Jonathan @ICLR

I’m a founding member @AsariAILabs and PhD researcher @Caltech @RPI @NEC working in LLM-agents, reasoning, RL, test-time scaling, and computer use agents

[Jonathan @ICLR @lightetal]

Post-training LLMs is like mixing a cocktail:

Too much easy data → no learning
Too much hard data → instability
Wrong balance → collapse

And today, we mix it by hand.

What if the data mixture could be learned instead of hand-tuned?

https://t.co/2EntaqvO2G

🧵👇

[Jonathan @ICLR @lightetal]

@aaliya_va @sarahookr It’s always fun connecting with curious minds!

[Sara Hooker @sarahookr]

Our second continual learning dinner + community gathering is coming up next Thursday evening.

We cutoff the guestlist on Tuesday to give names to security. Sign up before then.

Will be very fun.

https://t.co/uVVMYjDUAu

[Jonathan @ICLR @lightetal]

@Matthewagi Thanks!!

[Jonathan @ICLR @lightetal]

Accepted at #ICML2026 recently!

[Jonathan @ICLR @lightetal]

AI runs on data.

But… data is hard to buy.

❌ How much is a dataset worth?
❌ Will it actually help your model?
❌ What about privacy / trust?

What if we never priced data directly…
and instead priced what actually matters: model improvement?

https://t.co/7L2j7GMS31

🧵👇

[Jonathan @ICLR @lightetal]

The real AGI was the researchers Anthropic recruited along the way.

[Andrej Karpathy @karpathy]

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

[Jonathan @ICLR @lightetal]

@violazhouyi @restofworld First time getting quoted in a feature article 😂

[Jonathan @ICLR @lightetal]

@tw_killian @BYU @BYUCS Super exciting, congrats Taylor!

[Jonathan @ICLR @lightetal]

@atharva_sehgal This is so cool!

[Jonathan @ICLR @lightetal]

@michaelryan207 @lateinteraction @KnightHennessy Congrats Michael!!

[Jonathan @ICLR @lightetal]

🚀 Super excited to present at the continual learning meetup!

[Nilou Salehi @nilou_salehi]

📣 The May speaker of our researcher meetup on continual learning will be Caltech's @lightetal Jonathan Li! 📣

Jonathan will share recent work on structured, language-native search methods such as DISC and SFS, which enable agents to explore diverse reasoning paths, strategies, and solutions beyond what is directly represented in their training data.

Space is limited and last time we had a waitlist of 200+ so please register soon if you'd like to join.

@sarahookr @mralbertchun @NikzadAfshin
https://t.co/lm1pkfdAk0

[Jonathan @ICLR @lightetal]

@MimeeXu Wow this is very cool! Thanks for sharing

[Jonathan @ICLR @lightetal]

AI runs on data.

But… data is hard to buy.

❌ How much is a dataset worth?
❌ Will it actually help your model?
❌ What about privacy / trust?

What if we never priced data directly…
and instead priced what actually matters: model improvement?

https://t.co/7L2j7GMS31

🧵👇

[Jonathan @ICLR @lightetal]

Grateful to amazing collaborators on this work, including:

Minbiao Han, Steven Xia,
@haifengxu0 @SainyamGalhotra @raulcfernandez

@Cornell @UChicago

Paper: https://t.co/3VRxAl7VZP

🧵 9/n

[Jonathan @ICLR @lightetal]

The big takeaway:

👉 Don’t sell data.
👉 Sell what data does.

By pricing model improvement, we can:

• unlock data markets
• align incentives
• make ML systems more accessible
• ensure privacy

This is a step toward real AI economies.

🧵 8/n