Automatic (meta-) discovery and continual learning

Core objective

This research topic reimagines search and optimization within high-dimensional, non-differentiable, and structurally complex environments. Instead of relying on traditional black-box solvers or hyperparameter optimization, the project utilizes a population of “English-driven agents”—equipped with the reasoning and generative capabilities of foundation models—to collaboratively explore and discover high-performing solutions.

The language-mediated search paradigm

By shifting from purely numerical exploration to language-mediated search, the system leverages the vast general world knowledge embedded within foundation models. This enables the agents to propose intelligent search strategies based on known algorithms, architectural heuristics, system design best practices, and prior scientific literature, bypassing the inefficiencies of blind search.

Primary Use-Cases

1. Co-Design of Kernel Code and Hardware Architectures

This use-case focuses on jointly optimizing software kernels (such as CUDA code or dataflow graphs) alongside the underlying hardware substrate, including memory hierarchies and compute layouts. Agents collaborate by generating candidate implementations, critiquing each other’s proposals for expected bottlenecks or hardware compatibility, and iteratively refining the solutions through dialogue. The ultimate goal is to build a language-driven co-design loop capable of uncovering innovative hardware-software combinations that conventional search methods would likely miss.

2. Discovery of Photonics Architectures

In photonic chip design, traditional simulation-driven optimization is highly expensive and typically lacks gradients. This project frames photonics design as a multi-agent exploration problem where domain-specialized agents suggest new chip configurations, waveguide geometries, and coupling strategies by describing physical constraints and optical behaviors in natural language. Because foundation models encode substantial physics knowledge, the agents can analyze simulated feedback collectively to create a self-improving photonics design loop that accelerates the discovery of novel architectures.

3. Automatic Scientific Discovery

The final use-case aims to assist human researchers in formulating new methodologies, hypotheses, and experimental strategies, particularly in domains involving sensitive or private data. An ensemble of “scientist-like entities” reviews available datasets and domain knowledge, exchanges natural-language arguments, and proposes new analytical techniques. By critically evaluating each other’s ideas and identifying logical flaws, this multi-step debate process emphasizes creative generation while reducing hallucinations and increasing the robustness of the chosen research direction.