A new inference-time alignment method works on diffusion and flow models without ever looking inside them.
A team of researchers has published TRS, a trust-region search algorithm that tunes only the source noise fed into a pre-trained generative model. Unlike most alignment approaches, TRS treats both the generative model and the reward model as black boxes — no gradients required, no architectural assumptions, no expensive memory overhead from backpropagating through long sampling chains. The method balances broad exploration with local refinement and claims minimal hyperparameter tuning. Code is publicly available.
Most inference-time alignment work assumes you can either differentiate through your reward model or afford the compute to backpropagate through hundreds of diffusion steps. TRS sidesteps both constraints, which matters because real-world reward signals — human preference scores, wet-lab binding assays, safety classifiers — are rarely differentiable. A method that treats everything as a black box is easier to slot into existing pipelines and harder to accidentally break when a model is updated.
The authors test TRS on text-to-image generation, molecule design, and protein design, reporting improvements over both the base models and competing noise-optimization methods. That breadth is notable, though independent replication in any one of those domains would be more convincing than three benchmarks in a single paper. Inference-time steering is a crowded field right now, and "significantly improved" is doing a lot of work until someone else runs the numbers.