Jack Cai

CS PhD student at Princeton PLI, Research Intern at Microsoft. Previously Masters at UW-Madison. Working towards goal generating long horizon agents.

[Jack Cai @jackcai1206]

My experience after playing with a lot of auto-research agents is its hard to optimize for simple, elegant ideas without human in the loop. Instead agents are good at squeezing out gains by increasing complexity (and they can tolerate a lot of complexity!)

[Jack Cai @jackcai1206]

Paper + code 👇
💻 Code: https://t.co/oKPm7RZjIh 📄 arXiv: https://t.co/Hj04PlEL3d
By @jackcai1206 *, @XingyuZhu_ *, @YiheS5 , @yinghui_he_ , @prfsanjeevarora

[Jack Cai @jackcai1206]

Want the benefits of looping but not the inference overhead? Or train latent reasoning with dense supervision?
Meet T²MLR — Transformers with Temporal Middle-Layer Recurrence. Latent reasoning that persists across decoding steps, at essentially the same inference cost. 🧵

[Jack Cai @jackcai1206]

We also retrofit the recurrent pathway into an existing pretrained 1.7B Transformer though an SFT stage, and show gains on math reasoning.

[Jack Cai @jackcai1206]

@Kangwook_Lee @marksaroufim @CoreAutoAI @tilderesearch Thank you Kangwook!

[Jack Cai @jackcai1206]

Found out I can clear cookies every minute to watch World Cup for free using the limited time preview feature but Claude is refusing to write a script to automate it for me lol, finally a use case for kimi

[Jack Cai @jackcai1206]

@MarcHoelle @KushaSareen @rish2k1 @agarwl_ @Devvrit_Khatri @LakshyAAAgrawal @inderjit_ml @profjoeyg @KurtKeutzer Yeah I am also very interested in this follow up

[Jack Cai @jackcai1206]

@KushaSareen @rish2k1 @agarwl_ @Devvrit_Khatri @LakshyAAAgrawal @inderjit_ml @profjoeyg @KurtKeutzer wait but correct me if I am wrong -- you are using GPT-5.2 in the prompt optimizer, so the strong model is teaching the Qwen model you are training? Vanilla RL does not have an external model providing prefix.

[Jack Cai @jackcai1206]

@HeMuyu0327 Could it be the model size and data is small scale? Though I would be a bit surprised if that is the case

[Jack Cai @jackcai1206]

Joint work with @BehlHarkirat. He has been an awesome mentor during my time at Mircrosoft 😀

[Jack Cai @jackcai1206]

New (working) paper: AI Scientist via Synthetic Task Scaling

We have seen the first sparks of automated AI research done by LLMs, but how do we train them to be better auto-researchers?

[Jack Cai @jackcai1206]

Paper link: https://t.co/UGDB6W7hzd

[Jack Cai @jackcai1206]

Per-task breakdown — improvements across most of the 13 MLGym tasks including CIFAR-10, Fashion MNIST, language modeling, game playing, and more.