Reference-Mean Guidance

• Reference-Mean Guidance is a methodological principle that aggregates the average from multiple reference entities to guide predictions and regularize decision-making.
• It is applied in generative modeling, risk calibration, portfolio optimization, and clinical trial imputation to boost stability, fairness, and interpretability.
• Practical implementations balance computational efficiency and hyperparameter tuning while addressing challenges like noisy references and late-time instability.

Reference-Mean Guidance is a methodological principle and set of concrete algorithms for steering inference, learning, or decision-making by aggregating information or conditioning from a collection of reference entities. The defining feature is the use of the mean or expectation across a selected set of references—instances, classes, or examples—to inform predictions, regularization, adaptation, or imputation. This concept arises independently in domains including generative modeling, statistical calibration, portfolio optimization, and clinical trial analysis, where it provides a robust and interpretable alternative to methods that depend on single-reference or group decisions.

1. Conceptual Foundations

Reference-Mean Guidance (RMG) unites a collection of strategies wherein the mean value, effect, or score over a set of reference points is used to guide model predictions, transform target distributions, or regularize decisions. Unlike approaches anchored in a single group, anchor, or class, RMG explicitly blends information from multiple relevant references or conditions, thereby mitigating the instabilities, ambiguities, and fairness faults inherent to single-reference logic (Hu, 2024).

RMG principles appear in:

• Distribution matching for generative models via endpoint means or distributional moments (Curvo et al., 11 May 2026, Sani et al., 13 Jan 2026).
• Risk scoring and calibration through averaging conditional probabilities across multiple groupings (Hu, 2024).
• Mean-variance optimization with regularization to a reference portfolio policy (Deng et al., 31 May 2026).
• Reference-based rules for missing data imputation in longitudinal clinical settings (Wolbers et al., 2021).

This aggregation protects against over-specialization and the fallacy of "single-class inference," ensuring robustness to data or reference variability, and it often fosters alignment with domain-specific or normative desiderata.

2. Formalizations and Mathematical Structures

The technical formulation of Reference-Mean Guidance varies across applications:

Flow-Based Generative Modeling

In deterministic flow matching, the velocity field admits direct control by shifting the endpoint conditional mean. For a frozen model with endpoint mean $\mu_t^\theta(x)$, steering toward a mean defined by a reference bank, $\mu_t^{(\rho)}(x)$, yields the guided velocity: $$u_t^\pi(x) = u_t^\theta(x) + \frac{\beta_t}{1-t} (\mu_t^{(\rho)}(x) - \mu_t^\theta(x))$$ where the empirical reference mean is: $$\mu_t^{(\rho)}(x) = \sum_{n=1}^M w_t^n(x) r^n,\quad w_t^n(x) = \mathrm{Softmax}_n\left[-\frac{\|x - t r^n\|^2}{2(1-t)^2}\right]$$ (Curvo et al., 11 May 2026).

Diffusion Model Distribution Adaptation

In MMD Guidance, alignment with a reference distribution $Q$ is enforced by augmenting the reverse diffusion step with the gradient of Maximum Mean Discrepancy: $$g_t^{(i)} = \frac{2}{B^2}\sum_{j=1}^B \nabla_x k(x_t^{(i)}, x_t^{(j)}) - \frac{2}{B N_r} \sum_{j=1}^{N_r} \nabla_x k(x_t^{(i)}, x_j^{(r)})$$ with updates

$$x_{t-1}^{(i)} = \widehat{x}_{t-1}^{(i)} - \lambda_t g_t^{(i)}$$

(Sani et al., 13 Jan 2026).

Statistical Calibration and Risk Estimation

For scoring functions $S(X)$ and a collection of reference classes $\{G_r\}_{r\in R}$: $$S^*(X) = \frac{1}{|R(X)|} \sum_{r\in R(X)} \mu_r(S(X)),\qquad \mu_r(s) = P(Y=1 \mid S(X)=s, G_r(X)=1)$$ This blends group-conditional frequencies, providing a multicalibrated risk estimate (Hu, 2024).

Portfolio Optimization

Mean-variance optimization with a reference-regulation penalty incorporates a mean-shift toward a reference policy $\mu_t^{(\rho)}(x)$0 at each period: $\mu_t^{(\rho)}(x)$1 (Deng et al., 31 May 2026).

Clinical Reference-Based Mean Imputation

In deterministic conditional mean imputation with reference-based (e.g., jump-to-reference, copy-reference) rules, the imputation mean for missing values post-intercurrent event is formulated as: $\mu_t^{(\rho)}(x)$2 where $\mu_t^{(\rho)}(x)$3 encodes the reference-based mean trajectory (Wolbers et al., 2021).

3. Algorithmic Instantiations

Practical RMG algorithms are tailored to their context but share core motifs:

• Generative flow and diffusion models utilize closed-form or gradient-based updates derived from reference or mixture means, with guidance weights scheduled (often quadratically or clipped) to modulate adaptation intensity (Curvo et al., 11 May 2026, Sani et al., 13 Jan 2026).
• Calibration and scoring systems aggregate group-conditional probabilities for each instance, optionally using sample-size or reliability-weighted means (Hu, 2024).
• Portfolio optimization implements reference penalties in feedback policy computations, often via backward dynamic programming recursion or solution of regularized Riccati equations (Deng et al., 31 May 2026).
• Imputation for longitudinal data involves deterministic conditional mean fill-in using reference-based trajectory means combined with jackknife variance estimation for valid inference (Wolbers et al., 2021).

