cs.LGOct 12, 2026

Learning to Detect Shifts in Unstructured Data Streams via Monitoring-by-Betting

Authors: Siyu Zhou, Jialin Ouyang, Fan Xia, Adarsh Subbaswamy, Drew Prinster, Suchi Saria, Augusto Garcia-Agundez, Sarah Pollet, 

Organizations: University of California, San Francisco · University of Maryland · Johns Hopkins University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Continuous monitoring of generative AI systems requires detecting distribution shifts in the unstructured data they ingest and generate. Existing monitoring methods typically control false alarms by requiring the discrepancy measure used to detect shifts to be pre-specified. However, for unstructured data, pre-specification creates a sensitivity--breadth trade-off: a narrowly defined measure is sensitive to the anticipated shift but may miss unexpected changes, whereas a broadly defined measure may detect a wider range of shifts but with reduced sensitivity. Although adaptively hypothesizing the shift direction can overcome this limitation, such data-dependent adaptation generally invalidates conventional false-alarm guarantees. We reconcile adaptivity with statistical validity through testing-by-betting, reframing sequential changepoint detection as a betting game in which the bettor continually updates its hypothesized shift direction, or witness function, as new observations arrive. We introduce asymptotic e-detectors, which approximate the pre-change distribution using a calibration set of size mm, and prove that they asymptotically control the average run length to a false alarm as mm increases. Empirically, calibration sets as small as m=30m=30 provide effective control. We then develop adaptive monitoring strategies ranging from black-box bettors that learn from embeddings to interpretable bettors based on LLM-proposed concepts. Across three real-world datasets, including data from a clinical AI scribe deployment, adaptive monitors detect shifts faster than fixed monitors at the same false-alarm level. Further, when the true shifts are sparse and interpretable, LLM-concept monitors provide the fastest and most interpretable detection, while also guiding root-cause diagnosis for complex shifts.

