Tag
This paper introduces critique refinement and DISH to enhance the realism of alignment evaluations by mitigating evaluation awareness, demonstrating that combining both techniques yields significant improvements over using either alone.
Miles Brundage comments on OpenAI's statement clarifying that the computation depth of frontier models like Astra is within a factor of two of GPT-4, advocating for continuous embedded auditing rather than reactive measures.
Counterfactual Fragility Certificates (CFC) introduce a model-agnostic audit protocol to detect high-confidence brittleness in machine learning models under structured evidence failure scenarios, improving over existing methods.
This study examines how large language models judge urban safety based on neighborhood names versus coordinates, revealing that names carry both crime signal and demographic bias, with bias scaling with geographic knowledge.
EXPL-FR is a method that explains face recognition models by aligning vision-language embeddings to the recognition space, enabling label-free auditing of semantic attributes and model comparison without requiring architecture access.
Proposes J-Access, an inference-time audit using the Jacobian lens to measure residual knowledge accessibility in unlearned LLMs, finding that accessibility predicts recovery speed but that directly minimizing it fails to promote genuine deletion.
This paper investigates whether time-series forecasters use the correct historical inputs, separating recoverability of delays, model reporting, and functional use. It proves that accurate forecasts and correct delay reports can still hide the use of wrong lags, and demonstrates this issue empirically in N-HiTS and TCN models.
This paper challenges the conflation of dynamic parameterization with dynamic inference, introducing Frozen-Controller Auditing to show that input-dependent coefficients do not imply computational savings. Experiments on Transformers reveal static layerwise profiles retain near-full performance despite no conditional execution.
This paper introduces a reference-based method to detect whether an LLM was distilled from a specific teacher model, using membership inference. The approach achieves near-perfect accuracy in controlled settings and provides new evidence about potential distillation relationships involving QwQ, DeepSeek-R1, and GPT-OSS.
MIT researchers developed a technique to audit AI models for their capability to generate child sexual abuse material without producing illegal outputs, achieving 100% accuracy in tests. This method examines hidden model representations to infer whether a model has been fine-tuned for harmful content, providing a scalable way for platforms and law enforcement to detect unsafe models.
This paper introduces SAR, a lightweight LoRA adapter that enables fine-tuned language models to self-report hidden behaviors like backdoors, outperforming existing methods by detecting all implanted behaviors and halving hallucination rates.
Cyclic denoising is introduced as a novel extraction attack that reveals ultrastable memorized training images in diffusion models by repeatedly noising and denoising samples. The technique requires no gradients or weight inspection and has implications for privacy auditing.
This paper introduces idSCD, a white-box method that uses semantic correlation descriptors to identify whether a dataset was used in training a model, outperforming existing baselines across multiple settings.
This paper introduces I-SAFE, a post-hoc distributional auditing framework for scientific AI models using Wasserstein Coherence Metrics, which reveals structural differences in model outputs that accuracy-based evaluation fails to capture. Demonstrated on drug-target interaction prediction, the framework is model-agnostic and applicable to any domain with structured inputs and external priors.