{"id":"arxiv:2608.00146","title":"DiffusionGemma Technical Report","date":"2026-07-31","section":"LLMs: architectures, context, training","quality":0.286144,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/","source_url":"https://arxiv.org/abs/2608.00146","text_available":true,"authors":[" DiffusionGemma Team","Adrien Ali Taïga","James Assiene","Daniele Calandriello"],"outgoing":[{"source":"arxiv:2608.00146","target":"arxiv:2310.04378","evidence":"title","id":"arxiv:2310.04378","title":"Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/16b9865215d555f17a2d1abc00c0915667686f20c49898f63872327a416690db/"},{"source":"arxiv:2608.00146","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/"},{"source":"arxiv:2608.00146","target":"arxiv:2512.15745","evidence":"id","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":[],"text_parts":[{"start":0,"end":15828,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-1.html"},{"start":15828,"end":30937,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html"},{"start":30937,"end":46420,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html"},{"start":46420,"end":62308,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html"},{"start":62308,"end":77774,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html"},{"start":77774,"end":84408,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-6.html"}],"contents":[{"title":"DiffusionGemma Technical Report","offset":10,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-1.html#text"},{"title":"Abstract","offset":45,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-1.html#text"},{"title":"1 Introduction","offset":1588,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-1.html#text"},{"title":"Open-weights release.","offset":7382,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-1.html#text"},{"title":"Outline.","offset":9023,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-1.html#text"},{"title":"2 Generative Text Modeling with Discrete Diffusion","offset":9914,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-1.html#text"},{"title":"2.1 Probability Paths and Denoising","offset":14229,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-1.html#text"},{"title":"2.1.1 The Forward Process","offset":15796,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-1.html#text"},{"title":"2.1.2 The Backward Denoising Process","offset":16816,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"3 The DiffusionGemma Architecture","offset":20014,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"3.1 Block-Autoregressive Generation","offset":21209,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"KV cache initialization.","offset":21597,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"Canvas denoising conditioned on KV cache.","offset":22200,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"3.2 The Denoising Framework at Inference","offset":24185,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"Canvas initialization.","offset":24231,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"The decoder forward pass.","offset":24598,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"The denoising iteration.","offset":25428,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"Multinomial diffusion.","offset":26827,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"Remark on interpretability.","offset":27330,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"3.3 The DiffusionGemma Sampler","offset":27719,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"Entropy-bounded token refinement.","offset":28775,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"Temperature annealing.","offset":29588,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"Adaptive stopping heuristic.","offset":30327,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-2.html#text"},{"title":"3.4 Inference Efficiency Metrics","offset":31837,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"3.5 Retained Autoregressive Capability","offset":33996,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"4 Supervised Finetuning","offset":34594,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"5 Sampler Distillation & Reinforcement Learning","offset":36492,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"Training setup.","offset":38149,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"Training dynamics & implicit curriculum effect.","offset":38920,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"Improved speed-to-intelligence Pareto frontier.","offset":40149,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"SD ⋅ \\cdot RL specialization in the few-step regime.","offset":41408,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"Emergent conciseness.","offset":42026,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"6 Inference Optimizations","offset":42809,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"Low batch size serving.","offset":43662,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"GPU time breakdown.","offset":44645,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-3.html#text"},{"title":"Eliminating CPU-GPU synchronization.","offset":47460,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"Throughput for batch size of 1 serving.","offset":48143,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"Multi-user throughput.","offset":49187,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"Toward real traffic throughput.","offset":50049,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"7 Experimental Results","offset":50928,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"8 Open-Source Downstream SFT","offset":55203,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"Finetuning procedure.","offset":55629,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"LoRA for parameter-efficient finetuning.","offset":57275,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"Case study: Sudoku puzzle solving.","offset":57669,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"9 Practical Advantages of Text Diffusion","offset":58326,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"9.1 Bidirectional Reasoning and Self-Correction","offset":59087,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"9.2 Dynamic and Adaptive Computation","offset":61227,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-4.html#text"},{"title":"9.3 Structured and Constrained Outputs","offset":62885,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"10 Limitations & Known Issues","offset":64156,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"11 Conclusion","offset":66990,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Appendix A Related work","offset":67875,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"A.1 Continuous Diffusion for Text","offset":67903,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"A.2 Speculative Decoding","offset":70027,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Appendix B Samples Before and After SD ⋅ \\cdot RL Training","offset":71509,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Appendix C Additional Open-Source Downstream Finetuning Results","offset":71886,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Sudoku solving trace summary.","offset":72163,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Practical recipe summary.","offset":72290,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Case study: PubMedQA.","offset":72619,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Appendix D Prompt Formatting","offset":73198,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"D.1 BOS and EOS Special Tokens","offset":73886,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"D.2 Conversational Structuring","offset":74451,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"D.3 Thinking Channels","offset":75114,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"D.4 Tool and Function Calling Serialization","offset":75764,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Appendix E Mercury 2 Speed Estimation","offset":76643,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Estimation data.","offset":76685,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Estimation model.","offset":77749,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-5.html#text"},{"title":"Speed estimates.","offset":79041,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-6.html#text"},{"title":"Appendix F Denoising trajectory sampled from DiffusionGemma","offset":81068,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-6.html#text"},{"title":"Appendix G More on Practical Advantages of Text Diffusion","offset":81132,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-6.html#text"},{"title":"G.1 Bidirectional Reasoning and Self-Correction","offset":81194,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-6.html#text"},{"title":"G.2 More on Dynamic and Adaptive Computation","offset":81864,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-6.html#text"},{"title":"G.3 Structured and Constrained Outputs","offset":83400,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-6.html#text"},{"title":"Instructions for reporting errors","offset":83444,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/part-6.html#text"}],"text_code_points":84408,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/90a7b084496cd65968627c115230fc3a5e78005a66b5ebb34547e874754bc82e/source.md"}