HalluTracer: Pre-Decoding Truthfulness Prediction via Depth-Averaged Probe-Logit

Zhihao Guo; Zonghan Wu; Huan Huo; DaYong Ye; Junwei Zhang; Weiran Yao; Zhiwei Liu; Qingsong Wen

cs.CLAug 17, 2026

HalluTracer: Pre-Decoding Truthfulness Prediction via Depth-Averaged Probe-Logit

Authors: Zhihao Guo, Zonghan Wu, Huan Huo, DaYong Ye, Junwei Zhang, Weiran Yao, Zhiwei Liu, Qingsong Wen,

Organizations: East China Normal University, China · University of Technology Sydney, Australia · City University of Macau, Macau · Meta, USA ·

Abstract

Internal-state probes enable truthfulness prediction before a large language model generates an answer. When detectors change both the layers they read and the rules used to combine them, the source of improved prediction becomes difficult to identify. We separate these choices and find that retaining more layers improves prediction even under fixed equal weighting. An exact Fisher-ratio decomposition explains why the additional benefit of linear reweighting is limited on these probe scores: information inlayer-wise differences largely overlaps with that captured by the depth mean. Estimating additional weights can then offset this small benefit when the data used to fit them are limited. These findings motivate our proposed method HalluTracer, which averages layer-wise probe logits to predict truthfulness before decoding. Across six models and four benchmarks, including TruthfulQA, HalluTracer achieves the highest area under the receiver operating characteristic curve (AUROC) in 23 of 24 model--benchmark pairs among the compared methods. The results support using evidence from across the network without requiring a correspondingly more flexible aggregation rule, clarifying the distinct roles of layer selection and weighting in pre-decoding detection.

Figures & tables

Figure 1: Three readouts of the same examples . A single layer retains one coordinate. The plug-in Fisher score estimates a direction from the mean gap and equal-class pooled covariance. The depth mean uses fixed equal weights. All three share the same training-fitted probes.

		Pooled SNR			AUROC (%)			Selection
Model	Dataset	Single	Fisher	Mean	Single	Fisher	Mean	$j^{\star}$	$n_{j}$
Llama-2-7B	HaluEval2	1.108	1.230	1.247	78.98	80.42	80.85	18	3
	TrueFalse	2.330	2.548	2.479	94.29	95.23	95.19	18	1
	HELM	1.389	1.687	1.664	83.55	87.89	87.70	12	3
Llama-3.1-8B	HaluEval2	1.335	1.431	1.415	82.38	83.83	83.60	28	2
	TrueFalse	2.841	3.147	3.038	96.27	97.19	97.30	21	2

Table 1: Single-layer selection and full-support aggregation. Held-out pooled SNR and AUROC of the training-selected single layer (Single), the plug-in Fisher score (Fisher), and the depth mean (Mean).

j^{\star}

is the modal selected layer and

n_{j}

the number of distinct layers selected across the five folds. Bold marks the largest AUROC in each row and underline the second largest.

Figure 2: Estimation cost of the unshrunk plug-in Fisher score. The score is re-estimated on a fraction of the training fold, with probes, held-out folds, and the depth mean held fixed. Top: held-out AUROC difference to the depth mean. Bottom: held-out SNR ratio (solid) and training-fold

\widehat{R}_{I}

(dashed). Horizontal axis: covariance parameters per minority-class training example, log scale.

		Qwen2.5-7B		Qwen2.5-14B		Qwen2.5-32B		Mistral-Small-24B		Llama-2-7B		Llama-3.1-8B
Dataset	Method	AUROC	AUPRC	AUROC	AUPRC	AUROC	AUPRC	AUROC	AUPRC	AUROC	AUPRC	AUROC	AUPRC
TruthfulQA	HalluGuard	60.57	31.95	59.84	35.33	52.25	18.21	53.82	34.05	54.98	46.25	51.60	42.96
	Fact-Probe	67.26	48.40	60.30	22.00	75.53	40.14	69.23	47.79	69.20	62.98	66.07	59.33
	ITI-Probe	64.46	44.71	69.92	42.95	74.57	42.72	65.95	46.82	64.90	56.09	63.43	57.04
	IRIS	68.44	50.10	72.54	43.51	74.29	40.63	65.86	45.59	71.12	60.64	65.04	58.17
	SIVR-PD	73.69	59.73	78.95	60.34	81.29	53.62	69.74	51.60	72.77	72.23	69.72	72.96

Table 2: Truthfulness prediction results (AUROC / AUPRC, %) across four benchmarks and six LLMs. Bold : largest point estimate; underline : second largest. Where shown,

