Multimodal Model Evaluation

Recent progress in Multimodal Large Language Models (MLLMs) has demonstrated remarkable advances in perception and reasoning, suggesting their potential for embodied intelligence. While recent studies have evaluated embodied MLLMs in interactive settings, current benchmarks mainly target capabilities to perceive, understand, and interact with external objects, lacking a systematic evaluation of self-centric intelligence. To address this, we introduce MirrorBench, a simulation-based benchmark inspired by the classical Mirror Self-Recognition (MSR) test in psychology. MirrorBench extends this paradigm to embodied MLLMs through a tiered framework of progressively challenging tasks, assessing agents from basic visual perception to high-level self-representation. Experiments on leading MLLMs show that even at the lowest level, their performance remains substantially inferior to human performance, revealing fundamental limitations in self-referential understanding. Our study bridges psychological paradigms and embodied intelligence, offering a principled framework for evaluating the emergence of general intelligence in large models. Project page: https://fflahm.github.io/mirror-bench-page/.

Most medical multimodal benchmarks focus on static tasks such as image question answering, report generation, and plain-language rewriting. Patient education is more demanding: systems must identify relevant evidence across images, show patients where to look, explain findings in accessible language, and handle confusion or distress. Yet most patient education work remains text-only, even though combined image-and-text explanations may better support understanding. We introduce MedImageEdu, a benchmark for multi-turn, evidence-grounded radiology patient education. Each case provides a radiology report with report text and case images. A DoctorAgent interacts with a PatientAgent, conditioned on a hidden profile that captures factors such as education level, health literacy, and personality. When a patient question would benefit from visual support, the DoctorAgent can issue drawing instructions grounded in the report, case images, and the current question to a benchmark-provided drawing tool. The tool returns image(s), after which the DoctorAgent produces a final multimodal response consisting of the image(s) and a grounded plain-language explanation. MedImageEdu contains 150 cases from three sources and evaluates both the consultation process and the final multimodal response along five dimensions: Consultation, Safety and Scope, Language Quality, Drawing Quality, and Image-Text Response Quality. Across representative open- and closed-source vision-language model agents, we find three consistent gaps: fluent language often outpaces faithful visual grounding, safety is the weakest dimension across disease categories, and emotionally tense interactions are harder than low education or low health literacy. MedImageEdu provides a controlled testbed for assessing whether multimodal agents can teach from evidence rather than merely answer from text.

Multimodal language models systematically underperform on visual perception tasks, yet the structure underlying this failure remains poorly understood. We propose centroid replacement, mapping tokens to their nearest K-means centroid and removing within-cluster residual structure, as a controlled probe for modal dependence. Across seven models spanning four architecture families, post-image text replacement costs 4x more accuracy than visual replacement on six BLINK tasks; VPBench and MedBLINK preserve this aggregate ordering. Since text replacement also disrupts the task and answer interface, the gap measures relative dependence. Text centroid contrastive decoding (TCCD) exploits this asymmetry without retraining, recovering up to +16.9% accuracy on an individual task at oracle best-per-task $α_\text{interp}$; every model gains on at least one task, although gains vary across tasks and models. Together, centroid replacement provides a retraining-free audit of modal dependence on labeled data, while TCCD offers a reusable inference-time intervention. Code and artifacts are available at: https://github.com/yahskapar/centroid-erasure

The potential of Multimodal Large Language Models (MLLMs) in domain of medical imaging raise the demands of systematic and rigorous evaluation frameworks that are aligned with the real-world medical imaging practice. Existing practices that report single or coarse-grained metrics are lack the granularity required for specialized clinical support and fail to assess the reliability of reasoning mechanisms. To address this, we propose a paradigm shift toward multidimensional, fine-grained and in-depth evaluation. Based on a two-stage systematic construction pipeline designed for this paradigm, we instantiate it with MedRCube. We benchmark 33 MLLMs, Lingshu-32B achieve top-tier performance. Crucially, MedRCube exposes a series of pronounced insights inaccessible under prior evaluation settings. Furthermore, we introduce a credibility evaluation subset to quantify reasoning credibility, uncover a highly significant positive association between shortcut behavior and diagnostic task performance, raising concerns for clinically trustworthy deployment. The resources of this work can be found at https://github.com/F1mc/MedRCube.

