Defensive driving scores are useful only when they preserve distinctions between policies that observe surrounding actors and those that do not. Re-simulation benchmarks may use reference-conditioned forgiveness, under which an agent receives credit when the logged human reference fails a compliance channel.
arXiv:2608. 04896v1 Announce Type: new Abstract: Defensive driving scores are useful only when they preserve distinctions between policies that observe surrounding actors and those that do not.
The paper introduces a claim‑safe protocol for evaluating closed‑loop AI systems, consisting of three actions: Refuse, Decompose, and Refresh. It demonstrates the protocol in a simulator with 24 policy components and 1,440 held‑out cases, showing that abstention and stable false admission rates are low while providing detailed statistical diagnostics. The approach emphasizes that evaluation results should be tied to observable support and statistical calibration rather than a single PASS/FAIL label.
The paper introduces NAQD‑Env, a synthetic benchmark designed to test language agents’ ability to selectively withdraw and resume actions when new evidence, permissions, or stop instructions arise. It evaluates models against a deterministic reference policy across eleven dependency families, measuring policy agreement, task value, withdrawal, resumption, and event reporting. Experiments on 350 scenarios show low withdrawal recall, no valid resumption, and limited policy alignment, highlighting the need to treat selective withdrawal as a distinct reliability component.
arXiv:2609.22582v1 Announce Type: new Abstract: End-to-end and vision-language-action (VLA) driving policies are compared by leaderboard rank, but a rank reports an outcome, not the behaviour behind...
The paper introduces the Alignment Flywheel, a governance‑centric hybrid multi‑agent system (MAS) that separates decision generation from safety governance. It defines a Proposer that generates candidate trajectories, a Safety Oracle stack that evaluates safety, and an Enforcement layer that applies risk policies at runtime. A governance MAS oversees monitoring, red‑teaming, verification, and versioned release management, enabling patch‑local fixes to safety failures without retraining the Proposer. The architecture is implementation‑agnostic and is demonstrated in two scenarios: a learned spatial Oracle and a clinical GenAI proxy. The authors provide open‑source code at https://github.com/decide-ugent/Alignment-Flywheel.
arXiv:2607. 12469v1 Announce Type: cross Abstract: Many agent-safety evaluation results are not yet load-bearing evidence: identical nominal outcomes (task success, attack success, monitor scores) may sit atop materially different evidence regimes.
The paper investigates the reliability of language‑model judges used as measurement instruments on shared endpoints. Through two preregistered audits of 52,988 requests, the authors found that repeat rankings and byte‑identical replays fell far short of required thresholds, revealing significant instability. They identify three mechanisms—label‑to‑meaning bias, candidate gaps below the noise floor, and input permutation noise—that explain the gap, and propose a snapshot‑identity ladder, design rules, and a reporting checklist to mitigate such failures.
The paper argues that cross‑view correspondence, commonly used in agent evaluation and trace‑based learning, functions as a measurement intervention. Removing or altering this correspondence can create artificial sensitivity or invariance, and multiple optimal correspondences can obscure mechanism labels and learning credit. The authors propose a validity theory with two‑sided validation, all‑optima identification, and uncertainty propagation, and demonstrate through experiments that unvalidated correspondences can misattribute credit and erase harmful responses.
The paper reports a preregistered audit of language‑model judges used as measurement instruments, revealing that the assumption that a model’s responses remain stable over time is invalid. Across nearly 53,000 audited requests, repeat rankings and byte‑identical replays fell far below required reliability thresholds, with three identified mechanisms—label‑to‑meaning bias, candidate gaps below the noise floor, and input permutation noise—explaining the discrepancy. The study proposes a three‑level snapshot‑identity framework, eight design rules, and a reporting checklist to prevent such reliability failures in future evaluations.
arXiv:2608. 19760v1 Announce Type: cross Abstract: Audited against causal ground truth from executed replay in a single-agent tool environment (ALFWorld), none of the step-level credit signals used to train LLM agents -- LLM-judge scores, outcome-conditioned logprob ratios, or the policy's own confidence -- identifies which steps causally matter better than chance.
The paper audits runtime failure monitors that use a model’s internal representations to predict failures in autonomous driving tasks. Across two tasks—online vectorized map generation with LaneSegNet and end‑to‑end planning with VAD—the authors find that frame‑level errors can be predicted with high AUROC scores using supervised latent probes. However, adding latent features to baseline monitors that use only observable inputs and outputs does not yield statistically significant improvements, suggesting that internal representations may not provide additional predictive value beyond what is already observable.
arXiv:2609.13714v1 Announce Type: new Abstract: An updated model can improve an aggregate metric while degrading a slice that matters to a downstream user. We study checkpoint selection subject to no...