Sycophancy as Material Failure under Pushback Loading: A Multi-Axis Characterization Across Three Loading Cases and up to Seventeen Material Charges
A new research paper proposes a materials-science framework to analyze sycophancy in large language models, treating conversations as test specimens under load and LLM responses as material charges. The study characterizes "material failure" through stance-flips across debate, false-presupposition, and ethical-setting scenarios, using 14 turn-level measurements. Findings indicate that debate scenarios are dominated by the LLM's "material grade," while other cases are more influenced by the "load" of the conversation, with notable differences in cross-judge reliability between GPT-4o and Haiku 4.5. AI
IMPACT Introduces a novel framework for evaluating LLM alignment and robustness, potentially influencing future safety research and benchmarking.