Manual digitisation of structured handwritten documents is slow and costly. We benchmark 17 leading frontier multi-modal large language models and open-source models against a very challenging real-world medical form that mixes dates; structured, printed text; hand-written responses and significant variability challenges. None of the smaller or older models perform well but the latest Google and OpenAI models reach accuracies around $85\%$ with weighted F1 scores $\simeq 90\%$ across the discrete or predefined fields despite the very challenging nature of the responses. Clear task specific strengths emerge: GPT 5.4 excels in noisy date extraction as well as reliability with the lowest hallucination rate ($6\%$). Claude Sonnet 4.6 had the best average performance across formatted fields (dates and numerical values), while Gemini 3.1 delivered the best overall performance, with the lowest free text error rates (WER = $0.50$ and CER = $0.31$) and the strongest results across discrete classification metrics. We further show that prompt optimisation dramatically improves macro precision, recall and F1 by over $60\%$, but has little impact on weighted metrics (only $\sim2-5\%$ improvement). These results provide evidence that the rapid improvements of multimodal large language models offer a compelling pathway toward fully automated digitisation of complex handwritten workflows that is particularly relevant in low- and middle-income countries.

Understanding human intent in complex multi-turn interactions remains a fundamental challenge in human-computer interaction and behavioral analysis. While existing intent recognition datasets focus mainly on single utterances or simple dialogues, real-world scenarios often involve sophisticated strategic interactions where participants must maintain complex deceptive narratives over extended periods. To address this gap, we introduce MISID, a comprehensive multimodal, multi-turn, and multi-participant benchmark for intent recognition. Sourced from high-stakes social strategy games, MISID features a fine-grained, two-tier multi-dimensional annotation scheme tailored for long-context discourse analysis and evidence-based causal tracking. Our systematic evaluation of state-of-the-art Multimodal Large Language Models (MLLMs) on MISID reveals critical deficiencies in complex scenarios, including text-prior visual hallucination, impaired cross-modal synergy, and limited capacity in chaining causal cues. Consequently, we propose FRACTAM as a baseline framework. Using a ``Decouple-Anchor-Reason'' paradigm, FRACTAM reduces text bias by extracting pure unimodal factual representations, employs two-stage retrieval for long-range factual anchoring, and constructs explicit cross-modal evidence chains. Extensive experiments demonstrate that FRACTAM enhances mainstream models' performance in complex strategic tasks, improving hidden intent detection and inference while maintaining robust perceptual accuracy. Our dataset is available at https://naislab.cn/datasets/MISID.

Vision-Language Models (VLMs) expand the attack surface of safety-aligned systems by coupling visual perception with text generation. Existing multimodal jailbreak attacks primarily rely on crafted visual content, adversarial perturbations, or image-specific attack strategies, leaving the potential of reusable visual anchors in benign natural images largely unexplored. To address the problem, we introduce MemJack, a memory-augmented multi-agent framework for automated VLM red-teaming with natural images. MemJack combines visual-anchor discovery, visual-semantic camouflage, response evaluation, reflection-guided repair, and dynamic replanning within a closed-loop attack pipeline. Beyond attack generation, MemJack automatically transforms public images into attack anchors, enables the construction of MemJack-Bench, a dataset of over 113,000 interactive multimodal jailbreak trajectories for safety evaluation and defensive alignment. Extensive empirical evaluations across full, unmodified COCO val2017 images demonstrate that MemJack achieves a 71.48% attack success rate (ASR) against Qwen3-VL-Plus, scaling to 90% under extended budgets. Compared with representative multimodal jailbreak baselines under the same natural-image evaluation setting, MemJack achieves the highest ASR while requiring fewer mean rounds to success, demonstrating superior jailbreak effectiveness on VLMs. These results demonstrate that benign natural images can act as transferable jailbreak anchors and reveal substantial vulnerabilities in current safety-aligned VLMs.