\pm

denotes the standard deviation over five folds.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Model	$m{\times}H$	AUC CV	$\boldsymbol{\bar{r}_{\mathrm{head}}}$	Eff. Rank	Rank/ $H$
Qwen2.5-7B	$28{\times}28$	0.047	0.75	3.5	12.4%
Llama-2-7B	$32{\times}32$	0.073	0.65	5.7	17.8%
Llama-3.1-8B	$32{\times}32$	0.061	0.74	3.7	11.5%

Appendix

Table 3: Head-level probe statistics on pooled HaluEval2. AUC CV: within-layer AUROC coefficient of variation;

\bar{r}_{\mathrm{head}}

: mean pairwise Pearson correlation of head probabilities; Eff. Rank: exponential entropy of normalised correlation eigenvalues. CV, correlation, and effective rank are averaged over layers.

Model	Dataset	Top-1	Top-5	Bottom-1	Random-1	$\boldsymbol{\Delta}$ (pp)
Llama-2-7B	HaluEval2	80.8	80.9	78.7	80.4 $\pm$ .013	$-0.4$
	TrueFalse	95.4	95.4	91.7	94.7 $\pm$ .004	$-0.7$
	HELM	88.3	88.4	87.0	88.0 $\pm$ .016	$-0.3$
Llama-3.1-8B	HaluEval2	83.9	83.9	82.0	83.3 $\pm$ .012	$-0.5$
	TrueFalse	97.5	97.5	96.2	97.2 $\pm$ .004	$-0.3$
	HELM	89.2	89.3	88.1	88.9 $\pm$ .016	$-0.4$

Appendix

Table 4: Head selection with probability averaging within and across layers. Top-

K

and Bottom-1 follow the per-layer training AUROC ranking. Random-1 samples one head per layer uniformly (mean

\pm

std over 50 draws). AUROC means are in percent;

\Delta

is Random-1 minus Top-1 in AUROC points.

Model	Dataset	Selected Head	Residual Stream	$\boldsymbol{\Delta}$
Qwen2.5-32B	HaluEval2	0.852	0.856	$+.005$
	TrueFalse	0.985	0.986	$+.001$
	HELM	0.887	0.907	$+.020$
Qwen2.5-14B	HaluEval2	0.847	0.847	$\phantom{+}.000$
	TrueFalse	0.983	0.984	$+.001$
	HELM	0.884	0.907	$+.023$

Appendix

Table 5: Readout mechanism ablation (AUROC): selected-head readout (

d{=}d_{h}

) vs. residual-stream (

d{=}d_{\mathrm{model}}

, no head selection), both evaluated via the depth mean

\bar{L}

\Delta

: residual-stream minus selected-head. The two readouts differ by at most

0.005

on HaluEval2 and TrueFalse, and the residual-stream value is higher on HELM.

	HaluEval2				TrueFalse				HELM
Model	Mean	Traj.	Incr.	Exp.	Mean	Traj.	Incr.	Exp.	Mean	Traj.	Incr.	Exp.
Mistral-Small-24B	84.7	84.7	84.2	83.7	98.2	98.4	98.4	98.4	89.5	89.5	89.0	89.2
Qwen2.5-32B	84.0	84.6	84.3	83.8	98.5	98.6	98.6	98.3	88.5	88.8	88.4	88.7
Qwen2.5-14B	83.6	83.8	83.8	83.1	98.3	98.4	98.3	98.0	88.3	88.4	88.5	88.3
Qwen2.5-7B	82.0	82.4	82.2	82.4	97.4	97.5	97.5	97.2	87.9	87.7	87.6	87.6
Llama-3.1-8B	83.6	83.8	83.3	83.8	97.3	97.4	97.5	97.1	88.8	88.7	88.3	88.9

Appendix

Table 6: Aggregation inputs on a fixed probe-logit trajectory under one shared full-support, five-fold run (AUROC, %). Mean: depth mean (1D); Traj.: trajectory statistics (5D); Incr.: increment statistics (13D); Exp.: expanded trajectory (

2m{+}5

D). The depth mean stays within 1.1 AUROC points of every tested richer representation. Bold marks the largest displayed point estimate within each benchmark.

Model	Dataset	$\boldsymbol{\rho}$	$\boldsymbol{\gamma}$	$\boldsymbol{R_{\text{plugin}}}$	$\boldsymbol{R_{\text{emp}}}$	$\boldsymbol{R_{\text{emp}}/R_{\text{plugin}}}$
Mistral-Small-24B	HaluEval2	0.666	8.50	0.568	0.567	0.999
	TrueFalse	1.212	6.07	0.823	0.824	1.000
	HELM	0.197	6.83	0.188	0.188	1.002
Qwen2.5-32B	HaluEval2	0.945	8.67	0.700	0.706	1.008
	TrueFalse	1.777	6.25	1.017	1.017	1.000
	HELM	0.144	7.90	0.148	0.148	0.999

Appendix

Table 7: SNR measurements across six models and three dataset families.

