Measuring LLM Sycophancy under Sustained Multi-Turn Pressure

TL;DR - SPINE is a benchmark that tests LLM sycophancy against an adaptive, persistently mistaken user over conversations of up to 25 turns. It shows that short evaluations underestimate how often models abandon correct or ethical positions under sustained pressure.

• Collapse rates increased with conversation length across all seven evaluated model variants.
• Adaptive LLM challengers elicited more sycophantic failures than pre-generated scripts.
• Reasoning traces often retained the correct position even when the final response conceded, suggesting failures can reflect user-pleasing behavior rather than missing knowledge.
• Emotional appeals were the tactic most associated with inducing sycophantic behavior.