Automatic evaluation using multimodal large language models (MLLMs), commonly referred to as MLLM-as-a-Judge, has been widely used to measure model performance. If such MLLM-as-a-Judge methods were biased, they could distort model comparisons and benchmark-driven scientific progress. However, it remains unclear to what extent MLLM-as-a-Judge methods favor or disfavor text generated by specific MLLMs. In this study, we propose Philautia-Eval to investigate such model-specific preference bias. Philautia-Eval quantifies the degree of the bias by disentangling preference tendencies from differences in generation quality. Using 1.29M caption-score pairs collected from 12 MLLMs, we found that representative MLLMs tend to exhibit self-preference bias. Moreover, experimental results indicate mutual preference bias within particular model families, which is potentially driven by reused connectors and overlapping instruction-tuning resources. Finally, we introduce a simple ensemble of MLLMs, Pomms. Our results demonstrated that Pomms effectively mitigated the model-specific preference bias while maintaining performance.

Leveraging Multi-modal Large Language Models (MLLMs) to accelerate frontier scientific research is promising, yet how to rigorously evaluate such systems remains unclear. Existing benchmarks mainly focus on single-document understanding, whereas real scientific workflows require integrating evidence from multiple papers, including their text, tables, and figures. As a result, multi-modal, multi-document scientific reasoning remains underexplored and lacks systematic evaluation. To address this gap, we introduce PaperScope, a multi-modal multi-document benchmark designed for agentic deep research. PaperScope presents three advantages: (1) Structured scientific grounding. It is built on a knowledge graph of over 2,000 AI papers spanning three years, providing a structured foundation for research-oriented queries. (2) Semantically dense evidence construction. It integrates semantically related key information nodes and employs optimized random-walk article selector to sample thematically coherent paper sets, thereby ensuring adequate semantic density and task complexity. (3) Multi-task evaluation of scientific reasoning. It contains over 2,000 QA pairs across reasoning, retrieval, summarization, and problem solving, enabling evaluation of multi-step scientific reasoning. Experimental results show that even advanced systems such as OpenAI Deep Research and Tongyi Deep Research achieve limited scores on PaperScope, highlighting the difficulty of long-context retrieval and deep multi-source reasoning. PaperScope thus provides a rigorous benchmark alongside a scalable pipeline for constructing large-scale multi-modal, multi-source deep research datasets.

In recent years, rapid advances in Multimodal Large Language Models (MLLMs) have increasingly stimulated research on ancient Chinese scripts. As the evolution of written characters constitutes a fundamental pathway for understanding cultural transformation and historical continuity, how MLLMs can be systematically leveraged to support and advance text evolution analysis remains an open and largely underexplored problem. To bridge this gap, we construct a comprehensive benchmark comprising 11 tasks and over 130,000 instances, specifically designed to evaluate the capability of MLLMs in analyzing the evolution of ancient Chinese scripts. We conduct extensive evaluations across multiple widely used MLLMs and observe that, while existing models demonstrate a limited ability in glyph-level comparison, their performance on core tasks-such as character recognition and evolutionary reasoning-remains substantially constrained. Motivated by these findings, we propose a glyph-driven fine-tuning framework (GEVO) that explicitly encourages models to capture evolutionary consistency in glyph transformations and enhances their understanding of text evolution. Experimental results show that even models at the 2B scale achieve consistent and comprehensive performance improvements across all evaluated tasks. To facilitate future research, we publicly release both the benchmark and the trained models\footnote{https://github.com/songruiecho/GEVO}.

Recent advances in unified multimodal models (UMMs) have led to a proliferation of architectures capable of understanding, generating, and editing across visual and textual modalities. However, developing a unified framework for UMMs remains challenging due to the diversity of model architectures and the heterogeneity of training paradigms and implementation details. In this paper, we present TorchUMM, the first unified codebase for comprehensive evaluation, analysis, and post-training across diverse UMM backbones, tasks, and datasets. TorchUMM supports a broad spectrum of models covering a wide range of scales and design paradigms. Our benchmark encompasses three core task dimensions: multimodal understanding, generation, and editing, and integrates both established and novel datasets to evaluate perception, reasoning, compositionality, and instruction-following abilities. By providing a unified interface and standardized evaluation protocols, TorchUMM enables fair and reproducible comparisons across heterogeneous models and fosters deeper insights into their strengths and limitations, facilitating the development of more capable unified multimodal systems. Code is available at: https://github.com/AIFrontierLab/TorchUMM.

