---
title: Retrieval-based Prior Selection (RPS)
url: https://www.emergentmind.com/topics/retrieval-based-prior-selection-rps
type: topic
---

# Retrieval-based Prior Selection (RPS)

Retrieval-based Prior Selection (RPS) denotes a family of procedures in which a model does not fix its prior arbitrarily, but instead derives, selects, or reweights prior information from an external evidence source or from a library of prior-like candidates. In the literature considered here, the retrieved object varies by domain: a multivariate normal prior over response-propensity coefficients in responsive survey design, a shrinkage penalty interpreted as a prior in linear models, a repertoire-specific Gaussian-process prior in robotics, or a prompt strategy selected from a predefined pool in information elicitation dialogue. The unifying structure is the use of externally grounded or precompiled prior information to improve inference, prediction, or control when current observations are limited or evolving [1907.06560] [2201.02244] [1907.07029] [2604.13817].

## 1. Scope and conceptual definition

Across these works, RPS is not a single standardized algorithm. This suggests that it is better understood as a cross-domain design pattern: identify candidate priors or prior-like objects, retrieve or score them using task-relevant evidence, and then use the selected object to regularize prediction or action selection.

| Setting | Prior-like object | Selection or construction signal |
|---|---|---|
| Responsive survey design | \( \beta \sim N(\mu,\Sigma) \) | historical quarters or screened literature |
| Linear models | shrinkage penalty / implied prior | perturbation-based predictive stability or GA fitness |
| Robotics | repertoire-specific GP prior | posterior over policy and repertoire, with UCB over repertoires |
| Dialogue elicitation | prompt from a prompt pool | normalized information gain reward |

A terminological complication arises because the acronym **RPS** is itself overloaded. In the dialogue paper, **RPS** stands for **Reinforcement Prompt Selection**, not Retrieval-based Prior Selection. However, that work explicitly frames the prompt pool as a library of prior elicitation behaviors and treats selection as adaptive choice among predefined strategies, making it conceptually adjacent to retrieval- or prior-based selection rather than a purely generative prompting method [2604.13817].

## 2. Informative-prior retrieval in responsive survey design

A direct and explicit instantiation of retrieval-style prior construction appears in Bayesian prediction of **daily response propensity** for **Responsive Survey Design (RSD)**. The target quantity is the probability that sampled unit \(i\) responds at contact attempt \(t\), modeled through a discrete-time hazard / logistic regression,
$$
\log\left(\frac{P(Y_{it}=1\mid X_{it})}{1-P(Y_{it}=1\mid X_{it})}\right)=\beta^\top X_{it},
$$
with predictors drawn from fielding day of the quarter, paradata, sampling frame information, linked commercial data, and manually retained variables; after variable selection, the final model contained **72 coefficients**. The Bayesian formulation places a multivariate normal prior on the coefficient vector,
$$
\beta \sim N(\mu,\Sigma).
$$

The paper evaluates three informative or near-informative strategies plus a baseline. The principal historical-data strategy is **PWP** (precision-weighted prior), which fits the same discrete-time logit model to each of the **eight prior quarters** and combines quarter-specific coefficient estimates and covariance matrices so that more precise quarters receive more weight. The prior mean for coefficient \(c\) is
$$
\mu_c=\frac{\sum_{q=1}^{8}\widehat{\mathrm{Var}(\hat\beta_{qc})}^{-1}\hat\beta_{qc}}{\sum_{q=1}^{8}\widehat{\mathrm{Var}(\hat\beta_{qc})}^{-1}},
$$
and the covariance is based on the inverse of the average precision, with a ridge-type stabilization to ensure invertibility. The simpler **LAST** strategy uses only the immediately preceding quarter,
$$
\mu=\hat\beta_{q-1}, \qquad \Sigma=\widehat{\mathrm{Var}(\hat\beta_{q-1})},
$$
while **LASTZ** removes off-diagonal covariances and uses independent normal priors. The literature-based strategy, **LIT**, searches methodological literature and official documentation, retains **eight documents** after screening, maps analogous variables across studies, and sets
$$
\mu_c=\frac{1}{S_c}\sum_{s=1}^{S_c}\hat\beta_{sc}, \qquad
\Sigma_{cc}=\frac{1}{S_c}\sum_{s=1}^{S_c}\widehat{\mathrm{Var}(\hat\beta_{sc})},
$$
with a diagonal \(\Sigma\) because cross-study covariance information was generally unavailable. If no usable literature evidence exists for coefficient \(c\), the prior is set to \(\mu_c=0\) and \(\Sigma_{cc}=10\). One probit-based study is converted to the log-odds scale by multiplying coefficients and standard errors by **1.61** [1907.06560].

