Research map / Large language models / AI and society
Adaptive auditing of AI systems with anytime-valid guarantees
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Frames adaptive AI auditing as two dueling hypothesis tests (model's null vs auditor's null) and builds anytime-valid e-process tests (LR, UI, SR variants) that control type-I error under adaptive sampling and optional stopping.
Problem
Adaptive, small-sample testing of AI failure modes violates classical assumptions (pre-specified sampling and stopping, large n), so standard inference for robustness audits is invalid or undefined.
Approach
Formulates robustness audits as dueling nulls (no failure mode vs. auditor will find one), proves they are asymptotic complements under asymptotically consistent auditors, and tests both via testing-by-betting e-processes (likelihood ratio, universal-inference-based adaptive bets, and Shiryaev-Roberts changepoint mixtures) with Ville's inequality guarantees.
Key results
- SR-LR-UI achieved the highest power among adaptive testing methods in semi-synthetic GPQA experiments, with the Oracle auditor rejecting in as few as 15-25 observations.
- Under the true model null, all methods correctly rejected the auditor's null at high rates, with type-I error for the model's null controlled.
- On the SDoH real-data audit, sampling frequency was strongly negatively correlated with category accuracy (Spearman rho = -0.84).
Surprising findings
- Pre-specified auditing strategies (Stratified, Pre-learned) frequently reject the auditor's null due to low power, which risks being misread as evidence of robustness.
- Stronger adaptive auditors need more observations to reject the auditor's null, which the authors frame as desirable.
Limitations
- Auditing strategies are treated as black boxes and the framework relies on asymptotic consistency of the auditor for the duality result.
- Experiments focus on automated auditors and simulated/semi-synthetic labels rather than human-in-the-loop audits.
- Results are for binary scores in the main text; continuous or multiclass extensions are only asserted.
Experimental setup
Datasets: GPQA CUB-200-2011 SDoH clinical notes dataset (81 notes, 2268 comparisons)
Benchmarks: GPQA (semi-synthetic) SDoH clinical note extraction CUB-200-2011
Metrics: rejection rate of model's null rejection rate of auditor's null (audit passage) number of observations to decision type-I error
Baselines: Stratified sampling Pre-learned sampler Oracle auditor LR LR-UI SR-LR
Builds on: Safe Anytime-Valid Inference Universal Inference Shiryaev-Roberts procedure Upper Confidence Bound Exponentially Weighted Averaging Forecaster
Closest papers by meaning
- AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems (NeurIPS 2026 · similarity 0.86)
- Robust ML Auditing using Prior Knowledge (ICML 2025 · similarity 0.86)
- Auditing Sabotage Bench: A Benchmark for Detecting and Fixing Research Sabotage in ML Codebases (NeurIPS 2026 · similarity 0.86)
- Position: AI Competitions Provide the Gold Standard for Empirical Rigor in GenAI Evaluation (ICML 2025 · similarity 0.85)
- Improving Audit Realism with Inference-Time Compute and Deployment Scaffolds (NeurIPS 2026 · similarity 0.85)
- PROBE: Learning to Audit Policy Compliance in Tool-Using LLM Agents (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.