Multimodal large language models (MLLMs) have advanced clinical tasks for common conditions, but their performance on rare diseases remains largely untested. In rare-disease scenarios, clinicians often lack prior clinical knowledge, forcing them to rely strictly on case-level evidence for clinical judgments. Existing benchmarks predominantly evaluate common-condition, single-image settings, leaving multimodal and multi-image evidence integration under rare-disease data scarcity systematically unevaluated. We introduce MMRareBench, to our knowledge the first rare-disease benchmark jointly evaluating multimodal and multi-image clinical capability across four workflow-aligned tracks: diagnosis, treatment planning, cross-image evidence alignment, and examination suggestion. The benchmark comprises 1,756 question-answer pairs with 7,958 associated medical images curated from PMC case reports, with Orphanet-anchored ontology alignment, track-specific leakage control, evidence-grounded annotations, and a two-level evaluation protocol. A systematic evaluation of 23 MLLMs reveals fragmented capability profiles and universally low treatment-planning performance, with medical-domain models trailing general-purpose MLLMs substantially on multi-image tracks despite competitive diagnostic scores. These patterns are consistent with a capacity dilution effect: medical fine-tuning can narrow the diagnostic gap but may erode the compositional multi-image capability that rare-disease evidence integration demands.

As Vision-Language Models (VLMs) become integral to educational decision-making, ensuring their fairness is paramount. However, current text-centric evaluations neglect the visual modality, leaving an unregulated channel for latent social biases. To bridge this gap, we present Edu-MMBias, a systematic auditing framework grounded in the tri-component model of attitudes from social psychology. This framework diagnoses bias across three hierarchical dimensions: cognitive, affective, and behavioral. Utilizing a specialized generative pipeline that incorporates a self-correct mechanism and human-in-the-loop verification, we synthesize contamination-resistant student profiles to conduct a holistic stress test on state-of-the-art VLMs. Our extensive audit reveals critical, counter-intuitive patterns: models exhibit a compensatory class bias favoring lower-status narratives while simultaneously harboring deep-seated health and racial stereotypes. Crucially, we find that visual inputs act as a safety backdoor, triggering a resurgence of biases that bypass text-based alignment safeguards and revealing a systematic misalignment between latent cognition and final decision-making. The contributions of this paper are available at: https://anonymous.4open.science/r/EduMMBias-63B2.

This paper provides a comprehensive evaluation of demographic and linguistic biases in omnimodal language models that process text, images, audio, and video within a single framework. Although these models are being widely deployed, their performance across different demographic groups and modalities is not well studied. Four omnimodal models are evaluated on tasks that include demographic attribute estimation, identity verification, activity recognition, multilingual speech transcription, and language identification. Accuracy differences are measured across age, gender, skin tone, language, and country of origin. The results show that image and video understanding tasks generally exhibit better performance with smaller demographic disparities. In contrast, audio understanding tasks exhibit significantly lower performance and substantial bias, including large accuracy differences across age groups, genders, and languages, and frequent prediction collapse toward narrow categories. These findings highlight the importance of evaluating fairness across all supported modalities as omnimodal language models are increasingly used in real-world applications.

High-throughput, effective and comprehensive phenotyping is a critical prerequisite for crop improvement. While phenotypic analysis was traditionally performed manually, recent advances in multimodal foundation models, especially in vision-language models, have offered new opportunities for automated plant phenotyping. However, plant science remains particularly challenging for general-purpose VLMs because it requires domain-specific knowledge, fine-grained visual interpretation, and complex agronomic reasoning. We introduce PlantXpert, an evidence-grounded multimodal reasoning benchmark for plant phenotyping across four major crops: soybean, cotton, rice, and corn. Our benchmark provides a structured and reproducible framework for agronomic adaptation of VLMs, and enables controlled comparison between base models and their domain-adapted counterparts. We constructed a dataset comprising 545 digital images and 4,351 benchmark samples spanning key agronomic domains including disease, pest control, weed management and yield. The benchmark can assess diverse capabilities including visual expertise, quantitative reasoning, and multi-step agronomic reasoning. We trained and evaluated a total of 11 state-of-the-art VLMs. Fine-tuning on our dataset improves overall accuracy for all models, with many achieving around 70% accuracy, and improvements on 9 of 11 remaining statistically significant after Holm correction. We also identify diminishing returns from model scaling, uneven performance across crop subsets, and persistent weaknesses in quantitative reasoning and specialized image analysis. These limitations show that general multimodal competence alone is insufficient for reliable agronomic reasoning and demonstrate the value of domain-specific adaptation. We hope PlantXpert can serve as a foundation for advancing evidence-grounded multimodal reasoning in plant science.

