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

Himal Badu

cs.CLAug 27, 2026

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

Authors: Himal Badu

Abstract

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.

Figures & tables

Family	Signal source	PoP comparison question
Output-only	Logits, entropy, perplexity	Do transitions add information beyond output uncertainty?
Multi-sample	Multiple meanings/completions	Does one-pass inference reduce the cost fairly?
Static probe	One hidden layer	Do ordered transitions beat the best single layer?
Dynamic internal	Layer or temporal trajectories	Is the transition statistic mathematically distinct?
Attention/ graph	Attention, spectra, routing, graphs	Does PoP improve the matched-cost frontier?

Table 1: Mechanism families and the comparison boundary for PoP.

Figure 1: Compact top-to-bottom architecture of PoP. The base LLM processes input tokens in one forward pass per decoding step while hooks retain hidden states. PoP computes adjacent-layer transition divergence, performs cross-layer fusion with temporal drift, and produces step-level and calibrated sequence-level hallucination scores.

Component	Specification
Model families	Llama-3-8B-Instruct; Qwen2.5-7B-Instruct; Mistral-7B-Instruct-v0.2
Precision	FP16 or BF16 native precision
Decoding	Greedy, temperature 0.0; nucleus, temperature 0.7, $p=0.9$
Maximum length	256 generated tokens
Datasets	TruthfulQA; HaluEval 2.0; FaithDial
Split	70% train, 15% validation, 15% test, stratified by factual label

Table 2: Model and evaluation configuration.

Family	Methods included
Output-only heuristics	Token-logit entropy; sequence predictive perplexity; final-layer vocabulary-logit margin
Multi-sample and post-hoc	Semantic entropy with $K=5$ generations; external DeBERTa-v3-large NLI entailment scoring
Static internal probing	LDA and logistic regression on isolated mid-layer and final-layer hidden states
Dynamic internal and spectral	Residual-stream norm growth; attention-head entropy dispersion; layer-wise spectral geometry distance

Table 3: Baseline families included in the manuscript.

Method	Type	TQ AUC	TQ AP	TQ P@R90	HE AUC	HE AP	HE P@R90
Logit entropy	Output	54.2%	51.8%	22.4%	52.1%	49.3%	18.7%
Perplexity	Output	56.1%	53.4%	24.1%	55.0%	52.6%	21.0%
Static probe, $l=16$	Static	63.8%	61.2%	35.6%	61.4%	58.9%	31.2%
Static probe, $l=32$	Static	66.1%	64.0%	39.8%	64.2%	62.1%	36.5%
Spectral geometry	Dynamic	68.4%	66.7%	43.1%	67.0%	65.2%	40.8%
Semantic entropy, $K=5$	Multi-sample	74.1%	72.8%	51.3%	72.8%	71.0%	48.6%

Table 4: Detection performance comparison. TQ denotes TruthfulQA and HE denotes HaluEval.

Configuration	ECE	Brier	Pre- cision	Recall	F1
Uncalibrated PoP	0.142	0.188	68.4%	76.2%	0.721
Platt-calibrated PoP, $\alpha=1.24$ , $\beta=-0.12$	0.031	0.124	73.1%	74.5%	0.738

Table 5: Calibration and operating-point results.

Method	Extra calls	Latency	VRAM overhead	Through- put penalty
Unhooked base LLM	0	24.1 ms/token	0.0 MB	0.0%
PoP (ours)	0	24.4 ms/token	18.4 MB	1.2%
Spectral geometry	0	31.8 ms/token	142.0 MB	31.9%
Semantic entropy, $K=5$	4 auxiliary generations	120.5 ms/token	0.0 MB	400.0%
External DeBERTa NLI	1 auxiliary pass	48.2 ms/token	1740.0 MB	100.0%

Table 6: Cost and latency under the stated benchmark protocol.

