arXiv:2601.05478v1 Announce Type: new Abstract: To reliably assist human decision-making, LLMs must maintain factual internal beliefs against misleading injections. While current models resist explicit misinformation, we uncover a fundamental vulnerability to sophisticated, hard-to-falsify evidence. To systematically probe this weakness, we introduce MisBelief, a framework that generates misleading evidence via collaborative, multi-round interactions among multi-role LLMs. This process mimics subtle, defeasible reasoning and progressive refinement to create logically persuasive yet factually deceptive claims. Using MisBelief, we generate 4,800 instances […]

Ver mais

arXiv:2601.08316v1 Announce Type: cross Abstract: Deep double descent is one of the key phenomena underlying the generalization capability of deep learning models. In this study, epoch-wise double descent, which is delayed generalization following overfitting, was empirically investigated by focusing on the evolution of internal structures. Fully connected neural networks of three different sizes were trained on the CIFAR-10 dataset with 30% label noise. By decomposing the loss curves into signal contributions from clean and noisy training data, the […]

Ver mais

arXiv:2601.08599v1 Announce Type: cross Abstract: We study the free energy for pure and mixed spherical $p$-spin models with i.i.d. disorder. In the mixed case, each $p$-interaction layer is assumed either to have regularly varying tails with exponent $alpha_p$ or to satisfy a finite $2p$-th moment condition. For the pure spherical $p$-spin model with regularly varying disorder of tail index $alpha$, we introduce a tail-adapted normalization that interpolates between the classical Gaussian scaling and the extreme-value scale, and we […]

Ver mais

Artificial intelligence (AI) is increasingly integrated into scientific discovery processes, such as protein design, gene analysis, and materials research, significantly enhancing the efficiency of discoveries in these fields. While much recent literature emphasizes fully automated pipelines, it is crucial to acknowledge that scientific discovery is inherently a creative and high-stakes endeavor. Therefore, it relies heavily on human expertise for judgment and guidance, especially in the face of uncertainty. Despite rapid growth in human-in-the-loop and collaborative systems, the field […]

Ver mais

arXiv:2601.00021v1 Announce Type: new Abstract: We present a physical theory of intelligence grounded in irreversible information processing in systems constrained by conservation laws. An intelligent system is modelled as a coupled agent-environment process whose evolution transforms information into goal-directed work. To connect information to physical state, we introduce the Conservation-Congruent Encoding (CCE) framework, in which encodings correspond to metastable basins of attraction whose separability is enforced by conservation laws. Within this framework, intelligence is defined as the amount […]

Ver mais

Rough working notes, musing out loud. Much effort in machine learning and AI research is focused on a few broad classes of problem. Three examples of such classes are: Classifiers, which do things like classify images according to their category, generalizing from their training data so they can classify previously unseen data in the wild; Generative models, which are exposed to data from some distribution (say, images of houses), and then build a new model which can generate […]

Ver mais

Deep learning, despite its remarkable successes, is a young field – perhaps ten years old. While models called artificial neural networks have been studied for decades, much of that work seems only tenuously connected to modern results. It’s often the case that young fields start in a very ad-hoc manner. Later, the mature field is understood very differently than it was understood by its early practitioners. It seems quite likely that deep learning is in this ad-hoc state… […]

Ver mais

arXiv:2512.22909v2 Announce Type: replace-cross Abstract: In this paper we study a nonconvex-strongly-concave constrained minimax problem. Specifically, we propose a first-order augmented Lagrangian method for solving it, whose subproblems are nonconvex-strongly-concave unconstrained minimax problems and suitably solved by a first-order method developed in this paper that leverages the strong concavity structure. Under suitable assumptions, the proposed method achieves an operation complexity of $O(varepsilon^{-3.5}logvarepsilon^{-1})$, measured in terms of its fundamental operations, for finding an $varepsilon$-KKT solution of the constrained minimax […]

Ver mais

With the goal to generate more scalable algorithms with higher efficiency and fewer open parameters, reinforcement learning (RL) has recently moved towards combining classical techniques from optimal control and dynamic programming with modern learning techniques from statistical estimation theory. In this vein, this paper suggests to use the framework of stochastic optimal control with path integrals to derive a novel approach to RL with parametrized policies. While solidly grounded in value function estimation and optimal control based on […]

Ver mais