Towards AI Co-Scientists: Agentic AI for Scientific Discovery

TL;DR

• The framing: not a chatbot, but a research partner that can reason inside a physical universe and lead to open-ended discovery.
• Three assumptions the physical world breaks: (1) there are laws of physics, symmetries, and conservation laws that must act as guardrails; (2) simulations are so expensive they run for days or weeks — and yet they're exactly the world models you'd use to verify an agent; (3) the interesting events are rare and carry enormous uncertainty, so when an agent says it's 100% sure it will rain, that had better be true.
• The recipe is physics-guided AI, described in her PNAS paper: ground agents in differential equations (the building blocks of simulations) and symmetries (the governing principles of the physical universe), then loop that grounding into open-ended agentic reasoning.
• Results across four areas: 2× more accurate climate question answering; #1 on PutnamBench using Lean and Isabelle verifiers, lifting solve rate from 50% to 70% and beating ByteDance's proprietary solution; 18–35% better binding energies in drug design under molecular dynamics evaluation; and the first agentic weather scientist, compressing a couple of weeks of analysis into a couple of hours.

Key Points

Three broken assumptions (~01:12–01:15)

Yu directs the spatiotemporal machine learning lab at UC San Diego, building machine learning algorithms for the physical world. The question that's been nagging at her for years: how do you build an AI co-scientist that genuinely helps with research — not as a chatbot, but as a partner that can reason in a physical universe and lead to open-ended discoveries?

Answering it means being clear about what makes AI for science distinctive once the physical universe is involved:

1. Laws of physics, symmetries, and conservation laws. These are hard constraints in the physical world, and the question is how to build that kind of guardrail into today's agentic systems.
2. Simulations are expensive. Physical sciences lean on simulations built from first-principles mathematical models — how climate changes over 50 years, how a disease spreads through billions of people, how atoms interact in materials and biological systems. These take days or weeks to return feedback, and yet they're precisely what you'd want as the world model for verifying an agent. How does an agent interact with an environment like that?
3. Rare events and uncertainty. The events worth caring about are rare and carry enormous uncertainty. How do you calibrate that uncertainty so that when an agent says it's 100% sure it will rain tomorrow, it's actually 100% correct?

All three, she argues, break the common assumptions behind agentic systems.

What follows is a requirements list: agents that adaptively interleave internal reasoning with simulations that may run for days; laws of physics baked into the reasoning engine; properly calibrated uncertainty. In practice that means adaptive tool use plus integrating formal reasoning tools into a verifiable environment.

Physics-guided AI: the recipe and the results (~01:15–01:18)

She calls the recipe her group has built up over the years physics-guided AI, and wrote a PNAS paper laying out what she takes to be the right approach: ground agents in the laws of physics. Two grounding substrates — differential equations, the building blocks of simulations, and symmetries, the governing principles of the physical universe — looped into open-ended agentic reasoning.

The recipe works, and she ran through where:

• Climate science: an AI scientist for understanding climate change impacts uses an adaptive reasoning loop so the agent can trade fast retrieval from memory against expensive climate model projections, producing 2× more accurate question-answering results on climate reasoning tasks.
• Formal mathematics: verifiers leveraging formal methods in Lean and Isabelle ground the reasoning engine for theorem proving. They ranked #1 on PutnamBench, one of the hardest formal theorem-proving benchmarks, lifting solve rate from 50% to 70% and beating ByteDance's proprietary solution.
• Drug design: designs evaluated by molecular dynamics simulators showed 18–35% more favorable binding energies.

The concrete example: an agentic weather scientist. She showed an agent — transcribed as "Deferris," spelling unconfirmed — billed as the first agentic weather scientist. It reads huge volumes of highly numerical weather data, generates and runs code in a parallel code execution environment, orchestrates a wide range of tools from climate simulators to weather forecasters, and writes a report on the impact of extreme events like wildfires and earthquakes. Analysis that typically takes weather scientists a couple of weeks now takes a couple of hours.

From the lab to the physical economy (~01:18–01:20)

Having done a lot of research on physics-grounded agents for physical science, they now want to take it to the real world. She has launched a startup — she's the CEO — applying the same recipe outside the lab: deploying agents that automatically produce forecasting, simulation, what-if scenarios, and actionable insights to support supply chain and operations, which she frames as the critical infrastructure of the physical economy.

The demo query is plain natural language. Imagine you're a chip manufacturer: if a typhoon hits Southeast Asia, how should we understand the impact on shipping lanes and lead times, and how do we optimize and mitigate the potential disruptions to the business? The agent runs adaptive research to find every affected lane, estimates the disruption to lead-time arrival, and produces a mitigation plan in a couple of hours.

Quotes

"How can we build an AI co-scientist that actually helps you with research, not as chatbot but as partners that can reason in a physical universe and then lead to open-ended discoveries?" (~01:12:48)

The framing for the whole talk: a partner, not a chat interface.

"…when the agent says I'm 100% sure that tomorrow is going to rain, then we know it's actually 100% correct." (~01:14)

The physical world's extra demand: be right about how confident you are, not just about the answer.