Posterior updating begins after **day 7** of the quarter, using current data through day \(d\) and **PROC MCMC**. Daily predictive probabilities are estimated by Monte Carlo averaging over **5000** posterior draws, and performance is summarized by daily mean difference from a final benchmark prediction, its standard error, and
$$
\mathrm{RMSE}_d=\sqrt{B_d^2+\mathrm{se}(B_d)^2}.
$$
Evaluation uses **five recent NSFG quarters** as target collections and the **eight preceding quarters** as historical evidence. The field period is divided into **Early (days 7–30)**, **Middle (days 31–60)**, and **Late (days 61–84)** windows, and the final hazard models are additionally described with **Nagelkerke pseudo-\(R^2\)**, **Hosmer–Lemeshow** goodness-of-fit, and **AUC**.

The empirical pattern is specific and operationally important. **PWP** generally performed best or among the best, with lower mean differences from the final benchmark earlier in the quarter and faster convergence toward zero. Improvements were especially notable in the **middle period**, when RSD interventions are often considered. **RMSE** also tended to be lower under Bayesian methods than under the standard approach. **LAST** was unstable and often poor, plausibly because a single quarter yielded noisy covariance estimates; **LASTZ** improved stability by removing covariances. **LIT** was competitive in some settings and often better than the non-informative baseline, though less consistently strong than PWP. The paper further notes that the **NSFG** is unusually well suited to the PWP strategy because quarterly design and response behavior are stable over time. In this setting, RPS is most literal: prior evidence is retrieved either from internal longitudinal history or from published external studies and mapped into a prior that improves early and middle-field prediction when current data are still sparse [1907.06560].

## 3. Prior selection through shrinkage equivalence in linear models

A broader theoretical framing comes from linear-model shrinkage, where prior selection is treated as mathematically equivalent to penalty selection. The starting point is a penalized optimization problem,
$$
\hat{\beta}=\arg\min_{\lambda,\beta}\sum_{i=1}^n L_1(y_i-x_i^\top\beta)+\lambda\sum_{j=1}^p L_2(\beta_j),
$$
together with an adaptive version using weights \(w_j\). When the first term is a negative log-likelihood, exponentiation yields a posterior-mode problem with density
$$
\rho(\beta\mid Y^n)\propto e^{-\lambda\sum_{j=1}^p L_2(\beta_j)}\,e^{-\sum_{i=1}^n L_1(y_i-x_i^\top\beta)}.
$$
Thus \(e^{-\lambda L_2(\beta_j)}\) acts as a prior factor for \(\beta_j\), so the penalty is the negative log-prior up to constants. Ridge corresponds to a Gaussian prior, LASSO to a Laplace prior, and nonconvex penalties such as SCAD or MCP to more complicated priors [2201.02244].

The paper argues that asymptotic **oracle property (OP)** considerations are useful but insufficient for practical prior or penalty selection. Its OP formulation requires, for null coefficients \(\beta_2=0\), both \(P(\hat\beta_2=0)\to 1\) and asymptotic normality for active coefficients. Regularity conditions imply familiar scaling requirements: for active coefficients, \(a_n\) must be small enough that \(\sqrt{n}a_n\to 0\); for inactive coefficients, \(b_n\) must be large enough that \(\sqrt{n}b_n\to\infty\). Yet the paper’s core claim is that OP is not decisive because it is asymptotic, many penalties satisfy it under mild conditions, and methods with OP can still predict poorly in finite samples, while some methods without OP can predict well.

The proposed finite-sample criterion is **predictive instability under data perturbation**. Responses are perturbed as
$$
y_i \mapsto y_i+\eta_i,\qquad \eta_i\sim N(0,\tau^2),
$$
followed by train-test splitting, refitting, and evaluation of
$$
S(\hat Y)_\tau=\sqrt{\frac{1}{n_{\text{test}}}\sum_{i\in \mathcal D_{\text{test}}(\tau)}(y_i-\hat Y_\tau(x_i))^2}.
$$
This is averaged over many random splits and perturbations, with a grid \(\tau_k=0.2k\) for \(k=1,\dots,8\). A good method has low instability values, smooth curves, and curves that increase slowly with \(\tau\); a decreasing curve under increasing perturbation is treated as a warning sign of fundamental instability or misspecification.