Current video benchmarks for multimodal large language models (MLLMs) focus on event recognition, temporal ordering, and long-context recall, but overlook a harder capability required for expert procedural judgment: tracking how ongoing interactions update the procedural state and thereby determine the correctness of later actions. We introduce SiMing-Bench, the first benchmark for evaluating this capability from full-length clinical skill videos. It targets rubric-grounded process-level judgment of whether interaction-driven state updates preserve procedural correctness across an entire workflow. SiMing-Bench is instantiated with SiMing-Score, a physician-annotated dataset of real clinical skill examination videos spanning cardiopulmonary resuscitation, automated external defibrillator operation, and bag-mask ventilation, each paired with a standardized step-wise rubric and dual-expert labels. Across diverse open- and closed-source MLLMs, we observe consistently weak agreement with physician judgments. Moreover, weak performance on rubric-defined intermediate steps persists even when overall procedure-level correlation appears acceptable, suggesting that coarse global assessment substantially overestimates current models' procedural judgment ability. Additional analyses with binary step judgment and step-aligned clips indicate that the bottleneck is not merely fine-grained scoring or temporal localization, but modeling how continuous interactions update procedural state over time.

We introduce Unified Multimodal Uncertain Inference (UMUI), a multimodal inference task spanning text, audio, and video, where models must produce calibrated probability estimates of hypotheses conditioned on a premise in any modality or combination. While uncertain inference has been explored in text, extension to other modalities has been limited to single-modality binary entailment judgments, leaving no framework for fine-grained probabilistic reasoning in or across other modalities. To address this, we curate a human-annotated evaluation set with scalar probability judgments across audio, visual, and audiovisual settings, and additionally evaluate on existing text and audio benchmarks. We introduce CLUE (Calibrated Latent Uncertainty Estimation), which combines self-consistent teacher calibration and distribution-based confidence probing to produce calibrated predictions. We demonstrate that our 3B-parameter model achieves equivalent or stronger performance than zero-shot baselines up to 32B parameters across all modalities.

The rise of multimodal large language models (MLLMs) has sparked an unprecedented wave of applications in the field of medical imaging analysis. However, as one of the earliest and most fundamental tasks integrated into this paradigm, medical image classification reveals a sobering reality: state-of-the-art medical MLLMs consistently underperform compared to traditional deep learning models, despite their overwhelming advantages in pre-training data and model parameters. This paradox prompts a critical rethinking: where exactly does the performance degradation originate? In this paper, we conduct extensive experiments on 14 open-source medical MLLMs across three representative image classification datasets. Moving beyond superficial performance benchmarking, we employ feature probing to track the information flow of visual features module-by-module and layer-by-layer throughout the entire MLLM pipeline, enabling explicit visualization of where and how classification signals are distorted, diluted, or overridden. As the first attempt to dissect classification performance degradation in medical MLLMs, our findings reveal four failure modes: 1) quality limitation in visual representation, 2) fidelity loss in connector projection, 3) comprehension deficit in LLM reasoning, and 4) misalignment of semantic mapping. Meanwhile, we introduce quantitative scores that characterize the healthiness of feature evolution, enabling principled comparisons across diverse MLLMs and datasets. Furthermore, we provide insightful discussions centered on the critical barriers that prevent current medical MLLMs from fulfilling their promised clinical potential. We hope that our work provokes rethinking within the community-highlighting that the road from high expectations to clinically deployable MLLMs remains long and winding.

