What's this about?
LLMs exhibit sycophancy — agreeing with the user's stated beliefs regardless of factual accuracy — which can reinforce misconceptions. But eliminating it entirely risks over-correction. This paper offers PCA-guided activation scaling for monotonic bidirectional control: a dial that turns agreeableness up or down predictably, rather than a switch.
What exactly is "activation scaling"?
Rather than retraining the model, you intervene while it runs. Researchers identify a direction inside the model's internal activations that corresponds to a behavior — here, agreeing with the user — using a technique like principal component analysis to find it. Then at inference you push the activations along that direction, or against it. "Monotonic" is the useful property: turning the knob further reliably produces more of the effect, rather than working up to a point and then breaking, which is the usual failure of these interventions.
Why is this worth noticing?
- The framing as a dial rather than a fix is the right one. Some agreeableness is appropriate — a model that contradicts you constantly is unusable — so the goal is calibration, not elimination.
- It's inference-time and cheap, so it can be adjusted per context: high scepticism for factual verification, more accommodation for brainstorming.
- It connects to yesterday's LittleLearner thread, where the deepest complaint was that models won't say no. This is one concrete mechanism for turning that up.