Backbone	Dataset/ domain	Zero-shot AUROC	$\boldsymbol{\Delta}$
Llama-3-8B-Instruct	TruthfulQA, in-domain	75.5%	–
Llama-3-8B-Instruct	FaithDial, dialogue	72.1%	-3.4%
Qwen2.5-7B-Instruct	TruthfulQA	73.8%	-1.7%
Mistral-7B-Instruct-v0.2	TruthfulQA	72.4%	-3.1%
Mistral-7B-Instruct-v0.2	HaluEval, QA/dialogue	71.2%	-4.3%

Table 7: Cross-model and cross-domain transfer.

Pertur- bation	Control	Base	Pert.	$\boldsymbol{\Delta}$
Entity swap	Factual validity flips; syntax retained	75.5%	74.8%	-0.7%
Style transformation	Register changes; factual status constant	75.5%	75.2%	-0.3%
Paraphrase	Surface form changes; meaning constant	75.5%	74.9%	-0.6%
Distractor context	Irrelevant facts appended	75.5%	73.8%	-1.7%
Temperature 0.7	Nucleus sampling increases variation	75.5%	73.1%	-2.4%

Table 8: Perturbation robustness.

Figure 2: Step-wise uncertainty trajectory during entity hallucination.

Configuration	AUROC	$\boldsymbol{\Delta}$
Full PoP framework	75.5%	–
Best single layer, $l=24$	66.8%	-8.7%
Static features only	67.2%	-8.3%
Transition features only	70.1%	-5.4%
No layer normalization	68.9%	-6.6%
Shuffled layer order	58.3%	-17.2%

Table 9: Ablation results on Llama-3-8B-Instruct and TruthfulQA.

Figure 3: Layer-transition divergence across transformer depth.

Language Model Error Detection

Language model error detection identifies unreliable answers, reasoning steps, and actions so systems can flag, verify, or repair them.

See all 91 papers

Detecting Errors in Large Language Model Responses: The Incremental Predictive Value of Hidden States
Oct 20, 2026Ningzhi Li, Xiaojun Ma, Xiaofei WangLanguage Model Error Detection LLM Hallucination Detection
Beijing Institute of Computer Technology and Application, Beijing, China
Look Before You Leap: Factual Decoding with Internal Attribution Signals
Sep 14, 2026Hayeong Ryu, JungMin Yun, Byeonggeuk Lim +2Language Model Decoding Language Model Error Detection
Department of Artificial Intelligence, Chung-Ang University · Graduate School of Advanced Imaging Sciences, Multimedia and Film, Chung-Ang University
Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits
May 7, 2026Erik Nielsen, Elia Cunegatti, Marcus Vukojevic +1Probabilistic Circuits LLM Hallucination Mitigation
Department of Information Engineering and Computer Science University of Trento Trento, Italy

Oct 20, 2026J/K move · Enter open · S save

Detecting Errors in Large Language Model Responses: The Incremental Predictive Value of Hidden States

Ningzhi Li, Xiaojun Ma, Xiaofei Wang

Beijing Institute of Computer Technology and Application, Beijing, China

Language Model Error Detection LLM Hallucination Detection

Open paper AskPDF arXiv

Although large language models support a wide range of tasks, hallucinations remain a major obstacle to user trust. When a model generates an answer, both its internal processing and final output vary with the question. Internal states are often expected to offer richer signals for hallucination detection. However, studies using these states typically report their standalone discrimination performance, while output-side comparisons often use simple baselines with weaker modeling capacity. We integrate output-side information available at inference time into a combined baseline and compare detection performance before and after adding hidden states under identical data splits and evaluation rules. Experiments cover open-ended generation and constrained multiple-choice tasks across different models. Adding hidden states improves AUROC by approximately 0.004 and 0.039 on the two tasks, respectively. Hidden states thus provide additional predictive information, but strong standalone detection performance can still yield only limited additional gains.

Abstract

Language Model Error Detection

Explore similar work

Detecting Errors in Large Language Model Responses: The Incremental Predictive Value of Hidden States

Look Before You Leap: Factual Decoding with Internal Attribution Signals

Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits