HalluField Framework for LLM Hallucination Detection

• HalluField Framework is a thermodynamics-inspired method that models LLM token generation as an energy landscape to detect hallucinations.
• It computes per-token free energy and entropy using discrete variational principles, enabling robust quantification of semantic stability.
• Empirical tests across datasets show that HalluField and its variant achieve state-of-the-art detection with minimal computational overhead.

HalluField is a framework for detecting hallucinations in LLMs grounded in field-theoretic modeling and thermodynamics. It quantifies the stability of an LLM’s response with respect to temperature perturbations, interpreting token generation as an energy landscape and applying a discrete variational principle to derive a semantic stability criterion. HalluField operates directly on output logits, requires no fine-tuning or auxiliary models, and achieves state-of-the-art hallucination detection across models and tasks, leveraging physically motivated metrics such as per-token free energy and entropy to separate reliable from hallucinated generations (Vu et al., 12 Sep 2025).

1. Field-Theoretic Formalism and Variational Principle

An LLM’s response to a query $Q$ is represented as a path in token-rank–temperature space, $(r,T)\in\mathbb T^*$, where $r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$ tracks, for each token position $i$, the rank $r_i$ of the sampled token under the model’s softmax logits $\ell_i=f_\theta(Q,\tau_{<i})$. Each path corresponds to a token sequence $\{\tau_i(r_i,T)\}_{i=1}^N$.

The HalluField framework treats both free energy and entropy of a token sequence as “action” functionals on the curve $(r,T)$. In the continuous case, the action $\mathbb A[c]=\int A(c)\,ds$ varies as $$\delta\mathbb A=\int\frac{\delta A}{\delta c}\,\delta c\,ds$$. In the discrete, parametrized setting, variation with respect to a temperature shift $(r,T)\in\mathbb T^*$0 is expressed as:

$(r,T)\in\mathbb T^*$1

Choosing $(r,T)\in\mathbb T^*$2 and uniform $(r,T)\in\mathbb T^*$3 yields $(r,T)\in\mathbb T^*$4.

2. Thermodynamic Quantities in Token Generation

HalluField defines thermodynamic analogies for LLM token-generation:

• Per-token free energy (“energy”):

$(r,T)\in\mathbb T^*$5

Here, $(r,T)\in\mathbb T^*$6.

• Per-token entropy:

$(r,T)\in\mathbb T^*$7

with $(r,T)\in\mathbb T^*$8.

• Partition function and Helmholtz free energy (by analogy):

If discrete token paths are viewed as microstates of energy $(r,T)\in\mathbb T^*$9:

$r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$0

HalluField does not compute $r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$1 explicitly; instead, its functional $r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$2 serves as the relevant free energy action.

3. Semantic Stability and Internal-Energy Variation

By analogy to the first law and Legendre transform, HalluField defines the total variation of “internal energy” as:

$r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$3

where:

$r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$4

$r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$5

$r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$6 captures the change in free energy along the base path $r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$7, while $r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$8 quantifies the free-energy shift arising from path changes at $r=(r_1,\dots,r_N)\in(\mathbb Z^+)^N$9.

Instability interpretation: Low-energy (correct) responses exhibit high sensitivity of $i$0 to temperature shifts, whereas hallucinated (high-energy) sequences show minimal internal-energy variation. This empirical separation in $i$1 and its components is observed across multiple datasets and temperature perturbations.

4. Algorithmic Workflow and Computational Complexity

The HalluField algorithm (Algorithm 1) computes semantic stability via the following steps:

1. Input: LLM, query $i$2, base temperature $i$3, perturbation set $i$4.
2. Generate base response $i$5; record per-step token probabilities.
3. Initialize $i$6, $i$7.
4. For each $i$8:
   • Query LLM at $i$9; collect token probabilities.
   • Compute $r_i$0, $r_i$1 and accumulate into $r_i$2.
   • Compute $r_i$3 and accumulate into $r_i$4.
5. Return $r_i$5.

Computational properties: The approach requires $r_i$6 LLM calls (one per perturbation), with $r_i$7 post-processing per call (practically reduced via top-$r_i$8 logit selection). No model fine-tuning or external neural architectures are invoked.

5. Hallucination Detection Criteria

Detection is realized by thresholding the semantic stability statistic:

$r_i$9

The HalluFieldSE variant incorporates the semantic entropy $\ell_i=f_\theta(Q,\tau_{<i})$0 (as in Equation 1) into the score:

$\ell_i=f_\theta(Q,\tau_{<i})$1

with thresholding at chosen values of $\ell_i=f_\theta(Q,\tau_{<i})$2 and $\ell_i=f_\theta(Q,\tau_{<i})$3, determined via optimization on a held-out validation set (e.g., by Youden’s index).

6. Empirical Validation and Comparative Results

Experiments utilize SQuAD, TriviaQA, Natural Questions, and BioASQ with 500 examples per set. Baseline methods compared include Semantic Entropy (SE), Kernel Language Entropy (KLE), $\ell_i=f_\theta(Q,\tau_{<i})$4, cluster entropy (CE), and regular entropy (RE). Metrics reported are area under the curve (AUC) and raw accuracy.

Representative results for LLaMA-2 7B are as follows:

Method	AUC	Accuracy
HalluField	~0.82	~0.64
HalluFieldSE	~0.84	~0.68
KLE/SE	~0.83	~0.68

HalluFieldSE shows consistent improvements of 1–3 AUC points across models and tasks relative to baselines. Runtime per query for HalluField is approximately $\ell_i=f_\theta(Q,\tau_{<i})$5 s (no extra LLM invocations) vs. $\ell_i=f_\theta(Q,\tau_{<i})$641 s for SE/KLE approaches, which require additional LLM calls.

7. Theoretical and Practical Significance

HalluField introduces a thermodynamic and field-theoretic perspective on LLM hallucination detection by translating logits into an energy–entropy landscape and quantifying semantic instability via temperature perturbations. The analogy to the free energy $\ell_i=f_\theta(Q,\tau_{<i})$7 yields a method that is both theoretically well-founded and computationally practical, requiring only access to output logits with no architectural or training changes. The physically motivated internal-energy variation $\ell_i=f_\theta(Q,\tau_{<i})$8 provides a robust separation between hallucinated and non-hallucinated generations, as substantiated by empirical evaluation (Vu et al., 12 Sep 2025).