The simulation recommendations are regime dependent. For \(p=100\) and \(n=40\), where \(n\ll p\), variable selection is poor for all methods and **RR** and **EN** are generally best; under Toeplitz correlation and high sparsity, **SCAD2** and **MCP** can become preferable. For \(n=75\), **EN** is typically best for low or medium sparsity. For \(n=150\), **LM** can do well when sparsity is low and **SCAD1** often becomes very competitive or best. At \(n=500\), many methods are nearly indistinguishable, OP methods are often near-perfect in variable selection, but several non-OP methods still predict well. The methods most often recommended against are **RR**, **LAD-LASSO**, and **ASCAD1**, and sometimes **LM** when sparsity is high. The paper’s real-data superconductivity example shows that these recommendations can shift under model misspecification, where a less sparse linear approximation can outperform a sparser one.

The paper’s second contribution is a **genetic-algorithm (GA)** procedure for learning a custom penalty, hence a custom prior. It uses a shifted penalty,
$$
Q(\beta)=L(\beta\mid x^n)+n\sum_{j=1}^p \lambda_j f_j(\beta_j-\hat\beta_j^*),
$$
with \(\hat\beta^*\) a \(\sqrt n\)-consistent pilot estimate, and searches over a polynomial penalty class
$$
f_j(\beta_j)=\sum_{k=1}^6 \alpha_k\,|\beta_j-\hat\beta_j^*|^k.
$$
In the high-dimensional case \(p\ge n\), the construction reduces to a penalty on \(|\beta_j|^k\). In the \(n=40\) example, the GA returns \(\hat\alpha=(0,1,0,0,0,0)\), exactly recovering ridge regression; in the \(n=150\) example, it returns \(\hat\alpha=(2,0,0,20,0,0)\), a custom penalty that produced the best predictive error among compared methods. Within an RPS interpretation, this work shifts attention from retrieving external empirical summaries to selecting among, or constructing, prior families using predictive stability as the operative criterion rather than asymptotic optimality alone [2201.02244].

## 4. Adaptive prior retrieval in repertoire-based robotics

In robotics, prior selection appears as adaptive choice among multiple **repertoires** learned offline in simulation. The setting is online adaptation under an unknown situation \(c\), such as robot damage, changes in floor friction, or unknown object properties, with task-space dynamics
$$
s_{t+1}=s_t+F(s_t,\pi_\theta,c)+w,
$$
where \(w\) is Gaussian noise and the robot does not know either \(F\) or the current situation \(c\). The robot does, however, have access to a low-fidelity simulator \(\hat F\) and to a subset of likely situations. The paper’s contribution is **APROL** (**Adaptive Prior selection for Repertoire-based Online Learning**), which relaxes the single-repertoire assumption and instead lets the system select the most useful prior online [1907.07029].

APROL has three phases. In offline repertoire generation, **MAP-Elites** is run for each probable situation \(c\in C\), producing repertoires \(\Pi\) containing tuples
$$
(\pi_{\boldsymbol\theta}, \Delta \mathbf{s}_\theta, r_{\boldsymbol\theta}, cell\_id),
$$
and forming a repertoire set
$$
\mathcal{R}=\{\Pi_0,\Pi_1,\dots,\Pi_n\}.
$$
In online learning, each repertoire supplies the prior mean-function for a Gaussian process that maps simulated transitions to real transitions. For dimension \(d\),
$$
P(f_d(\Delta \mathbf{s}_\theta)\mid D_{1:t}^d)=\mathcal{N}(\mu_d(\Delta \mathbf{s}_\theta),\sigma_d^2(\Delta \mathbf{s}_\theta)),
$$
with prior mean
$$
M_d(\Delta \mathbf{s}_\theta)=\Delta \mathbf{s}_\theta[d]
$$
and squared-exponential kernel
$$
k_d(\Delta \mathbf{s}_\theta,\Delta \mathbf{s}'_\theta)=\sigma_{se}^2\exp\left(-\frac{\lVert\Delta \mathbf{s}_\theta-\Delta \mathbf{s}'_\theta\rVert^2}{l^2}\right).
$$