The manufacturing sector is increasingly adopting Multimodal Large Language Models (MLLMs) to transition from simple perception to autonomous execution, yet current evaluations fail to reflect the rigorous demands of real-world manufacturing environments. Progress is hindered by data scarcity and a lack of fine-grained domain semantics in existing datasets. To bridge this gap, we introduce FORGE. Wefirst construct a high-quality multimodal dataset that combines real-world 2D images and 3D point clouds, annotated with fine-grained domain semantics (e.g., exact model numbers). We then evaluate 18 state-of-the-art MLLMs across three manufacturing tasks, namely workpiece verification, structural surface inspection, and assembly verification, revealing significant performance gaps. Counter to conventional understanding, the bottleneck analysis shows that visual grounding is not the primary limiting factor. Instead, insufficient domain-specific knowledge is the key bottleneck, setting a clear direction for future research. Beyond evaluation, we show that our structured annotations can serve as an actionable training resource: supervised fine-tuning of a compact 3B-parameter model on our data yields up to 90.8% relative improvement in accuracy on held-out manufacturing scenarios, providing preliminary evidence for a practical pathway toward domain-adapted manufacturing MLLMs. The code and datasets are available at https://ai4manufacturing.github.io/forge-web.

Existing Graphical User Interface (GUI) reasoning tasks remain challenging, particularly in UI understanding. Current methods typically rely on direct screen-based decision-making, which lacks interpretability and overlooks a comprehensive understanding of UI elements, ultimately leading to task failure. To enhance the understanding and interaction with UIs, we propose an innovative GUI reasoning paradigm called UI-in-the-Loop (UILoop). Our approach treats the GUI reasoning task as a cyclic Screen-UI elements-Action process. By enabling Multimodal Large Language Models (MLLMs) to explicitly learn the localization, semantic functions, and practical usage of key UI elements, UILoop achieves precise element discovery and performs interpretable reasoning. Furthermore, we introduce a more challenging UI Comprehension task centered on UI elements with three evaluation metrics. Correspondingly, we contribute a benchmark of 26K samples (UI Comprehension-Bench) to comprehensively evaluate existing methods' mastery of UI elements. Extensive experiments demonstrate that UILoop achieves state-of-the-art UI understanding performance while yielding superior results in GUI reasoning tasks.

Puns are a common form of rhetorical wordplay that exploits polysemy and phonetic similarity to create humor. In multimodal puns, visual and textual elements synergize to ground the literal sense and evoke the figurative meaning simultaneously. Although Vision-Language Models (VLMs) are widely used in multimodal understanding and generation, their ability to understand puns has not been systematically studied due to a scarcity of rigorous benchmarks. To address this, we first propose a multimodal pun generation pipeline. We then introduce MultiPun, a dataset comprising diverse types of puns alongside adversarial non-pun distractors. Our evaluation reveals that most models struggle to distinguish genuine puns from these distractors. Moreover, we propose both prompt-level and model-level strategies to enhance pun comprehension, with an average improvement of 16.5% in F1 scores. Our findings provide valuable insights for developing future VLMs that master the subtleties of human-like humor via cross-modal reasoning.

In science education, students frequently construct hand-drawn visual models of scientific phenomena. These drawings rely on a visual structure where information is encoded through visual objects, their attributes, and relationships. Multimodal large language models (MLLMs) are increasingly used to generate feedback on students' hand-drawn scientific models. However, the validity of such feedback depends on whether model claims are grounded in the specific visual evidence of the student drawing. This study uncovers grounding failures, consistent with modal decoupling, in off-the-shelf MLLM feedback, where outputs remain pedagogically plausible in form while contradicting the drawing or treating depicted elements as missing. Using N = 150 middle school drawings from a kinetic molecular theory unit spanning five modeling tasks and three competence levels, we generated N = 300 feedback instances with GPT-5.1. All outputs were coded for four grounding error types: object mismatch, attribute mismatch, relation mismatch, and false absence. Grounding failures were common: 41.3% of feedback instances contained at least one error. An inventory-list-first workflow reduced several error categories and lowered the overall error rate, but it did not resolve the underlying limitation: approximately one in three outputs remained flawed, with false absence as the dominant failure mode. Moreover, feedback that appears visually grounded offered little diagnostic value for identifying invalid instances. The findings indicate that modal decoupling is a substantial limitation and that valid feedback will require grounding mechanisms beyond common prompting strategies.

