PaperlayerView exact revision on arXiv

Source-linked AI summary

A Plausible Explanation Is Enough to Mislead Multimodal Judges

Anglin Cai, Zhiqi Wang, Xiaozhi Yang, Yurun Lin, Dongwon Lee, Yuchen Yang

arXiv:2610.22631v1cs.CLcs.AI

TL;DR

The paper asks whether ordinary plausible explanations can sway MLLM judgments that require visual verification. Using JUPE, a controlled benchmark that varies explanations while fixing candidate answers, it finds shifts toward acceptance across judges, with effects from ordinary explanations comparable to persuasive arguments.

  • Problem

    It is unclear whether ordinary plausible explanations sway multimodal judgments that require visual verification, despite MLLMs increasingly serving as judges in training and evaluation pipelines.

    Source: §1 · p. 1

  • Method

    JUPE is a controlled VQA-style benchmark that fixes candidate answers while varying explanation presence and type to isolate explanation effects.

    Source: §1 · p. 1, §2.1 · p. 2

  • Results

    Across tested settings, plausible explanations shift judgments toward accepting correct and incorrect answers; plain justifications reduce rejection of incorrect answers by 12.7 percentage points.

    Source: §1 · p. 2, §5 · p. 8

  • Takeaways & Limitations

    Reference-free evaluation of answer-explanation outputs risks over-accepting incorrect answers and inflating performance estimates; judgments on correct and incorrect answers should be reported separately with criterion and discriminability.

    Source: §5 · p. 8

    The reported effect sizes vary across judges, datasets, and protocols, and the study examines plausibility capture in a controlled evaluation setting.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Multimodal large language models (MLLMs) increasingly serve as judges in training and evaluation pipelines, often assessing answers together with explanations. Although AI judges are sensitive to textual factors, it remains unclear whether ordinary plausible explanations can sway judgments requiring visual verification. We introduce JUPE, a visual-question-answering (VQA)-style benchmark that fixes answers while varying explanations. Across 14 MLLM judges, ordinary explanations that sound plausible mislead judges, with effects comparable on average to deliberately persuasive arguments. Plain justifications and chain-of-thought rationales reduce rejection of incorrect answers by 12.7 and 13.8 percentage points, respectively. We call this failure plausibility capture. Using signal detection theory, we find a significant mean shift in decision criterion toward acceptance, without a clear change in mean discriminability. Explanations sway judgments across domains, explanation generators, and judging protocols, while tested mitigations offer limited, model-dependent protection. Such shifts can inflate performance estimates when judges evaluate answers with explanations.

1 Introduction

The paper asks whether plausible explanations can sway MLLM judgments that should be grounded in visual evidence, and introduces JUPE to isolate explanation effects. Across judges, explanations shift decisions toward acceptance, with ordinary explanations producing substantial effects and tested mitigations offering limited or inconsistent protection.

1 Introduction

  • The study addresses whether plausible explanations sway judgments requiring visual verification, a question made salient by MLLMs’ use as judges in training and evaluation pipelines.The paper identifies plausibility capture and uses signal detection theory to distinguish shifts in decision criteria from changes in discriminability.Source: §1 · pp. 1–2
  • JUPE fixes candidate answers while varying explanation presence and type, isolating explanation effects in a controlled VQA-style benchmark.Its answer-supporting explanation conditions include jointly generated CoT, post-hoc justifications, and persuasion-guided arguments across medical, chart, and hallucination settings.Source: §1 · pp. 1–2
  • Plain justifications and CoT reduce rejection of incorrect answers by 12.7 and 13.8 percentage points, comparable to the 12.5-point drop from persuasive arguments.These effects are observed across 14 MLLM judges from five model families; Figure 1 illustrates a verdict changing from definitely incorrect to definitely correct when an explanation is added.Source: Figure 1 · p. 1, §1 · p. 2
  • Signal detection analysis finds a significant shift in decision criterion toward acceptance, without a clear change in mean discriminability.Overall accuracy falls by less than two percentage points on average despite a 12.5–13.8-point drop in rejection of incorrect answers.Source: §1 · p. 2
  • Plausibility capture persists across domains, an out-of-distribution grounding set, explanation generators, image degradation, and judging protocols, while tested mitigations are limited or inconsistent.Persuasive explanations also induce errors on easy questions that judges answer correctly on their own.Source: §1 · p. 2
  • The findings suggest that judge pipelines may over-accept wrong answer-explanation outputs and inflate model-performance estimates.The authors call for mitigation methods designed specifically for plausibility capture.Source: §1 · p. 2

