{"id":"arxiv:2512.15745","title":"LLaDA2.0: Scaling Up Diffusion Language Models to 100B","date":"2025-12-10","section":"LLMs: architectures, context, training","quality":0.154862,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/","source_url":"https://arxiv.org/abs/2512.15745","text_available":true,"authors":["Tiwei Bie","Maosong Cao","Kun Chen","Lun Du"],"outgoing":[{"source":"arxiv:2512.15745","target":"arxiv:2407.21783","evidence":"id","id":"arxiv:2407.21783","title":"The Llama 3 Herd of Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de3cebb4c28bc5d41a843279f219dca14476284991ea5c7c0aeab7068a40e306/"},{"source":"arxiv:2512.15745","target":"arxiv:2502.09992","evidence":"title","id":"arxiv:2502.09992","title":"Large Language Diffusion Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a4863ed6f639560bd2feb6c66c13d1e3ac4cf42f8919a65c9818625cc6b6cde9/"}],"incoming":[{"source":"arxiv:2602.08676","target":"arxiv:2512.15745","evidence":"title","id":"arxiv:2602.08676","title":"LLaDA2.1: Speeding Up Text Diffusion via Token Editing","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1b6dd6e64e69da1fff43a183af57a467878db80d54f550bceefa049d7bb18878/"},{"source":"arxiv:2604.08302","target":"arxiv:2512.15745","evidence":"id","id":"arxiv:2604.08302","title":"DMax: Aggressive Parallel Decoding for dLLMs","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/22790263bd218aac28626fe3dab2fcd0ed73ceef94aa8a948bef2cb8bc007f7a/"},{"source":"arxiv:2608.00146","target":"arxiv:2512.15745","evidence":"id","id":"arxiv:2608.00146","title":"DiffusionGemma Technical Report","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/"}],"text_parts":[{"start":0,"end":15776,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-1.html"},{"start":15776,"end":31300,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html"},{"start":31300,"end":46521,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html"},{"start":46521,"end":51040,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-4.html"}],"contents":[{"title":"LLaDA2.0: Scaling Up Diffusion Language Models to 100B","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-1.html#text"},{"title":"Abstract","offset":58,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-1.html#text"},{"title":"1 Introduction","offset":1176,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-1.html#text"},{"title":"2 Related Work","offset":6597,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-1.html#text"},{"title":"2.1 Train dLLMs from scratch","offset":6616,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-1.html#text"},{"title":"2.2 Scaling dLLMs with AR initialization","offset":9085,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-1.html#text"},{"title":"2.3 dLLMs post-training","offset":11352,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-1.html#text"},{"title":"3 LLaDA2.0 Training Paradigm","offset":14627,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-1.html#text"},{"title":"4 Continual Pre-training via Warmup-Stable-Decay (WSD)","offset":16433,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"4.1 Warmup-Stable-Decay Conversion Strategy","offset":18435,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"Phase-1: Progressive Block Size Warmup","offset":18746,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"Phase-2: Large Scale Stable Training","offset":19661,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"Phase-3: Block Size Decay","offset":20354,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"Overall Training Objective","offset":20973,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"4.2 Document-level Attention Mask","offset":22790,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"4.3 Top-k Checkpoint Merge","offset":26656,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"5 Post-training","offset":27938,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"5.1 Supervised Fine-Tuning with Block Diffusion","offset":27958,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"Padding strategies & Mask ratio bandwidth","offset":29243,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"Complementary Masking","offset":30558,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-2.html#text"},{"title":"Data Recipe Curation","offset":31887,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"5.2 Confidence-Aware Parallel Training","offset":32607,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"5.3 DPO","offset":33996,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"5.4 Inference","offset":36621,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"6 Evaluation","offset":37392,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"6.1 Setup","offset":37409,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"6.2 Results","offset":39574,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"6.3 Analysis","offset":42113,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"Analysis of Inference Hyper-parameters","offset":42131,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"Analysis of Context Length","offset":44172,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"7 Training & Inference Infrastructure","offset":45420,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"7.1 Pretraining","offset":45462,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"Efficient Block Diffusion Training","offset":45975,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"Numerical Stability","offset":46495,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-3.html#text"},{"title":"7.2 Post-Training","offset":47408,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-4.html#text"},{"title":"7.3 Inference Engine","offset":48248,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-4.html#text"},{"title":"Inference speed","offset":49112,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-4.html#text"},{"title":"8 Conclusion","offset":49997,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/part-4.html#text"}],"text_code_points":51040,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a25474fdf076c254ea3ddbbc7ddbc1f2f49fb98065168d66b5ad9d8f7dafaa/source.md"}