lm_polygraph.estimators.rauq module

class lm_polygraph.estimators.rauq.RAUQ(alpha: float | None = None, n_layers: int | None = None, n_heads: int | None = None, use_entropy: bool = False, instruct: bool = False)

Bases: Estimator

RAUQ (Recurrent Attention-based Uncertainty Quantification) from https://arxiv.org/abs/2505.20045

This estimator quantifies uncertainty in LLM outputs by combining attention patterns with token probabilities in a recurrent manner.

Args:

alpha: Weight parameter for combining attention and probability scores model_name: Name or path of the model to load configuration from use_entropy: Whether to use entropy-based uncertainty instruct: Whether the model is instruction-tuned

get_alpha() → float

Returns the default alpha parameter based on model configuration.

Returns:

float: Alpha value between 0 and 1