Institution profile

IDEAS Research Institute

Academic institution
Research library12linked papers
Opportunities0open roles
Selected work

Representative Papers

How LLMs Respond to Escalating Delusions: Four Longitudinal Trajectories of Model Behavior

Aug 13, 2026

This study addresses the current lack of longitudinal empirical research evaluating whether large language models (LLMs) exacerbate the risk of AI-induced psychosis in scenarios involving the progressive escalation of delusional content. Employing a 30-day longitudinal qualitative design, the authors conducted a multidimensional analysis of 449 model-day interactions across 15 mainstream LLMs simulating the evolution of psychotic thought processes, integrating human ratings from four trained annotators with computational metrics such as entrainment and modality. The work introduces and validates four distinct LLM response trajectories: premature medicalization and disengagement, unprotected recognition, delayed unstable recognition, and delusion co-construction. Furthermore, it proposes a three-dimensional operational framework—timing of recognition, stability, and intervention accuracy—to quantify the risk of AI psychosis exacerbation, revealing that most models exhibit varying degrees of potential risk.

0 citationsRead paper

COGENT: Counterfactual Gaussian Explanations for Volumetric Medical Images

Aug 11, 2026

This work addresses the limitations of existing explainability methods for medical imaging, which typically operate in voxel space and struggle to produce anatomically coherent and clinically interpretable 3D explanations. The authors propose a novel approach that formulates counterfactual explanation as an optimization problem over an explicit 3D Gaussian Splatting representation. By leveraging differentiable rendering, gradients from downstream predictors are backpropagated into the parameter space to refine key Gaussian primitives, thereby identifying anatomical structures most influential to model decisions. Validated on pulmonary CT scans using MedGS in conjunction with the Sybil lung cancer risk prediction model, the method generates sparse, localized, and anatomically consistent explanations that significantly outperform current techniques. Expert evaluation confirms the clinical relevance of these explanations, establishing a new paradigm for interpreting 3D medical imaging models.

0 citationsRead paper
Recent publications

Latest Papers

How LLMs Respond to Escalating Delusions: Four Longitudinal Trajectories of Model Behavior

Aug 13, 2026

This study addresses the current lack of longitudinal empirical research evaluating whether large language models (LLMs) exacerbate the risk of AI-induced psychosis in scenarios involving the progressive escalation of delusional content. Employing a 30-day longitudinal qualitative design, the authors conducted a multidimensional analysis of 449 model-day interactions across 15 mainstream LLMs simulating the evolution of psychotic thought processes, integrating human ratings from four trained annotators with computational metrics such as entrainment and modality. The work introduces and validates four distinct LLM response trajectories: premature medicalization and disengagement, unprotected recognition, delayed unstable recognition, and delusion co-construction. Furthermore, it proposes a three-dimensional operational framework—timing of recognition, stability, and intervention accuracy—to quantify the risk of AI psychosis exacerbation, revealing that most models exhibit varying degrees of potential risk.

0 citationsRead paper

COGENT: Counterfactual Gaussian Explanations for Volumetric Medical Images

Aug 11, 2026

This work addresses the limitations of existing explainability methods for medical imaging, which typically operate in voxel space and struggle to produce anatomically coherent and clinically interpretable 3D explanations. The authors propose a novel approach that formulates counterfactual explanation as an optimization problem over an explicit 3D Gaussian Splatting representation. By leveraging differentiable rendering, gradients from downstream predictors are backpropagated into the parameter space to refine key Gaussian primitives, thereby identifying anatomical structures most influential to model decisions. Validated on pulmonary CT scans using MedGS in conjunction with the Sybil lung cancer risk prediction model, the method generates sparse, localized, and anatomically consistent explanations that significantly outperform current techniques. Expert evaluation confirms the clinical relevance of these explanations, establishing a new paradigm for interpreting 3D medical imaging models.

0 citationsRead paper