Figures & tables

Figure 1: Adaptive monitoring procedure overview: At each time step, an adaptive bettor uses past observations to bet on how the data may have shifted. When the next observation arrives, the bet’s return is computed by comparing it with a calibration set, and an alarm is raised once the accumulated wealth crosses a threshold.
Figure 2: LLM-concept bettor for estimating when and what shifted, using LLM reasoning, unsupervised learning, and statistical changepoint estimation.
Figure 3: Experimental results. Top: EDD against nominal ARL0\mathrm{ARL}_{0} , where lower is better. Bottom: Concepts learned by the LLM-concept bettor, plotted by prevalence before vs after the changepoint. Concepts in (e-g) are colored by their agreement with the true shifted attribute (stars). Concepts in (h) are colored by themes (Table C.3 ), as true shifts are unknown for real-world AI deployments. The circled concept in (h) asks whether the draft includes unneeded non-clinical information, such as a review of systems or introductory sentences.
Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure B.1: Keyphrase extraction prompt (abbreviated) for CEBaB experiment.
Figure B.2: Concept proposal prompt (abbreviated) for CEBaB experiment. Top keyphrase clusters (topics) are shown from an example replicate.
Figure B.3: Concept annotation prompt for the CEBaB experiment.
 CEBaBCUB: yellow billCUB: white crownAI scribe
DataRestaurant reviews from CEBaBBird images from CUB-200-2011Bird images from CUB-200-2011HPI section of Rheumatology clinical notes
Shifted attributeWhether review mentions ambianceWhether bird has yellow bill in imageWhether bird has white crown in imageClinical notes from 60 days before vs after model update on 2026-02-25
Prevalence pre →\to post0.52 →\to 0.880.17 →\to 0.640.30 →\to 0.95N/A
Embedding modelMiniLMCLIP ViT-B/32CLIP ViT-B/32MiniLM
Embedding dim dd384512512384
npren_{\text{pre}}505050100
Appendix
Table C.1: Summary of experimental settings across the three datasets.
DatasetDirectionIC textOC text
CEBaBInformeda typical restaurant review describing the customer’s experience of the visita restaurant review that comments on the ambiance or decor of the restaurant
Unrelateda typical restaurant review describing the customer’s experience of the visita restaurant review that mentions parking the car near the restaurant
CUB (yellow bill)Informeda photo of a birda photo of a bird with a yellow bill
Unrelateda photo of a birda photo of a bird perched on a wire or fence post
CUB (white crown)Informeda photo of a birda photo of a bird with a white crown
Unrelateda photo of a birda photo of a bird perched on a wire or fence post
Appendix
Table C.2: Text exemplars used to define pre-specified drift directions for fixed monitors. Each direction is defined by an in-class (IC) and out-of-class (OC) text description, embedded to form a unit drift vector.
Figure C.1: Empirical ARL0\mathrm{ARL}_{0} against nominal ARL0\mathrm{ARL}_{0} on the 100100 CEBaB ambiance null streams. Shaded regions are simultaneous 95%95\% bands over streams, and the dotted line marks empirical ARL0\mathrm{ARL}_{0} equal to nominal ARL0\mathrm{ARL}_{0} .
Figure C.2: Empirical ARL0\mathrm{ARL}_{0} against nominal ARL0\mathrm{ARL}_{0} on the CUB yellow bill (left) and white crown (right) null streams, over 100100 null streams per setting. Shaded regions are simultaneous 95%95\% bands over streams, and the dotted line marks empirical ARL0\mathrm{ARL}_{0} equal to nominal ARL0\mathrm{ARL}_{0} .
ThemeCodebook definition# ConceptsBefore →\to afterExample concepts (before →\to after)
Unneeded contentDraft content that does not belong in the note because it is unnecessary, redundant, verbose, non-clinical, introductory or identifying, or the removal, condensing or restructuring of such content, whether clinical or not460.79 →\to 0.60Does the initial draft include unnecessary or irrelevant non-clinical information such as review of systems, introductory sentences, or personal details? ( 0.93→0.600.93\to 0.60 ) Does the edit remove redundant or unnecessary note sections (including review of systems or initial summary)? ( 0.86→0.540.86\to 0.54 )
Clinical correctionClinical content (diagnoses, symptoms, medications, treatment) that the clinician adds, corrects or clarifies, or that the draft is missing210.56 →\to 0.61Does the edit correct or improve diagnostic information? ( 0.38→0.590.38\to 0.59 ) Does the initial draft miss or have incorrect medication, treatment, or procedure details? ( 0.59→0.430.59\to 0.43 )
Non-clinical correctionNon-clinical information (patient or social history, follow-up, administrative or demographic details) that the clinician adds, corrects or clarifies, or that the draft is missing50.51 →\to 0.67Does the edit improve the accuracy and completeness of patient, social history, or provider demographic information? ( 0.50→0.680.50\to 0.68 )
Appendix
Table C.3: Themes of the AI scribe concepts in Figure 3 (h), over the four concepts active at the end of each of 1818 streams. An LLM judge ( gpt-4.1-mini ) assigns each concept to a theme from its text alone, using a fixed codebook based on the vendor’s stated purpose for the update and the clinicians’ reports of editing less. The first two themes come from these reports; the third theme is named after reading concepts that fit neither. Before →\to after shows the mean fraction of notes for which a concept is annotated as present, before and after the changepoint.
StreamConceptThemeBefore →\to after
1Does the edit add or correct procedural or diagnostic details based on inaccuracies or omissions in the initial draft?Clinical correction0.56→0.670.56\to 0.67
Does the edit clarify or correct medication, management, or demographic information based on inaccuracies or omissions in the initial draft?Clinical correction0.61→0.720.61\to 0.72
Does the edit add or correct missing or inaccurate patient history and administrative details (such as functional status, anatomical location, provider identification, or lab data) in the initial draft?Non-clinical correction0.63→0.760.63\to 0.76
Does the edit add or correct missing treatment and clinical history details based on inaccuracies or omissions in the initial draft?Clinical correction0.53→0.640.53\to 0.64
2Does the edit remove unnecessary or irrelevant patient-related information (such as identification, introduction, background, history, or personal details) from the initial draft?Unneeded content0.82→0.660.82\to 0.66
Does the edit improve procedural and clinical detail documentation by clarifying, correcting, or adding missing information based on the initial draft?Clinical correction0.53→0.720.53\to 0.72
Appendix
Table C.4: Concepts active at the end of the first four AI scribe streams in Figure 3 (h), with the theme each is assigned in Table C.3 and the fraction of notes for which it is annotated as present before and after the update.
Figure D.1: Empirical ARL0\mathrm{ARL}_{0} as a multiple of the nominal ARL0=1/α\mathrm{ARL}_{0}=1/\alpha with a fixed score (left, Beta(2,2)\mathrm{Beta}(2,2) scores) and with a learned score from a classifier refit on the stream history (right). Dashed colored lines reuse a calibration dataset of size m∈{10,30,100,1000}m\in\{10,30,100,1000\} at every step, the solid black line uses the oracle expectation EP0[f^t(X)]\mathbb{E}_{P_{0}}[\hat{f}_{t}(X)] , and the solid grey line redraws 1010 reference observations at every step. The dotted line marks the nominal ARL0\mathrm{ARL}_{0} . Each point summarizes 2,0002{,}000 null streams, and streams without an alarm after 100,000100{,}000 steps are counted at 100,000100{,}000 , so each value is a lower bound.
Figure D.2: Distribution of the null alarm time τm\tau_{m} at 1/α=3001/\alpha=300 with a fixed score (left) and a learned score (right). Every null stream draws its own calibration dataset, so each curve estimates P(τm≤t)P(\tau_{m}\leq t) averaged over calibration datasets from 2,0002{,}000 null streams. The black curve is the oracle detector, which uses the exact pre-change expectation.
More in this topic

Distribution Shift Detection

Distribution shift detection identifies when data or model behavior changes between reference and deployment settings.

