Multi-Crit: Benchmarking Multimodal Judges on Pluralistic Criteria-Following

AI summary

Overview

Research area: Multimodal AI evaluation — specifically benchmarking large multimodal models (LMMs) that act as automated judges ("LMM-as-a-Judge").

Technical level: Intermediate. Readers need familiarity with preference-based evaluation, reward models, and multimodal benchmarks, but the paper's framing and metrics are explained with clear formal definitions.

Scope: The paper introduces Multi-Crit, a benchmark of 425 multimodal prompts and 1,425 criterion-level human judgments designed to test whether multimodal judges can follow several distinct evaluation criteria per response pair, and evaluates 25 LMMs against it.

What This Paper Is About

Existing multimodal judge benchmarks give each pair of model responses a single overall preference label, which hides the fact that two responses can trade off against each other — one may be factually correct but visually ungrounded, while another is well-grounded but slightly hallucinated. Multi-Crit asks whether LMM judges can follow multiple, fine-grained evaluation criteria at once and correctly report criterion-level preferences, including cases where the criteria disagree with each other. The goal is to measure how reliable and steerable current multimodal judges really are when evaluation criteria are pluralistic rather than holistic.

Key Contributions

1. A new benchmark with multi-criterion human annotations. Multi-Crit covers 425 multimodal prompts across open-ended generation (299 prompts) and verifiable reasoning (126 prompts), with 1,425 criterion-level human judgments and 782 conflicting criterion pairs. It is described as the first benchmark capable of assessing whether LMM judges follow pluralistic evaluation criteria and deliver reliable criterion-level judgments.

2. Three novel pluralistic metrics. The paper introduces Pluralistic Accuracy (PAcc), which requires correct judgments across all criteria for an instance; Trade-off Sensitivity (TOS), which checks whether a judge perceives criterion-level trade-offs within an instance; and Conflict Matching Rate (CMR), a stricter metric requiring the judge to resolve ground-truth criterion conflicts consistently with humans.

3. A rigorous data curation and annotation pipeline. Candidate responses were generated by 11 LMMs (both cross-model and intra-model pairs), narrowed from 3,538 generated pairs to 707 selected pairs via three-stage filtering, then annotated by nine computer science Ph.D. students over 289 hours, achieving Cohen's κ of 0.718 (open-ended) and 0.805 (reasoning).

4. A comprehensive evaluation of 25 LMMs spanning proprietary models, leading open-source LMMs, and fine-tuned judge/critic models, plus additional analyses of reasoning fine-tuning, test-time scaling, and the upper bounds of criterion-level consistency.

Main Findings

• Pluralistic judging is very hard, especially open-ended. The top open-ended PAcc is 32.78% (o4-mini), followed by Claude-3.7-Sonnet at 31.77%. On verifiable reasoning, o4-mini reaches 53.17% and GPT-5 reaches 45.24%.

• No single model dominates all criteria. On reasoning, o4-mini leads on Logic (75.58%) and Efficiency (88.07%), but o3 leads on Hallucination (84.21%) and Exploration (84.62%), and Gemini-2.5-Pro leads on Grounding (79.01%). On open-ended tasks, GPT-4o reaches 76.17% on Expressiveness but 68.14% on Completeness and 65.75% on Hallucination, while o3 reaches 72.93% on Hallucination but 63.21% on Expressiveness.

• Detecting trade-offs and resolving conflicts are separate skills. GPT-4o achieves the best open-ended Trade-off Sensitivity at 84.40%, yet its Conflict Matching Rate (55.16%) is weaker than o4-mini's (65.84%).

• Open-source models lag further behind on conflict handling than on accuracy. The conflict matching rate gap between proprietary and open-source models is about 9.4 points on open-ended tasks (43.1 to 33.7) and 18.1 points on reasoning tasks (65.8 to 47.7), exceeding the 4-point and 11-point declines in criterion-level accuracy.

• Critic fine-tuning helps visual grounding but not pluralistic criteria. Qwen-based fine-tuned judges all improved on visual grounding across both splits: R1-Reward (51.7 to 60.7 open-ended, 46.9 to 58.0 reasoning), LLaVA-Critic-R1 (51.7 to 57.6, 46.9 to 53.1), and UnifiedReward (51.7 to 52.2, 46.9 to 64.2). R1-Reward was the only Qwen-based fine-tuned judge to beat the base Qwen2.5-VL-7B-Instruct on PAcc in both splits (9.41 to 17.73 open-ended; 16.67 to 19.05 reasoning), but its trade-off and conflict gains were inconsistent. LLaVA-Critic-7B (LLaVA-OV) improved on reasoning (40.7 to 43.2) but dropped marginally on open-ended evaluation (48.8 to 47.3).

• Reasoning fine-tuning weakens trade-off recognition. Three GRPO-finetuned models trained on holistic accuracy rewards showed clear reasoning gains but reduced sensitivity to trade-offs and conflict matching in reasoning judgments.

• Test-time scaling helps the strongest model, inconsistently elsewhere. With majority voting over K runs, o4-mini improved from 32.78 to 37.12 on open-ended judgments and from 53.17 to 57.94 on reasoning; GPT-4o showed similar gains, while other models showed inconsistent trends and large variance.

• Proprietary models' upper bound tracks human agreement; open-source models do not. Correlating each group's strongest criterion accuracy with human inter-annotator agreement (excluding reasoning efficiency) gives r=0.73 (p=0.024) for proprietary models and r=0.36 (p=0.344) for open-source models.

