Prediction-Powered Risk Monitoring of Deployed Models for Detecting Harmful Distribution Shifts

TL;DR

This work tackles the problem of detecting harmful distribution shifts in deployed models when labeled feedback is scarce. It introduces Prediction-Powered Risk Monitoring (PPRM), a semi-supervised extension of supervised risk monitoring that uses prediction-powered inference (PPI) to combine labeled data with unlabeled data via synthetic labels, yielding unbiased risk estimates and tighter, anytime-valid bounds. The framework provides rigorous probability-of-false-alarm guarantees and adaptively tunes reliance on unlabeled data through a data-driven hyperparameter $\eta_t$, demonstrated across image-classification, LLM monitoring, and telecom tasks. Practically, PPRM enables earlier and more reliable alarms with reduced labeling burden, supporting on-device maintenance and safer deployment in dynamic environments, while offering a clear path for future integration with unsupervised adaptation techniques.

Abstract

We study the problem of monitoring model performance in dynamic environments where labeled data are limited. To this end, we propose prediction-powered risk monitoring (PPRM), a semi-supervised risk-monitoring approach based on prediction-powered inference (PPI). PPRM constructs anytime-valid lower bounds on the running risk by combining synthetic labels with a small set of true labels. Harmful shifts are detected via a threshold-based comparison with an upper bound on the nominal risk, satisfying assumption-free finite-sample guarantees in the probability of false alarm. We demonstrate the effectiveness of PPRM through extensive experiments on image classification, large language model (LLM), and telecommunications monitoring tasks.

Figures (7)

• Figure 1: Over a discrete-time index, $t=1,2,...,$, a deployed system is monitored to detect harmful data distribution shifts that cause the running risk $\bar{R}_t$ to exceed the nominal risk $R_0$ by more than a maximum tolerated value $\epsilon_0$. Supervised risk monitoring (SRM) assumes access to labeled calibration dataset $\mathcal{D}_t$ for $t=0,1,...$ (with $t=0$ corresponding to nominal behavior) of the form $(x,y)$, where $x$ is the input and $y$ is the true label. In contrast, the proposed prediction-powered risk monitoring (PPRM) leverages both labeled data and unlabeled data $\tilde{\mathcal{D}}_1,\tilde{\mathcal{D}}_2,...$, including inputs $\tilde{x}$ only, by integrating an auxiliary pre-designed function $f_{\text{p}}(\cdot)$, yielding synthetic labels $\tilde{y}=f_{\text{p}}(\tilde{x})$.
• Figure 2: Risk estimates as a function of time $t$ and average time to alarm for an image classification task under increasing shift severity: (a) binary loss monitored with an external predictor; (b) squared loss monitored using labels produced by the deployed model itself.
• Figure 3: Performance for the LLM QA task under prompt shifts introduced as time $t=200$: (a) Running risk lower bounds, using Qwen2-VL-7B as the predictive model for imputation; (b) Average time to alarm.
• Figure 4: Average time to alarm for the LLM QA task using different predictors $f_{\textrm{p}}(\cdot)$ to produce synthetic labels.
• Figure 5: Risk estimates as a function of time $t$ and average time to alarm for an image classification task: (a) under increasing shift severity (squared loss); (b) under a non-monotonic shift severity (binary loss).

Theorems & Definitions (5)

• Lemma 2.1: Theorem 4, Conjugate-mixture empirical Bernstein (CM-EB) adapted from howard2021time
• Lemma 2.2: PFA Guarantee of SRM Tracking2022Arxiv
• Lemma 3.1: Unbiasedness of PPRM
• Theorem 3.2: PFA Guarantee of PPRM
• Lemma 3.3: Optimal Choice of $\eta_t$ for Variance Reduction