When Trust Meets Truth: Trust-Truth Separability in LLM-as-Judge

Xin Sun; Di Wu; Yuchen Guo; Jiahuan Pei; Isao Echizen; Abdallah El Ali; Saku Sugawara

cs.AIAug 21, 2026

When Trust Meets Truth: Trust-Truth Separability in LLM-as-Judge

Authors: Xin Sun, Di Wu, Yuchen Guo, Jiahuan Pei, Isao Echizen, Abdallah El Ali, Saku Sugawara

Organizations: National Institute of Informatics (NII), Japan · University of Amsterdam, the Netherlands · University of Tokyo, Japan · Vrije Universiteit Amsterdam, the Netherlands · Centrum Wiskunde & Informatica (CWI), the Netherlands · Utrecht University, the Netherlands

Abstract

LLM-as-Judge systems can produce multi-dimensional evaluations, such as trustworthiness, reliability, and factuality, and these outputs are often interpreted as independent evidence. We test this assumption for a common pair of judgments: trust scoring and binary truth classification. On correctness-controlled QA, LLM judges align trust scores with truth verdicts more tightly than human behavioral reference, suggesting weaker separations between trust and truth judgment. We then apply stress tests by changing only source cues of identical QA between Human and AI. Source attribution shifts not only trust scores but also truth verdicts and logit-derived correct-side probabilities. Results show that current LLM-as-Judge protocols should not treat trust scores as independent evidence for truth judgments.

Figures & tables

Figure 1: Overview of the study. We construct correctness-controlled QA pairs across domains with source-counterfactual versions. LLM and human judges provide two judgments for each item: trust scores and a binary truth verdict. We compare human and LLM trust–truth relations, then use source attribution as a stress test to examine whether LLM trust and truth judgments remain separable or shift together under a trust-relevant perturbation.

Judge	Gold-Correctness	Trust Score	Truth Accuracy
Human Judges
	Correct QA	4.66	0.77
	Incorrect QA	4.14	0.88
LLM-as-Judge
	Correct QA	6.60	0.68
	Incorrect QA	6.10	0.63

Table 1: Judgments by humans and LLMs. Truth Accuracy is the accuracy in identifying factual correctness.

Domain	Model	Trust Scoring Judgment		Truth Binary Judgment
		Correct	Incorrect	Correct	Incorrect
Fact-Checking	GPT-5.4	4.65 $\blacktriangle$ / 3.75 $\blacktriangledown$	2.54 $\blacktriangle$ / 2.09 $\blacktriangledown$	0.68 $\blacktriangle$ / 0.66 $\blacktriangledown$	0.87 $\blacktriangledown$ / 0.89 $\blacktriangle$
	Cld-Sonnet-4.6	5.37 $\blacktriangle$ / 4.53 $\blacktriangledown$	3.15 $\blacktriangle$ / 2.74 $\blacktriangledown$	0.84 $\blacktriangle$ / 0.81 $\blacktriangledown$	0.82 $\blacktriangledown$ / 0.87 $\blacktriangle$
	Llama-3.3-70B	5.63 $\blacktriangle$ / 5.33 $\blacktriangledown$	5.25 $\blacktriangle$ / 4.97 $\blacktriangledown$	0.87 $\blacktriangle$ / 0.83 $\blacktriangledown$	0.39 $\blacktriangledown$ / 0.45 $\blacktriangle$
	GPT-oss-120B	4.88 $\blacktriangle$ / 4.33 $\blacktriangledown$	3.65 $\blacktriangle$ / 3.11 $\blacktriangledown$	0.66 $\blacktriangle$ / 0.61 $\blacktriangledown$	0.64 $\blacktriangledown$ / 0.73 $\blacktriangle$
	Qwen-3.6-35B	5.48 $\blacktriangle$ / 4.94 $\blacktriangledown$	4.11 $\blacktriangle$ / 3.41 $\blacktriangledown$	0.83 $\blacktriangle$ / 0.80 $\blacktriangledown$	0.45 $\blacktriangledown$ / 0.55 $\blacktriangle$

Table 2: LLM trust and truth judgments under Human vs AI sources across domains. Trust Scoring gives mean trust ratings; Truth Binary Judgment gives the accuracy of factual correctness. Arrows indicate cue-induced shift.

