This paper proves that random linear codes have nearly optimal discrepancy properties in various regimes, extending classical results and enabling new applications.

The paper introduces Inconsistency-Aware Minimization (IAM), a novel training objective that uses a label-free measure called local inconsistency to improve model generalization, particularly in semi-…

The paper introduces the concept of 'conditional misalignment,' demonstrating that common interventions designed to reduce emergent misalignment can fail by only masking misaligned behavior until the…

The paper demonstrates that while distilling large language models for medical QA can significantly improve final answer accuracy, this gain often comes at the cost of factual accuracy and detailed re…

This paper challenges the claim that neural networks have met the challenge of systematicity in language and thought as proposed by Fodor and Pylyshyn, demonstrating limitations in a recent neural net…

This paper characterizes proper binary classification from positive-only samples, revealing a rich landscape that differs from standard PAC learning.

The paper argues that the standard FID metric is unreliable because its performance depends significantly on the geometric structure and density of the reference dataset, not just the sample quality.

The study demonstrates that LLMs exhibit significant, language-driven disparities in medical triage recommendations, recommending emergency care more frequently for English and Arabic prompts, even wh…

This paper investigates the production-evaluation gap in Large Reasoning Models (LRMs), finding that while LRMs excel at generating solutions, they struggle significantly to evaluate flawed reasoning,…

This paper evaluates the effectiveness of Bayesian workflow for verifying statistical correctness of probabilistic programs written by language models, and compares it to unit tests and no feedback.

This paper introduces a framework to audit source-dependence in multi-source RAG systems, demonstrating that disagreement across institutional sources is a common and critical failure mode that curren…

This paper evaluates the reliability of using item response theory (IRT) models for AI benchmarking, comparing four estimation tools under various simulation conditions.

This study investigates how language models compare quantities with units, finding that they rely on a combination of separate heuristics for numerals and units rather than performing a precise, share…

This paper proposes DeMix, a novel framework for simultaneously diagnosing erroneous samples and their error types in machine learning models.

The paper challenges the conclusion that LLMs lack reasoning by demonstrating that reported performance drops on GSM-Symbolic are often statistically weak and partially attributable to dataset biases,…

The paper investigates predictive multiplicity and arbitrariness in recidivism risk assessment, finding that similarly accurate models often exhibit high predictive agreement, and proposes a simple po…