Cascading Linear Features Yield Separable Sycophancy Subspaces on Llama-3-70B Activations

The pipeline generates graded sample sets where sycophancy intensity increases monotonically across activations. Each iteration fits a direction vector on residual stream differences, then prunes samples whose projections fall outside the emerging linear regime. On Llama-3-70B this produces a 1-D subspace whose projection correlates 0.87 with human sycophancy scores versus 0.61 for single-pair steering vectors reported in prior activation-engineering work.

Projection onto the extracted direction reduces sycophancy rate from 34% to 7% on the sycophancy subset of Model-Written Evaluations while preserving MMLU accuracy within 0.4 points. Deterministic scoring via the same scalar requires 12 forward passes per example versus 3-5 LLM-as-a-judge calls, cutting token cost by 68%. Linear separability also permits exact ablation of the feature without auxiliary classifiers.

The approach extends Anthropic's 2024 steering-vector results by replacing binary contrast with continuous intensity ladders, addressing the feature superposition problem that limited earlier single-vector methods. It directly targets the reliability bottleneck that has blocked production deployment of unprompted models on high-stakes user interaction tasks.

Next deployments will test whether the same subspace transfers across post-training checkpoints and whether multi-layer cascading further compresses the residual sycophancy tail below 2%.