Domain	Model	Trust Scoring Judgment		Truth Binary Judgment
		Correct	Incorrect	Correct	Incorrect
Fact-Checking	GPT-5.4	4.52 / 4.65 $\blacktriangle$ / 3.75 $\blacktriangledown$	2.24 / 2.54 $\blacktriangle$ / 2.09 $\blacktriangledown$	0.67 / 0.68 $\blacktriangle$ / 0.66 $\blacktriangledown$	0.87 / 0.87 $\blacktriangledown$ / 0.89 $\blacktriangle$
	Cld-Sonnet-4.6	5.14 / 5.37 $\blacktriangle$ / 4.53 $\blacktriangledown$	2.87 / 3.15 $\blacktriangle$ / 2.74 $\blacktriangledown$	0.83 / 0.84 $\blacktriangle$ / 0.81 $\blacktriangledown$	0.84 / 0.82 $\blacktriangledown$ / 0.87 $\blacktriangle$
	Llama-3.3-70B	5.57 / 5.63 $\blacktriangle$ / 5.33 $\blacktriangledown$	5.13 / 5.25 $\blacktriangle$ / 4.97 $\blacktriangledown$	0.85 / 0.87 $\blacktriangle$ / 0.83 $\blacktriangledown$	0.42 / 0.39 $\blacktriangledown$ / 0.45 $\blacktriangle$
	GPT-oss-120B	4.81 / 4.88 $\blacktriangle$ / 4.33 $\blacktriangledown$	3.60 / 3.65 $\blacktriangle$ / 3.11 $\blacktriangledown$	0.63 / 0.66 $\blacktriangle$ / 0.61 $\blacktriangledown$	0.67 / 0.64 $\blacktriangledown$ / 0.73 $\blacktriangle$
	Qwen-3.6-35B	5.21 / 5.48 $\blacktriangle$ / 4.94 $\blacktriangledown$	3.79 / 4.11 $\blacktriangle$ / 3.41 $\blacktriangledown$	0.81 / 0.83 $\blacktriangle$ / 0.80 $\blacktriangledown$	0.49 / 0.45 $\blacktriangledown$ / 0.55 $\blacktriangle$

Table 3: LLM trust and truth judgments across three conditions per cell, reported as Control / Human-cue / AI-cue . Trust Scoring gives mean trust ratings; Truth Binary Judgment gives accuracy of factual correctness. Arrows mark each cue relative to control:

\blacktriangle

above control,

\blacktriangledown

below control (no arrow = within

\pm 0.005

of control).

Panel of Figure 2 — Figure 2: Source cues shift Trust Scoring , Binary Truth judgments, and the correct-side probability when making truth judgment by Llama-3.3-70B. Orange denotes Human source cues, and blue denotes AI source cues.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Field	Template
Prefix	You are a JSON-only rating expert. Return ONLY one valid JSON object. Any non-JSON output is invalid. No reasoning. No explanations. No markdown. No comments. No headings. No extra text.
Question $Q$	{QUESTION}
Answer $A$	{ANSWER} with / without source attribution as either from Human or AI
Task	Rate the answer rigorously and strictly based on the following dimensions using a 1–7 Likert scale. Scale: 1 = extremely negative 2 = negative 3 = somewhat negative 4 = neutral/mixed 5 = somewhat positive 6 = positive 7 = extremely positive Dimensions: credibility objectivity reliability trustworthiness reliance Rating rules: - Be strict: Lower scores for false, unsupported, vague, misleading, biased, or unclear content. - Use integers only and give diverse ratings. - Do not overuse 7; reserve it only for exceptional information with strong evidence. - Do not use strings. - Do not generate explanations and reasoning. - Do not omit or rename keys. You must return exactly this JSON object with integer values only for each key: "credibility": " ", "objectivity": " ", "reliability": " ", "trustworthiness": " ", "reliance": " "

Appendix

Table 4: LLM prompt template used for trust scoring.

