PaperlayerView exact revision on arXiv

Source-linked AI summary

Do MLLM Judges Judge the Edit? Auditing Bias in Image Editing Evaluation with Verified Quality Preservation

Yuan Huang, Zirui Song, Xiuying Chen

arXiv:2610.01670v1cs.CVcs.LG

TL;DR

MLLM judges are increasingly used to evaluate image edits and as reward signals, but audits have not established whether quality-preserving cues can shift their decisions beyond ordinary variability. EditJudgeBias tests this with verified interventions and finds that all five judges shift beyond their own noise, while different measures capture different aspects of robustness.

  • Problem

    Audits leave open whether presentation changes that preserve an edit can move judges’ decisions beyond ordinary response variability, an issue that matters when judges are used as reward models.

    Source: §1 · p. 1, §2.1 · p. 3

  • Method

    EditJudgeBias tests 1,196 real editing samples with 13 cues across four evaluation sites, verifies image-side cue preservation, and evaluates invariance, human agreement, and pairwise stability against judge-specific noise.

    Source: §1 · p. 2, §4.1 · p. 7

  • Results

    Across all five judges, quality-preserving cues shift decisions beyond their own noise; display-order swaps reverse up to 60.9% of pairwise decisions.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Judge reliability should include whether decisions remain invariant when underlying quality is unchanged, because agreement with humans alone does not capture all observed behavior.

    Source: §6 · pp. 9–10

    INSET is borderline for quality preservation, so its results are reported separately and excluded from aggregate counts.

    Source: §5.4 · p. 9

Abstract

from arXiv · show

Multimodal large language models (MLLMs) are increasingly used as automated judges for instruction-based image editing and as reward signals for model training. However, systematically auditing whether these judges are influenced by cues irrelevant to editing quality is challenging because visual interventions may themselves alter the quality being evaluated. A judgment shift can therefore be attributed to bias only when the intervention is verified to preserve the underlying editing quality. To address this challenge, we introduce EditJudgeBias, a counterfactual benchmark with verified quality preservation, comprising 1,196 real editing samples and 13 cues injected across four evaluation sites. We verify quality preservation for the requested edit using calibrated multimodal validators, controls, and human inspection. We then audit five MLLM judges along three complementary dimensions: invariance to quality-preserving cues, agreement with human judgments, and stability of pairwise preferences. Importantly, observed shifts are evaluated against each judge's own zero-dose and re-query noise floors rather than against zero. Experiments show that quality-preserving cues move every judge beyond its own noise. Fabricated majority opinions increase ratings, irrelevant visual elements cause larger shifts than whole-image manipulations, and swapping candidate order reverses up to 60.9% of pairwise decisions. Edit-region cues also tend to reduce human agreement. The three measures characterize judges differently, showing that robustness cannot be captured by a single metric.

1 INTRODUCTION

MLLM judges are used to evaluate image edits and support model training, but their decisions may respond to cues unrelated to editing quality. EditJudgeBias tests these effects with verified quality-preserving interventions and finds that all five judges shift beyond their own response noise.

1 INTRODUCTION

  • Because human evaluation is costly, multimodal judges are increasingly used for image-editing evaluation, output selection, and reinforcement learning.This makes their sensitivity to the evaluation procedure consequential.Source: §1 · p. 1
  • A cue-induced shift indicates bias only if the intervention preserves edit quality, since visual changes can alter the requested edit itself.The introduction notes that region annotations can obscure objects and brightness changes can affect an instructed attribute.Source: §1 · pp. 1–2
  • EditJudgeBias tests cue sensitivity with 1,196 real editing triplets, 13 cues at four evaluation sites, and checks of image-side cue preservation.Observed shifts are interpreted relative to each judge’s SHAM and re-query variability.Source: §1 · p. 2
  • Quality-preserving cues move all five judges beyond their own response noise; display-order swaps reverse up to 60.9% of pairwise decisions.Fabricated majority opinions raise all three ratings, while a small sticker causes larger shifts than whole-image manipulations; edit-region cues also tend to reduce human agreement.Source: §1 · p. 2

2 RELATED WORK

Prior work studies image-editing judges, general judge biases, and quality verification, but leaves open how verified quality-preserving cues affect editing judgments. EditJudgeBias addresses that gap with calibrated validators and human checks for image-side cues.

