Technical Summary: D-Score for Hallucination Detection

Problem Statement

Large Language Models (LLMs) frequently generate fluent text containing statements that are false, unsupported by evidence, or inconsistent with the model's internal knowledge, a phenomenon known as hallucination. Current detection methods often rely on external resources (retrieval-based systems) or multiple generation passes (consistency-based methods), which increase inference costs and complexity. There is a need for detection mechanisms that operate directly on the internal states of a model during a single forward pass, without requiring external verifiers or repeated sampling.

Methodology: The D-Score

The paper introduces the D-Score, a spectral statistic derived from the geometry of hidden activations within a transformer model. The core intuition is that a well-supported text induces a coherent hidden trajectory concentrated along a few dominant directions. Conversely, when a model processes a text containing an unsupported claim that conflicts with its internal knowledge, the hidden representation may encode both the asserted content and traces of counter-evidence, uncertainty, or lack of support. This conflict causes the hidden trajectory to spread across additional singular directions, increasing the "effective" dimensionality of the activation matrix.

Mathematical Formulation

Given a fixed model $m$, a specific layer $j$, and an input text $\mathbf{x} = (x_1, \dots, x_T)$, the method constructs a hidden activation matrix $H^{(j)}(\mathbf{x}; m) \in \mathbb{R}^{d \times T}$, where columns represent the hidden states of tokens at layer $j$.

Let $\sigma_1 \ge \sigma_2 \ge \dots \ge \sigma_r > 0$ be the non-zero singular values of $H^{(j)}(\mathbf{x}; m)$. The D-Score is defined as the relative numerical rank of this matrix:
$$D^{(j)}_{\tau}(\mathbf{x}; m) = \max \{ n \le r : \frac{\sigma_1(H)}{\sigma_n(H)} \le \tau \}$$
where $\tau > 1$ is a tolerance parameter. The score counts how many singular directions remain significant relative to the leading singular value $\sigma_1$.

Detection Rule

The detector classifies an input as hallucinated if the D-Score exceeds a calibrated decision level $\bar{D}$:
$$\hat{Y}(\mathbf{x}) = 1 \quad \text{if} \quad D_{\tau}(\mathbf{x}) > \bar{D}$$
This approach requires only a single forward pass and the computation of the leading singular values (up to the point where they fall below the relative threshold), which can be efficiently estimated using iterative methods like the power method on the Gram matrix $H^\top H$, avoiding the computational cost of a full SVD for high-dimensional models.

Theoretical Justification

The paper defines model-internal hallucination as an unsupported claim to which the model assigns low internal truth belief. It posits that such claims introduce "conflict directions" in the hidden state space. A theoretical proposition demonstrates that if unsupported content introduces $q$ sufficiently strong directions orthogonal to the coherent representation, the D-Score increases by at least $q$. Thus, a high D-Score serves as a proxy for the presence of these conflict directions.

Key Contributions

1. Introduction of the D-Score: A simple spectral statistic representing the relative numerical rank of the hidden activation matrix, designed to measure the spread of the hidden-state trajectory.
2. Operational Definition: A formal definition of model-internal hallucination based on unsupported claims with low internal truth belief, distinguishing them from errors the model cannot internally detect.
3. Geometric Argument: A theoretical justification showing that hallucinated texts can increase the D-Score by spreading the hidden trajectory across additional conflict directions (uncertainty, counter-evidence).
4. Empirical Evaluation: Comprehensive evaluation demonstrating that the D-Score is a strong signal for hallucination detection across multiple models and datasets.

Experimental Results

The D-Score was evaluated on two benchmarks: FAVA-Annotation (factual errors in open-domain generation) and RAGTruth (hallucinations in retrieval-augmented generation). Experiments were conducted on three open-weight models: Llama-2-7B, Llama-3-8B, and Vicuna-7B.

• Performance: The D-Score consistently outperformed state-of-the-art baselines, including Hidden Score (a prior spectral method), perplexity-based scores, and entropy-based metrics.
  • On FAVA-Annotation, D-Score achieved the highest AUROC across all models (e.g., 66.98% for Llama-3-8B vs. 57.10% for Hidden Score).
  • On RAGTruth, D-Score maintained robust performance (e.g., 61.53% AUROC for Llama-3-8B), whereas Hidden Score performance dropped significantly on Vicuna-7B (near random chance), while D-Score remained effective (60.20%).
• Model Capacity Correlation: The performance gap between D-Score and baselines was more pronounced in more capable models (e.g., Llama-3-8B), suggesting the method effectively captures hallucinations that the model internally recognizes as non-factual.
• Subset Analysis: When evaluated on a subset of data where the model's explicit self-assessment aligned with human annotations (filtering out cases where the model failed to recognize its own errors), the D-Score achieved even higher AUROC scores (up to 72.51% on Llama-3-8B), confirming that the method relies on the model's internal representation of the error.

Significance and Claims

The paper claims that the D-Score provides a strong hidden-state signal for hallucination detection that operates under strict efficiency constraints:

• Single-Pass Efficiency: It requires only one forward pass through the model.
• No External Dependencies: It does not require external retrieval, external verifiers, or a second model.
• No Multiple Generations: Unlike consistency-based methods, it does not require sampling multiple outputs.

The authors position the D-Score not as a universal characterization of truth, but as a calibrated risk score. It is particularly effective for detecting "model-internal hallucinations"—errors where the model possesses the necessary information to recognize the inconsistency but fails to suppress it in the final output. The method highlights that the topology of the hidden-state trajectory (specifically, the spread of singular values) contains diagnostic information about the model's confidence and the presence of conflicting internal signals.

Limitations

The authors acknowledge several limitations:

• White-Box Requirement: The method requires access to hidden activations, making it inapplicable to closed-source models unless such access is provided.
• Internal Visibility: It can only detect hallucinations that leave a trace in the model's internal state; if the model lacks the internal knowledge to recognize an error, the D-Score may not flag it.
• Complexity Confusion: High D-Scores can also result from long, diverse, or semantically complex non-hallucinated texts, necessitating careful calibration.
• Parameter Sensitivity: While robust, the method depends on the choice of layer, tolerance $\tau$, and decision threshold $\bar{D}$.