MMErroR: A Benchmark for Erroneous Reasoning in Vision-Language Models

AI summary

Overview

Research area: Multimodal vision-language model evaluation, specifically process-level reasoning verification and error diagnosis.

Technical level: Intermediate. The core ideas are accessible without deep mathematics, but familiarity with vision-language models, chain-of-thought reasoning, and benchmark design helps.

Scope: This paper introduces MMErroR, a 1,997-sample benchmark that tests whether vision-language models can detect and categorize errors embedded in multimodal reasoning chains.

What This Paper Is About

Vision-language models (VLMs) can answer many image-and-text questions correctly, but it is unclear whether they genuinely understand the material or are just pattern-matching. Existing benchmarks only check whether the final answer is right; they do not ask whether a model can look at a flawed reasoning process and say what kind of mistake was made. MMErroR fills that gap by supplying reasoning chains with exactly one deliberate error and asking models to diagnose the error's type rather than simply produce an answer.

Key Contributions

1. A process-level, error-centric benchmark. MMErroR contains 1,997 curated samples across 24 subdomains and six top-level domains, each built around a reasoning chain containing exactly one verified root-cause error. This shifts evaluation from answer correctness to diagnostic reasoning.

2. A four-category error taxonomy. Errors are classified as Visual Perception Error (misreading images, charts, or symbols), Knowledge Deployment Error (applying wrong facts or formulas), Question Comprehension Error (misunderstanding what was asked), and Reasoning Error (arithmetic or logical mistakes). Each category points to a distinct weakness in the multimodal pipeline.

3. Two complementary evaluation tasks. Error Type Classification (ETC) tells the model an error exists and asks which type it is. Error Presence Detection (EPD) requires the model to first decide whether to invoke an error label, then classify it.

4. A comprehensive empirical study. Twelve representative VLMs are evaluated with domain-level breakdowns, plus auxiliary experiments on reasoning consistency, cross-modal alignment (via logit lens visualization), the effect of error awareness on answer correction, and few-shot in-context learning.

Main Findings

• Top models still fail often. The strongest model, Gemini-3-Pro-Preview, classifies the error correctly only 66.65% of the time under ETC, versus 89.52% for high-performing human experts and 76.22% for lower-performing experts. Random guessing yields 23.45%.

• Proprietary models lead, but no single model dominates every domain. Gemini-3-Pro-Preview leads in Mathematics & Logic and Chemistry & Materials, Doubao-Seed-2.0-pro in Physics & Engineering, Qwen-VL-Max in Biology & Healthcare, Kimi-K2.5 in Earth & Environment, and GPT-5.2 (xhigh) in Data & Analytics. This pattern suggests error diagnosis draws on several distinct capabilities at once.

• Error Presence Detection is harder than Error Type Classification. Every model scores lower under EPD, with Gemini-3-Pro-Preview dropping from 66.65% to 61.39%.

• Open-weight models are competitive but trailing. Kimi-K2.5 (60.19% ETC) and Qwen3-VL-32B-Thinking (59.29%) approach some proprietary systems in select domains but stay below the leader overall.

• Knowledge Deployment Error is the most common category. It accounts for 44.07% of samples (880), followed by Visual Perception Error (31.85%, 636), Question Comprehension Error (12.37%, 247), and Reasoning Error (11.72%, 234).

• Correct error diagnosis correlates with correct answering. When a model previously diagnosed an error correctly, its accuracy on the original visual question for those same samples was much higher (e.g., Gemini: 85.5% vs. 74.5%). Diagnosis ability appears tied to underlying comprehension.

• Failed detection aligns with broken visual grounding. Logit lens analysis shows that when the model correctly identifies an error, text tokens attend to the right image regions; when it fails, that alignment is disrupted.

• Telling a model the error type helps more than telling it the step. Merely adding an erroneous chain to the prompt barely changes accuracy, identifying the faulty step helps modestly, but supplying the exact error type produces the largest gains (Gemini rises from 81.0% to 90.5%).

• Few-shot prompting gives only small improvements. Across 0-, 1-, 2-, and 4-shot settings, ETC accuracy rises only a few points (Gemini: 66.5% to 68.5%).

Methodology in Plain English

The researchers started with image-question-answer triplets drawn from established benchmarks (MMMU, MathVista, MathVerse, ScienceQA, AI2D). They applied stratified sampling to balance domains and a complexity filter that favors multi-step, cross-modal questions over simple ones.

