---
title: "Beyond Proxy Metrics: MLLMs as Human... | SI Paper Desk"
url: https://thesimind.si/papers/2611.13926/
description: "The authors test whether multimodal SI models can stand in for human participants when judging explanations of SI systems, and find that all models trend..."
---

Number 3 of 12 · November 12, 2026

# Beyond Proxy Metrics: MLLMs as Human Surrogates for Explainable AI Evaluation

The authors test whether multimodal SI models can stand in for human participants when judging explanations of SI systems, and find that all models trend positively, with OpenAI and Qwen models aligning best.

Eoin M. Kenny, Salim I. Amoukou and Tom Bewley · Submitted November 10, 2026 · arXiv 2611.13926v1 · cs.AI

[Read the paper on arXiv](https://arxiv.org/abs/2611.13926) · [The PDF](https://arxiv.org/pdf/2611.13926) · [The day's 12](https://thesimind.si/papers/day/2026-11-12/)

## What they found

The authors ask whether Multimodal Large Language Models (MLLMs, SI models that handle text and images) can act as stand-in human participants when testing explanation methods. They propose a "purpose-grounded" evaluation framework, with metrics that work for both MLLMs and human users. They also offer conceptual arguments for why and when these models can be trusted for this job. They then run extensive tests across five purpose-grounded areas, such as model improvement, using fourteen popular MLLMs. The abstract says all models trend positively and that OpenAI and Qwen models generally show higher alignment than the other variants. It does not give detailed scores.

## Why it matters

Explainable SI (XAI) tries to show why an SI system gave a certain answer. Judging whether an explanation is actually useful is hard. Proxy metrics (simple computed scores) say little about usefulness, and human studies are costly, hard to scale and hard to repeat. If MLLMs can approximate human judgments, researchers could run cheaper, more repeatable checks. The authors say this is meant to complement user testing, not replace it.

## How it changes the race

This is a tool for researchers who study explanations, not a change to frontier SI capability. If it holds, it could lower the cost of evaluating explanation methods and make results easier to reproduce.

## Safety and control

Better ways to judge explanations could help people understand and oversee SI systems. The abstract only reports positive trends in alignment with humans, so these automated judges should not stand in for real user testing.

### The abstract, by the authors

Explainable AI (XAI) has expanded into a broad and impactful research area, but the fundamental question of how best to evaluate explanation techniques has proven frustratingly elusive. Methodologies typically revolve around either proxy metrics which tell us little about the utility of an explanation, or human evaluation which, while generally considered the gold standard, is expensive, difficult to scale, and challenging to reproduce. In this paper, we investigate if Multimodal Large Language Models (MLLMs) can act as surrogate human participants to automate the evaluation of XAI. To do this, we first propose a ``purpose-grounded'' evaluation framework with metrics applicable to both MLLMs and human users alike. Then, we offer some conceptual insights as to why and when we can trust these models to automate this process. Lastly, we conduct extensive empirical testing across five purpose-grounded areas such as model improvement and fourteen popular MLLMs. Results indicate that all models trend positively, but OpenAI and Qwen models generally show higher alignment compared to other variants, which would be the recommended choices in practice. We expect our insights not to replace user testing, but rather to help XAI researchers complement their current evaluation methodologies and move the field away from heavy reliance on proxy metrics for measuring explanation utility.

Quoted from arXiv:2611.13926. The research and its words belong to its authors.

