Frozen models need no retraining: structured re-decoding lifts verified FLOWR delivery from 419 to 1,309
Synopsis
Pretrained generative models are usually screened only over complete outputs already produced, wasting feasible combinations visible in local scores. Structured Re-Decoding (SRD) keeps model parameters fixed while jointly selecting local candidates from scores or samples under task constraints, reconstructing and verifying the result. On FLOWR molecular generation at the same generation budget, single-target composition-and-quality success rises from 419 to 1,309 of 1,800 requests over best-of-4, and dual-target from 119 to 615; the quality pass rate falls from 99.833% to 94.389%.
Interpretation
When a pretrained model must deliver objects under a new structural constraint, it no longer has to pick only among complete outputs already generated; local candidates can be recombined into new complete outputs. The prior route screened a finite pool of complete model outputs, so delivery was limited to objects the model had already produced, and combinations satisfying a new task specification in local scores or observed states could not be delivered. Across 1,800 FLOWR requests (225 pockets, eight seeds, four complete generations per request), single-target composition-and-quality success rises from 419 under best-of-4 to 1,309, and dual-target from 119 to 615, differences of 49.444 and 27.556 percentage points with 95% pocket-cluster intervals [45.167, 53.611] and [24.056, 31.056].
The method splits an output into editable local units, lists candidate states for each unit, then jointly selects them under task constraints and reconstructs the whole object. Candidates come from the frozen model's terminal scores (atom, charge, bond), from aligned sample states, or from external chemical definitions, with model parameters fixed throughout. Selection is written as a quadratic unconstrained binary optimization (QUBO, a 0-1 problem whose pairwise preferences and constraint penalties are quadratic terms); FLOWR compares local and full-domain search on one objective, and their success rates differ by only 0.111 and 0 percentage points.
Reconstructed objects pass through verification, and failures fall back to the original output, so what is delivered has passed chemical and geometric checks. Verification separates encoded feasibility, domain validity and downstream utility; a task policy ranks accepted candidates, and when none is accepted the original output returns with its actual validity flag. Full-domain re-decoding with a first-output fallback passes the quality checks on 94.389% of single-target and 94.111% of dual-target requests, against 99.833% for best-of-4 screening.
Perspective
The demonstrated scope is inference-time re-decoding on frozen model outputs: 1,800 FLOWR requests, six four-site ProteinMPNN domains, 7,215 UFold RNA records and 3,600 nested cases from 50 DiffAb complexes, plus capability studies on noncanonical amino acids, cyclization and peptide grammar. The intended reader is a generative-modeling researcher who wants to impose new structural constraints without retraining.
The FLOWR gain mixes terminal-score candidate construction, composition guidance and verified selection, so attributing it separately needs a target-directed comparator with matched edit permissions; whole-sample screening remains stronger on SemlaFlow. RNA training and development overlap and unresolved DiffAb checkpoint membership limit conclusions about independent generalization, and near-native pose recovery remains unresolved in the cyclization and thioether settings.