GenoTrace uses codon-aware watermarks so distilled genome models stay detectable, reaching 94.5% detection in a three-seed experiment
Related research and updates
Synopsis
The work introduces GenoTrace, a codon-aware extension of green-list watermarking that modulates a teacher model's generation bias with two token-level factors—codon position and organism-specific codon usage—so that the resulting synthetic sequences remain auditable after distillation into a smaller student; in a three-seed GenomeOcean-500M-to-100M experiment the joint configuration achieves a mean audit score of 17.88 and 94.5% detection at a fixed threshold, retaining 49.0% detection after key-aware token substitution (versus 0% for the available single-seed plain-watermark comparator) and 47.0% after combined mechanism-targeted nucleotide edits.
Figure 1: GenoTrace framework. A teacher generates watermarked DNA, a smaller student learns from these sequences, and the auditor scores student outputs without an inference-time watermark processor. The surrounding panels identify output-editing tests and computational quality assays. The schematic’s preservation and constraint labels indicate design objectives; measured outcomes are reported in Section 5 .
arXivInterpretation
It proposes GenoTrace, which extends green-list watermarking to the codon level by modulating the teacher's generation bias with two token-level factors, codon position and organism-specific codon usage, embedding an inheritable signal in synthetic sequences. Relative to plain green-list watermarking, the construction brings codon structure of genomic sequences into watermark generation, aligning the watermark with genome-specific token organization. The abstract describes the method-level construction and reports a mean audit score of 17.88 and 94.5% detection at a fixed threshold for the joint configuration in a three-seed GenomeOcean-500M-to-100M distillation experiment.
The watermark signal transfers through distillation to a smaller student model, and student outputs can be audited without an active watermark processor. Prior watermarking work largely focuses on generation-side detection; here detectability is extended to auditing outputs of a distilled downstream model. The three-seed distillation experiment reports 94.5% detection; additional experiments establish inherited signal across five organism-conditioned datasets and teacher-student size ratios up to 40.
The watermark is somewhat robust to key-aware token substitution and mechanism-targeted nucleotide edits. Where the available single-seed plain-watermark comparator shows 0% detection, GenoTrace retains 49.0% detection after key-aware token substitution and 47.0% after combined mechanism-targeted nucleotide edits. The abstract provides these comparator numbers and identifies the comparator as the available single-seed plain-watermark comparison.
Component ablations and computational sequence-quality assays reveal distinct operating points for detection strength and coding coverage. The work characterizes detection strength and coding coverage as separately tunable operating points rather than a single metric. The abstract reports component ablations and computational sequence-quality assays, and notes that calibration and biological utility are treated as separate evaluation requirements.
Perspective
The work targets shared-tokenizer distillation and the tested editing procedures; in that setting GenoTrace keeps synthetic sequences generated by the teacher auditable after training a smaller student, and retains partial detection under key-aware token substitution and mechanism-targeted nucleotide edits. It suits research and engineering settings that need to trace synthetic genomic sequences through a distillation chain, and inherited signal has been established across five organism-conditioned datasets and teacher-student size ratios up to 40. Calibration and biological utility are listed by the authors as separate evaluation requirements, so applicability of the construction presupposes those separate evaluations.
The abstract does not give the distribution of audit scores, the basis for threshold selection, or statistical uncertainty, nor the operational details of key-aware token substitution and mechanism-targeted nucleotide edits; the specific metrics of component ablations and sequence-quality assays are not expanded in the abstract. Calibration and biological utility are listed as separate evaluation requirements whose conclusions await those evaluations. These are directions for continued observation rather than a negative judgment of the work.