2.1 JUDGES FOR INSTRUCTION-BASED IMAGE EDITING

  • Earlier editing-evaluator studies compare judges with humans or catalog failure modes, but leave open whether presentation changes preserving the edit move decisions beyond ordinary variability.This matters when judges serve as reward models, since optimization can amplify exploitable shortcuts.Source: §2.1 · pp. 2–3

2.2 BIASES IN MLLM-AS-A-JUDGE

  • Text-only audits cover position, verbosity, and self-preference, while multimodal audits address visual biases in other tasks but leave edit-region cues largely unexplored.Image-editing evaluation must assess whether an edit follows an instruction given the source image, including the edit’s location and extent.Source: §2.2 · p. 3

2.3 QUALITY VERIFICATION OF INJECTED CUES

  • Perturbation audits are interpretable only when interventions preserve the evaluated construct, yet prior work handles preservation inconsistently.EditJudgeBias evaluates preservation relative to the source image and instruction using calibrated validators and human checks for every image-side cue.Source: §2.3 · p. 3

3 EDITJUDGEBIAS: BENCHMARK DESIGN

EditJudgeBias defines judge bias as sensitivity to a cue that leaves editing quality unchanged, and tests it across multiple evaluation protocols and cue sites. Its benchmark combines human-supervised editing samples, controlled interventions, and checks of quality preservation.

3 EDITJUDGEBIAS: BENCHMARK DESIGN

  • EditJudgeBias measures invariance, agreement with human judgments, and stability under quality-preserving cues, using human supervision to assess agreement and verify edit quality.Its 1,196 editing triplets span five public benchmarks and three independent content pools.Source: Figure 2 · p. 4

3.1 PROBLEM FORMULATION

  • A cue is a transformation that preserves underlying quality, and bias arises when a judge’s decision depends on that cue despite quality preservation.The framework assesses rating changes, agreement changes, and preference preservation under display-order swaps.Source: §3.1 · p. 4

3.2 EDITING SAMPLES

  • EditJudgeBias uses five published image-editing benchmarks spanning three independent content pools, with a 611-sample breadth block for invariance analysis and two complete-turn anchor blocks for judge–human comparison.The anchor blocks preserve complete editing turns and contain 384 EBench-18K samples and 240 ImagenHub samples.Source: §3.2 · pp. 4–5

3.3 BIAS CUES

  • Thirteen cues enter at protocol, prompt, pixel, or content sites; pixel and content cues are applied deterministically to edited images without re-rendering the underlying edit.Content cues use benchmark source masks when available, otherwise an estimated edit region; SHAM supplies a zero-dose control.Source: §3.3 · pp. 5–6

3.4 QUALITY-PRESERVATION VERIFICATION

  • Quality preservation is assessed using image similarity, calibrated MLLM validators, and human checks, with validators calibrated against SHAM false positives and known quality changes.Human checks cover all ten image-side cues; only a “Yes” label counts as a preservation failure under the predefined criterion.Source: §3.4 · p. 6

4 EXPERIMENTAL SETUP

The setup evaluates five primary judges using separate measures of cue sensitivity, human agreement, and pairwise stability, with cue effects assessed against judge-specific noise baselines. A sixth judge is evaluated separately as a pre-declared replication.

4.1 EVALUATION METRICS

  • Invariance measures paired rating changes and signed shifts, while cue-induced changes are compared with judge-specific SHAM and retest noise baselines.Bootstrap intervals and Wilcoxon signed-rank tests assess uncertainty and significance.Source: §4.1 · p. 6
  • Table 1 reports scaled mean absolute per-item rating changes by dimension, marks cells clearing both noise floors, and indicates significant agreement changes with human ratings.Its dimensions are instruction adherence, editing quality, and detail preservation.Source: Table 1 · p. 7
  • Agreement is measured by Spearman correlation between judge scores and human labels, and pairwise accuracy against human preferences; intervals are clustered by editing turn.McNemar’s test assesses pairwise changes.Source: §4.1 · p. 7
  • Pairwise stability includes preference preservation after an A/B order swap, consistency across repeated evaluations, and systematic preference for a presentation slot.The evaluation applies Benjamini–Hochberg correction within pre-defined comparison families.Source: §4.1 · p. 7

4.2 JUDGES AND EVALUATION SETUP

  • Five primary judges comprise two closed frontier models, two open-weight-family models, and the task-specific VIEScore; a sixth judge is analyzed separately as a pre-declared replication.Temperature 0 is requested when configurable, and retest noise is estimated separately for each judge.Source: §4.2 · p. 7

5 RESULTS AND DISCUSSION

The audit finds that judge robustness depends on the measure: quality-preserving cues shift ratings and preferences, can reduce human agreement, and display-order swaps produce reversals beyond re-query variability. Preservation checks support most image-side cues, while INSET remains borderline.

