Expanding the Decision Space of Pretrained Models through Structured Re-Decoding
Summary
- Researchers introduced Structured Re-Decoding (SRD), a training-free framework that constructs and jointly optimizes local candidates from pretrained model scores, samples, or external specifications.
- SRD expands feasible support and improves attainable reward under a fixed KL divergence budget without altering native decoders.
- The framework enables improved structured inference across molecular generation, protein design, RNA structure prediction, and antibody modeling.
- External chemical templates extend SRD to noncanonical amino acids, peptide cyclization, and grammar-guided molecular construction.
- Preliminary tests on a physical coherent Ising machine demonstrate the potential for hardware-based optimization in structured inference.
Why it matters
SRD provides a training-free method to improve structured inference in pretrained generative models, advancing computational biology tasks like molecular generation and protein design.