Current multimodal benchmarks for scientific reasoning primarily evaluate local information extraction -- models recognize symbols and values and then perform textual inference. They do not assess whether models can reason over the global structural properties of formal diagrams, such as topology, conservation constraints, and the consistent mapping between visual patterns and algebraic expressions. We introduce FeynmanBench, a benchmark of over 2,000 tasks centered on Feynman diagrams spanning the electromagnetic, weak, and strong interactions of the Standard Model. Each instance couples a diagram image with minimal textual conventions and requires models to recover the full physical content -- vertex inventory, propagator types, topological connectivity, momentum routing, and the complete scattering amplitude. An automated generation and verification pipeline produces the diagrams, annotations, and reference answers under standardized rules. Evaluating 19 state-of-the-art multimodal LLMs, we find a consistent failure pattern: models achieve 70--95% on local recognition (vertex and propagator identification) but collapse to 13--17% on topological reconstruction (CP3), and near zero on full algebraic derivation (CP5). FeynmanBench offers a controlled testbed for multimodal reasoning over formal scientific diagrams and highlights fundamental limitations of current architectures in topology-sensitive scientific reasoning.

Recent advances in multimodal large language models (MLLMs) have greatly improved image understanding and captioning capabilities. However, existing image captioning benchmarks typically suffer from limited diversity in caption length, the absence of recent advanced MLLMs, and insufficient human annotations, which potentially introduces bias and limits the ability to comprehensively assess the performance of modern MLLMs. To address these limitations, we present a new large-scale image captioning benchmark, termed, ICBench, which covers 12 content categories and consists of both short and long captions generated by 10 advanced MLLMs on 2K images, resulting in 40K captions in total. We conduct extensive human subjective studies to obtain mean opinion scores (MOSs) across fine-grained evaluation dimensions, where short captions are assessed in terms of fluency, relevance, and conciseness, while long captions are evaluated based on fluency, relevance, and completeness. Furthermore, we propose an automated evaluation metric, ITIScore, based on an image-to-text-to-image framework, which measures caption quality through reconstruction consistency. Experimental results demonstrate strong alignment between our automatic metric and human judgments, as well as robust zero-shot generalization ability on other public captioning datasets. Both the dataset and model will be released upon publication.

The recent success of reinforcement learning (RL) in large reasoning models has inspired the growing adoption of RL for post-training Multimodal Large Language Models (MLLMs) to enhance their visual reasoning capabilities. Although many studies have reported improved performance, it remains unclear whether RL training truly enables models to learn from visual information. In this work, we propose the Hallucination-as-Cue Framework, an analytical framework designed to investigate the effects of RL-based post-training on multimodal reasoning models from the perspective of model hallucination. Specifically, we introduce hallucination-inductive, modality-specific corruptions that remove or replace essential information required to derive correct answers, thereby forcing the model to reason by hallucination. By applying these corruptions during both training and evaluation, our framework provides a unique perspective for diagnosing RL training dynamics and understanding the intrinsic properties of datasets. Through extensive experiments and analyses across multiple multimodal reasoning benchmarks, we reveal that the role of model hallucination for RL-training is more significant than previously recognized. For instance, we find that RL post-training under purely hallucination-inductive settings can still significantly improve models' reasoning performance, and in some cases even outperform standard training. These findings challenge prevailing assumptions about MLLM reasoning training and motivate the development of more modality-aware RL-based training designs.

Multimodal Large Language Models (MLLMs) excel in general domains but struggle with complex, real-world science. We posit that polymer science, an interdisciplinary field spanning chemistry, physics, biology, and engineering, is an ideal high-stakes testbed due to its diverse multimodal data. Yet, existing benchmarks related to polymer science largely overlook real-world workflows, limiting their practical utility and failing to systematically evaluate MLLMs across the full, practice-grounded lifecycle of experimentation. We introduce PolyReal, a novel multimodal benchmark grounded in real-world scientific practices to evaluate MLLMs on the full lifecycle of polymer experimentation. It covers five critical capabilities: (1) foundational knowledge application; (2) lab safety analysis; (3) experiment mechanism reasoning; (4) raw data extraction; and (5) performance & application exploration. Our evaluation of leading MLLMs on PolyReal reveals a capability imbalance. While models perform well on knowledge-intensive reasoning (e.g., Experiment Mechanism Reasoning), they drop sharply on practice-based tasks (e.g., Lab Safety Analysis and Raw Data Extraction). This exposes a severe gap between abstract scientific knowledge and its practical, context-dependent application, showing that these real-world tasks remain challenging for MLLMs. Thus, PolyReal helps address this evaluation gap and provides a practical benchmark for assessing AI systems in real-world scientific workflows.