5.1 INVARIANCE UNDER QUALITY-PRESERVING CUES

  • Excluding INSET, half of instruction-adherence cells and 34 of 55 editing-quality and 32 of 55 detail-preservation cells clear both noise floors; sticker shifts exceed every pixel-site cue in each judge–dimension cell.BANDWAGON clears both floors for most judge–dimension combinations, whereas SATURATION never does; a cell can still show a significant directional shift without clearing both floors.Source: §5.1 · p. 8
  • BANDWAGON increases the cued candidate’s pairwise win rate under every judge, while the sticker and text overlay tend to disadvantage it; other cues have little pairwise effect.Cue effects depend on both the judge and whether evaluation uses absolute scores or comparative preferences.Source: §5.1 · p. 8

5.2 AGREEMENT WITH HUMAN JUDGMENTS

  • 33 of 34 significant judge–human agreement changes are decreases, concentrated descriptively among edit-region comparisons; fabricated majority opinions can lower or raise agreement for every judge depending on which candidate they favor.The reported concentration is not a formal comparison across sites because pixel and content groups include comparisons against both SHAM and uncued images.Source: §5.2 · p. 8

5.3 PAIRWISE STABILITY UNDER DISPLAY-ORDER SWAPS

  • For qwen3.5, 60.9% of order-swapped pairs reverse versus 10.5% under identical re-query; the two-order rule filters order-sensitive pairs at a cost to coverage.Pairs whose two orders select opposite winners remain unresolved under the described lenient variant.Source: §5.3 · p. 8

5.4 PRESERVATION EVIDENCE

  • Most image-side cues are supported as quality-preserving at the cue level by validator and human checks, but INSET is borderline and excluded from aggregate counts.Only INSET significantly departs from paired SHAM images among the ten image-side cues; none of five evaluated cues differs significantly from SHAM in the sham-controlled human set.Source: §5.4 · pp. 8–9, Figure 4 · p. 9

5.5 ROBUSTNESS ACROSS MEASURES

  • Judge rankings differ across invariance, agreement, and stability, so a single aggregate score would obscure which failure mode underlies apparent weakness.kimi-k2.5 ranks first on invariance and agreement but third on stability; gemini-3.5 moves from fifth on invariance to first on stability.Source: §5.5 · p. 9, Figure 4 · p. 9

6 CONCLUSION

EditJudgeBias finds that quality-preserving cues can shift ratings and pairwise preferences beyond ordinary response variability, and that human agreement alone does not capture judge reliability. The authors argue that reliability also requires decisions to remain invariant when evaluated quality is unchanged.

6 CONCLUSION

  • Across five primary judges, quality-preserving cues shift ratings and pairwise preferences beyond ordinary response variability.Human agreement does not fully predict this behavior; display-order swaps can reverse preferences, and judges rank differently across invariance, agreement, and stability.Source: §6 · p. 9
  • The authors argue that judge reliability should include invariance when the underlying quality being evaluated is unchanged, not just agreement with humans.They frame reliable judging as measuring the intended construct rather than responding to incidental presentation features.Source: §6 · p. 10

AI DISCLOSURE

  • The authors report that generative AI tools assisted with drafting, language editing, figure preparation, and software tasks, while research and interpretation were carried out by the authors.The listed author-led work includes research questions, literature selection, experimental design and deployment, data analysis, and interpretation.Source: AI DISCLOSURE · p. 11

ETHICS STATEMENT

  • Fabricated social information is confined to the evaluation protocol and is not treated as factual benchmark annotation.Human annotation is used only to check preservation of image-side cues.Source: ETHICS STATEMENT · p. 11

REPRODUCIBILITY STATEMENT

  • The open-source release omits benchmark images, human annotation records, judge and validator responses, and author-generated I2EBench outputs, preventing exact reconstruction of all evaluated inputs or results.Repeated API calls can also differ; reproducibility is therefore framed around statistical findings and qualitative conclusions under expected API variability, not exact model responses.Source: REPRODUCIBILITY STATEMENT · p. 11

A LIMITATIONS

The study’s scope is limited by the tested cues, judge versions, and mitigation analyses; preservation checks focus on the requested edit.