Next, they used GPT-5 to inject a single, contextually coherent error into an otherwise plausible reasoning chain, constrained to exactly one of the four error categories. Keeping the error singular was deliberate: multiple errors would make attribution ambiguous. Apart from the injected mistake, the rest of the chain had to remain locally valid, so the sample looks like a realistic failure rather than obvious nonsense.

Quality control was heavy. Twenty experts (6 professors, 14 doctoral students) reviewed an initial 10,000 samples over 23 days in three independent rounds, discarding any chain that was incoherent, mislabeled, or plausibly attributable to more than one category. Only unanimously approved samples survived, leaving 3,148 after round three. A further scoring stage had linguistics experts rate each chain on coherence, step clarity, error localizability, and semantic consistency using a -1/0/1 scale; only chains averaging above 0.5 were kept, yielding the final 1,997. A pilot study on 300 instances achieved a Cohen's Kappa of 0.796, indicating consistent annotation.

For evaluation, models answer multiple-choice prompts at temperature 0 for reproducibility, and results are reported per domain, as a macro average, and as overall weighted accuracy. Because the current release contains only erroneous chains, EPD is described as a controlled stress test of error sensitivity rather than a full calibration benchmark.

Why This Matters

The paper argues that a model's ability to spot and name its own mistakes is a better litmus test of genuine multimodal understanding than answer accuracy alone. A model that scores well on standard VQA may still be relying on shallow pattern matching; asking it to diagnose flawed reasoning probes deeper.

Real-world applications:

• Educational assistants: A tutoring system that can tell a student why a solution is wrong (misread a chart vs. applied the wrong formula) gives far more useful feedback than one that only flags an incorrect answer.
• Medical imaging analysis: Distinguishing a perceptual misreading of a scan from an incorrect application of clinical knowledge directs human reviewers to the actual source of a failure.
• Autonomous systems: Vehicles and robots operating on visual input benefit from self-checks that catch visual grounding failures before they cascade into unsafe actions.
• Scientific and engineering workflows: Systems that check reasoning about charts, formulas, and physical laws can flag errors in data interpretation pipelines and technical documentation.

Industry relevance: The results show a substantial gap between generative fluency and introspective verification. Companies deploying VLMs in high-stakes settings need reliable self-oversight, and MMErroR provides a measurable target for improving it. The finding that error-type information is far more actionable than step localization also suggests concrete design guidance for correction and repair interfaces.

Future Directions

• Multi-error and cascading failures. The current benchmark isolates one error per chain. Real reasoning failures often involve interacting mistakes, and extending the dataset to model those would be more realistic, though harder to attribute.

• Verified clean chains. The present release contains only erroneous chains, so EPD cannot measure false positives. Adding human-verified correct chains would turn EPD into a genuine calibration benchmark.

• Reducing generator bias. The erroneous chains are model-assisted, which may introduce stylistic or structural patterns specific to the generating model. Independent or multi-model generation pipelines could mitigate this.

• Scaling self-oversight. Few-shot prompting yielded only marginal gains, leaving open the question of whether training-based approaches, better prompting, or architectural changes can meaningfully close the gap to human expert diagnosis.

Target Audience

This paper is most useful for multimodal and vision-language researchers who design benchmarks or study model reliability, as well as for machine learning engineers evaluating VLMs for deployment in education, healthcare, scientific analysis, or autonomous systems. It also suits anyone interested in model interpretability, hallucination analysis, and self-verification, since it provides both a diagnostic dataset and evidence about where current models break down.

Authors’ abstract

Recent advances in Vision-Language Models (VLMs) have improved performance in multi-modal learning, raising the question of whether these models truly understand the content they process. Crucially, can VLMs detect when a reasoning process is wrong and identify its error type? To answer this, we present MMErroR, a multi-modal benchmark of 1997 samples, each embedding a single coherent reasoning error. These samples span 24 subdomains across six top-level domains, ensuring broad coverage and taxonomic richness. Unlike existing benchmarks that focus on answer correctness, MMErroR targets a process-level, error-centric evaluation that requires models to detect incorrect reasoning and classify the error type within both visual and linguistic contexts. We evaluate 12 representative VLMs, and even the best model, Gemini-3-Pro-Preview, classifies the error correctly in only 66.65\% of cases, underscoring the challenge of identifying erroneous reasoning. Furthermore, the ability to accurately identify errors offers valuable insights into the capabilities of multi-modal models. Project Page: https://mmerror-benchmark.github.io