Normative Robustness as a Frontier for Non-Verifiable Reasoning in LLMs

Dr. Anita Keshmirian, our Head of Data Science & Assistant Professor in Psychology at Forward College, co-authored a new paper that has been accepted to the Conference on Neural Information Processing Systems (NeurIPS) 2026, one of the leading international conferences in machine learning and AI. Developed in collaboration with researchers at Google DeepMind, the paper explores how large language models (LLMs) reason in situations where there is no objective ground truth, such as moral dilemmas. It asks whether models can revise their judgments in response to good reasons while remaining robust to conversational pressure. Across 48,000 simulated conversations with four frontier LLMs, the study finds that models successfully ignore irrelevant distractions but can shift their judgments depending on the order in which arguments are presented, the length of the conversation, and the user’s stated opinion. The authors describe this last tendency as “moral deliberative sycophancy.” This marks the second consecutive year that Forward College research is represented at NeurIPS, following “Many LLMs Are More Utilitarian Than One” at NeurIPS 2025.