4arXiv cs.LG (Machine Learning)·
KAISEN: A reproducible five-phase fairness audit pipeline for clinical risk models, stress-tested to failure
KAISEN is a five-phase audit pipeline for subgroup fairness in clinical risk models, covering stratification, disparity measurement, mechanism diagnostics, post-hoc mitigation, and drift monitoring. The authors evaluate it on a synthetic benchmark of 16 disease tasks and 15 social-determinant axes, deliberately stress-testing each component to identify failure modes. Key findings include that per-group threshold optimization reliably reduces equalized-odds disparity while group-wise Platt scaling behaves near-randomly on that metric, and that the mechanism diagnostic fails silently under proxy misspecification. All code and artifacts are released for full reproducibility.
Related guides (2)
Related events (8)
Eticas AI Risk Taxonomy v2.0.0: Open infrastructure for operationalizing AI audits
Eticas presents a structured AI auditing framework that bridges risk cataloging to executable audit methodology, demonstrated end-to-end on PII leakage testing against GPT-4-0314. The taxonomy organizes 76 active subcategories across 10 categories with mappings to 18 external frameworks, and is published under CC BY 4.0 with SKOS/JSON-LD distributions. The key contribution is an operationalization layer that converts named risks into measurable, severity-graded findings — addressing a gap the authors identify across at least 74 existing AI risk taxonomies. The PII leakage demonstration shows disclosure rates ranging from 0% to 84% under adversarial conditioning, graded as SYSTEMIC severity.
Five-layer framework for hidden safety-critical failures in deployed AI systems
A preprint from arXiv argues that AI safety discourse over-indexes on visible, model-centric failures while neglecting quieter systemic risks in deployed socio-technical systems. The authors propose a five-layer diagnostic framework covering epistemic, control, temporal, organizational, and ecosystem integrity. The paper identifies under-recognized risk patterns including uncertainty laundering, prompt injection, memory poisoning, evaluation deception, and model collapse, and calls for a shift from model-centric evaluation toward socio-technical reliability. It concludes with design, governance, and research agenda recommendations.
MedFailBench: Clinician-built benchmark for medical AI safety boundary failures
MedFailBench is a new open-source benchmark designed to identify which safety boundary a medical AI system failed, rather than measuring correctness. The v0.2.1 release contains 44 clinician-reviewed synthetic cases annotated by severity (1–5) and six safety gate types including missed urgent escalation, unsafe dosing, and evidence fabrication. The benchmark includes a HuggingFace leaderboard preview, a safety gate taxonomy, and an automated pipeline for archiving model-response screening runs, released under Apache-2.0 and CC-BY-4.0.
KINA: 899-item knowledge benchmark across 261 disciplines with formal representativeness and annotation incentive guarantees
KINA (Knowledge Index of Noah's Ark) is a new 899-item LLM benchmark spanning 261 fine-grained disciplines, addressing three methodological weaknesses in existing knowledge benchmarks: poor disciplinary representativeness, flat-payment annotation incentives, and unaudited ranking instability. The authors provide formal results: a (1-1/e) greedy approximation for disciplinary coverage and a proof that bonus-on-bar tournament payment weakly dominates flat payment for annotation quality. Evaluating 42 models from 13 labs, the top performer Gemini-3.1-Pro-Preview reaches 53.17%, with Claude-Opus-4.6 and GPT-5.4 close behind, revealing a tiered rather than smooth leaderboard structure with substantial headroom below saturation.
Computational audit finds ClinicalBERT amplifies demographic bias beyond training data distributions
Researchers present a systematic audit of representational bias in ClinicalBERT, a BERT-based model pretrained on MIMIC-III clinical discharge summaries, using two probing methodologies: Log Probability Bias Analysis and Masked Language Model probing across 98 clinical sentence templates and eight intersectional race-gender combinations. Of 32 statistically significant findings, 65.6% contradict observed corpus distributions, rising to 80% for Black patients and 87.5% for agency attribution under MLM probing. The key finding is that bias in ClinicalBERT operates predominantly through model-internal amplification rather than simple inheritance from training data, which has direct implications for clinical AI safety and deployment. This challenges the assumption that auditing training corpora is sufficient to characterize model bias.
Bayesian audit framework for public AI evaluation archives challenges frontier model claims
A new arXiv preprint proposes a Bayesian inference and decision-audit framework for interpreting public AI evaluation archives (LiveBench, Open LLM Leaderboard v2, LMArena, GAIA, tau-bench) as longitudinal time series rather than terminal leaderboards. The paper demonstrates that a single terminal snapshot is compatible with multiple distinct performance histories, yielding ambiguous timing estimates for reaching capability ceilings. A candidate selection-aware frontier model is shown to fail synthetic recovery, objective-archive prediction, preference transfer, and uncertainty calibration, with fixed audit gates rejecting its stronger claims. The work proposes an archive-and-adjudication protocol to reconstruct evaluation histories and falsify unsupported frontier capability claims.
Automated Benchmark Auditing for AI Agents and Large Language Models (ABA)
The paper introduces Auto Benchmark Audit (ABA), an agentic framework that systematically audits AI benchmark tasks for issues such as ambiguous specifications, environment conflicts, and incorrect ground truths. Applied to 168 benchmarks across nine domains including NeurIPS publications, ABA identifies critical issues in over 25.7% of evaluated tasks. The authors demonstrate that filtering out flawed tasks materially shifts model rankings and improves average performance on SWE-bench Verified and Terminal-Bench 2 by 9.9% and 9.6% respectively, indicating that current benchmark scores are significantly distorted by task quality problems. The agentic tool and annotations are released publicly.
AIriskEval-edu: Platform for auditing pedagogical risks in AI-generated educational explanations
Researchers present AIriskEval-edu Demo, a platform that audits instructional explanations across five pedagogical risk dimensions including factual accuracy, ideological bias, and student-level appropriateness. The system integrates GPT-5.5 via API alongside a fine-tuned self-hosted Llama 3.1 8B evaluator, with the local model outperforming GPT-5.5 on most metrics. The platform targets K-12 educational contexts and supports both automated auditing of AI-generated explanations and real-time auditing of human-written content, offering institutions a privacy-preserving deployment option.