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 now centers on Autoresearch v2: Qwen-guided config search over executable Proteina-Complexa inference programs, with a live pipeline diagram, run ledger, failure-handling notes, baseline reuse policy, and static tree-search reports.
The older audited baseline/results map is preserved as a historical archive, but it is no longer the main project entry point.