The Piggyback Hypothesis explains how chat-template tokens can cause emergent misalignment in LLMs by generalizing finetuned behavior to out-of-domain queries. Token-Regularized Finetuning (TReFT) is proposed to mitigate this issue, preserving in-domain learning while reducing misalignment across models and datasets.
This study investigates cross-prompt generalization in detecting AI-generated fake news using interpretable linguistic features like lexical diversity and readability. A random forest classifier achieved consistently high performance (AUC 0.988-1.000) across various train-test combinations, demonstrating robustness against different prompting strategies.
This paper introduces the Langevin Gradient Descent (LGD) algorithm for convex regression problems, proving that optimal hyperparameter configurations achieve the Bayes' optimal solution. The work also provides generalization guarantees for meta-learning LGD's optimal hyperparameters, with a pseudo-dimension bound of O(dh).
This survey reviews deep learning methods for cross-subject EEG decoding, addressing the challenge of high inter-subject variability and domain shift. It categorizes current literature into methodological families like feature alignment and contrastive learning, emphasizing rigorous evaluation and theoretical considerations.
This paper challenges the conventional view that flat minima inherently lead to better generalization, showing that function-preserving reparameterization can drastically alter a minimum's perceived sharpness. It introduces "weakness"—a reparameterization-invariant measure based on what the network does—as the actual driver of generalization, proving its minimax optimality and correlation with PAC-Bayes bounds.
This paper identifies normalized spectral entropy as a scalar order parameter for the grokking transition, where models generalize long after memorization. The research shows that entropy collapse precedes generalization, and causal interventions confirm its critical role, providing a predictive model for grokking onset.
This research explores a differential-privacy based approach to improve generalization and prevent overfitting in Deep Neural Networks. Overfitting, where models learn noise and perform poorly on unseen data, is a growing challenge in modern AI systems.
This paper introduces predictable history-adaptive virtual perturbations to enhance information-theoretic generalization bounds for Stochastic Gradient Descent. This new approach allows perturbation covariances to dynamically depend on past SGD history, addressing limitations of existing methods that require fixed covariances.
This research introduces a novel delayed per-step reward attribution method for training language model agents in multi-agent strategic interactions. It addresses the challenge of entangled outcomes by computing rewards at episode end and backpropagating them, enabling stable and sample-efficient reinforcement learning.
MAVEN (Modular Agentic Verification and Execution Network) is a lightweight symbolic reasoning scaffold designed to improve generalization in agentic tool-calling environments. It has been evaluated across established benchmarks, and introduces MAVEN-Bench, a new stress-test benchmark for multi-step mathematical and physical reasoning.
This paper introduces the Configurable Safety Reward Model (CSRM) to address the challenge of aligning LLMs with heterogeneous and rapidly evolving safety requirements. CSRM substantially improves generalization to previously unseen safety configurations by being jointly optimized for calibrated safety compliance and reward modeling, achieving state-of-the-art performance on benchmarks.
O texto aborda as limitações de generalização do Reinforcement Learning (RL), onde políticas aprendidas falham fora da distribuição de treinamento. Propõe uma nova taxonomia de contextos (alógenos e autógenos) e identifica direções de pesquisa cruciais para desenvolver uma verdadeira inteligência contextual.