Source-linked AI summary
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
TL;DR
The paper asks whether LLM judges keep trust scoring separate from binary truth classification, an assumption underlying multi-dimensional evaluation. It compares human and LLM judgments on correctness-controlled QA and stress-tests identical answers with Human versus AI source cues. LLM trust and truth judgments align more tightly than human references, while source cues shift trust, truth verdicts, and logit-derived correct-side probabilities.
Problem
The paper examines whether trust-relevant factors such as source can shift factual truth judgments, despite trust and truth representing different evaluations.
Method
The study compares human and LLM trust–truth judgments on correctness-controlled QA, then changes only Human versus AI source attribution while measuring trust, truth verdicts, and logit-derived probabilities.
Results
LLM judges align trust scores with truth verdicts more tightly than human references, and source cues shift both judgments plus logit-derived correct-side probabilities.
Takeaways & Limitations
Trust scores should not be treated as independent support for truth judgments unless factual verdicts remain stable under content-invariant trust-relevant perturbations.
Takeaways & Limitations
Evidence is limited to the evaluated QA domains, source cues, and judge models, so the same trust–truth dependence is not claimed for all LLMs or factual-evaluation tasks.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper examines whether LLM judges keep trust scoring separate from binary factual truth judgments. It compares human and LLM trust–truth relations, then tests source attribution as a trust-relevant perturbation.
- LLM-as-Judge systems commonly treat trustworthiness, reliability, and factuality judgments as independent evidence.
- Trust scores and binary truth judgments should be distinguishable because correctness and perceived trustworthiness can diverge.
- The study asks whether trust-relevant factors such as source can shift factual truth judgments.
- Humans and LLMs judge correctness-controlled QA to compare LLM trust–truth alignment with human behavioral references.
- The stress test keeps QA content fixed while changing Human versus AI source attribution to test whether source cues shift trust alone or also truth verdicts.
- LLM judges show stronger trust–truth association than human references, while Human sources raise both trust scores and CORRECT verdict rates.
2 Experiments
The experiments use correctness-controlled QA with matched Human and AI source-counterfactuals, collecting separate trust and truth judgments from humans and LLM judges. Source effects are evaluated on identical content, with local-model logits providing an additional truth signal.
- The dataset spans HealthQA, GeneralQA, and Fact-Checking, with each answer paired across unchanged Human and AI source attributions.
- 54 human participants provide baseline trust–truth judgments, while commercial and local models are evaluated as LLM judges.
- Each judge separately rates trust-related dimensions and classifies the answer as CORRECT or INCORRECT.
- The separability test compares trust scores and truth classifications under Human and AI sources while holding QA content identical.
- The procedure first compares human and LLM trust–truth judgments, then measures source-induced changes in trust, truth verdicts, and local-model correct-side probabilities.
3 Results
Results show that LLM trust and truth judgments align more closely than human judgments and shift together under source cues. Human attribution increases trust and CORRECT acceptance, with corresponding effects on model confidence.
- Preliminary human–LLM comparison: Humans give higher trust to correct than incorrect answers, but their truth-decision accuracy is lower on correct answers than incorrect answers.
- Preliminary human–LLM comparison: LLM judges assign higher trust and more CORRECT verdicts to correct answers, showing tighter behavioral alignment between trust and truth.
- Source-cue stress test: Across domains and LLMs, source attribution produces consistent shifts in trust and truth judgments.
- Source-cue stress test: Human attribution raises trust ratings for both correct and incorrect QA and increases CORRECT verdicts, improving correct acceptance but increasing false acceptance.
- Interpretation: The results indicate source-induced non-separability: a content-invariant trust cue shifts trust ratings and factual truth judgments together.
- Source-cue stress test: Higher trust ratings track higher logit-derived correct-judge confidence for Llama-3.3-70B under both source cues and for both factual statuses.
4 Discussion
The paper argues that trust and truth are not reliably separable in LLM judges: source-sensitive trust cues can leak into factual evaluation, making multiple outputs non-independent evidence.
- LLM judges may blur the boundary between content-based truth assessment and trust-like contextual evaluation.
- Source cues can shift truth judgments for identical content, not merely trust ratings.
- Agreements among multiple evaluation fields can be partly circular because one dependent assessment may be re-expressed across labels.
- Separability matters alongside accuracy: average correctness does not ensure that judgment fields remain distinct and meaningful.
- Multi-dimensional judging should test separability by holding content fixed while varying heuristic cues such as source or authority.
5 Related Work
Prior work identifies broad reliability and non-content-bias concerns for LLM-as-Judge systems. This paper focuses on the under-tested risk that trust and truth outputs move together rather than serving as independent evidence.