Methodology in Plain English

The researchers assembled prompts from eight sources, split between open-ended tasks (ImageInWords, DOCCI, WildVision-Bench and -Battle) and verifiable reasoning tasks (MathVerse, MM-K12, EMMA-mini, VisualPuzzles). They generated candidate responses from 11 LMMs, forming both cross-model pairs (from two different models) and intra-model pairs (from the same model sampled five times at temperature 1.0, choosing the pair with the largest MiniLM-L6 embedding distance). From 3,538 response pairs, a three-stage filter — length normalization within roughly [0.7, 1.4], correctness-matching on reasoning tasks using GPT-4o-mini, and ensemble difficulty filtering with GPT-4o, Gemini-2.5-Flash, and Claude-3.7-Sonnet — retained 707 pairs.

Each pair was then annotated against five criteria for open-ended tasks (Completeness and Coverage; Visual Grounding and Details; Factuality / No Hallucination; Creativity and Expressiveness; Clarity and Coherence) and five for reasoning (Visual Grounding; Logic Coherence and Consistency; Factuality / No Hallucination; Reflection and Exploration; Conciseness and Efficiency). Annotators judged one criterion at a time, could declare ties in under 10% of cases, and three annotators cross-validated each sample after a calibration seed set of 10 open-ended and 10 reasoning examples.

At inference, each judge model was asked to evaluate one criterion at a time for a prompt and response pair. Sampling followed original repository settings, defaulting to temperature 0.6 and top-p 0.95 when unspecified.

Why This Matters

Impact on research. The paper argues that the standard single-preference formulation of multimodal judge benchmarks cannot expose the failure modes that matter for steering model behavior — criterion-level conflicts and trade-offs. By releasing criterion-level annotations and metrics that separate detection (TOS) from resolution (CMR), it gives the field a way to measure whether judges can be steered by fine-grained instructions, not just whether they agree with an aggregate human preference. The finding that holistic critic fine-tuning mainly improves visual grounding challenges the assumption that existing reward-model pipelines transfer to pluralistic evaluation.

Real-world applications:

• Automated evaluation of multimodal assistants, where a judge must score an answer on factuality, grounding, and clarity separately rather than giving one overall verdict.
• Reinforcement learning from AI feedback and reward modeling, where criterion-level signals provide denser, more controllable reward shaping than a single preference label.
• Test-time scaling and multi-sample selection, where intra-model response pairs (42.6% of Multi-Crit's pairs) reflect the realistic setting of choosing among several samples from one model.
• Arena-style and in-the-wild evaluation (the WildVision-Bench and -Battle sources), where cross-model pairs mirror real deployment comparisons.

Industry relevance. Companies deploying LMM judges for model evaluation, alignment, or red-teaming need to know when their judge is overconfident or criterion-agnostic. The results show that open-source judges in particular fail to capture criterion-level conflicts, and that critic fine-tuning as currently practiced does not close that gap — a direct signal about where training investment is needed.

Future Directions

• Pluralistic, criterion-aware training. The authors explicitly call for training pipelines that move beyond holistic preference signals toward criterion-aware supervision, since current critic fine-tuning generalizes poorly to conflicting evaluation dimensions.

• Scaling high-quality human annotations for open-source models. The correlation analysis (r=0.36, p=0.344 for open-source) suggests open-source judges have not internalized human-consistent criteria, and the paper proposes scaling annotation as a remedy.

• Surpassing human-level evaluation alignment for proprietary models. The proprietary correlation (r=0.73, p=0.024) aligns closely with human inter-annotator agreement, raising the open question of what comes next once models match human consistency.

• Understanding why reasoning fine-tuning harms trade-off recognition. The GRPO-finetuned models gained reasoning accuracy while losing sensitivity to criterion-level conflicts — a tension the paper does not resolve.

Target Audience

Researchers and engineers working on multimodal evaluation, LMM-as-a-Judge systems, reward modeling, and preference alignment will get the most from this paper. It is also relevant to practitioners building automated evaluation pipelines who need to decide whether a single-preference judge is sufficient for their use case, and to dataset builders interested in multi-criterion human annotation protocols and inter-annotator agreement practices. Readers without background in multimodal benchmarks may find the metric definitions and results tables dense, since the paper assumes familiarity with preference-pair formulations and reward-model training.

Authors’ abstract

Large multimodal models (LMMs) are increasingly adopted as judges in multimodal evaluation systems due to their strong instruction following and consistency with human preferences. However, their ability to follow diverse, fine-grained evaluation criteria remains underexplored. We develop Multi-Crit, a benchmark for evaluating multimodal judges on their capacity to follow pluralistic criteria and produce reliable criterion-level judgments. Covering both open-ended generation and verifiable reasoning tasks, Multi-Crit is built through a rigorous data curation pipeline that gathers challenging response pairs with multi-criterion human annotations. It further introduces three novel metrics for systematically assessing pluralistic adherence, criterion-switching flexibility, and the ability to recognize criterion-level preference conflicts. Comprehensive analysis of 25 LMMs reveals that 1) proprietary models still struggle to maintain consistent adherence to pluralistic criteria--especially in open-ended evaluation; 2) open-source models lag further behind in flexibly following diverse criteria; and 3) critic fine-tuning with holistic judgment signals enhances visual grounding but fails to generalize to pluralistic criterion-level judgment. Additional analyses on reasoning fine-tuning, test-time scaling, and boundary consistency between open-source and proprietary models further probe the limits of current multimodal judges. As a pioneering study, Multi-Crit lays the foundation for building reliable and steerable multimodal AI evaluation.