Prospector Fits: Bayesian & ML Applications

• Prospector fits are probabilistic modeling pipelines using Bayesian inference and rigorous sampling to estimate physical parameters via forward modeling.
• They integrate nonparametric SFH, informed priors, and nested sampling to resolve parameter degeneracies in astrophysical SED analysis.
• Applications extend to ML attributions and LLM data selection, featuring efficient kernel fittings and probabilistic data scoring pipelines.

Prospector fits refer to the set of fitting methodologies, numerical implementations, and practical pipelines associated with the Prospector software framework. This term encompasses the Bayesian inference of physical parameters via forward modeling across domains including galaxy spectral energy distributions (SEDs), the selection of high-quality data for LLM fine-tuning, and modality-agnostic interpretable learning heads for attribution tasks. Prospector fits typically denote fully probabilistic modeling, prior specification, and rigorous sampling strategies designed for moderate- to high-dimensional parameter spaces. The software variants (Prospector-α, Prospector-β, Prospector Heads, SuperNUGGETS) demonstrate the breadth of application, from astrophysical SEDs to machine learning datasets.

1. Bayesian Forward Modeling in Astrophysical SED Fitting

Prospector fits in galaxy SED inference are based on on-the-fly forward modeling within a fully Bayesian framework. Physical and nuisance parameter vectors θ—including stellar mass, star formation history (SFH), metallicity, dust attenuation, AGN contribution, and nebular emission—are mapped to a synthetic rest-frame spectrum $L_\lambda(\theta)$ using underlying stellar population synthesis codes such as FSPS. Model photometry or spectra are projected into observed data space via filter convolutions or instrumental calibrations (Johnson et al., 2020).

The joint posterior is constructed as

$P(\theta\,|\,D) \propto P(D\,|\,\theta)\,P(\theta)$

with a Gaussian likelihood for photometry and/or spectroscopy, and flexible, domain-informed priors (uniform, log-uniform, Student-t, Beta) per parameter. Prospector supports parametric and non-parametric SFH representations, e.g. piecewise constant bins with continuity priors, and two-component dust laws (Charlot & Fall 2000). Nebular emission lines are generated and optionally analytically marginalized.

Sampling uses both ensemble MCMC (emcee) and nested sampling (dynesty), with full chain diagnostics and convergence monitoring. Output products include marginalized credible intervals for physical parameters, evidence estimates, and posterior predictive SED envelopes (Leja et al., 2016).

2. Nonparametric SFH and Model Component Parameterization

In Prospector-α, SFH is modeled non-parametrically via N-bin piecewise-constant SFRs, with bin weights $f_n$ subject to a Dirichlet prior. Mass formed, SFR, and sSFR are derived per bin, with priors on $f_n$ equivalent to a uniform Dirichlet in N=6 dimensions, yielding approximately Gaussian priors on log sSFR. Strong constraints are recovered for the youngest SFH bins and total $M_\ast$, while ancient bins are prior-dominated in the absence of data (Leja et al., 2016).

Dust attenuation combines birth-cloud and diffuse ISM screens with a variable-slope law and Drude UV bump, while emission-line predictions use FSPS+Cloudy models. Metallicity is fitted as a single, time-independent Z*, with quadratic interpolation to avoid spurious spectral features. Inference is fully Bayesian, yielding error bars on physical parameters and synthetic observables (e.g. H$\alpha$ luminosity, Balmer decrements, D$_{\mathrm{n}}$4000, PAH mass fractions). Overall, Prospector-α achieves unbiased SED fits and realistic uncertainties across UV–MIR, subject to caveats at $z>1$ or in the presence of strong AGN (Leja et al., 2016).

3. Fit Quality, Constraints, and Degeneracy Resolution

In combined photometry + spectroscopy fits, Prospector achieves reduced $\chi^2_\nu\approx1.15$ for high-S/N galaxy spectra (LEGA-C z∼1), demonstrating excellent fit quality (Nersesian et al., 5 Feb 2025). Bayesian evidence stabilizes upon convergence in nested sampling fits. Spectroscopy substantially tightens constraints compared to photometric-only runs: age and metallicity uncertainties are reduced by factors ∼1.5–3.

When photometry alone is fitted with a flat log$Z_\ast$ prior, metallicities are biased downward by $\Delta\approx-0.47$ dex. Switching to a linear $Z_\ast$ prior largely eliminates this bias ($\Delta\approx-0.03$ dex, scatter $\approx$ 0.16 dex), a direct fit result (Nersesian et al., 5 Feb 2025). Posteriors are diagnostic of parameter degeneracies (age–dust–metallicity), best resolved with spectroscopic constraints.

