Neural networks increasingly combine data across populations, time periods, and operating conditions to improve generalization. This raises a reliability question: whether a model refitted on pooled data preserves an action ordering supported by both sources.
arXiv:2607. 27255v1 Announce Type: cross Abstract: Neural networks increasingly combine data across populations, time periods, and operating conditions to improve generalization.
arXiv:2607. 11347v1 Announce Type: new Abstract: Neural networks increasingly guide decisions in high-stakes domains such as medical diagnosis, credit approval, and energy bidding.
arXiv:2602. 24266v2 Announce Type: replace-cross Abstract: Which internal mechanisms of a neural network can be replaced while preserving the computation it performs?
Matryoshka Attribution (MAttr) is a mask‑learning method that identifies nested subsets of a language model’s internal components by minimizing downstream loss. It uses a differentiable sigmoid top‑k operator and randomizes sparsity during training to produce an attribution ordering of components. MAttr tops the Mechanistic Interpretability Benchmark leaderboard and can be applied via reinforcement learning to pinpoint weight changes that control behaviors such as refusal in Llama 3.1 8B Instruct, where restoring just 1% of weights removes refusals while preserving capabilities.
arXiv:2606. 10632v1 Announce Type: cross Abstract: Lipschitz-style individual fairness formalizes the idea that semantically similar examples should receive similar predictions, but its evaluation in multi-task learning (MTL) can be confounded by method-induced representation scales.
arXiv:2607. 20201v1 Announce Type: cross Abstract: Additive models buy interpretability by forbidding feature interactions, a constraint that neural instantiations enforce architecturally.
arXiv:2604. 27733v2 Announce Type: replace Abstract: Aligning Large Language Models (LLMs) with human intent, whether through explicit reward modeling or direct methods such as DPO, fundamentally relies on minimizing a surrogate loss as a proxy for the true pairwise ranking objective.
The paper investigates the consistency of surrogate loss methods for classification and policy learning when the set of admissible classifiers is constrained, such as by interpretability or fairness requirements. It shows that hinge loss is the only surrogate that preserves consistency when constraints limit only the prediction set, but consistency can fail if constraints also restrict the functional form. The authors derive conditions guaranteeing consistency for hinge-risk-minimizing classifiers and use these results to design efficient hinge-loss-based procedures for monotone classification problems.
Additive models buy interpretability by forbidding feature interactions, a constraint that neural instantiations enforce architecturally. We introduce the quadrilateral loss, a differentiable penalty that treats additivity as a measurable behavior instead: a second-order mixed difference on pairs of training points swapping one coordinate, which vanishes if and only if the coordinate carries no interaction, remains informative for piecewise-linear networks, and equals in expectation the per-coordinate interaction mass of the interventional Shapley-GAM.
The paper investigates whether neural networks retain a case-based structure in their learned representations, enabling the decomposition of decision margins into contributions from individual training cases. By linking neural networks to Case-Based Decision Theory (CBDT), the authors identify conditions under which this recovered structure can be interpreted within CBDT and demonstrate that the resulting interpretation is consistent with the network’s original decisions. Experiments on a controlled CBDT setting and three real-world decision tasks confirm the viability of this approach.
arXiv:2606. 03904v1 Announce Type: new Abstract: Multi-objective optimization (MOO) underlies many machine learning problems, yet MOO solvers across the loss-balancing, gradient-balancing, and Pareto-based families almost universally hand their reconciled directions to Adam~\cite{kingma2015adam}.
The paper presents a new scaling law for reward optimization in AI alignment, showing that performance scales as Θ(√min{log(M), K}), where M is the number of preference comparisons used to train a proxy reward model and K is the KL‑divergence budget relative to a reference policy. The authors derive this law using an information‑theoretic model, prove its tightness, and validate it with extensive experiments involving a 70B gold reward model and smaller proxy models (0.6B–4B). The empirical results demonstrate a strong fit (R² 97–99 %) across different model sizes, noise levels, and optimization methods, suggesting that reward optimization behaves like a simple selection task over IID Gaussian variables with noisy feedback.