- LLM-as-Judge systems have been reported as unreliable, inconsistent, and sensitive to tone, source, and position.
- The paper studies whether trust scores and truth verdicts may move together instead of providing independent evidence.
- The concern is operationalized by testing whether source-sensitive trust scoring can be separated from content-grounded truth classification.
6 Conclusion
Current LLM judges often collapse trust and truth in practice, so multi-dimensional outputs should not be assumed to be independent evidence without empirical separability checks.
- Current LLM judges often collapse trust and truth in practice, with trust tracking truth judgments closely under source-cue stress tests.
Limitations
The evidence is bounded by the evaluated QA domains, source cues, judge models, controlled perturbation design, and behavioral human comparison. The study also does not identify the internal mechanism behind trust–truth dependence.
- Scope of models and domains: Evidence is limited to the evaluated QA domains, source cues, and judge models, without claiming the same dependence for all LLMs or factual-evaluation tasks.
- Source cues as controlled perturbations: Human and AI attribution is a controlled stress test, whereas real platforms use richer provenance signals that may carry valid evidence.
- Human comparison: The human comparison is a behavioral baseline, not a normative gold standard for the optimal judgment policy.
- Mechanism: The results establish a behavioral dependence in this setting but do not identify its internal mechanism.
Ethical Statement
The study audits whether LLM-as-Judge outputs provide separable trust and factual-correctness signals, while reporting aggregate results and framing the work as an evaluation audit rather than deployment guidance.
- The study audits whether trust scores and factual correctness verdicts function as meaningfully separate signals.
- The HealthQA examples are controlled evaluation items, not medical advice, diagnoses, or deployment validation of health claims.
- Aggregate reporting and an evaluation-audit framing are used to reduce misuse risk.
- Human evaluations received ethics-board approval, informed consent, compensation, and anonymization procedures.
AI Usage Disclosure
AI tools were used in a limited supportive role for language editing, while the authors conducted and verified the literature review, analysis, writing, findings, and figures.
- GPT-5.5 was used for language editing, including improving clarity and conciseness.
- The authors conducted and verified the literature review, data analysis, and writing.
- The reported findings and figures are based on the authors’ own data and results.
A.1 Dataset Construction
The study constructs correctness-controlled QA pairs and source-counterfactual variants, then separately elicits trust and truth judgments from commercial and local LLM judges using structured prompts and metrics.
- Dataset Construction: QA items span HealthQA, GeneralQA, and Fact-Checking, with each question paired to one factually correct and one plausibly incorrect answer.
- Source-Counterfactual Design: Human and AI source variants preserve question, answer content, cue position, and prompt format while changing only the source label.
- Human Baseline: The human baseline included 54 participants who evaluated four correct and four incorrect fact-checking QA examples without source cues.
- Judgment Tasks: Trust scoring rates credibility, objectivity, reliability, trustworthiness, and reliance on a 1–7 scale, summarized by their average.
- Model and Prompt Setup: Trust and truth tasks use separate stateless prompts, while proprietary models provide black-box judgments and local models provide logits.
- Judgment Tasks: Truth judgments independently return CORRECT or INCORRECT, with accuracy computed against the gold correctness label.
- Metrics: P(Correct) measures the probability of accepting an answer as correct and is reported separately for correct and incorrect QA.
- Metrics: Local-model Correct−JudgeConfidence normalizes CORRECT and INCORRECT logits with a two-way softmax, producing a 0-to-1 correct-side tendency.
C Supplementary Source-Effect Results
Across evaluated domains and judges, Human source cues raise trust and correct-side judgments relative to AI cues even when QA content is identical, supporting source-induced non-separability.
- Interpretation: The matched-pair results show source sensitivity in both final binary verdicts and continuous correct-side tendencies under identical QA content.
- Pooled Source Effects: +0.570 trust points and +3.92 percentage points in P(Correct) are the pooled Human-minus-AI shifts across judges and domains.
- Pooled Source Effects: +4.16 percentage points for correct QA and +3.67 percentage points for incorrect QA quantify the P(Correct) shift by factual status.For incorrect QA, the positive shift represents more false acceptance rather than higher accuracy.
- Domain-Stratified Effects: Human attribution increases trust ratings in Fact-Checking (+0.598), HealthQA (+0.613), and GeneralQA (+0.502).
- Domain-Stratified Effects: False-acceptance shifts range from +1.77 percentage points in GeneralQA to +6.16 percentage points in Fact-Checking.
- Logit-Derived Confidence: +5.35 percentage points is the average Human-cue increase in logit-derived Correct−JudgeConfidence for models with available logits.