Mass-weighted and light-weighted ages, as well as metallicities, are derived from SFH and luminosity integrations:

$$\langle t \rangle_{MW} = \frac{\int t\,\mathrm{SFR}(t)\,dt}{\int \mathrm{SFR}(t)\,dt}$$

$$\langle t \rangle_{LW} = \frac{\int t\,L(t)\,dt}{\int L(t)\,dt}$$

Scaling relations for ages and metallicity versus stellar velocity dispersion were robustly established, with detailed empirical fit coefficients (Nersesian et al., 5 Feb 2025).

4. Prospector Fits in Photometric Redshift Estimation

Prospector-β extends Prospector fits to galaxies with unknown redshift by introducing empirically motivated joint priors:

• $P(M_\ast\mid z)$ based on the double Schechter stellar mass function,
• $P(z)\propto N(z)\,dV/dz$ with mass-completeness cut,
• Dynamic SFH prior tied to cosmic SFRD, shifted by galaxy mass for downsizing (Wang et al., 2023).

Bayesian inference proceeds via nested sampling (dynesty), returning the full joint $P(z, \theta)$ posterior, from which marginalized and joint distributions of all physical parameters can be derived. Mean bias error in mass and age is reduced (from 0.3 to 0.1 dex, and 0.6 to 0.2 dex respectively), and photo-z outlier fractions decrease versus uniform priors or standard codes (EAzY) (Wang et al., 2023). Non-Gaussian photo-z uncertainties are propagated into all population summaries.

5. Prospector Fits for Data Selection and Feature Attribution in ML

In ML, "Prospector Fits" denote the training and inference steps involving Prospector Heads, specialized attribution modules for modalities including text, images, and graphs (Machiraju et al., 2024). The fitting procedure consists of:

• K-means quantization of encoder token embeddings,
• Rollup of monogram and skip-bigram concept counts per receptive-field,
• Fitting a kernel via elastic-net logistic regression or fold-change scoring,
• Efficient O(T) per datum evaluation, superior to SHAP/DASP for high-dimensional attributions.

Hyperparameters include concept count ($k$), neighborhood radius ($r$), regularization strength ($\lambda$), and fold-change thresholds ($\tau$, $\alpha$). Fit selection is driven by training set precision, Dice coefficient, MCC, and AUPRC. Numeric gains measured on WikiSection, Camelyon16, and MetalPDB datasets reach $+8.5$ to $+49$ points over baselines (Machiraju et al., 2024). The learned kernel $\omega$ provides direct visualization of class-driving concepts and patterns.

6. Efficient Data Prospecting with SLM-based Scoring

In LLM tuning, SuperNUGGETS applies a small LLM (SLM) as a Data Prospector to rank instruction data for efficient selection. The fit pipeline includes:

• Test set refinement by reward score and k-center_greedy diversity,
• Golden Score calculation comparing SLM's zero-shot and one-shot log-probabilities,

$\mathrm{GS}(z_k)=\frac{1}{m}\sum_{j=1}^m \mathbbm{I}[s^j_{one}(z_k)>s^j_{zero}]$

• Sorting by GS and thresholding to select top $n\%$ for LLM fine-tuning,
• Downstream utility measured as win_rate change on Alpaca-Eval (Ni et al., 2024).

Compared to the earlier NUGGETS system (LLM-based), SuperNUGGETS achieves a $\sim$58-fold efficiency gain with only a 1–2% utility drop (win_rate). Fitting uses off-the-shelf reward models (DeBERTa-v3-large), SLMs (OPT-125M/350M/Llama2-7B), and standard LLM fine-tuning hyperparameters (Ni et al., 2024).

7. Legacy: PROSPECTOR Fit Accuracy in Uncertain Inference

Historically, the PROSPECTOR system for uncertain reasoning used ad-hoc piecewise-linear update rules with explicit combination formulas (AND, OR, Independence). In fits to simulated evidence networks, PROSPECTOR's Independence rule yielded average errors of $0.014$ (independent evidence) and $0.022$ (associated evidence), with best-case performance in the majority of normative tests (Yadrick et al., 2013). However, errors grow sharply with strong evidence-conclusion coupling, and accuracy rapidly deteriorates for irregular conditional patterns or endpoint evidence reports. The Independence rule is empirically safest, but full Bayesian updating is needed for quantitative accuracy (Yadrick et al., 2013).

Summary Table: Prospector Fit Modalities and Domains

Modality	Fit Methodology	Primary Outputs/Results
Galaxy SED	Bayesian, Nested/MCMC	Posterior $M_*$, SFH, $Z_*$, dust, AGN
Photometric Redshift	Nested sampling + informed priors	Joint $P(z, M_*, \mathrm{SFH})$, bias, NMAD
ML Attribution	K-means + regression/fold-change	Attribution maps, kernel $\omega$, AUPRC gain
Data Prospecting for LLM	SLM-based GS scoring	Top-$n\%$ selections, utility, efficiency

Prospector fits thus denote a rigorous, probabilistic approach to physical property inference, data selection, and feature attribution, with domain-specific priors and highly efficient sampling strategies yielding validated and unbiased results across disparate scientific and data-driven domains.