src/methods/linear_probe.py

"""M5 — LinearProbe: Logistic regression weight vector as steering direction."""

import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler

from src.methods.base import SteeringMethod


class LinearProbe(SteeringMethod):
    """LinearProbe — Logistic regression weight vector."""

    @property
    def name(self) -> str:
        return "LinearProbe"

    @property
    def method_id(self) -> str:
        return "M5"

    def extract_vector(
        self,
        h_pos: np.ndarray,
        h_neg: np.ndarray,
        **kwargs,
    ) -> np.ndarray:
        """Compute logistic regression weight vector.

        Args:
            h_pos: (N_pos, d) positive activations
            h_neg: (N_neg, d) negative activations

        Returns:
            (d,) weight vector direction
        """
        X = np.concatenate([h_pos, h_neg], axis=0)
        y = np.concatenate([
            np.ones(len(h_pos)),
            np.zeros(len(h_neg)),
        ])

        scaler = StandardScaler()
        X_scaled = scaler.fit_transform(X)

        C = kwargs.get("C", 1.0)
        max_iter = kwargs.get("max_iter", 5000)

        lr = LogisticRegression(C=C, max_iter=max_iter, solver="lbfgs")
        lr.fit(X_scaled, y)

        # Get weight vector in original space
        w = lr.coef_[0] / scaler.scale_
        w = w / (np.linalg.norm(w) + 1e-8)
        return w