Sihan Tan

Sihan Tan

谭斯涵/タンシカン

Ph.D. student at Science Tokyo

About me

Hi, there! I am a Ph.D. student at Institute of Science Tokyo (formerly Tokyo Tech), advised by Nakadai Kazuhiro and affiliated with Nakadai Lab, Systems and Control Engineering. I work on multimodal LLMs and vision-language models (VLMs): how models see, read, and generate across modalities.

My recent work spans the multimodal model development cycle:

  • Post-training VLMs. At SB Intuitions, I contributed to the release of Sarashina Vision, a Japanese VLM, through SFT and GRPO post-training, including reward design and training-data distillation.
  • Building VLMs. At Rakuten, I extended a text-only LLM into a Japanese VLM (SigLIP encoder, Megatron-LM) and benchmarked open VLMs (Qwen, Mistral) with vLLM/SGLang.
  • Video ↔ text. In my Ph.D., I build LLM-based models for video-to-text translation and text-to-pose generation, using sign language as a challenging real-world testbed, most recently with reference-free preference optimization to reduce hallucinations (EMNLP’26, ACL’25, COLING’25).

Before that, I was a visiting researcher at University of Zürich with Prof. Sarah Ebling, and at NHK Science & Technology Research Laboratories.

I did my master’s at Tokyo Tech where I graduated as the Best Graduate from Systems and Control Engineering, and I did my undergrad at Dalian University of Technology advised by Fan Xin and graduated as one of the outstanding graduates.

I am from China and on an adventure in Japan and Switzerland. I am a tarot lover (my favorite card is the world) and a newbie in Pilates.
I welcome collaboration; contact me, please :)

I am interested in, but not limited to, the following topics:

  • Multimodal LLMs and vision-language pre-training / post-training.
  • Video understanding and cross-modal generation (video ↔ text ↔ motion).
  • Faithful and reliable multimodal generation (reducing hallucinations).
  • Sign language processing as a real-world multimodal challenge.



    • Recent News

    • 📺10/ 2026: Our sign language translation & generation model SignFLIP was featured on TV Tokyo's World Business Satellite (WBS)!
    • 😊08/ 2026: I joined SB Intuitions as a research intern, working on post-training for Sarashina Vision, a Japanese VLM.
    • 🎉08/ 2026: One paper has been accepted to EMNLP26 Findings! See you in Budapest!
    • 😊03/ 2026: I joined Rakuten as a research scientist intern, building a Japanese vision-language model.
    • 😊11/ 2025: I just gave the first lecture on SLT in my life at UZH!
    • 🎉09/ 2025: I attended the SLTAT workshop in Berlin! So nice to meet sign language researchers worldwide!
    • 🎉05/ 2025: Our paper about multilingual sign language translation has been accepted to ACL2025 Main! Let's catch up in Vienna!
    • 😊05/ 2025: I will join the Language, Technology and Accessibility Group at UZurich as a visiting researcher from 08.2025!
    • 🎉12/ 2024: Our survey paper: A Review of Deep Learning-Based Approaches to Sign Language Processing is out!
    • 🎉12/ 2024: Our paper about CTC reordering for sign language translation has been accepted to COLING25! See you in Abu Dhabi!
    • 😊08/ 2024: I started working as a visiting researcher at NHK STRL.
    • 🎉05/ 2024: I presented two of our works for sign language translation at Signlang workshop, LREC-Coling 2024, Torino, Italy.
    • 🎉03/ 2024: I graduated as the best graduate from the Department of Systems and Control Engineering, Tokyo Tech!