2 JUPE: Isolating Explanation Effects

JUPE uses a within-item VQA design to vary explanations while holding the image, question, candidate answer, and correctness fixed. Its construction combines reference-free judging, selected queries from three visual domains, and multiple explanation conditions designed to test item-specific support and persuasion.

2.1 Task Setup

  • JUPE fixes the image, question, candidate answer, and correctness label while varying only the explanation, so judgment changes can be associated with explanation conditions.Judges receive the visual query without a reference answer and return one of four ordinal judgments, which are grouped into ACCEPT and REJECT for binary metrics.Source: §2 · p. 2, §2.1 · p. 2

2.2 Query Selection

  • 250 queries comprise 180 main-set items and 70 induced-wrong items, sampled after profiling and filtering queries with failures, low-quality images, ambiguous questions, or erroneous labels.The benchmark draws from medical VQA, chart reasoning, and hallucination-prone visual-question datasets.Source: §2.2 · pp. 2–3

2.3 Explanation Conditions

  • For the main set, GPT-5.4 mini generates candidate answers and CoT traces; ground-truth matching yields 96 incorrect and 84 correct candidates.The correctness mix allows analysis of judgment criteria separately from discrimination between correct and incorrect answers.Source: §2.3 · p. 3
  • PLAIN adds a post-hoc justification and CoT adds its jointly generated reasoning trace; VACUOUS and MIS-MATCH controls test explanation presence and item-specific support.MIS-MATCH uses a CoT trace from another item in the same domain, preserving domain style and vocabulary while breaking alignment with the current item.Source: §2.3 · p. 3
  • PERSUASION uses 41 techniques in vision-informed and text-only variants, while Table 1 describes condition counts before judge-model replication.The image-access comparison tests whether visual access during explanation construction changes its effect; judges receive the image in both variants.Source: §2.3 · p. 3, Table 1 · p. 3

3 Evaluation

The evaluation tests whether plausible explanations change MLLM judgments, how they shift decisions, and whether the effects persist across settings or respond to mitigations. Across judges, explanations reduce rejection of incorrect answers, shift criteria toward acceptance without a clear mean discriminability change, and the effects generalize while mitigation results vary by judge.

3 Evaluation

  • The evaluation uses 14 MLLM judges, binarized four-point verdicts, and CRR and IRR to measure rejection of incorrect and correct candidates.It organizes analysis around reliability, decision shifts, factors modulating capture, and generalization, then tests mitigations.Source: §3 · p. 4

3.1 Effects on Judge Reliability

  • 12.7, 13.8, and 12.5 pp: PLAIN, COT, and PERSUASION reduce incorrect-answer rejection by comparable margins across 14 judges.Open-source judges show larger average effects, while four proprietary judges show smaller drops and two rise under all three conditions.Source: §3.1 · pp. 4–5

3.2 How Explanations Shift Judgments

  • ∆c shifts toward acceptance under all three explanation conditions, while mean ∆d′ shows no clear change in discriminability.The criterion shifts are −0.39, −0.45, and −0.39 for PLAIN, COT, and PERSUASION; overall accuracy falls by less than 2 pp despite a 12.5–13.8 pp CRR decline.Source: §3.2 · p. 5

3.3 Factors Modulating Plausibility Capture

  • Rhetoric ratings do not correlate with CRR drops across 41 techniques, while capture persists with text-only rewriting and when judges can answer the questions.Removing images reduces d′ in 40 of 44 judge-condition cells, indicating that visual input still carries discriminative signal.Source: §3.3 · p. 6