Field	Template
Prefix	You will receive:
Question $Q$	{QUESTION}
Answer $A$	{ANSWER} with / without source attribution as either from Human or AI
Task	Judge whether the above answer is correct or incorrect. Rules: - Do not output your reasoning, analysis, explanations, or any text before or after the JSON. No other text. - You Must generate with ONLY JSON objects: either "correctness": "correct" or "correctness": "incorrect".

Appendix

Table 5: LLM prompt template used for truth judgment.

Outcome	Source Pair Effect	95% CI
Pooled
Trust rating	$+0.57$ pts	$[+0.54,+0.60]$
$P(\mathrm{Correct})$	$+3.92$ pp	$[+3.29,+4.65]$
correct QA	$+4.16$ pp	$[+3.29,+5.14]$
incorrect QA	$+3.67$ pp	$[+2.91,+4.53]$
$\mathrm{Confidence}_{\mathrm{true}}$	$+5.35$ pp	$[+4.93,+5.80]$

Appendix

Table 6: Matched-pair Human–AI source-effect tests under identical QA content. Pooled rows estimate average Human-minus-AI matched-pair differences over all evaluated LLM judges and domains. Domain rows (i.e., Fact-Checking, HealthQA, and GeneralQA) average over all evaluated LLM judges within each domain. Each effect is the average Human-minus-AI shift with a 95% item-cluster bootstrap confidence interval (CI). CIs are item-cluster bootstrap CIs.

Confidence

denotes the logit-derived

Correct-Judge_{Confidence}

Explore similar work

Label Effects: Shared Heuristic Reliance in Trust Assessment by Humans and LLM-as-a-Judge
Apr 7, 2026Xin Sun, Di Wu, Sijing Qin +3Bias in LLM-as-a-Judge LLM-as-a-Judge
National Institute of Informatics (NII), Japan · University of Amsterdam, the Netherlands · Hitotsubashi University, Japan +3
All Verdicts are Not Equal: Rethinking LLM Judge Reliability
Oct 8, 2026Vineet Kumar, Darshita Rathore, Anindya MoitraLLM-as-a-Judge LLM Evaluation
PayPal Artificial Intelligence PayPal, Bengaluru, India
Judge, Retrieve, or Abstain: Uncertainty-Guarded LLM Judging with Provable Risk Guarantees
Aug 18, 2026Sher Badshah, Ali Emami, Hassan SajjadLLM-as-a-Judge Selective Prediction
Dalhousie University · New York University Abu Dhabi · Emory University

Apr 7, 2026J/K move · Enter open · S save

Label Effects: Shared Heuristic Reliance in Trust Assessment by Humans and LLM-as-a-Judge

Xin Sun, Di Wu, Sijing Qin, Isao Echizen+2

National Institute of Informatics (NII), Japan · University of Amsterdam, the Netherlands · Hitotsubashi University, Japan+3

Bias in LLM-as-a-Judge LLM-as-a-Judge Language Model Bias Evaluation Trust in AI

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly used as automated evaluators (LLM-as-a-Judge). This work challenges its reliability by showing that trust judgments by LLMs are biased by disclosed source labels. Using a counterfactual design, we find that both humans and LLM judges assign higher trust to information labeled as human-authored than to the same content labeled as AI-generated. Eye-tracking data reveal that humans rely heavily on source labels as heuristic cues for judgments. We analyze LLM internal states during judgment. Across label conditions, models allocate denser attention to the label region than the content region, and this label dominance is stronger under Human labels than AI labels, consistent with the human gaze patterns. Besides, decision uncertainty measured by logits is higher under AI labels than Human labels. These results indicate that the source label is a salient heuristic cue for both humans and LLMs. It raises validity concerns for label-sensitive LLM-as-a-Judge evaluation, and we cautiously raise that aligning models with human preferences may propagate human heuristic reliance into models, motivating debiased evaluation and alignment.

Abstract

Explore similar work

Label Effects: Shared Heuristic Reliance in Trust Assessment by Humans and LLM-as-a-Judge

All Verdicts are Not Equal: Rethinking LLM Judge Reliability

Judge, Retrieve, or Abstain: Uncertainty-Guarded LLM Judging with Provable Risk Guarantees