XAI-Arena: Can LLMs Assess the Quality of XAI Explanations?

URL SCAN: XAI-Arena: Can LLMs Assess the Quality of XAI Explanations?
FIRST LINE: # Computer Science > Artificial Intelligence

The Dissection

This paper converts subjective XAI assessment into an LLM-mediated scoring pipeline. The advertised product is not merely better evaluation; it is scalable, reproducible automation of evaluative cognition. Human judgment becomes calibration data, validation labor, and eventually a dispensable bottleneck.

The reported Spearman correlation of .693 shows that the LLM tracks human ratings. It does not show that the LLM discovers explanation quality independently, resolves subjectivity, or identifies truth.

The Core Fallacy

The central error is conflating agreement with humans with validity. If human ratings are noisy, biased, or conceptually incomplete, an LLM that reproduces them simply industrializes those defects. Reproducibility makes a judgment repeatable; it does not make it correct.

The abstract also moves from correlation to broad claims about quality assessment. That evidence does not establish robustness against persuasive but unfaithful explanations, unfamiliar formats, distribution shifts, prompt changes, or systems optimized to satisfy the judge. The evaluator may reward fluency, confidence, and familiar explanatory style while mistaking them for faithfulness.

Hidden Assumptions

• Human ratings are an adequate ground truth rather than a disputed proxy.
• A correlation of .693 is sufficient for consequential evaluation.
• Stakeholder personas can substitute for the cognition, incentives, expertise, and risks of actual stakeholders.
• Simplicity, clarity, faithfulness, actionability, trust calibration, and transparency can be scored reliably and meaningfully combined.
• The LLM receives enough context to judge whether an explanation is genuinely faithful.
• Reproducibility compensates for the loss of accountable human judgment.
• The benchmark will remain valid when models, prompts, datasets, and evaluation targets change.

Social Function

Primary classification: transition management, with elements of prestige signaling and partial truth.

The partial truth is that LLM evaluation can reduce cost and increase throughput. The anesthetic is the word “reproducible.” It hides a transfer of authority from human evaluators to whoever controls the judging model, prompts, benchmarks, and deployment rules. Subjectivity is not eliminated; it is packaged into an automated institution.

This is P1 in miniature: cognitive labor is converted into a model call. The evaluator does not survive as a mass economic role merely because the new system still needs a few people to design, audit, or govern it.

The Verdict

XAI-Arena is a useful automation layer, not a solution to the epistemic problem of explanation quality. It demonstrates that an LLM can approximate human ratings at scale—not that it can establish truth, guarantee faithfulness, or make judgment objective. Under the Discontinuity Thesis, its deeper significance is terminal for routine evaluative labor: authority and employment move upward to the owners of the evaluation infrastructure, while everyone else becomes benchmark data or a replaceable model call.