A LIMITATIONS

  • The study tests a finite set of controlled cues at specified strengths; interactions among co-occurring cues and responses across a broader strength range remain uncharacterized.The authors identify both as areas requiring further study.Source: A LIMITATIONS · p. 16
  • Results reflect the evaluated judge versions and configurations, so effect magnitudes and relative judge rankings may change as models and API services evolve.The authors identify periodic re-evaluation with the same audit protocol as a response to this boundary.Source: A LIMITATIONS · p. 16
  • The experiments diagnose cue sensitivity, while mitigation receives only limited aggregation analyses and a preliminary prompting pilot.Systematic mitigation evaluation under the main experimental protocol remains future work; preservation checks also focus on the requested edit rather than rubric-aligned human ratings of both versions.Source: A LIMITATIONS · p. 16

B BENCHMARK AND EVALUATION PROTOCOL

The benchmark draws editing samples from five published benchmarks representing three independent content pools, then evaluates judges under scoring and pairwise protocols with controlled cues. Its protocol specifies cue operations, edit-region handling, and judge versions and collection procedures.

B.1 EDITING SAMPLES AND POOLS

  • The 1,196 unique triplets come from five published benchmarks but represent three independent content pools; breadth and anchor blocks support invariance, human-agreement, and leaderboard analyses.The pairwise protocol uses 616 same-request edit pairs from 1,086 edited images, with decisive human preferences for 438 pairs.Source: §B.1 · p. 16

B.2 CUES, OPERATORS AND DOSES

  • Image cues include global pixel operations and edit-region content cues, while prompt cues add text and POSITION swaps pair display order without changing images or prompts.Examples include saturation and brightness scaling by 1.2, a sticker outside the edit region covering at most 4% of the frame, and a seeded majority-preference claim in pairwise evaluation.Source: §B.2 · p. 16

B.3 THE EDIT REGION

  • Content cues use a benchmark source mask when available; otherwise, the edit region is estimated from thresholded source–edited differences, with a centered-box fallback if no component is found.BOX and INSET operate on the region, STICKER is placed away from it, and CAPTION avoids covering it.Source: §B.3 · p. 17

B.4 JUDGES, RUBRICS AND COLLECTION

  • Scoring judges return three 1–10 ratings summed to a 3–30 score; the study re-queries 200 uncued items for retest noise and identifies the primary and replication judge versions.VIEScore uses its original rubric with GPT-4o; prompts and human annotation instructions are reproduced in Appendix G.Source: §B.4 · p. 17

C.1 VALIDATOR CALIBRATION AND RESULTS

The authors calibrate preservation validators, check image-side cues with human annotation, and use a validated-image subset to test whether effect directions persist.

C.1 VALIDATOR CALIBRATION AND RESULTS

  • INSET alone had significantly lower validator pass rates than SHAM after correction (gemini-3.5 q = 0.020; glm-4v q < 0.001), and validators differed in degradation sensitivity.The gate uses gemini-3.5 and gpt-4o-mini, while glm-4v is an independent cross-check; Table 2 reports cue pass rates and controlled-degradation detection.Source: §C.1 · pp. 17–18, Table 2 · p. 18

C.2 HUMAN VERIFICATION

  • Under the prespecified criterion, no sham-controlled image cue differed significantly from SHAM in human-rated preservation failures; counting “Slightly” post hoc changed the result only for INSET.The post-hoc INSET comparison was 11/30 versus 0/30 failures (q = 0.002); the second annotator had 98.3% raw agreement and Cohen’s κ = 0.893.Source: §C.2 · p. 18

C.3 SENSITIVITY ANALYSIS ON VALIDATED IMAGES

  • All 24 effects significant in both full and validator-filtered analyses retained the same direction, with 47–57 items per cue–judge cell in the filtered analysis.The authors treat this as a direction check within the validated subset, not as a second estimate of effect magnitude or evidence of preservation for every image.Source: §C.3 · p. 18

D.1 DETAILED SCORING RESULTS

D.1 reports signed rating shifts and agreement changes, while D.2 defines the aggregation schemes used to compare judges’ robustness rankings.

