Hallucination Span Detection with Input-Side Evidence Alignment

Miyu Yamada; Yuki Arase

cs.CLAug 16, 2026

Hallucination Span Detection with Input-Side Evidence Alignment

Authors: Miyu Yamada, Yuki Arase

Organizations: Institute of Science Tokyo

Abstract

Hallucinations remain a major obstacle to the reliable use of large language models (LLMs) in conditional text generation. Existing methods primarily assess the factuality of an entire generated text, providing limited insight into which output spans are hallucinated or how they relate to the input. We introduce the task of hallucination span detection with input-side evidence alignment, which jointly identifies hallucinated spans and aligns output tokens with the corresponding input evidence. Our approach is based on the observation that faithful output tokens are predictable from the input, whereas hallucinated tokens are not. We therefore train an encoder-based model to predict masked output tokens from the input representation, using prediction confidence for hallucination detection while naturally producing alignments to the input. Experiments show that the proposed method effectively detects hallucinated spans and identifies meaningful input-side evidence. Human evaluation confirms the quality of the predicted alignments.

Figures & tables

Figure 1: Example of hallucination span detection with input-side evidence alignment.

	QA	Summary	Overall
Train (FT)	$4,134$ ( $8.6$ %)	$3,858$ ( $3.0$ %)	$7,992$ ( $5.9$ %)
Dev	$400$ ( $9.0$ %)	$400$ ( $3.3$ %)	$800$ ( $6.1$ %)
Dev ( $\lambda$ )	$500$ ( $9.9$ %)	$500$ ( $3.5$ %)	$1,000$ ( $6.7$ %)
Test	$900$ ( $5.4$ %)	$900$ ( $2.9$ %)	$1,800$ ( $4.1$ %)

Table 1: Number of samples in RAGTruth (the parentheses indicate the proportion of hallucinated characters)

Instruction

Your task is to determine whether the answer contains either or both of the following two types of hallucinations:

1

. conflict: instances where the answer presents direct contradiction or opposition to the passages;

2

. baseless info: instances where the answer includes information which is not substantiated by or inferred from the passages. Then, compile the labeled hallucinated spans into a JSON dict, with a key “hallucination list” and its value is a list of hallucinated spans. If there exist potential hallucinations, the output should be in the following JSON format: {{“hallucination list”: [hallucination span

1

, hallucination span

2

, …]}}. Otherwise, leave the value as a empty list as following: {{“hallucination list”: []}}. Output:

Table 2: Prompt for Llama-SFT

Method	QA			Summary
	P	R	F 1	P	R	F 1
Llama-SFT	$45.0$	$35.6$	$39.8$	64.1	$37.9$	47.6
LettuceDetect	66.9	$62.1$	64.4	$60.2$	$35.5$	$44.6$
OTAlign	$8.9$	$45.3$	$14.9$	$4.5$	$34.8$	$7.9$
Proposed	$40.9$	82.7	$54.8$	$42.8$	38.0	$40.3$

Table 3: Performance of each method in hallucination span detection (P: Precision, R: Recall, F 1 : F 1 score)

Score	Criterion
$3$	Predicted input token is exactly the evidence.
$2$	Predicted input token is in the sentence where the evidence is described, or in the relevant sentence.
$1$	Predicted input token is irrelevant to the true evidence.
$0$	Hallucination judgment is incorrect (false-positive or false-negative).

Table 4: Criteria for manual assessment of input-side evidence alignment

Method	Top- $k$	Type	# of words	$3$	$2$	$1$	$0$	Avg.
OTAlign	Top- $1$	Faithful	$588$	$255$	$57$	$91$	$185$	$1.65$
OTAlign	Top- $3$	Faithful	$588$	$295$	$54$	$54$	$185$	$1.78$
Proposed	Top- $1$	Faithful	$588$	$188$	$267$	$59$	$74$	$1.97$
		Baseless	$71$	$40$	$0$	$19$	$12$	$1.96$
		Conflict	$41$	$2$	$3$	$4$	$32$	$0.39$
	Top- $3$	Faithful	$588$	$268$	$216$	$30$	$74$	$2.15$

