reasoning streams at every decoding step.

The goal is to reduce AR-like decoding bias while preserving (or improving) reasoning accuracy.

Key Diagnosis

• Common corpora (FineWeb, OpenR1-Math) are strongly sequential.
• AO (confidence-based arbitrary-order) decoding in standard DLMs still shows high ARness.
• Long-CoT SFT further increases ARness.
• Existing fast DLM decoding methods often improve speed by following an AR-like critical path.

TODOs

We will try our best to achieve

• [✅] Training code of NAP
• [🚀] Datasets and Model weights
• [🚀] Inference and evaluation code

Method

NAP uses a structured output canvas:

[<think #1>, R^(1), <think #2>, R^(2), ..., <think #m>, R^(m), <summary>, S]

• R^(j) are independent reasoning paths.
• <summary> aggregates evidence from all paths and outputs the final answer.

Decoding policy:

• Macro-parallel: distribute unmasking budget across all reasoning blocks each step.
• Micro-confidence: within each block, commit tokens by confidence (not strict left-to-right order).

Main Quantitative Results

GSM8K (Accuracy, %)

Steps	Tok/Step	LLaDA Long-CoT	NAP-LLaDA	Gain	Dream Long-CoT	NAP-Dream	Gain
256	4	54.1	56.1	+2.0	46.5	60.9	+14.4
336	3	60.9	63.3	+2.4	56.9	70.9	+14.0
512	2	82.0	82.6	+0.6	66.8	79.2	+12.4
1024	1	83.5	84.1	+0.6	78.0	83.6	+5.6

MATH-500 (Accuracy, %)

Steps	Tok/Step	LLaDA Long-CoT	NAP-LLaDA	Gain	Dream Long-CoT	NAP-Dream	Gain
256	4	21.4	26.6	+5.2	16.2	23.8	+7.6
336	3	26.6	35.4	+8.8	25.6	31.4	+5.8
512	2	41.2	43.0	+1.8	40.0	43.0	+3.0
1024	1	45.0	47.0	+2.0	47.4	49.6	+2.2

GPQA (Accuracy, %)

Steps	Tok/Step	LLaDA Long-CoT	NAP-LLaDA	Gain	Dream Long-CoT	NAP-Dream	Gain
336	3	15.4	19.0	+3.6	7.3	10.5	+3.2
512	2	21.2	25.9	+4.7	19.4	22.5	+3.1
1024	1	23.0	28.6	+5.6	28.6	29.5	+0.9

ARness Findings

• AO decoding ARness (Global-ARness@1):
  • LLaDA-8B: 0.73
  • Dream-7B: 0.92
• After Long-CoT SFT:
  • LLaDA-8B: 0.73 -> 0.81 (+0.08)
  • Dream-7B: 0.92 -> 0.93 (+0.01)

These numbers indicate that standard supervision tends to increase autoregressive decoding behavior.

Citation