Searcharxiv⌕ Search

arXiv subjects

Kate Thomas

Publications and source records attributed to Kate Thomas.

3 recordsLinked to original sources

Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety Auditing

AI safety evaluations often involve sensitive prompts and proprietary models. Evaluators want to keep their evaluation prompts secret from model owners to prevent purposeful or incidental data leakage, ensuring longer usefulness of their benchmark. Model owners in turn wish to keep their proprietary model weights and inference code within their own secure infrastructure to protect their IP. We present a way forward for "double-blind evals", which provide cryptographic guarantees that an evaluator's prompts are not shared with the model owner, while simultaneously maintaining the confidentiality of the model weights and inference code. To accomplish this, we use secure enclaves that encrypt data in RAM at the hardware level. In a world-first pilot, we test this approach by evaluating a proprietary Google DeepMind model (Gemini 2.5 Flash Lite) against a selection of the private AILuminate benchmark corpus, using an NVIDIA H100 secure enclave on Google Cloud Platform, and using OpenMined's PySyft to maintain data and model privacy. We believe this effort marks the first use of this technology to provide mutual secrecy for one organization's frontier-class proprietary model weights and another organization's closed benchmark.

cs.CR↗

Niven numbers are an asymptotic basis of order 3

A base-$g$ Niven number is a natural number divisible by the sum of its base-$g$ digits. We show that, for any $g\geq 3$, all sufficiently large natural numbers can be written as the sum of three base-$g$ Niven numbers. We also give an asymptotic formula for the number of representations of a sufficiently large integer as the sum of three integers with fixed, close to average, digit sums.

math.NT↗

Leaf-inspired rain-energy harvesting device

We study a rain-powered energy-harvesting device inspired by the natural impact of raindrops on leaves. In nature, a raindrop striking a leaf at high speed causes it to deform and vibrate. Inspired by this, our device uses an elastic beam coupled with a piezoelectric material to convert mechanical vibrations from droplet impacts into electrical energy. We conduct experiments to analyze how beam length, droplet impact location, and residual droplet mass affect energy conversion. Our results, supported by a theoretical model, show strong agreement when the beam length exceeds 5 cm. Beyond this length, the energy conversion becomes independent of further increases, suggesting that 5 cm is optimal for maximizing output. To validate practical applicability, we also test the device under real rain conditions, demonstrating consistent performance. Understanding the interplay between raindrop dynamics and energy conversion can guide the design of efficient, scalable rain-powered energy-harvesting systems for environmental applications.

physics.ins-det↗