DriftLite: Lightweight Drift Control for Inference-Time Scaling of Diffusion Models

ICLR 2026 Conference SubmissionAnonymous Authors
Diffusion ModelInference-Time ScalingVariance ReductionSequential Monte CarloGuidance
Abstract:

We study inference-time scaling for diffusion models, where the goal is to adapt a pre-trained model to new target distributions without retraining. Existing guidance-based methods are simple but introduce bias, while particle-based corrections suffer from weight degeneracy and high computational cost. We introduce DriftLite, a lightweight, training-free particle-based approach that steers the inference dynamics on-the-fly with provably optimal stability control. DriftLite exploits a fundamental degree of freedom in the Fokker-Planck equation between the drift and particle potential, and yields two practical instantiations: Variance- and Energy-Controlling Guidance (VCG/ECG) for approximating the optimal drift with modest and scalable overhead. Across Gaussian mixture models, particle systems, and large-scale protein-ligand co-folding problems, DriftLite consistently reduces variance and improves sample quality over pure guidance and sequential Monte Carlo baselines. These results highlight a principled, efficient route toward scalable inference-time adaptation of diffusion models.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

DriftLite proposes a lightweight particle-based framework for inference-time adaptation of diffusion models, exploiting a degree of freedom in the Fokker-Planck equation to derive Variance- and Energy-Controlling Guidance (VCG/ECG). The paper sits within the 'Kernel Density and Ensemble-Based Steering' leaf of the taxonomy, which contains only two papers total: DriftLite itself and one sibling work on kernel density steering. This sparse population suggests the specific approach of ensemble-based steering with provable stability control represents a relatively underexplored niche within the broader particle-based methods landscape.

The taxonomy reveals that DriftLite's parent branch, 'Particle-Based and Sequential Monte Carlo Methods', contains four distinct leaves spanning Sequential Monte Carlo frameworks, Feynman-Kac steering, Particle Gibbs sampling, and the kernel density approaches. Neighboring branches include Search-Based methods (exploring noise trajectories or latent spaces) and Guidance-Based alignment (using reward gradients or flow maps). DriftLite's focus on lightweight ensemble aggregation with stability guarantees distinguishes it from heavier SMC frameworks with importance weighting and from gradient-based guidance methods that introduce known biases, positioning it at the intersection of statistical rigor and computational efficiency.

Among the three identified contributions, the literature search examined nineteen candidates total. The 'Fundamental degree of freedom in Fokker-Planck equation' contribution examined two candidates with zero refutations, suggesting this theoretical insight may be novel within the limited search scope. The 'DriftLite framework with VCG/ECG instantiations' examined ten candidates with no refutations, indicating the specific algorithmic design appears distinct from prior ensemble methods. The 'Provably optimal stability control' contribution examined seven candidates with zero refutations, though the limited search scale means comprehensive prior work on stability guarantees in particle-based diffusion steering may exist beyond these candidates.

Based on the limited search scope of nineteen semantically similar papers, DriftLite appears to occupy a relatively sparse research direction within particle-based inference-time scaling. The absence of refutations across all contributions suggests novelty, though the small candidate pool and narrow taxonomy leaf population mean this assessment reflects top-K semantic matches rather than exhaustive field coverage. The work's positioning between theoretical guarantees and practical efficiency may represent a genuine gap, but broader literature beyond these candidates could reveal additional related efforts.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
19
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: inference-time scaling of diffusion models. The field explores how to improve diffusion model outputs by investing additional computation at test time rather than solely during training. The taxonomy reveals a rich landscape organized around distinct strategic approaches. Search-Based methods (e.g., Classical Search[1], Latent Beam Search[12]) explore multiple candidate trajectories or noise sequences to identify high-quality samples. Particle-Based and Sequential Monte Carlo techniques leverage ensemble diversity and resampling to refine generation. Guidance-Based alignment methods steer outputs toward desired properties using reward models or classifiers (RL Guidance[19], DiffPO[50]). Iterative Refinement and Resampling branches focus on multi-pass correction, while Adaptive and Dynamic approaches adjust computational budgets on the fly. Domain-Specific branches address specialized settings like video (Video-t1[7]) or 3D (ITS3D[45]), and Efficiency techniques (KV Caching[27], Distrifusion[41]) aim to reduce overhead. Theoretical work on Scaling Laws[18] and emerging Diffusion Language Models[36] round out the taxonomy, alongside Verifier-Free intrinsic reasoning and Higher-Resolution self-cascade methods (Self-cascade[30]). Recent activity highlights contrasts between explicit search over noise trajectories versus implicit ensemble steering. Search methods like Noise Trajectory Search[14] and Dynamic Search[9] systematically explore the latent space, trading off exploration and exploitation, while particle-based approaches such as Particle Gibbs[8] and Kernel Density Steering[34] maintain diverse sample pools and use statistical reweighting. DriftLite[0] sits within the Particle-Based branch under Kernel Density and Ensemble-Based Steering, closely neighboring Kernel Density Steering[34]. Its emphasis on lightweight ensemble aggregation contrasts with heavier search frameworks like Classical Search[1] or the iterative correction loops in Iterative Refinement[35]. This positioning suggests DriftLite prioritizes efficient, statistically grounded steering over exhaustive trajectory enumeration, aligning with broader trends toward balancing quality gains and computational cost in inference-time scaling.

Claimed Contributions

Fundamental degree of freedom in Fokker-Planck equation

The authors identify and formalize a degree of freedom in the Fokker-Planck equation that allows introducing a control drift to offset the reweighting potential. This trade-off can be exploited to actively minimize particle weight variance, addressing the weight degeneracy problem in particle-based methods.

2 retrieved papers
DriftLite framework with VCG and ECG instantiations

The authors propose DriftLite, a training-free inference-time scaling framework that computes control drift dynamically to stabilize diffusion sampling. Two practical algorithms, VCG and ECG, are derived that require only solving a small linear system at each time step, making the approach computationally efficient.

10 retrieved papers
Provably optimal stability control for inference-time scaling

The method provides a principled approach to inference-time adaptation of pre-trained diffusion models with theoretical guarantees on stability. It addresses the weight degeneracy problem in particle methods while maintaining mathematical rigor and avoiding the bias of pure guidance methods.

7 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Fundamental degree of freedom in Fokker-Planck equation

The authors identify and formalize a degree of freedom in the Fokker-Planck equation that allows introducing a control drift to offset the reweighting potential. This trade-off can be exploited to actively minimize particle weight variance, addressing the weight degeneracy problem in particle-based methods.

Contribution

DriftLite framework with VCG and ECG instantiations

The authors propose DriftLite, a training-free inference-time scaling framework that computes control drift dynamically to stabilize diffusion sampling. Two practical algorithms, VCG and ECG, are derived that require only solving a small linear system at each time step, making the approach computationally efficient.

Contribution

Provably optimal stability control for inference-time scaling

The method provides a principled approach to inference-time adaptation of pre-trained diffusion models with theoretical guarantees on stability. It addresses the weight degeneracy problem in particle methods while maintaining mathematical rigor and avoiding the bias of pure guidance methods.