Classical data mining treats data as generated outside the learner. Emerging AI-native infrastructure invalidates this assumption: as agents observe system state, act, update retrieval stores and models, and promote new artifacts into the serving path, their own decisions reshape the future logs, traces, corpora, and evaluation records used for later training and release. We call this new research object a self-rewriting system and the resulting system-induced distribution shift endogenous drift. Our central, falsifiable claim is that cross-surface causes are not identifiable in general from marginal per-surface monitoring: different coupled failures can produce the same individual signals while differing in their joint behavior, requiring joint analysis for reliable attribution. This agenda comprises four foundational research problems: fusing heterogeneous drift signals into calibrated gating and attribution (C1); mining agent traces whose activity alphabet evolves (C2); maintaining knowledge fabrics through online reindexing, faithfulness, and compression decisions (C3); and safely promoting concurrent learning streams through one shared evaluation gate (C4). These remain fragmented across concept drift, performative prediction, process mining, streaming retrieval-augmented generation, and MLOps, without a unified data-mining agenda. We outline public benchmarks and evaluation criteria, using AI-native telecom only as a motivating exemplar with analogues in finance, software CI/CD, and clinical decision support.
Figures & tables
Fig. 1: Classical data mining versus a self-rewriting loop.
focuses on one model’s decisions; not a multi-artifact infrastructure loop with heterogeneous, incommensurable signals
MLOps [ 16 ]
the model deployment and release lifecycle
operationalizes models but defines no mining task over endogenous traces and coupled promotion interactions
AIOps [ 17 ]
operational telemetry (metrics, logs, alerts)
mines system signals, not the agent’s own structured reasoning/action trace as the primary data object
Process mining; business-process drift [ 18 , 19 ]
event logs, workflow discovery/conformance, and detecting/localizing control-flow change in a running process
handles drift among a roughly fixed activity set; not open-vocabulary alphabet growth where tools, roles, and skills are added and retired by the same loop being mined
Streaming RAG [ 20 , 6 ]
retrieval over changing corpora
does not unify citation contracts, summary invalidation, reindex decisions, multimodal updates, and bounded compression
TABLE I: Comparison with adjacent research areas.
Benchmark
Data object
Perturbation / update
Labels / oracle
Metrics
Drift-arena (C1)
multi-surface drift over corpus, models, tools, policies, serving
controlled updates with changing cross-surface dependence
Machine learning models increasingly generate their own training data -- online bandits, reinforcement learning, and post-training pipelines for language models are leading examples. In these adaptive settings, a single training observation both updates the learner and shifts the distribution of future data the learner will collect. Standard attribution methods, designed for static datasets, ignore this feedback. We formalize occurrence-level attribution for finite-horizon adaptive learning via a conditional interventional target, prove that replay-side information cannot recover it in general, and identify a structural class in which the target is identified from logged data.
Amit Kiran Rege
Department of Computer Science, University of Colorado Boulder, Boulder, Colorado 80309
Invariants, the relations expected to hold among measured signals of a network, underpin applications from verification to traffic generation, telemetry imputation, and input validation, yet writing them by hand demands rare expertise in both formal logic and networking. Automatic miners can help but fall short on two fronts: they still require the hardest input (the grammar of admissible invariants) and they learn only exact, hard'' rules, struggling with real-world approximation caused by inherent noise in data. LLMs are tools that can provide semantic reasoning over data, but are non-deterministic and opaque in their learning. Our key idea is to partition the invariant search problem into an AI-driven grammar discovery'' problem, followed by a statistics-driven ``search'' problem within the learned grammar. Taken together, this allows non-deterministic, hallucination-prone AI to help produce auditable invariants with formal guarantees. We design and implement such a system, Autogram, and evaluate it on both public and production telemetry data, recovering expert-derived invariants with high coverage and low false positives. We close with discussion on open problems on the path toward fully open-ended discovery.
Hongyu Hè, Alexander Krentsel, Sylvia Ratnasamy +1
Concept drift is a common phenomenon in real-world data streams, in which changes in the data-generating distribution can degrade predictive model performance. Most existing definitions characterize drift as changes in the joint distribution P(x,y), without distinguishing which component of the data-generating process has changed. In this work, we introduce a causal perspective on concept drift based on Structural Causal Models (SCMs). We propose a taxonomy that categorizes drift events by their causal origin, including changes in exogenous variables, endogenous mechanisms, confounders, and target-generating processes. Building on this framework, we develop an SCM-based data stream generator that simulates controlled mechanism-level drift events. Our experiments empirically characterize the distributional effects of each drift type and show that drifts with different causal origins induce distinct patterns of distribution shift and predictive behavior. Furthermore, by integrating causal discovery methods, we use our framework to construct data streams grounded in real-world dependency structures, enabling more realistic and informative evaluation scenarios. We also demonstrate that leveraging the generated data can improve downstream performance. These results highlight the importance of accounting for causal structure when studying and evaluating adaptive learning methods, and establish a foundation for causally-aware evaluation in non-stationary environments.
Eduardo V. L. Barboza, Jean Paul Barddal, Robert Sabourin +1
LIVIA, École de Technologie Supérieure · Graduate Program in Informatics (PPGIa), Pontifícia Universidade Católica do Paraná (PUCPR)