Table 5: Human evaluation results

Type	Output Text	Input Text	Conf.
Faithful	[…] and a commemorative Fiesta de la Flor event is planned for April $17$ and $18$ .	[…] The commemorative Fiesta de la Flor in Corpus Christi, Texas – which celebrates her life – is scheduled for April $17$ and $18$ […]	$0.782$
Faithful	[…] it drew the attention of multiple law enforcement agencies including Atlanta Police Department, FBI and Federal Homeland Security. This led to her arrest on August $9$ th, where a firearm and three computers were found in her home. […]	[…] authorities obtained a search warrant for her residence. They took her into custody while executing that warrant. “A firearm along with three computers was located during the search,” East Point police said […]	$0.723$
Faithful	[…] As you drive through tolls, the system deducts the toll amounts from your balance until it reaches a certain threshold (10$ for the TX-Tag). […]	[…] They deduct the tolls from that balance till the balance goes below a certain value (10$ for the TX-Tag) and they charge your credit card […]	$0.767$
Conflict	[…] we can see that the FEHA (Fair Employment and Housing Act) applies to all employers with four to fifteen employees , while the ADA (Americans with Disabilities Act) applies to employers with twenty or more employees. Additionally, the ADA distinguishes between private and public employers, whereas the FEHA does not. […]	[…] One difference between the FEHA (the Fair Employment and Housing Act) and the ADA (Americans with Disabilities Act) is that the ADA applies to all employers in the private sector that have four to fifteen employees, whereas the FEHA affects more employees, and it doesn’t distinguish between private and public employers like the ADA does. […]	$0.471$
Conflict	[…] The APhA has previously opposed the use of the term “drug” for chemicals used in lethal injection and has urged laws prohibiting pharmacists from participating in such cases. […]	[…] This bolsters the association’s previous positions to oppose the use of the term “drug” for chemicals used in lethal injection and to oppose laws that require or prohibit pharmacists from participation in lethal injection cases. […]	$0.787$

Table 6: Examples of input-side evidence alignment by the proposed method. Bold text indicates the gold hallucination span. Cyan and red indicate masked faithful and hallucinated output words, respectively. Green , orange , and purple indicate the top-

1

, -

2

, and -

3

predicted input-side words, respectively. “Conf.” colum shows the confidence values.

Type	QA	Sum
Evident Conflict	$2,349$	$5,812$
Subtle Conflict	$0$	$339$
Evident Baseless Info	$23,441$	$10,820$
Subtle Baseless Info	$5,545$	$1,020$

Table 7: Number of hallucinated characters for each hallucination type

Type	QA	Sum
Evident Conflict	$14.0$	$16.6$
Subtle Conflict	-	$18.3$
Evident Baseless Info	$87.6$	$50.1$
Subtle Baseless Info	$91.2$	$37.7$

Table 8: Recall of hallucination detection by type

Figure 3: Distribution of maximum confidence values by hallucination type. The red dashed line indicates the decision threshold $\lambda_{\mathrm{h}}=0.68$ .

Figure 4: Original maximum confidence distributions. Blue and orange correspond to faithful and hallucinated tokens, respectively. The red dashed line represents the threshold $\lambda_{\mathrm{h}}$ .

Figure 5: Maximum confidence distributions after fine-tuning on QA

Figure 6: Maximum confidence distributions after fine-tuning on Summarization

Hallucination Detection

Hallucination detection identifies unsupported, fabricated, or evidence-inconsistent content in language and multimodal model outputs.

See all 78 papers

External Observers May See More Clearly: Cross-Model Span-Level Hallucination Detection in Large Language Models via Hidden State Probing
Oct 1, 2026Kingshuk Gupta, Davide BuscaldiHallucination Detection LLM Hallucination Detection
LIX, École Polytechnique · School of Electrical and Electronic Engineering, Nanyang Technological University · LIPN, Sorbonne Paris Nord
PARALLAX: Separating Genuine Hallucination Detection from Benchmark Construction Artifacts
May 16, 2026Khizar Hussain, Murat KantarciogluLLM Evaluation Language Model Error Detection
Virginia Tech
Domain-Specific Hallucination Detection in Large Language Models
Sep 10, 2026Varun Teja Chundru, Debasmita BiswasLanguage Model Error Detection LLM Hallucination Mitigation
Department of Computer Science Purdue University Fort Wayne

Oct 1, 2026J/K move · Enter open · S save

External Observers May See More Clearly: Cross-Model Span-Level Hallucination Detection in Large Language Models via Hidden State Probing

Kingshuk Gupta, Davide Buscaldi

LIX, École Polytechnique · School of Electrical and Electronic Engineering, Nanyang Technological University · LIPN, Sorbonne Paris Nord

Hallucination Detection LLM Hallucination Detection Language Model Probing

Open paper AskPDF arXiv

As Large Language Models (LLMs) increasingly serve as foundational reasoning engines, their tendency to hallucinate remains a critical vulnerability. While recent internal state probes offer a promising alternative to slow external retrieval systems, they largely reduce hallucination detection to a token-wise binary classification task, failing to capture the structured, sequential boundaries of semantic drift. Here, we introduce an internal hidden state framework for fine-grained, span-level hallucination detection. By inspecting layer-wise activation patterns, we attempt to detect the exact hallucination onset and continuation tokens in an LLM generation. Our experiments show that this approach successfully isolates hallucination onsets, achieving substantial improvements in Precision-Recall AUC over random baselines despite extreme class imbalance. Ultimately, we propose a novel cross-model detection framework in which one model observes the internal representations elicited by another model's generation. We find that an external observer can match or exceed a generator's self-detection of its own hallucination onsets, including when the observer is the smaller model, suggesting that self-detection is not the ceiling for onset localisation.

Abstract

Hallucination Detection

Explore similar work

External Observers May See More Clearly: Cross-Model Span-Level Hallucination Detection in Large Language Models via Hidden State Probing

PARALLAX: Separating Genuine Hallucination Detection from Benchmark Construction Artifacts

Domain-Specific Hallucination Detection in Large Language Models