When Do Protein Designs Commit? Interpreting Intermediate States During Generation

Generative models can form different properties of an output at different stages of generation. Understanding when those properties become determined during the denoising trajectory can reveal when there is enough information to make a useful decision, and therefore when inference-time interventions may be most useful. We study this question in protein binder design using Proteína-Complexa, an atomistic flow model that generates candidate binders for a target protein. We analyze 19 target proteins at 10 points along its 400-step denoising trajectory. We define "commitment" as the fraction of a final design quality already determined by an intermediate state, and measure it across protein fold identity, structural confidence, designability, and interface quality. Averaged across targets, these qualities appear to settle in a common progression, but this ordering varies substantially across individual target proteins. Linear readouts from Complexa’s intermediate activations predict final outcomes better than readouts from the decoded partial structure. This shows that the model’s internal state carries more information about the final design than the partial structure does at the same stage. This raises the possibility that inference-time scaling could decide what it optimizes for over the course of generation. Our initial test of choosing different selection points for different target proteins did not show a clear improvement, but the broader results suggest that such methods could use different design qualities at different stages rather than the same objective throughout. Code and results are at https://anonymous.4open.science/r/complexa-quality-dynamics-61DE.