3.4 Generalization Across Settings

  • Explanations lower CRR across generators, out-of-distribution visual grounding, and degraded images; alternative judging protocols instead increase normalized ratings of incorrect candidates, not binary acceptance rates.CRR falls by 12.4 pp with Gemini 3 Flash versus 12.7 pp with GPT-5.4 mini; visual-grounding CRR drops by 15.5 pp under PLAIN and 11.6 pp under PERSUASION, while degraded-image CRR and IRR drop by 18–19 pp and 11–13 pp relative to no explanation. Under alternative protocols, incorrect candidates’ mean normalized ratings rise by 0.08–0.15 versus about 0.14 under the main protocol, and by 0.20–0.23 when answers and explanations are merged.Source: §3.4 · p. 7

3.5 Can Plausibility Capture Be Mitigated?

  • Conservative judging and self-verification increase wrong-answer rejection by 25% and 13%, respectively, relative to explanations alone.Effects vary by judge: conservative instruction raises CRR by 22.7 pp for Gemma 3-12B but lowers it by 1.0 pp for InternVL3.5-8B.Source: §3.5 · pp. 7–8

4 Related Work

Prior research documents textual and multimodal biases in AI judging, including effects of persuasive or fluent explanations. This work distinguishes item-specific plausible explanations from deliberate persuasion and finds comparable average judgment shifts.

4 Related Work

5 Conclusion

Across 14 MLLM judges, plausible explanations shift judgments toward accepting both correct and incorrect answers without a clear change in mean discriminability. The effects persist across evaluation settings, with risks for reference-free assessment.

5 Conclusion

  • Across 14 MLLM judges, plausible explanations shift judgments toward acceptance of correct and incorrect answers without a clear change in mean discriminability.Plain justifications and CoT produce shifts comparable to dedicated persuasive arguments; vacuous and mismatched controls suggest a role for item-specific answer support.Source: §5 · p. 8
  • The effect persists across datasets, explanation generators, and judging protocols, while the strongest prompt-level mitigations improve mean wrong-answer rejection with effects varying across judges.Source: §5 · p. 8
  • Reference-free evaluation of answer-explanation outputs risks over-accepting incorrect answers and inflating performance estimates; the authors recommend reporting correct and incorrect judgments separately with criterion and discriminability.Capture can occur even when judges answer questions correctly themselves; image input still aids discrimination, while boxes marking candidate objects can further sway judgments.Source: §5 · p. 8

Limitations

The study examines plausibility capture in a controlled evaluation setting, holding candidate answers and correctness fixed while varying explanations. Reported effect sizes vary across judges, datasets, and protocols.

Limitations

  • The study examines plausibility capture in a controlled evaluation setting.Source: Limitations · p. 9
  • Candidate answers and their correctness are held fixed while the accompanying explanations vary, allowing judgment changes to be associated with explanation conditions.Source: Limitations · p. 9
  • Reported effect sizes vary across judges, datasets, and protocols, with aggregate trends accompanied by per-judge results to contextualize that variation.Source: Limitations · p. 9

Ethics Statement

The paper reports safeguards around data and model access, and specifies experimental settings for reproducibility. Dataset, code, and experimental outputs are slated for public release upon publication.

Ethics Statement

  • The study uses publicly released open-source VQA-style datasets under their licenses, avoids private or restricted data, and accesses commercial models only through authorized interfaces.It does not extract model parameters, bypass access controls, or train competing models on model outputs.Source: Ethics Statement · p. 9

A.1 GPU, Hyperparameters, and Code

  • Open-source models use greedy decoding and API models use temperature 0, except Gemini 3 and GPT-5.6 Luna, which run at temperature 1 for stated API or performance reasons.Judge reasoning is set to the lowest level each API offers.Source: §A.1 · p. 12
  • The authors plan to release the dataset, code, and experimental outputs publicly upon publication.Source: §A.1 · p. 12

A.2 Benchmarked models

This appendix documents the benchmarked judges and source datasets, the instructions used to elicit judgments and explanations, and the procedures for selecting queries and constructing explanation conditions.