See all 62 papers

Explore similar work

CardsList
Jun 23, 2026

Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift

Deployed AI text detection approaches are supervised, relying on training-time access to labeled datasets of both human-written and AI-generated text. This approach is vulnerable to three types of distribution shifts that occur continually post-deployment, and for which labeled data is often unavailable: adversarial humanization, new LLMs being released, and unidentifiable temporal drift in human writing after widespread LLM usage. Simultaneously, existing approaches do not leverage a key signal of LLM usage: inference-time homogeneity. We propose a test-time adaptation (TTA) approach that adapts to distribution shifts by leveraging homogeneity among unlabeled samples observed at inference time. Empirically, we find that state-of-the-art supervised detectors systematically fail when they encounter distribution shifts in AI-generated and human writing, both adversarial and natural, while TTA with semi-supervised learning is largely robust; e.g., the commercial model Pangram detects just 27.0% of our adversarial AI-generated text, compared to 93.3% for our test-time approach. We establish that TTA is a promising framework for AI text detection in the wild.

Kevin Ren, Manish Raghavan, Nikhil Garg

Cornell Tech · MIT

AI-Generated Text DetectionDistribution ShiftContinual Test-Time AdaptationRobustness of Machine-Generated Text Detection
AskPDF
Jul 4, 2026

Rethinking AI-Generated Text Detection: A Strong Baseline and the Distribution-Shift Problem That Remains

Recent AI-generated text detection work often introduces a new benchmark together with a specialized detector tailored to it. We revisit this practice from a baseline-first perspective. Across several benchmarks, we show that a plain, fully fine-tuned RoBERTa matches or exceeds the specialized detectors those benchmarks are built around. This suggests that much of the recent architectural complexity is not what drives strong in-distribution detection. The remaining challenge is the distribution shift. The same strong baseline degrades sharply when the topic domain or generating model changes at test time, and simply adding more source data does not close the gap. We identify a key failure mode: under distribution shift, the detector can assign high-confidence machine labels to human-written text from unseen domains. We then study two lightweight domain adaptation methods to address this problem: KK-shot adaptation with first-order MAML over LoRA adapters, and a per-sample confidence-weighted ensemble built on top of the adapted detector. Overall, our results suggest that progress in AI-generated text detection should be measured not only by in-distribution performance, but also by robustness under distribution shift.

Zhuoer Shen, Mingyi Wang, Shaofeng Zou +1

Department of Computer Science, University of California, Santa Barbara · School of ECEE, Arizona State University

Distribution Shift RobustnessAI-Generated Text DetectionDomain AdaptationDistribution Shift DetectionRobustness of Machine-Generated Text Detection
AskPDF
Jun 10, 2026

Online Shift Detection and Conformal Adaptation for Deployed Safety Classifiers

Safety classifiers deployed in production operate under a stationarity assumption that fails silently: when input distributions drift, accuracy degrades with no error signal until ground-truth labels arrive. We present an online monitor that detects distributional shift in classifier scores via a sliding-window KS statistic with empirically calibrated alarm thresholds. In a pre-registered factorial evaluation (4 classifiers ×\times 5 shift conditions ×\times 20 seeds ×\times 2 window sizes; 800 cells), the monitor achieves 86.6% valid detection (mean latency 39.5 steps) across synthetic-onset, real-jailbreak, and adversarial regimes; a classifier ×\times shift interaction (η2=0.185η^2 = 0.185) shows that monitoring must be tuned per classifier. Attempting to recover post-detection coverage via weighted conformal prediction exposes a failure mode: density-ratio estimation collapses for generative classifiers because logistic regression separates source from target perfectly in 3584-4096-dimensional embedding space, clipping all importance weights to zero; projecting to ≤32\leq 32 dimensions restores coverage. We then extend the framework to gradient-based evasion and give the first threat-model characterisation of score-disagreement monitoring as a canary. We falsify three assumptions: that architectural diversity drives the signal (false, η2=0.011η^2 = 0.011), that it is generic out-of-distribution detection (false, GCG-specific, p<10−12p < 10^{-12}), and that an adaptive attacker can suppress it (false while the canary is confident). We derive the exact security boundary, a confidence-gated equilibrium at which a monitor-aware attacker stalls at gap =1/(2λ)= 1/(2λ), and provide a calibration-free scan martingale achieving false-alarm rate ≤1%\leq 1\% across all classifiers with no per-model tuning.

Jun Wen Leong

Distribution ShiftConformal PredictionDistribution Shift DetectionAI Safety EvaluationChange-Point DetectionAdversarial Example Detection
AskPDF