The online selection rule is a MAP decision over both policy and repertoire:
$$
\pi_{\theta_t}^*=
\arg\max_{\pi_\theta\in \Pi,\,\Pi\in\mathcal R}
P(s_{g_t}\mid \pi_\theta,obs_{0:t-1},\Pi)\,P(\pi_\theta)\,P(\Pi\mid obs_{0:t-1}).
$$
Here \(P(\pi_\theta)\) is either uniform over policies or restricted to policies whose simulated transition lies near the desired transition; \(P(s_{g_t}\mid \pi_\theta,obs,\Pi)\) is a Gaussian likelihood favoring policies whose repertoire-specific GP predicts motion toward the current sub-goal; and \(P(\Pi\mid obs)\) measures repertoire compatibility with reality. Compatibility begins with a closeness score
$$
\psi_{\Pi}=e^{-k\lVert \Delta \mathbf{s}_\theta-\Delta \mathbf{s}_{\theta,\text{real}}\rVert^2},\qquad k>0,
$$
then uses **UCB** to balance exploitation and exploration,
$$
\psi_{ucb\Pi}=\frac{\sum \psi_{\Pi}}{N_\Pi}+m\sqrt{\frac{\ln(n)}{N_\Pi}},
$$
followed by normalization across repertoires. The result is a prior-selection mechanism in which repertoire identity is itself a latent variable updated online.

The empirical evaluation covers object pushing with a robotic arm, goal reaching with a damaged hexapod in simulation, and a real damaged hexapod. Simulation comparisons use **40 replicates** and include **CP-L**, **SP-L**, **SP-NL**, and **APROL-NL**. APROL solves both simulated tasks in less interaction time than baselines, matches or approaches the close-prior baseline, and significantly outperforms wrong-single-prior and no-learning variants. In the real damaged-hexapod experiment, the robot reaches the goal with **100% success** within a maximum of **30 steps** and recovers up to **88%** of the intact robot’s capability in time to goal. The robotics case therefore instantiates RPS as retrieval over a discrete prior library, where the retrieved object is not a coefficient prior but a repertoire-specific transition prior for planning and adaptation [1907.07029].

## 5. Prompt-pool selection as a prior-selection analogue in dialogue elicitation

In open-ended dialogue, the paper titled **"RPS: Information Elicitation with Reinforcement Prompt Selection"** defines **RPS** differently: it is a reinforcement learning framework for selecting prompts from a predefined prompt pool to elicit concealed or incompletely expressed user information. The paper explicitly states that it is **not a classic retrieval system**, but it also states that the prompt pool acts like a **library of prior elicitation behaviors**, making the method conceptually related to prior-based prompt selection and retrieval over strategies [2604.13817].

The formalization is sequential. User information is divided into
$$
\mathcal{C}=\{\mathcal{I}^+,\mathcal{I}^0,\mathcal{I}^-\},
$$
representing positive, neutral, and negative information. At dialogue round \(t\), the state is
$$
s_t=(\hat{\mathcal I}_{t-1},\mathcal D_{1:t-1}),
$$
the action is a prompt \(a_t\in\mathcal A\), and the selected prompt conditions the query-generation process,
$$
q_t=\hat Q(a_t,\hat{\mathcal I}_{t-1},\mathcal D_{1:t-1}).
$$
After the user responds, the model extracts new information \(i_t=f_m(v_t)\) and updates \(\hat{\mathcal I}_t\). The reward is a normalized information gain,
$$
r_t=
\frac{d(\hat{\mathcal I}_{t-1},\mathcal I)-d(\hat{\mathcal I}_t,\mathcal I)}
{d(\hat{\mathcal I}_{t-1},\mathcal I)},
$$
designed to stabilize learning and avoid reward vanishing as the elicitation gap shrinks. The paper also gives the policy-gradient style objective
$$
\text{Loss}=-\mathbb E[A\cdot \log \pi_\theta(\mathbf z\mid s)].
$$

The synthetic validation uses a **Gaussian Mixture Model** user-disclosure environment with two settings: an unbiased user, where \(\mathcal P_u=\mathcal I\), and a biased user, where disclosure weights favor positive information. Performance is averaged over **10 independent runs**. The RL querying strategy consistently achieves lower KL divergence than a random querying baseline, remains stable around **0.01 KL divergence** in both settings, and the random baseline worsens under bias, with divergence increasing from **3.36** to **3.51**.