In conditional diffusion for text-to-image, "Reference-Mean Guidance" (Editor’s term; called "scheduled guidance" in (Kim et al., 2024)) decomposes denoiser predictions into independently-weighted contributions from text and reference modalities, scheduled over diffusion steps to prioritize shape/layout early and semantics late.

4. Empirical and Theoretical Assessment

Reference-Mean Guidance demonstrates both empirical gains and theoretical justification:

• In flow-matching generative models, RMG delivers compositional control (color, structure, identity) with no runtime or sample fidelity costs versus baselines, achieving +11.07 GenEval points overall and +28.75 in position attribute (Curvo et al., 11 May 2026).
• MMD Guidance in diffusion models yields lowest Fréchet and kernel distances for synthetic data and up to 50% FD reductions for real-image domain adaptation with preservation of sample coverage and density; it operates without any per-user fine-tuning or parameter storage (Sani et al., 13 Jan 2026).
• In calibration and risk scoring, RMG is the only general remedy to the "reference class problem"—guaranteeing no instance’s score is the product of a single group’s conditional, but rather a robust mean over all relevant groupings (multicalibration) (Hu, 2024).
• Reference-regulated mean-variance portfolios in high-dimensional regimes outperform standard and shrinkage Markowitz baselines, attaining higher out-of-sample Sharpe ratios (10–50 basis points per month) and lower turnover; the benefit intensifies for long investment horizons where multiperiod feedback amplifies estimation noise (Deng et al., 31 May 2026).
• In clinical trials, deterministic reference-mean imputation supplies effect estimates and confidence intervals coinciding with infinite-sample multiple imputation, but without Monte Carlo error, with frequentist validity, and perfect replicability (Wolbers et al., 2021).
• Reference-mean guidance in text-to-image diffusion (RefDiffuser) yields parameter-efficient adaptation, outperforming alternative methods in OCR accuracy (up to 64.6% for Latin and 29.2% for non-Latin scripts), and provides effective control for complex compositional and style-transfer prompts (Kim et al., 2024).

5. Applications Across Domains

RMG principles have been adapted and validated in a diverse array of settings:

• Image generation: Steering frozen generative models to exhibit user-specified attributes, styles, or compositional layouts using example banks, without retraining.
• Calibration/fairness: Ensuring equitable risk score interpretation by averaging across multiple relevant group memberships or reference classes.
• High-dimensional portfolio construction: Stabilizing multiperiod financial strategies by penalizing deviation from stable or normative reference allocations.
• Clinical trial missing data: Imputing post-intercurrent event outcomes using means drawn from reference hypothetical trajectories (copy-reference, jump-to-reference, washout), enabling precise estimand targeting.

Tables summarizing empirical impact and domain adaptation:

Domain	RMG Formulation	Key Empirical Result
Flow-Matching Generation	Endpoint mean shift	+11 GenEval; compositional ctrl
Latent Diffusion	MMD guidance gradient	-30–50% FD, high coverage
Calibration/Fairness	Group-conditional mean aggregation	Reduces reference class bias
Portfolio Optimization	Mean-variance reference penalty	+10–50 bps Sharpe improvement
Clinical Trial Imputation	Reference-mean conditional mean	Frequentist validity, no MC err

A plausible implication is that domains facing instability, error amplification, or fairness ambiguity due to the arbitrariness of reference choice may benefit from adopting RMG logic, particularly for high-dimensional, dynamic, or compositional tasks.

6. Limitations, Recommendations, and Extensions

While RMG achieves robust adaptation, several practical and theoretical considerations are documented:

• Reference quality is crucial; noisy or biased reference banks propagate undesirable artifacts or distortions (Curvo et al., 11 May 2026).
• Computational burden can become significant at scale—computing pairwise means/gradients or group-conditional probabilities requires efficient caching, subsampling, or approximate nearest-neighbor retrieval (Curvo et al., 11 May 2026, Hu, 2024).
• Hyperparameter tuning, especially for step schedules and guidance strengths, is frequently cited, with recommended quadratic or decaying schedules and empirically set weight decay or softmax temperature (Kim et al., 2024, Curvo et al., 11 May 2026).
• Late-time instability, e.g., the $\mu_t^{(\rho)}(x)$4 factor, requires stepwise clipping or schedule design (Curvo et al., 11 May 2026).

Future extensions under active exploration include adaptive schedule learning, combined multi-modal or hybrid guidance systems, continual and streaming reference adaptation, online dynamic reference selection, and deeper analysis of convergence, finite-sample bias, and multi-attribute reference bank interactions (Curvo et al., 11 May 2026, Kim et al., 2024, Hu, 2024).

Reference-Mean Guidance thus constitutes a flexible, mathematically principled paradigm for incorporating reference data directly and robustly into inference, optimization, and decision-making mechanisms across modern machine learning and statistical settings.