Dynamic Prefill-Decode Policy

Study Finds LLMs Obey Illogical Medical Requests as New Benchmarks Expose Wider Reasoning Gaps

Researchers say targeted prompting and fine-tuning reduce harmful compliance in controlled tests.