Reproduce “LLaDA2.1: Speeding Up Text Diffusion via Token Editing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tiwei Bie, Maosong Cao, Xiang Cao, Bingsen Chen, Fuyuan Chen, Kun Chen, Lun Du, Daozhuo Feng and 42 more
LLaDA2.1 introduces a "Draft-and-Edit" paradigm for discrete diffusion Large Language Models (dLLMs) through token editing, enabling configurable decoding for either high-throughput generation (Speedy Mode with 892 TPS on HumanEval+) or enhanced output quality (Quality Mode, outperforming LLaDA2.0 by 1.11 average score). The framework integrates large-scale reinforcement learning and a novel editable state evolution to balance speed and fidelity.
Key insights
- Moving beyond the rigid absorbing-state framework to an editable state evolution (Draft-and-Edit) fundamentally resolves token-level inconsistencies and exposure bias in dLLMs.
- A configurable decoding strategy, managed by specific confidence thresholds, enables flexible optimization between high generation speed and high output quality within a single model architecture.
- Scalable reinforcement learning for dLLMs can be achieved through ELBO-based policy optimization, which is crucial for improving reasoning precision and instruction-following fidelity.
- Iterative cross-block refinement (Multi-Block Editing) effectively corrects local errors and improves global consistency in generated text.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.