Post-training & Searching RNA Foundation Models
Diagnosing diversity collapse and using structural feedback to turn pretrained RNA generation into exact design.
Learning to Search Beyond a Generative RNA Model
Pretrained RNA foundation models can generate plausible sequences, but scientific design asks for something more demanding: sequences that satisfy an exact target structure while remaining diverse and biologically meaningful.
The Problem
Reward-based post-training can improve structural success while collapsing a broad generative distribution into a small set of high-probability solutions. This project studies that tension between optimization, diversity, and biological validity.
From Post-training to Structured Search
I investigate reinforcement-learning post-training together with verifier-guided inference. A fold-back verifier compares each generated sequence with the target structure, identifies missing and spurious base pairs, and turns those errors into structured feedback for the next search step.
Broader Lesson
Scientific foundation models should not be treated as one-shot generators. Their pretrained representations can serve as a proposal mechanism inside a feedback loop that reasons over domain constraints at inference time.
Manuscript, code, figures, and complete collaborator information will be added when they are ready for public release.