- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
The paper studies how to train the property oracle used for inference-time steering of diffusion and flow models when labelled data are scarce. It proposes the Dynamics-Aware Predictor (DAP), which combines a supervised loss on the small labelled set of clean samples with a loss on unlabelled reverse trajectories of the pretrained model. The second loss requires the oracle’s prediction at each noisy state to match its prediction one reverse step later. To avoid the bias of the single-sample estimate, the target is computed with an exponential-moving-average copy of the network. Convergence is analyzed under linear function approximation. Experiments on a 2-D toy dataset, MNIST, and molecular benchmarks report better steering quality than oracles trained in clean space, on forward-noised data, or on pseudo-labelled trajectories.
Critical strengths
The paper addresses a practically relevant question: how to train the oracle that inference-time steering relies on when labels are scarce, a component that is usually taken for granted. The inter-step consistency constraint is a correct property of the ideal intermediate oracle, and the paper correctly identifies the bias of the single-sample estimator. The evaluation covers classification and regression targets, continuous and discrete generative models, and both particle-based and gradient-based steering, albeit on small-scale benchmarks.
Critical weaknesses
First, the methodological novelty is limited:
The inter-step identity presented as the key insight, , is the standard Doob -transform / Bellman recursion underlying twisted SMC, FK steering, value-based decoding, and diffusion-as-MDP formulations. Several of these works are cited, but the connection is not made. Estimating this fixed point by regression is standard value learning, and DAP is a textbook instance of it: semi-gradient TD(0) with a Polyak-averaged target network, i.e., fitted value evaluation. The paper proposes neither a new target nor a new estimator. The related work omits semi-supervised learning, to which the setting (abundant unlabelled samples, scarce labels) belongs. It also describes SVDD only through its posterior-mean variant, although, to my knowledge, SVDD also learns value functions by Monte Carlo regression on rollouts.
Second, the method and theory do not exploit the structure of the problem:
The reverse chain is finite-horizon and acyclic, with supervision only at . The fixed point can therefore be obtained by backward induction or by Monte Carlo regression onto the endpoint prediction, without bootstrapping. The transition can be sampled repeatedly from the generative model, which makes double sampling or multi-sample targets readily available to remove the bias of Eq. (6). App. A.5 dismisses double sampling by citing RL literature, without an experiment. Theorem 1 restates TD convergence under linear function approximation. It bounds the distance to the solution of the constraint, not to . Because labels exist only on , this solution is not unique on unlabelled trajectory endpoints and depends on the function class. The inter-step loss carries no label information: at convergence, DAP approximates for a clean-space predictor fit on . For the regression tasks, where the reward is the likelihood of a target value, enforcing consistency on the predicted mean does not yield the required potential .
Third, the experiments do not support the claims:
The oracle is trained on unguided prior trajectories but deployed along guided trajectories. Prediction quality is evaluated only on unguided trajectories, leaving this distribution shift untested, and no analysis of error propagation over 200–1000 bootstrapped steps is given. Key ablations are missing: the non-anchored estimator (discussed only theoretically in App. A.3), double sampling, Monte Carlo regression onto DAP’s own endpoint prediction, and backward induction. The necessity of the anchored-target scheme, the main technical component, is therefore not established. The comparison with IDP is confounded by the different endpoint predictors, and the gap shrinks as grows (Fig. 3). Standard baselines are absent: semi-supervised methods, and a noisy classifier trained with all labels as in standard classifier guidance. The “universe” oracle is applied through the posterior mean , which makes it a weak upper reference, and it also serves as the evaluator for reward and SWD. All benchmarks are low-dimensional (2-D data, MNIST, SMILES with 2-layer MLPs, a 1024-dimensional VAE latent), with at most 3000 trajectories. Label scarcity is simulated: MNIST is fully labelled, and the molecular properties are computed with RDKit.
Minor:
Eq. (3) uses “” for a constraint and the same for all time steps. “Upper-bounds” is used throughout to mean “performs best”.
Questions for authors
None