This paper introduces Cognivia, an AI therapist designed to assist with Cognitive Behavioral Therapy by identifying cognitive distortions and generating evidence-based rational responses. It addresses the scalability limitations of traditional CBT and the shortcomings of existing LLM-based mental health applications by grounding its approach in authoritative CBT literature and an expert-validated evaluation framework.

TIGA introduces a novel, training-free framework for generating detector-evasive images against black-box AI-generated content (AIGC) detectors. It directly steers the latent Denoising Diffusion Implicit Model (DDIM) trajectory during generation, avoiding post-hoc perturbations or model retraining, while maintaining high perceptual quality and robustness.

This paper explores how competitive pressure in AI development influences safety choices. Through a behavioral experiment and an evolutionary model, it demonstrates that unsafe development is primarily driven by fear of falling behind and strategic responses to competitors, rather than individual risk preferences alone.

This paper introduces a formal logic of inability as a first-class concept, extending Coalition Logic to systematically study what multi-agent coalitions cannot achieve. It establishes the modal properties of this explicit inability operator, highlighting its significance for reasoning about constraints and safety in AI systems.

This paper introduces a Mixture of Experts (MOE)-powered semi-supervised inference framework that enhances Prediction-Powered Inference (PPI) by leveraging multiple predictors. The framework adapts to unknown predictor performance, combines their collective power, and offers a best-expert guarantee, improving inferential efficiency with abundant unlabeled data.