arXiv subjects

Aleksandra Spyra

Publications and source records attributed to Aleksandra Spyra.

6 recordsLinked to original sources

arXiv2026

OpenAI GPT-5 System Card

This is the system card published alongside the OpenAI GPT-5 launch, August 2025. GPT-5 is a unified system with a smart and fast model that answers most questions, a deeper reasoning model for harder problems, and a real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent (for example, if you say 'think hard about this' in the prompt). The router is continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness, improving over time. Once usage limits are reached, a mini version of each model handles remaining queries. This system card focuses primarily on gpt-5-thinking and gpt-5-main, while evaluations for other models are available in the appendix. The GPT-5 system not only outperforms previous models on benchmarks and answers questions more quickly, but -- more importantly -- is more useful for real-world queries. We've made significant advances in reducing hallucinations, improving instruction following, and minimizing sycophancy, and have leveled up GPT-5's performance in three of ChatGPT's most common uses: writing, coding, and health. All of the GPT-5 models additionally feature safe-completions, our latest approach to safety training to prevent disallowed content. Similarly to ChatGPT agent, we have decided to treat gpt-5-thinking as High capability in the Biological and Chemical domain under our Preparedness Framework, activating the associated safeguards. While we do not have definitive evidence that this model could meaningfully help a novice to create severe biological harm -- our defined threshold for High capability -- we have chosen to take a precautionary approach.

cs.CL↗

arXiv2022

Inverse design of photonic devices with strict foundry fabrication constraints

We introduce a new method for inverse design of nanophotonic devices which guarantees that resulting designs satisfy strict length scale constraints - including minimum width and spacing constraints required by commercial semiconductor foundries. The method adopts several concepts from machine learning to transform the problem of topology optimization with strict length scale constraints to an unconstrained stochastic gradient optimization problem. Specifically, we introduce a conditional generator for feasible designs and adopt a straight-through estimator for backpropagation of gradients to a latent design. We demonstrate the performance and reliability of our method by designing several common integrated photonic components.

cs.ET↗

arXiv2022

Massively parallel pixel-by-pixel nanophotonic optimization using a Green's function formalism

We introduce an efficient parallelization scheme to implement pixel-by-pixel nanophotonic optimization using a Green's function based formalism. The crucial insight in our proposal is the reframing of the optimization algorithm as a large-scale data processing pipeline, which allows for the efficient distribution of computational tasks across thousands of workers. We demonstrate the utility of our implementation by exercising it to optimize a high numerical aperture focusing metalens at problem sizes that would otherwise be far out of reach for the Green's function based method. Finally, we highlight the connection to powerful ideas from reinforcement learning as a natural corollary of reinterpreting the nanophotonic inverse design problem as a graph traversal enabled by the pixel-by-pixel optimization paradigm.

physics.comp-ph↗

arXiv2019

A Generalized Framework for Population Based Training

Population Based Training (PBT) is a recent approach that jointly optimizes neural network weights and hyperparameters which periodically copies weights of the best performers and mutates hyperparameters during training. Previous PBT implementations have been synchronized glass-box systems. We propose a general, black-box PBT framework that distributes many asynchronous "trials" (a small number of training steps with warm-starting) across a cluster, coordinated by the PBT controller. The black-box design does not make assumptions on model architectures, loss functions or training procedures. Our system supports dynamic hyperparameter schedules to optimize both differentiable and non-differentiable metrics. We apply our system to train a state-of-the-art WaveNet generative model for human voice synthesis. We show that our PBT system achieves better accuracy, less sensitivity and faster convergence compared to existing methods, given the same computational resource.

cs.AI↗

arXiv2017

Logarithmic price of buffer downscaling on line metrics

We consider the reordering buffer problem on a line consisting of n equidistant points. We show that, for any constant delta, an (offline) algorithm that has a buffer (1-delta) k performs worse by a factor of Omega(log n) than an offline algorithm with buffer k. In particular, this demonstrates that the O(log n)-competitive online algorithm MovingPartition by Gamzu and Segev (ACM Trans. on Algorithms, 6(1), 2009) is essentially optimal against any offline algorithm with a slightly larger buffer.

cs.DS↗

arXiv2017

Online Tree Caching

We initiate the study of a natural and practically relevant new variant of online caching where the to-be-cached items can have dependencies. We assume that the universe is a tree T and items are tree nodes; we require that if a node v is cached then the whole subtree T(v) rooted at v is cached as well. This theoretical problem finds an immediate application in the context of forwarding table optimization in IP routing and software-defined networks. We present an elegant online deterministic algorithm TC for this problem, and rigorously prove that its competitive ratio is O(height(T) * k_ALG/(k_ALG-k_OPT+1)), where k_ALG and k_OPT denote the cache sizes of an online and the optimal offline algorithm, respectively. The result is optimal up to a factor of O(height(T)).

cs.DS↗