Jacob Si

CS PhD @ImperialCollege | previously @UCLA, @UofTCompSci, @VectorInst

[Jonathan Kao @JonathanCKao]

Excited to share flow control, a method that allows users to steer VLAs in real time with keyboard arrow presses. This method doesn't require any fine-tuning of a VLA and can run out-of-the-box.

[Ian Shi @ianshi3]

Excited to annouce our seed round!

[yingzhen @liyzhen2]

Our own answer: structured coupling
https://t.co/wCsuYCpLsM

- flow matching with VAE-based coupling
- VAE encoder & flow sharing networks
- VAE decoder init. + flow refinement for sampling

flow matching 🤝 VAEs -> good representation & sample quality🚀

[Lisa Alazraki @LisaAlazraki]

Scaling Small Agents Through Strategy Auctions accepted at #ICML2026 🇰🇷 🎉

Grateful to have worked with @akhilmathurs @williamfshen @yorambac

[Jerry Ji @jerryji2019]

Clinical ML has a generalization problem. The standard playbook: train a model, watch it fail at the next hospital, and retrain it with new tricks. We invert this! Don't change the model—change the input so any model will generalize. Introducing Record2Vec at #ICLR2026! 🔄🚀 1/5

[Chen-Hao (Lance) Chao @chenhao_chao]

(1/7) We introduce MDM-Prime-v2 which scales 21.8× better than autoregressive models (ARMs) in compute-optimal comparisons.

📎 Paper: https://t.co/VhBVo75abe
🌟 Blog: https://t.co/miWdTmcGtL
⌨️ Github: https://t.co/ac1eDV8O8Q

Here’s how we did it👇:

[yingzhen @liyzhen2]

A few months back I pushed @jacobyhsi88 to test how well GPT-5.2 did on this task. We found that open-source models can do even better when they are instructed to use the right tools. 😉

(I hope the Qwen team will continue leading the open-source LLM contributions 🫡)

[Jacob Si @jacobyhsi88]

🧵10/10 Thank you for your interest in our work 🙏! It has been a pleasure working with @mikezqu, @michelle_xioax, & my supervisors, @MarekRei and @liyzhen2 🙌. This work was done at @imperialcollege and @Columbia🎓.

https://t.co/KGHyGiLEZO

[Jacob Si @jacobyhsi88]

Ever queried RAG pipelines about tabular documents 𝄜 and find that answers are often incorrect ❌🤔?

🚀 Introducing TabRAG, an end-to-end parsing-based RAG framework designed to improve tabular document question answering via structured representations 👍!

📄 Paper: https://t.co/I90RvDPn9h
💻 GitHub: https://t.co/Pv93LkB945

[Jacob Si @jacobyhsi88]

🧵9/10 We analyze the effect of the number of self-generated ICL demonstrations on our generation performance. The performance improves sharply when moving from zero to a small number of demonstrations, with three ICL examples consistently yielding stable gains across most datasets.

[Jacob Si @jacobyhsi88]

@BalsellsRodas Vamos papito 💪💪

[Carles Balsells Rodas @BalsellsRodas]

New paper! We prove identifiability of Regime-Switching Models with arbitrary lagged dependencies.