src/methods/reps.py

"""M9 — RePS: Representation-based Prompt Steering.

Lightly-trained method that learns steering through representation optimisation.
"""

import numpy as np

from src.methods.base import SteeringMethod


class RePS(SteeringMethod):
    """RePS — Representation-based Prompt Steering."""

    def __init__(self, **kwargs):
        self._trained = False
        self._direction: np.ndarray = None

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

    @property
    def method_id(self) -> str:
        return "M9"

    @property
    def is_training_free(self) -> bool:
        return False

    def extract_vector(
        self,
        h_pos: np.ndarray,
        h_neg: np.ndarray,
        **kwargs,
    ) -> np.ndarray:
        if self._direction is not None:
            return self._direction
        # Fallback
        from src.methods.diffmean import DiffMean
        return DiffMean().extract_vector(h_pos, h_neg)

    def train(self, train_data: dict) -> None:
        """Train RePS steering.

        TODO: Implement representation-based prompt steering following
              AxBench catalogue.
        """
        raise NotImplementedError("RePS training not yet implemented.")