My new reflection: Language, Curiosity, and Life.
It is my attempt to put into words what has mattered most to me: language, family, curiosity, music, work, illness, and gratitude.
This is not a goodbye. I hope the conversation continues.
https://t.co/0i0N3GrteM
🔍 Key findings:
- Node and global information are linearly encoded within single image patches in the vision encoder.
- Edge information only becomes linearly encoded for the first time within text tokens in the language model.
2/N
🚀 New paper on arXiv!
https://t.co/wGSHaQ1kve
🤔 How do LVLMs internally form representations of nodes and edges?
💡 Node representations form early, whereas edge representations form late.
Feedback is welcome! 1/N
I’m currently attending #ICLR2025 and will be presenting our work today on sketch-to-diagram generation with vector code translation ✍️📐.
📍 See you at Poster Session 4, #82 (Friday 3-5:30pm)!
More details below 👇
Tired of the hassle of crafting figures and diagrams? Sketch2Diagram is our step toward turning hand-drawn sketches into structured, vector-based outputs—check it out if you're curious!
We’ll be at Poster #82 at 3PM — come say hi 👋 and check out our work! #ICLR2025
I’m currently attending #ICLR2025 and will be presenting our work today on sketch-to-diagram generation with vector code translation ✍️📐.
📍 See you at Poster Session 4, #82 (Friday 3-5:30pm)!
More details below 👇
I'm thrilled to share that our Delphi paper is officially published today at @NatMachIntell after almost four years of hard works from all my amazing collaborators (a quite insane timeline considering the rapid AI world)! Special thanks to the unwavering support of my advisor, the marvelous @YejinChoinka! This project has defined my PhD experience and profoundly shaped how I view and approach science. The challenges along the way have made this journey all the more rewarding; it has been a truly unique and transformative training experience—one that anyone could aspire to during their PhD! https://t.co/aT7wrGnRr8
Kotoba has been accepted to the Japanese government's GenAI program! Will be building speech foundation models with GPU support.
Speech is exciting. I feel like I'm back in my first year as a PhD student. Check our application form if you're interested:
https://t.co/MDq6VFZ1Lf
Kotoba Tech is launching the Open-GPT-4o project with a focus on speech foundation models. Let's make this happen fast with the power of open-source!
https://t.co/YncHZmS3aO
We've developed two main building blocks: Kotoba-Speech (an autoregressive speech generation model) and Kotoba-Whisper (fast, accurate Japanese ASR).
As an intro to speech models, Jungo (@jungokasai) gives a quick tutorial on the audio tokenization technique, which allows us to train models similarly to text LLMs.
https://t.co/RJjtutLRaz
Exciting announcements coming soon from Kotoba!
Kotoba-Speech v0.1: our first step towards speech, multimodal foundation model. Demo/HuggingFace Model are available. We'll be releasing our training code very soon!
@kotoba_tech#NLP2024
A Call for Clarity in Beam Search: beam search is a standard algorithm for text generation, but it turns out there's **no standard** for when the search stops, impacting the generation quality especially in summarization.
Paper accepted to @LrecColing.
https://t.co/E2mIVX0OZ0
Open-sourcing Kotomamba, our distributed training library for Mamba (a state space model that outperforms transformers). We'll release a tech blog post and HuggingFace models soon. Stay tuned!
Project led by @kazukifujii and @hiroto_kurita
https://t.co/v4DhF5m40Z
We're releasing the first-ever Japanese/English Mamba, trained on 200B tokens. The performance is competitive to transformers with similar sizes, with a 2x-3x inference speedup.
日英両言語対応で今話題のMamba事前学習済みモデルを公開しました!
https://t.co/vxLrlWeanQ
Many thanks to those who attended our New Year party in Tokyo. Amazing to see students/professors from universities, industrial researchers, and investors getting connected through Kotoba. Kotoba continues its mission of contributing to the AI community both in the US and Japan!