Guidance

Guidance changes the sampling process so generated outputs better satisfy a condition, classifier, score, property, or constraint.

A generic guided sampler modifies an update direction:

$$d_{\mathrm{guided}}=d_{\theta }(x_t,t,c)+\lambda g(x_t,c)$$

$d_{\theta }$ is the base model direction, $g$ is a guidance signal, and $\lambda$ controls guidance strength.

Common Forms

Form	Signal	Main Risk
classifier guidance	gradient or score from classifier/property predictor	guide exploits classifier shortcuts
classifier-free guidance	conditional and unconditional model predictions	high scale reduces diversity or creates artifacts
energy or score guidance	external energy, reward, docking score, or property score	proxy hacking and distribution shift
constraint guidance	syntax, graph, geometry, validity, or hard rule	narrows sample space and changes novelty claim
human or tool guidance	verifier, reranker, search, or external evaluator	hidden best-of-n and selection bias

Classifier-free guidance is often written as:

$${\widehat{ϵ}}_{\mathrm{cfg}}=(1+w){\widehat{ϵ}}_{\theta }(x_t,t,c)-w{\widehat{ϵ}}_{\theta }(x_t,t,\varnothing )$$

where $w$ is the guidance scale. Equivalent forms are used for denoised sample, score, or velocity prediction depending on the model.

Tradeoff

Guidance can improve condition satisfaction but reduce diversity or introduce predictor shortcuts. A generated sample can satisfy the guide while failing the true task.

Claim Boundary

Claim	Required Evidence
better controllability	condition satisfaction vs diversity curve over guidance scale
better utility	downstream utility measured after counting all rejected candidates
better sampler	same base model and matched NFE/candidate/filtering budget
better molecular generation	validity, novelty, diversity, property/interaction utility, and proxy bias checks
better structure generation	geometry validity, clash/constraint checks, coordinate-frame and symmetry handling

Leakage and Proxy Risks

Guidance is only as trustworthy as the guide. For a guide $r_{\psi }(x,c)$:

$$x^{\star }=\arg \underset{x\in \mathcal{C}}{\max }{r}_{\psi }(x,c)$$

can select artifacts if $r_{\psi }$ was trained on leaked labels, near-duplicates, template-derived structures, or a benchmark-specific proxy. A paper should state the guide training data, split boundary, and whether the final metric is independent of the guide.

Checks

• What signal guides sampling?
• Is the guide trained on data that leaks test labels or targets?
• Does increasing guidance improve validity but reduce diversity?
• Is the guide aligned with the final evaluation metric?
• Are rejected or filtered samples included in efficiency and validity reports?
• Is the same guide, scale, and filtering rule used for all baselines?
• Is the guide differentiable, black-box, learned, physics-based, or rule-based?
• Does the paper report an unguided baseline and a guidance-scale sweep?

Related

• Conditional generation
• Sampling
• Diffusion model
• Flow matching
• Generation evaluation
• Metric selection
• Robustness
• Scoring function