arXiv:2601.10740v1 Announce Type: new Abstract: Modern neural networks rely on generic activation functions (ReLU, GELU, SiLU) that ignore the mathematical structure inherent in scientific data. We propose Neuro-Symbolic Activation Discovery, a framework that uses Genetic Programming to extract interpretable mathematical formulas from data and inject them as custom activation functions. Our key contribution is the discovery of a Geometric Transfer phenomenon: activation functions learned from particle physics data successfully generalize to ecological classification, outperforming standard activations (ReLU, GELU, […]

Ver mais

Autonomous navigation in highly constrained environments remains challenging for mobile robots. Classical navigation approaches offer safety assurances but require environment-specific parameter tuning; end-to-end learning bypasses parameter tuning but struggles with precise control in constrained spaces. To this end, recent robot learning approaches automate parameter tuning while retaining classical systems’ safety, yet still face challenges in generalizing to unseen environments. Recently, Vision-Language-Action (VLA) models have shown promise by leveraging foundation models’ scene understanding capabilities, but still struggle with precise […]

Ver mais

arXiv:2603.17074v1 Announce Type: new Abstract: We present PRISM, a comprehensive empirical study of mid-training design choices for large language models. Through controlled experiments across seven base models spanning four families (Granite, LLaMA, Mistral, Nemotron-H), two architecture types (dense Transformer and attention-Mamba hybrid), and scales from 3B to 24B parameters, we show that mid-training on approximately 27B high-quality tokens yields consistent gains of +15 to +40 points on math, +5 to +12 points on code, and +6 to +13 […]

Ver mais

arXiv:2410.10258v2 Announce Type: replace-cross Abstract: Linear bandits have become a cornerstone of online learning and sequential decision-making, providing solid theoretical foundations for balancing exploration and exploitation. Within this domain, matrix sketching serves as a critical component for achieving computational efficiency, especially when confronting high-dimensional problem instances. The sketch-based approaches reduce per-round complexity from $Omega(d^2)$ to $O(dl)$, where $d$ is the dimension and $l

Ver mais

Van Rooij, Guest, Adolfi, Kolokolova, and Rich claimed to have proven that AGI via ML is impossible in Computational Brain & Behavior in 2024. The basic idea was to try to reduce a known NP-hard problem to the problem of learning a human-level classifier from data. The purported result, called “Ingenia Theorem” by the authors, made some noise on the internet, including here. My paper showing that the proof is irreparably broken is now also out in CBB […]

Ver mais

Let’s be honest: labeling data is the absolute worst part of computer vision. For years, the standard playbook was to scrape millions of images, pay an army of annotators to draw bounding boxes, and feed them into a supervised Convolutional Neural Network. It worked, but it didn’t scale. Then self-supervised learning (SSL) came along, promising a world where models could learn from raw, unlabeled images. But early contrastive methods like SimCLR and MoCo were notoriously heavy, requiring massive batch […]

Ver mais

Reinforcement learning algorithms are defined by their learning update rules, which are typically hand-designed and fixed. We present an evolutionary framework for discovering reinforcement learning algorithms by searching directly over executable update rules that implement complete training procedures. The approach builds on REvolve, an evolutionary system that uses large language models as generative variation operators, and extends it from reward-function discovery to algorithm discovery. To promote the emergence of nonstandard learning rules, the search excludes canonical mechanisms such […]

Ver mais

arXiv:2602.15986v1 Announce Type: new Abstract: We study convergence rates of random-order best-response dynamics in games on networks with linear best responses and strategic substitutes. Combining formal analysis with numerical simulations we identify phenomena that lead to slow convergence. One of the key such phenomena is convergence to stable strategy profiles in parts of the network neighboring sets of nodes which remain inactive until the dynamics is close to converging and then switch to activity, initiating convergence to profiles […]

Ver mais