{"id":"arxiv:2512.14856","title":"T5Gemma 2: Seeing, Reading, and Understanding Longer","date":"2025-12-16","section":"LLMs: architectures, context, training","quality":-0.289423,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/","source_url":"https://arxiv.org/abs/2512.14856","text_available":true,"authors":["Biao Zhang","Paul Suganthan","Gaël Liu","Ilya Philippov"],"outgoing":[{"source":"arxiv:2512.14856","target":"arxiv:2504.06225","evidence":"id","id":"arxiv:2504.06225","title":"Encoder-Decoder Gemma: Improving the Quality-Efficiency Trade-Off via Adaptation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/797af53c1331a8303ff731ef2763d562b29b9864407468ebbd7c1a89ab7c0eed/"},{"source":"arxiv:2512.14856","target":"arxiv:2510.26622","evidence":"id","id":"arxiv:2510.26622","title":"Encoder-Decoder or Decoder-Only? Revisiting Encoder-Decoder Large Language Model","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/dc73f87f798d016e2d83acd1d362b4b3525af826e68b055bb5196f120d23e080/"}],"incoming":[],"text_parts":[{"start":0,"end":15561,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html"},{"start":15561,"end":18033,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-2.html"}],"contents":[{"title":"T5Gemma 2: Seeing, Reading, and Understanding Longer","offset":10,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"Abstract","offset":66,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"1 Introduction","offset":1169,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"2 T5Gemma 2","offset":4421,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"Tied Embedding","offset":5157,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"Merged Attention","offset":5694,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"Rejected Ablation : Cross-Attention on Global Layers Only","offset":8468,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"3 Setup","offset":9212,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"3.1 Pretraining","offset":9224,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"Data","offset":9245,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"Optimization","offset":10341,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"3.2 Post-training","offset":10930,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"3.3 Evaluation","offset":11368,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"4 Results","offset":12752,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"PrefixLM+KD, UL2, vs. UL2+KD: Do distillation and training objective matter?","offset":12766,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"Text-only LLMs can be adapted into strong multimodal and long-context encoder-decoder models.","offset":13593,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"T5Gemma 2 achieves competitive pretraining performance and improved post-training performance than Gemma 3.","offset":14440,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"5 Conclusion","offset":15544,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-1.html#text"},{"title":"Acknowledgments","offset":17669,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/part-2.html#text"}],"text_code_points":18033,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e982edc00d6b6d9e20d11b62004bb8e91131e2d7f717eca944a9914897f273b6/source.md"}