Aug 2026· Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2· pp. 3198-3209· 0 citations· 28 references
TL;DR
counterfactual PoNR Auditing (CPA), a snapshot-replay framework that produces matched counterfactual rollouts and is accelerated by a neural operator surrogate (UNO) for a 3.09× end-to-end speedup, and serves as an offline supervision generator, allowing lightweight predictors to learn early inevitability signals beyond static observational baselines.
Abstract
Safety-critical failures in constrained dynamical environments are often detected only after a violation occurs, while outcome metrics (e.g., success rate, time-to-failure) conflate structural inevitability with decision-induced errors. We formulate failure diagnosis as recoverability boundary discovery and define the Point-of-No-Return (PoNR) as the last recoverable snapshot before violation becomes unavoidable under an audited intervention class. We introduce Counterfactual PoNR Auditing (CPA), a snapshot-replay framework that produces matched counterfactual rollouts and is accelerated by a neural operator surrogate (UNO) for a 3.09× end-to-end speedup. CPA yields Evidence Objects (EOs) that support structural--policy decomposition and enable Failure Atlas mining via contrastive analysis of event sequences within the policy-induced gap. CPA also serves as an offline supervision generator, allowing lightweight predictors to learn early inevitability signals beyond static observational baselines. Experiments on a high-fidelity hydrodynamic control testbed and a canonical bistable system show that PoNR precedes observable violation, remains stable under up to 30% multiplicative surrogate noise, and reduces auditing cost through surrogate-guided search.
SOWRC combines alert labels with randomized audits through Horvitz–Thompson losses and a martingale-mixture boundary and shows that population risk is non-identifiable when any silent region has zero labeling probability.
Si-Yang Bai, Zheng Fang, Jie Chen· Axioms· 0 citations
Language-model agents are increasingly asked to run experiments, yet benchmarks rarely measure whether those experiments improve the decisions they are meant to inform. We introduce CrashGym, a deterministic, replayable limit-order-book simulator in which an agent must choose between a deleveraging intervention and a l...
Hao-Cheng Lu, Hao-Xuan Cheng, Ming-Yang Chen et al.· 0 citations
Reasoning large language models (LLMs) benefit from test-time scaling, motivating adaptive policies that spend tokens when they help and stop early otherwise. Existing rules halt when a confidence or stability signal crosses a threshold, but such thresholds rarely transfer across models and datasets, and validation-cal...
An adaptive system can retain an earlier estimate while changing what that estimate is used to guarantee. The guarantee depends on the experiment that produced the evidence, including how its target was selected and which assumptions connect it to the current claim. We show that past-measurable selection preserves futu...
Ao Li, Weitong Chen Australian Institute for Machine Learning, Adelaide University et al.· 0 citations
Anytime-valid methods proposed for dependent data should be accompanied by null-calibration controls and mechanism traces, and Huber-style gating of the filter's own updates cuts isolated-spike degradation by an order of magnitude with no dataset specific tuning.
This work argues that update admission must be assessed through both error control and retained learning opportunities at a stated interaction budget, and identifies a concrete failure: a range-based confidence gate cannot certify unchanged old-task behavior within otherwise substantial budgets.
Qin-Zhen Ma, Rui-Hai Wu· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.