D.1 DETAILED SCORING RESULTS

  • Scoring analyses use 597–611 paired observations per judge, while agreement analyses use 361–384; cue comparisons use different reference arms.Bandwagon and several other cues are compared with SHAM, while luminance, framing, typographic, and attention guidance use the original uncued arm.Source: §D.1 · p. 19
  • Tables 3 and 5 report paired mean shifts on the 1–10 scale for instruction adherence and detail preservation, with Benjamini–Hochberg significance markers.Table 1 separately reports shift magnitudes irrespective of direction and marks changes in agreement with human ratings.Source: Table 3 · p. 19, §D.1 · p. 19, Table 5 · p. 20

D.2 ROBUSTNESS ACROSS AGGREGATION SCHEMES

  • The primary robustness summary averages absolute paired rating changes for invariance, absolute Spearman’s ρ changes for agreement, and RR for stability.Alternative summaries use worst-cell values or significant-cell counts, with CR for stability; Table 6 presents rankings and rank inversions.Source: §D.2 · pp. 19–20, Table 6 · p. 20

E.1 CONTROL AND CROSS-BENCHMARK ANALYSES

Controls show that key cue effects persist beyond instruction collisions and low-quality edits, while cross-benchmark and downstream analyses reveal consequences for agreement, rankings, and aggregation.

E.1 CONTROL AND CROSS-BENCHMARK ANALYSES

  • Significant negative luminance shifts remain on clean items for four of five judges, while BANDWAGON and sticker effects persist for all five judges on the highest human-MOS tertile.The controls do not support instruction collisions or poor edits as sufficient explanations for the strongest effects.Source: §E.1 · p. 21, Table 7 · p. 21, Table 8 · p. 21
  • Attention guidance most consistently reduces judge–human agreement across both anchors and is the only shared cue significant under both selection-corrected summaries.Panel A reports cue-level Δρ intervals; Panel B reports permutation tests with psel correcting for selection of the most extreme cue.Source: §E.1 · p. 22, Table 9 · p. 22
  • Pairwise bandwagon cues raise agreement by 4.2–25.5 points when favoring the human-preferred candidate and lower it by 19.2–28.6 points when favoring the rejected candidate.All five judges show significant decreases when the fabricated majority favors the human-rejected candidate.Source: §E.1 · p. 22

E.2 DOWNSTREAM CONSEQUENCES AND AGGREGATION

  • On EBench-18K, several cues change the top-ranked editor, with text overlay doing so under all five judges; no evaluated cue changes the top editor on ImagenHub.Table 10 reports cued-editor mean rank changes and marks changes to the top-ranked editor.Source: §E.2 · pp. 22–23, Table 10 · p. 23
  • Requiring pairwise agreement across display orders lowers coverage for every judge; accuracy rises for four among retained pairs, but not for qwen3.5.The strict rule removes order-sensitive pairs rather than resolving disagreements, trading coverage for a more selective set of verdicts.Source: §E.2 · p. 23, Table 11 · p. 23
  • The standardized-median ensemble remains significantly shifted for several cues, while model name and caption effects are near zero when individual directions are split.Table 12 standardizes each judge on its uncued distribution before aggregation; “Members” counts significant effects among the five primary judges.Source: §E.2 · p. 24, Table 12 · p. 24

F REPLICATION WITH QWEN3-VL-32B-INSTRUCT

A prespecified replication judge reproduces the strongest cross-judge cue effects and shows much greater reversal under order swaps than under re-queries. The appendices provide the prompts and preservation-annotation materials used in the evaluation.

F REPLICATION WITH QWEN3-VL-32B-INSTRUCT

  • All six cues significant in the same direction across the five primary judges are significant in that direction for qwen3-vl-32b-instruct.The replication judge reverses 45.5% of order-swapped verdicts versus 1.2% under identical re-query; its arm is analyzed separately from the primary panel.Source: F REPLICATION WITH QWEN3-VL-32B-INSTRUCT · pp. 24–25, Table 13 · p. 24

G PROMPT TEMPLATES AND ANNOTATION INSTRUCTIONS

G.1 JUDGE PROMPTS

  • The scoring prompt uses three dimension ratings individually or as a 3–30 sum; overall score and reason are required fields but are not analyzed.The shaded VIEScore paragraph appears only in the VIEScore scoring condition.Source: Figure 5 · p. 26

G.2 PROMPT-SITE INTERVENTIONS

  • Prompt-site interventions change only the indicated line, leaving the base judge prompt otherwise unchanged; the cue templates show the inserted strings.The pairwise bandwagon target is assigned with a fixed seed.Source: §G.2 · p. 28, Figure 7 · p. 28

G.3 PRESERVATION VERIFICATION

  • The same preservation-validator prompt template is used across all three validators, with evaluated images and editing instructions varying by instance.The human preservation instructions include an additional independence instruction for the second annotator.Source: §G.3 · p. 28, Figure 8 · p. 29, Figure 9 · p. 30

Related papers

Loading 2610.01670v1…