108 stories tagged Language Models · page 5 of 5
Researchers demonstrate that individual experts in sparse Mixture-of-Experts (MoE) models have causally meaningful identities. This discovery enables more precise control over model behavior through geometric routing techniques.
Researchers have developed a method to quantify exploration and exploitation in language model agents without accessing their internal policies. This breakthrough could improve AI decision-making in complex tasks.
A new study reveals that large language models, including GPT-4o, perform poorly on abstract meaning comprehension tasks. The findings highlight significant challenges in interpreting non-concrete, high-level semantics.
Researchers introduce GiantsBench, a dataset to evaluate AI's ability to anticipate scientific insights. The benchmark includes 17,000 examples across eight domains, challenging models to predict future discoveries from foundational papers.
Researchers introduce StaRPO, a reinforcement learning framework that improves logical consistency in language models. It augments traditional RL with stability constraints to capture internal reasoning structures.
Researchers developed a mathematical model to analyze how generated text shapes the public record, identifying two key forces: drift and selection. This work provides insights into the long-term evolution of language in AI systems.
Researchers found that discrete speech units (DSUs) struggle to reliably encode suprasegmental information like lexical tone. This poses challenges for tasks where prosody is crucial, such as text-to-speech and multimodal dialogue systems.
A new study explores how emotions affect decision-making in small language models (SLMs) by using activation steering for controlled emotional states. The research introduces a game-theoretic benchmark to evaluate these interactions.
Researchers introduce CAMO, an ensemble technique designed to improve language model performance on imbalanced datasets by dynamically boosting minority classes. The method uses vote distributions, confidence calibration, and inter-model uncertainty to enhance underrepresented class predictions.
Researchers propose a weak supervision framework to detect hallucinations in large language models. This approach enables hallucination detection from internal activations alone at inference time.
SELFDOUBT is a new framework for uncertainty quantification in reasoning language models. It addresses the difficulty of deploying uncertainty estimation in practice, particularly for proprietary APIs.
Researchers trained mRNA language models across 25 species for $165. This breakthrough has significant implications for bioinformatics and natural language processing.