Research map / Image generation / Fast diffusion sampling

Budgeted Quotient-Residual Guidance for Frozen Pocket-Conditioned Molecular Diffusion

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes budgeted quotient-residual guidance (QRG), an inference-time correction that lifts quotient covectors to metric-horizontal ambient directions and caps their scale by the frozen sampler's own step norm, improving validity of frozen TargetDiff on CBGBench fragment and scaffold tasks.

Problem

Inference-time guidance for frozen pocket-conditioned molecular diffusion samplers is often defined on quotient (invariant) features, whose raw local gradients are numerically dormant relative to the sampler's ambient step scale.

Approach

Lift quotient covectors to metric-horizontal ambient directions via M^{-1}J^T c, then deliver them with a trust budget equal to a fraction of the frozen sampler's own step norm (with a product budget splitting section and residual controls), so quotient geometry sets direction and sampler motion sets scale.

Key results

• Local-QRG improves TargetDiff fragment growing validity from 0.815 to 0.864 and scaffold hopping from 0.664 to 0.707 (official seed-0).
• Local-QRG improves linker design validity from 0.681 to 0.712, while PredNext-QRG is near-neutral (0.680).
• Novelty stays at 1.000 and diversity is preserved in the matched multi-seed molecular slice; control/base ratio around 0.19-0.27.

Surprising findings

• Raw local quotient gradients are dormant: dormant local guidance leaves success and quotient error essentially unchanged from the base sampler on toy tasks.
• Default ligand-generation budgets over-control side-chain decoration, reducing validity from 0.873 to 0.837-0.848.

Limitations

• Gains are task-dependent; PredNext-QRG is near-neutral on linker design and paired target intervals often cross zero.
• Docking results are mixed and treated only as a supporting diagnostic using raw-pocket AutoDock Vina.
• Side-chain decoration shows a budget-transfer failure where guidance hurts validity.

Experimental setup

Datasets: CBGBench

Benchmarks: CBGBench

Metrics: validity novelty diversity uniqueness QED seconds per sample control/base ratio AutoDock Vina score

Baselines: TargetDiff base sampler Section-only Sham-QRG Rollout-QRG Teacher Dormant local guidance

Builds on: classifier guidance TargetDiff CBGBench

Closest papers by meaning

• On scalable and efficient training of diffusion samplers (NeurIPS 2025 · similarity 0.86)
• Reward-Guided Iterative Refinement in Diffusion Models at Test-Time with Applications to Protein and DNA Design (ICML 2025 · similarity 0.85)
• Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.85)
• Unlocking hidden biomolecular conformational landscapes in diffusion models at inference time (NeurIPS 2025 · similarity 0.85)
• The Superposition of Diffusion Models Using the Itô Density Estimator (ICLR 2025 · similarity 0.85)
• Unified Guidance for Geometry-Conditioned Molecular Generation (NeurIPS 2024 · similarity 0.85)