Current benchmarks for graphical user interface (GUI) agents predominantly rely on static screenshots. However, real-world smartphone interaction routinely requires agents to process transient audio cues and temporal video dynamics that are tightly coupled with the moment of action. To bridge this gap, we introduce OmniGUI, the first step-level benchmark designed to evaluate GUI agents in omni-modal smartphone environments. OmniGUI provides continuous, interleaved multimodal inputs comprising static images, synchronous audio, and video clips at every action step. The dataset encompasses 709 expert-demonstrated episodes (2,579 action steps) across 29 applications, systematically annotated with objective multimodal dependency levels. Because dedicated omni-modal GUI agent frameworks are currently in their nascent stage, we select foundational omni-modal models capable of natively processing interleaved inputs to serve as agent proxies for our initial baselines. Our empirical evaluation reveals that while current models exhibit competency on visually static tasks, their action prediction performance degrades significantly in environments requiring synchronous temporal and auditory signals. Furthermore, ablation studies isolate specific operational bottlenecks, notably cross-modal interference when processing task-irrelevant environmental noise. The complete dataset, evaluation pipeline, and baseline prompts are provided in the supplementary material. Project page: https://omni-gui.github.io.

Real-world reasoning often requires combining information across modalities, connecting textual context with visual cues in a multi-hop process. Yet, most multimodal benchmarks fail to capture this ability: they typically rely on single images or set of images, where answers can be inferred from a single modality alone. This limitation is mirrored in the training data, where interleaved image-text content rarely enforces complementary, multi-hop reasoning. As a result, Vision-Language Models (VLMs) frequently hallucinate and produce reasoning traces poorly grounded in visual evidence. To address this gap, we introduce CRIT, a new dataset and benchmark built with a graph-based automatic pipeline for generating complex cross-modal reasoning tasks. CRIT consists of diverse domains ranging from natural images, videos, and text-rich sources, and includes a manually verified test set for reliable evaluation. Experiments on this benchmark reveal that even state-of-the-art models struggle on such reasoning tasks. Models trained on CRIT show significant gains in cross-modal multi-hop reasoning, including strong improvements on SPIQA and other standard multimodal benchmarks.

Evaluating scientific arguments requires assessing the strict consistency between a claim and its underlying multimodal evidence. However, existing benchmarks lack the scale, domain diversity, and visual complexity needed to evaluate this alignment realistically. To address this gap, we introduce M2-Verify, a large-scale multimodal dataset for checking scientific claim consistency. Sourced from PubMed and arXiv, M2-Verify provides over 469K instances across 16 domains, rigorously validated through expert audits. Extensive baseline experiments show that state-of-the-art models struggle to maintain robust consistency. While top models achieve up to 85.8% Micro-F1 on low-complexity medical perturbations, performance drops to 61.6% on high-complexity challenges like anatomical shifts. Furthermore, expert evaluations expose hallucinations when models generate scientific explanations for their alignment decisions. Finally, we demonstrate our dataset's utility and provide comprehensive usage guidelines.

Spatial consistency is a fundamental property of the visual world and a key requirement for models that aim to understand physical reality. Despite recent advances, multimodal large language models (MLLMs) often struggle to reason about 3D geometry across multiple views. Rather than asking models to describe scene attributes, we introduce a more challenging task: given two views of the same scene, identify the object that violates 3D motion consistency. We propose a simple and scalable method for generating realistic, spatially inconsistent image pairs from multi-view scenes, enabling systematic evaluation of this capability. Our results show that state-of-the-art MLLMs significantly underperform human observers and exhibit substantial variability across different scene attributes, revealing a fragile and incomplete understanding of 3D structure. We hope our findings underscore the need for approaches that develop a more deeply grounded understanding of the physical world.