The main benchmark is **IELegal**, derived from the **CAIL2018** Chinese criminal case corpus. It contains **2,000 cases** in two subsets, **IELegal-base** and **IELegal-augment**, each with **500 train / 500 test**. Case entries include case content, case type, plaintiff/defendant, key information, classified information, and negative/positive/neutral subsets. Evaluation uses semantic similarity between cumulative extracted information and a reference fact set via cosine similarity over **Sentence-BERT** embeddings. The baselines include five fixed prompt strategies—**Normal**, **Exploratory Information Gathering**, **Precise Evidence Confirmation**, **Corroborative Detail Verification**, and **Confrontational Disclosure Questioning**—plus **Adaptive Generative Questioning**, **GRIPS**, and **RLPrompt**. On IELegal-base and IELegal-augment, RPS consistently outperforms all baselines by the end of dialogue; **Normal** performs worst; and **GRIPS** and **RLPrompt** do not surpass RPS and often do no better than the original fixed strategies. In this setting, the “prior” is the prompt pool itself: a precompiled set of elicitation strategies from which the policy retrieves the most contextually useful option at each turn [2604.13817].

## 6. Shared structure, boundary conditions, and recurring misconceptions

Despite substantial heterogeneity, the four lines of work share a common architecture. First, there is a **candidate prior space**: earlier survey quarters and screened documents, a menu of penalties or a GA-search class, a library of repertoires, or a prompt pool. Second, there is a **selection signal**: coefficient precision, perturbation-based predictive stability, observed reality-versus-repertoire closeness with UCB, or normalized information gain. Third, the selected prior-like object directly affects downstream prediction or control: posterior response-propensity estimates, regularized linear prediction, policy choice for robotic adaptation, or dialogue strategy selection. This suggests that RPS is less a domain-specific method than a recurrent solution to a common problem: current data alone are not yet sufficient, so prior information must be chosen rather than assumed.

Several misconceptions are explicitly contradicted by the source materials. One is that a **single prior is enough**. APROL is motivated precisely by the failure mode of single-repertoire adaptation when the deployment situation differs too much from the assumed prior, and the survey paper shows that priors based on only the immediately preceding quarter can be unstable. Another is that **asymptotic optimality settles prior choice**. The shrinkage paper argues that the oracle property is too broad and too asymptotic to decide finite-sample predictive performance. A third is that **retrieval must mean document retrieval**. The survey paper uses both internal longitudinal history and published literature; robotics retrieves among repertoires; dialogue retrieves among strategy prompts. Retrieval, in these materials, is therefore object-dependent rather than document-specific [1907.06560] [2201.02244] [1907.07029] [2604.13817].

The same papers also define the principal boundary conditions. Historical-data priors are strongest when there is **substantial historical data from the same survey with stable design and response behavior**; when such history is absent, literature-based priors are a reasonable alternative. If behavior changes over time or there is seasonality, older history may need to be downweighted or replaced by more recent data. In linear models, the best penalty or implied prior depends strongly on the **sample-size-to-sparsity regime**, dependence structure, and model misspecification. APROL assumes access to a **low-fidelity simulator** and a subset of likely situations. Reinforcement Prompt Selection assumes that a **prompt pool is available**, that user disclosure is state-dependent and biased, and that reward can be approximated through divergence or semantic similarity. These are not incidental details; they define where each RPS instantiation is expected to work.

A plausible synthesis is that retrieval-based prior selection is most valuable in problems with **partial observability**, **nonstationary evidence accumulation**, and **meaningful prior heterogeneity**. The survey study shows the benefit when fieldwork is ongoing and managers need reliable early guidance; the shrinkage study shows why finite-sample predictive stability matters when asymptotics are uninformative; the robotics study shows that uncertainty over the correct prior can itself be modeled and updated online; and the dialogue study shows that a library of prior strategies can be adaptively selected to uncover concealed information. In that sense, RPS names a methodological orientation: prior information should be elicited, retrieved, or selected in a task-aware manner, rather than treated as fixed background structure.

Source: https://www.emergentmind.com/topics/retrieval-based-prior-selection-rps