\rho

: depth-gap ratio;

\gamma

: offset penalty;

R_{\text{plugin}}

: covariance-aware predicted ratio

\mathrm{SNR}(\hat{\beta})/\mathrm{SNR}(\bar{L})

;

R_{\text{emp}}

: empirically measured ratio. The final column reports

R_{\text{emp}}/R_{\text{plugin}}

Model	Dataset	$m$	$\frac{m(m+1)}{2N_{\min}}$	$\widehat{R}_{I}$	Held-out SNR ratio	$\widehat{\kappa}_{\mathrm{mean}}$	$\widehat{\kappa}_{\mathrm{Fisher}}$	$\widehat{\theta}_{I}$	$\widehat{\chi}_{I}$	$\widehat{t}_{I}$
Llama-2-7B	HaluEval2	32	0.41	$1.024_{\pm 0.004}$	0.986	0.25	0.36	0.479	0.696	0.458
	TrueFalse	32	0.21	$1.048_{\pm 0.006}$	1.028	0.83	1.19	0.944	0.998	0.924
	HELM	32	0.44	$1.044_{\pm 0.004}$	1.014	0.33	0.46	0.357	0.573	0.301
Llama-3.1-8B	HaluEval2	32	0.41	$1.043_{\pm 0.001}$	1.011	0.55	0.59	0.469	0.719	0.450
	TrueFalse	32	0.21	$1.050_{\pm 0.002}$	1.036	1.12	1.51	0.957	1.021	0.950
	HELM	32	0.44	$1.040_{\pm 0.004}$	1.001	0.44	0.60	0.366	0.611	0.329

Appendix

Table 8: Fisher-ratio diagnostics for Table 1 .

m

: number of layers;

N_{\min}

: minority-class training count.

m(m{+}1)/(2N_{\min})

measures covariance parameters per minority-class example.

\widehat{R}_{I}

\widehat{\theta}_{I}

\widehat{\chi}_{I}

, and

\widehat{t}_{I}

are training-fold diagnostics in Equation 7 . The held-out SNR ratio is Fisher SNR divided by depth-mean SNR.

\widehat{\kappa}

is the held-out Gaussian approximation error in AUROC points (Equation 63 ). Values are five-fold means;

\pm

denotes standard deviation.

		Full training fold		One sixteenth
Model	Dataset	Max. gain	$\alpha^{\star}$	Max. gain	$\alpha^{\star}$	$\alpha=0$
Llama-2-7B	HaluEval2	$+0.08$	0.80	$+0.08$	0.90	$-4.70$
	TrueFalse	$+0.18$	0.30	$+0.14$	0.55	$-0.88$
	HELM	$+0.34$	0.25	$+0.02$	0.75	$-4.06$
Llama-3.1-8B	HaluEval2	$+0.44$	0.15	$+0.15$	0.85	$-4.01$
	TrueFalse	$+0.04$	1.00	$+0.04$	1.00	$-0.49$

Appendix

Table 9: Covariance shrinkage across training sizes. AUROC of the shrunk Fisher score minus the depth mean, in AUROC points. For each condition and fraction,

\alpha^{\star}

maximises the mean difference over

\alpha\in\{0,0.05,\ldots,1\}

on the held-out folds. The final column reports the unshrunk score at one sixteenth of the training fold.

	HaluEval2			TrueFalse			HELM
Model	Raw	Standardised	Probability	Raw	Standardised	Probability	Raw	Standardised	Probability
Llama-2-7B	80.85	80.75	80.84	95.19	94.43	95.05	87.70	87.71	87.64
Llama-3.1-8B	83.60	83.42	83.53	97.30	96.96	97.15	88.80	88.79	88.83
Qwen2.5-7B	82.03	81.70	81.92	97.43	96.29	97.07	87.85	87.69	87.66
Qwen2.5-14B	83.56	83.18	83.53	98.34	97.89	98.13	88.34	88.40	88.44
Qwen2.5-32B	84.03	83.58	83.94	98.51	98.07	98.31	88.51	88.69	88.66

Appendix

Table 10: Probe-logit scale control. Five-fold mean AUROC in percent. Raw is the deployed depth mean. Standardised first uses the training-fold mean and standard deviation of each layer-wise logit. Probability averages the sigmoid-transformed logits.

