Improving GPT-4 performance for domain tasks

Hi, I wanted to share this ICML '24 paper ([2402.10980] ChemReasoner: Heuristic Search over a Large Language Model's Knowledge Space using Quantum-Chemical Feedback) that focuses on designing compound AI systems and shows an application for a scientific problem combining linguistic reasoning with 3D geometrical reasoning. Would appreciate hearing your feedback on the promise of such techniques.

ChemReasoner seems like really interesting use of LLM agents in scientific discovery. The topic is out of my core area of expertise though. I look forward to hearing future advances here. Thanks for sharing.

Thanks Vu - I appreciate you taking the time to read through. Basically two key lessons or insights.

Lesson 1) Learn how to probe a model and go beyond single step prompting for applications that can allow that latency. Everyone is focused on training a better model - and we are focused on eliciting information from the AI model. Knowing how to probe the model is an art, and heuristic search and using feedback are key tools in that art.

Lesson 2) Bridging language and geometry is a new qwest. Learning a correspondence between concepts and 3D structural orientation is a big deal for many scientific disciplines like biology and chemistry. SORA-like models are starting to focus on contours and colors, while designing proteins and materials require reasoning about what’s under the surface, how to pack molecules in a 3D space that leads to desired behavior. Most work before us bridging molecular structures and LLMs focused on string representation of molecules, not 3D ones.

1 Like