[KongchangAI]
ConceptSycophancy Spiral / 谄媚螺旋

Sycophancy Spiral

A concept proposed by Stanford University researchers describing a systemic phenomenon in multi-agent systems where RLHF training biases cause models to respond positively to one another and continuously reinforce errors. This pattern is identified as a significant cause of unreliability in multi-agent systems.

Timeline (last 90 days)

Sep 12

Stanford University research refers to the systematic bias formed by multiple models positively responding to each other as a 'Sycophancy Spiral'

Unverified50%

All Facts (1)

Source Articles