	Full	Layer-level	Increment	Update	Late layer
HaluEval2	83.4	83.3	82.3	65.3	82.0
TrueFalse	97.1	97.2	96.7	88.5	96.0
HELM	88.2	87.7	86.0	62.4	86.8

Appendix

Table 11: Increment-feature group ablation on Llama-3.1-8B (AUROC

\times 100

, averaged over five runs). Full uses all 13 statistics; Layer-level, Increment, Update, and Late layer retain only the indicated groups defined in Appendix C .

Figure 3: Depth-support ablation. AUROC for uniformly spaced supports, the first and second depth halves, and the full support across three models and three datasets.

Language Model Probing

Language model probing measures what information and capabilities models encode and how those signals relate to their behavior.

See all 275 papers

May 30, 2026

Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMs

We investigate whether open-source LLMs encode a linearly separable truthfulness signal in their hidden states, and at which network depth this signal is strongest. Across three

7

B--

8

B instruction-tuned models (Llama-3.1-8B, Mistral-7B, Qwen2.5-7B) loaded in

4

-bit NF4 quantization, we extract per-layer hidden states on four hallucination benchmarks (TruthfulQA, HaluEval-QA, FEVER, and a controlled synthetic set) and compare four detection approaches: linear and MLP probes, INSIDE EigenScore, self-consistency, and attention entropy. A linear probe on a single mid-network layer achieves

0.904

1.000

AUROC on held-out splits, while sampling-based detectors do not exceed

0.541

AUROC under the same protocol. The truthfulness signal is approximately linear: MLP probes rarely surpass linear probes by more than

0.01

AUROC. Peak probing layers fall in a consistent band across model families on natural-language benchmarks -- blocks~

13

18

of~

32

for Llama and Mistral, and blocks~

19

25

of~

28

for Qwen. First-block attention entropy provides a complementary signal in knowledge-grounded settings (

0.866

0.941

AUROC on HaluEval-QA) at no additional inference cost. The low discriminability of sampling methods under this protocol reflects a structural mismatch between paired-label evaluation and the information these methods access, rather than an inherent limitation of those methods. Code and data are released for full reproducibility on a single

8

,GB GPU.

Aizierjiang Aiersilan

University of Macau

Hallucination in Language Models LLM Hallucination Detection Language Model Probing Hidden-State Probing

Ask PDF

Apr 15, 2026

Linear Probe Accuracy Scales with Model Size and Benefits from Multi-Layer Ensembling

Linear probes can detect when language models produce outputs they "know" are wrong, a capability relevant to both deception and reward hacking. However, single-layer probes are fragile: the best layer varies across models and tasks, and probes fail entirely on some deception types. We show that combining probes from multiple layers into an ensemble recovers strong performance even where single-layer probes fail, improving AUROC by +29% on Insider Trading and +78% on Harm-Pressure Knowledge. Across 12 models (0.5B--176B parameters), we find probe accuracy improves with scale: ~5% AUROC per 10x parameters (R=0.81). Geometrically, deception directions rotate gradually across layers rather than appearing at one location, explaining both why single-layer probes are brittle and why multi-layer ensembles succeed.

Erik Nordby, Tasha Pais, Aviel Parrack

Georgia Institute of Technology, Atlanta, Georgia, USA · Independent Researcher · Stanford University, Stanford, California, USA

Language Model Probing Deception in Language Models Representation Geometry in Language Models Language Model Ensembles

Ask PDF

Aug 27, 2026

Prediction of Prediction (PoP): Inter-Layer Activation Fusion for Single-Pass Hallucination Detection in Large Language Models

Autoregressive large language models (LLMs) routinely generate factually incorrect outputs with high decoding confidence, limiting their deployment in high-stakes workflows. Existing output-stage uncertainty metrics can fail when models are overconfident on false assertions, while multi-sample verification pipelines introduce substantial memory and latency overhead. This work evaluates whether internal hidden-state transition dynamics during generation can signal factual errors without auxiliary decoding calls. We introduce Prediction of Prediction (PoP), a mechanism that captures layer-transition uncertainty by fusing intermediate hidden representations across depth during a single forward pass. Evaluated on the TruthfulQA benchmark using autoregressive transformer backbones, PoP achieves an area under the receiver operating characteristic curve (AUROC) of 75.5% for factual-correctness classification. The mechanism operates within the base forward pass, adding less than 1.2% runtime latency and requiring zero additional generation passes. The numerical results are reported from the author-verified experimental implementation and are bounded by the evaluation scope described below.

Himal Badu

Language Model Error Detection LLM Hallucination Detection Uncertainty Quantification Confidence Estimation in Language Models

Ask PDF

Abstract

Language Model Probing

Explore similar work

Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMs

Linear Probe Accuracy Scales with Model Size and Benefits from Multi-Layer Ensembling

Prediction of Prediction (PoP): Inter-Layer Activation Fusion for Single-Pass Hallucination Detection in Large Language Models