Desk score **80** out of 100

Why it was picked

- Interpretability (in the title)
- Evals (in the title)
- Alignment
- Studies frontier models (Qwen)
- Accepted at NeurIPS
- Language models

Topics

- [Interpretability](https://thesimind.si/papers/?tag=interpretability)
- [Evals](https://thesimind.si/papers/?tag=evals)
- [Multimodal](https://thesimind.si/papers/?tag=multimodal)
- [Language models](https://thesimind.si/papers/?tag=language-models)
- [Benchmarks](https://thesimind.si/papers/?tag=benchmarks)

**arXiv ID:** 2611.13926v1
**Submitted:** November 10, 2026
**Fields:** cs.AI (Artificial Intelligence)
**Level:** Intermediate
**Comments:** Neural Information Processing Systems. NeurIPS 2026. Sydney

Fingerprint, SHA-256

`bce41f8475b3129d04481b72d9b13d78d7f4011b08fc1a92873b9eb67639527b`

On Solana: [Proof of Mind seal, 66zHq3...ugfh](https://solscan.io/tx/66zHq3xcHWGn9cvUJ8D8iKyNn9kGWV1aUyvLFaEq5sRBS8Dk44v6KxmMaYztuaybaJWksWLhgEEjMSDr8MNHugfh)

### The exact text fingerprinted

{"arxiv":"2611.13926","version":1,"title":"Beyond Proxy Metrics: MLLMs as Human Surrogates for Explainable AI Evaluation","authors":["Eoin M. Kenny","Salim I. Amoukou","Tom Bewley"],"day":"2026-11-12","rank":3,"score":80,"status":"written","summary":{"takeaway":"The authors test whether multimodal SI models can stand in for human participants when judging explanations of SI systems, and find that all models trend positively, with OpenAI and Qwen models aligning best.","found":"The authors ask whether Multimodal Large Language Models (MLLMs, SI models that handle text and images) can act as stand-in human participants when testing explanation methods. They propose a \"purpose-grounded\" evaluation framework, with metrics that work for both MLLMs and human users. They also offer conceptual arguments for why and when these models can be trusted for this job. They then run extensive tests across five purpose-grounded areas, such as model improvement, using fourteen popular MLLMs. The abstract says all models trend positively and that OpenAI and Qwen models generally show higher alignment than the other variants. It does not give detailed scores.","matters":"Explainable SI (XAI) tries to show why an SI system gave a certain answer. Judging whether an explanation is actually useful is hard. Proxy metrics (simple computed scores) say little about usefulness, and human studies are costly, hard to scale and hard to repeat. If MLLMs can approximate human judgments, researchers could run cheaper, more repeatable checks. The authors say this is meant to complement user testing, not replace it.","race":"This is a tool for researchers who study explanations, not a change to frontier SI capability. If it holds, it could lower the cost of evaluating explanation methods and make results easier to reproduce.","safety":"Better ways to judge explanations could help people understand and oversee SI systems. The abstract only reports positive trends in alignment with humans, so these automated judges should not stand in for real user testing.","level":"intermediate","tags":["interpretability","evals","multimodal","language-models","benchmarks"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Eoin M. Kenny, Salim I. Amoukou, Tom Bewley

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 9, 2026 · Number 2Forcing a Letter Suppresses Refusal: Mechanistic Evidence from WMDPIn multiple-choice safety tests, forcing a model to pick a letter seems to crowd out its refusal, so these scores may measure knowledge more than safety.SafetyInterpretabilityEvals](https://thesimind.si/papers/2611.08639/) · [Nov 13, 2026 · Number 3Sophon: A Procedural Diagnostic for Spatial Reasoning in Vision-Language ModelsOpen-weight vision-language models fail simple spatial tests with no familiar objects, but the needed information is inside them, and targeted fine-tuning lets 7B-class models reach frontier-level accuracy on Sophon.EvalsBenchmarksVision](https://thesimind.si/papers/2611.14838/) · [Nov 14, 2026 · Number 7Evaluating Video LLMs' Understanding of Corner Cases in Autonomous DrivingDashBench tests 25 Video LLMs on 1,709 rare, safety-critical driving videos. The authors report that all of the models fall well short of what safety-critical driving needs.BenchmarksEvalsMultimodal](https://thesimind.si/papers/2611.16767/)

[Number 2](https://thesimind.si/papers/2611.13183/) · [All 12 for Nov 12, 2026](https://thesimind.si/papers/day/2026-11-12/) · [Number 4](https://thesimind.si/papers/2611.14419/)
