Sycophancy Hides Linearly in the Attention Heads
This work addresses the issue of sycophancy in AI systems for users relying on factual accuracy, though it is incremental as it builds on existing linear representation hypotheses and probing methods.
The study tackled the problem of sycophancy in language models by identifying that correct-to-incorrect sycophancy signals are most linearly separable in attention heads, and found that targeted linear interventions in these heads can effectively mitigate sycophancy, with probes trained on TruthfulQA transferring to other factual QA benchmarks.
We find that correct-to-incorrect sycophancy signals are most linearly separable within multi-head attention activations. Motivated by the linear representation hypothesis, we train linear probes across the residual stream, multilayer perceptron (MLP), and attention layers to analyze where these signals emerge. Although separability appears in the residual stream and MLPs, steering using these probes is most effective in a sparse subset of middle-layer attention heads. Using TruthfulQA as the base dataset, we find that probes trained on it transfer effectively to other factual QA benchmarks. Furthermore, comparing our discovered direction to previously identified "truthful" directions reveals limited overlap, suggesting that factual accuracy, and deference resistance, arise from related but distinct mechanisms. Attention-pattern analysis further indicates that the influential heads attend disproportionately to expressions of user doubt, contributing to sycophantic shifts. Overall, these findings suggest that sycophancy can be mitigated through simple, targeted linear interventions that exploit the internal geometry of attention activations.