Research projects in scientific machine learning, agentic protein design, and generative modeling.
Investigating whether a learned generation order (via OrderHead) can improve image quality in autoregressive models. Systematic evaluation across MAR (masked) and LlamaGen (causal) architectures, with 50+ experiments covering heuristic orderings, STE-based learned orderings, oracle supervision, and random-order training (RAR).
Current work centers on Autoresearch V7: an LLM scientific agent that runs budgeted de novo protein-complex binder-design campaigns across seven targets. The live hub tracks results, discovery curves, where successes come from, and the agent's recorded scientific reasoning at each decision point (results & reasoning only).
Earlier v2/v5 hubs and the original audited baseline map remain available as historical archives.