{"id":"arxiv:2602.08676","title":"LLaDA2.1: Speeding Up Text Diffusion via Token Editing","date":"2026-02-09","section":"LLMs: architectures, context, training","quality":-0.033511,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/","source_url":"https://arxiv.org/abs/2602.08676","text_available":true,"authors":["Tiwei Bie","Maosong Cao","Xiang Cao","Bingsen Chen"],"outgoing":[{"source":"arxiv:2602.08676","target":"arxiv:2512.15745","evidence":"title","id":"arxiv:2512.15745","title":"LLaDA2.0: Scaling Up Diffusion Language Models to 100B","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/"}],"incoming":[{"source":"arxiv:2604.08302","target":"arxiv:2602.08676","evidence":"id","id":"arxiv:2604.08302","title":"DMax: Aggressive Parallel Decoding for dLLMs","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/22790263bd218aac28626fe3dab2fcd0ed73ceef94aa8a948bef2cb8bc007f7a/"}],"text_parts":[{"start":0,"end":15574,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html"},{"start":15574,"end":19573,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-2.html"}],"contents":[{"title":"LLaDA2.1: Speeding Up Text Diffusion via Token Editing","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"Abstract","offset":58,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"1 Introduction","offset":1664,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"2 Configurable Decoding Scheme","offset":5139,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"3 Training Paradigm","offset":7549,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"3.1 Training Alignment for “Draft-and-Edit”","offset":7573,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"3.2 Reinforcement Learning Training","offset":8844,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"4 Infrastructure","offset":11548,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"4.1 Training Infrastructure","offset":11569,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"Continued Pre-Training and Supervised Fine-Tuning","offset":11602,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"RL Training","offset":12029,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"4.2 Inference Infrastructure","offset":12541,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"4.3 Decoding Algorithm at Inference","offset":13192,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"5 Evaluation","offset":13814,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-1.html#text"},{"title":"6 Outlook and Limitation","offset":16851,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-2.html#text"},{"title":"Tradeoff Between Inference Speed and Accuracy","offset":16880,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-2.html#text"},{"title":"Editable Enhanced dLLM","offset":17691,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-2.html#text"},{"title":"LLaDA2.1 remains in an experimental phase. Although rare, certain edge cases may occur.","offset":18500,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-2.html#text"},{"title":"Conclusion","offset":19188,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/part-2.html#text"}],"text_code_points":19573,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/source.md"}