@math_rachel
"In zero-shot settings, preference following accuracy falls below 10% at merely 10 turns (~3k tokens) across most evaluated models. "Even with advanced prompting and retrieval methods, preference following still deteriorates in long-context conversations." 8/