A.2 Benchmarked models

  • The benchmark covers 14 judges across Gemini 3, Gemma 3, GPT-5.4, GPT-5.6 Luna, InternVL3.5, and Qwen3-VL families.The listed family sizes are 3, 3, 2, 1, 2, and 3 models, respectively.Source: §A.2 · p. 12

A.3 Source Datasets

  • Six source datasets are used: VQA-RAD, PMC-VQA, ChartQA, ChartQAPro, HaloQuest, and PRE-HAL, with licenses listed for each.The licenses are CC0-1.0, CC BY-SA, GPL-3.0, MIT, Apache-2.0, and CC-BY-4.0, respectively.Source: §A.3 · p. 13

A.4 Instructions for MLLM Judges

  • Judges assess whether a candidate answer matches the image and question; tested prompt variants add conservative, self-verification, explanation-aware, persuasion-aware, or self-summarization instructions.The response format requires an A–D choice.Source: §A.4 · p. 13

A.5 Instructions for Explanation Generators

  • Plain explanations are written as if the generator arrived at the candidate answer itself, while persuasion rewrites aim to increase acceptance without changing the answer or adding fabricated visual details.The plain instruction requests an 80–100-word explanation; the persuasion instruction says to preserve the original inferential evidence.Source: §A.5 · pp. 13–14

B.1 Query Selection

  • Queries are profiled with VQA models, screened for quality, and labeled against source ground truth using dataset-specific rules before an LLM grader is used when needed.An induced-wrong set uses deliberately wrong but plausible answers on queries that profiling models generally answer correctly.Source: §B.1 · p. 14

B.2 Explanation Conditions

  • The PERSUASION conditions instantiate 41 techniques, with strategy and ethical labels and per-technique persuasion effects reported.The techniques draw on a source taxonomy, with Information-dense adapted from another work.Source: §B.2 · p. 15

B.3 Explanation Text-Quality Metrics

The text-quality analysis compares fluency, readability, lexical diversity, and answer entailment across explanation conditions, then distinguishes class-specific judging errors from aggregate metrics.

B.3 Explanation Text-Quality Metrics

  • PLAIN, COT, and persuasion variants occupy narrow surface-metric ranges, while entailment orders them: PLAIN .695, persuasion .595 and .597, and COT .489.The surface ranges are perplexity 45–60, reading level 9.9–11.2, and Distinct-1 .71–.74.Source: §B.3 · p. 15
  • Vacuous and Mismatch controls have near-zero answer entailment (.002 and .032), although Vacuous has the highest reading level (16.66).Mismatch falls within the plausible-condition range on surface metrics.Source: §B.3 · p. 15

B.4 Metrics Design

  • Accuracy can hide wrong-answer detection loss: under plausible explanations, mean accuracy falls by under 2 pp while CRR falls by 12.5–13.8 pp.CRR and IRR condition on the true label, allowing changes in discrimination to be distinguished from uniform criterion shifts.Source: §B.4 · pp. 15, 17

B.5 Human Evaluation

A human audit assesses item soundness and explanation readability and coherence, while accompanying examples and a rhetoric-rating procedure characterize the explanation materials.

B.5 Human Evaluation

  • Annotators judged 98.3% of sampled items sound and 93.2% of ground truths correct; COT and PERSUASION were coherent (about 4.5), unlike controls (about 1.9).The audit used four annotators on a stratified 60-item sample; the plausible-control coherence gap was 2.62, compared with 1.09 for readability.Source: §B.5 · p. 17, Table 6 · p. 17

B.6 Explanation Examples

  • The chart-based examples illustrate explanations that justify candidate answers through endpoint comparisons, including versions that add shared-standard framing or broader visual language.The examples include descriptions of the rightmost GDP-per-capita endpoints and identify Ethiopia as second-lowest.Source: §B.6 · pp. 17–18

C.1 Persuasive-Rhetoric Rating

  • Persuasive rhetoric is rated from 1 (neutral description) to 5 (rhetoric-dominated), but ratings do not predict the CRR drop across techniques.Three raters agree closely on rhetoric scores (ICC(3,1) 0.94), while each rater’s correlation with CRR drop is |ρ| ≤0.04.Source: §C.1 · p. 18

C.2 Additional Breakdowns

Additional analyses describe judge-specific signal-detection exceptions and provide breakdowns of verdict transitions and ratings on incorrect-candidate items.

C.2 Additional Breakdowns

  • Gemini 3.1 Pro and GPT-5.6 Luna are exceptions to the criterion shift toward acceptance, but for different reasons.Pro is least moved by explanations and shifts similarly under plausible and vacuous text; Luna’s discriminability rises under explanation-shaped text except MISMATCH.Source: §C.2 · pp. 18–19, Table 8 · p. 20
  • Table 7 reports CRR and IRR for control and persuasion conditions, plus accuracy and CRR on the induced-wrong set; Table 9 breaks down ratings for incorrect candidates by explanation condition.Table 9 distinguishes definitely from probably responses.Source: Table 7 · p. 19, Table 9 · p. 21
  • Figure 8 compares each judge’s no-explanation verdicts in rows with verdicts under explanation conditions in columns, with row-normalized percentages.Source: Figure 8 · p. 20

D.1 Datasets and Explanation Generators

The appendix probes grounding-set performance, image degradation, judging-protocol changes, prompt-level mitigations, and AI-assistant use. PLAIN lowers mean CRR across three judging settings, while blanking the image shrinks the rejection-rate gap for nearly every judge.

D.1 Datasets and Explanation Generators

  • Under Plain, the three strongest closed-source judges lose at most 4.2 pp of CRR, while open-source judges lose 20.3 pp on average.These results use the independent visual-grounding set; Table 10 reports per-judge CRR and IRR and cross-judge mean changes with query-level bootstrap confidence intervals.Source: §D.1 · p. 22, Table 10 · p. 23
  • The 120-query grounding set yields 240 balanced judgments with manually verified candidates; Plain+Box adds a red box on the candidate object and a visual-evidence prompt.Grok 4.6 generates the explanations, and 12 judges run the grounding evaluation.Source: §D.1 · p. 22

D.2 Image Degradation

  • Blanking the image shrinks the CRR-minus-IRR gap for nearly every judge, and d′ is lower in 40 of 44 judge-condition cells.Table 16 also reports CRR and IRR under JPEG compression at quality 5 and Gaussian noise with σ = 25.Source: §D.2 · p. 22, Table 16 · p. 26, Table 17 · p. 27

D.3 Judging Protocols

  • Without explanations, reasoning first raises incorrect-candidate ratings from 0.555 to 0.619 on the four-label scale and from 0.663 to 0.690 on the 0-3 scale.Table 15 reports mean normalized ratings across six open-source judges and the paired changes under each protocol.Source: §D.3 · p. 22, Table 15 · p. 26
  • PLAIN lowers mean CRR by 12.8 pp under the main, multi-turn, and multi-criteria settings, while mean IRR falls by 11.4, 11.5, and 12.1 pp, respectively.Gemini 3.1 Pro rejects more incorrect answers with the explanation in all three settings.Source: §D.3 · p. 23, Table 12 · p. 24

F Use of AI Assistants

  • AI assistants were used for language polishing, grammar correction, and readability; authors reviewed edits, and assistants did not generate scientific claims or conduct experiments or analysis.They also were not used to write substantive technical content.Source: F Use of AI Assistants · p. 23

E Prompt-Level Mitigation

  • Table 18 compares five one-sentence prompt add-ons under PERSUASION-VI and COT: Conservative, Self-Verify, Expl.-Aware, Pers.-Aware, and Self-Summ.The table reports CRR and IRR under each mitigation; each adds one sentence to the judging instruction.Source: E Prompt-Level Mitigation · p. 23, Table 18 · p. 28
Loading 2610.22631v1…