[Research Briefing] The Evolution of Multimodal AI Evaluation: From Recognition to Reasoning

Introduction

When deploying AI in enterprise environments, the most critical risk point is the gap between a model’s ‘benchmark score’ and its ‘actual field performance.’ Many companies decide on adoption based on SOTA (State-of-the-Art) leaderboard rankings, yet encounter reliability issues in actual workflows due to unexpected hallucinations or logical errors.

This phenomenon stems from ‘evaluation lag,’ where AI evaluation methodologies fail to keep pace with model training speeds. Models are in a state of data contamination, having memorized benchmark questions and answers while training on vast web data; this leads to the development of ‘test-taking skills’ rather than genuine intelligence.

Since multimodal AI must simultaneously process text, images, video, and audio, the complexity of evaluation increases further. To be integrated into actual business processes, a model must move beyond simply identifying ‘what is in the image’ to reasoning about the causal relationships of ‘why this situation occurred.’

Therefore, executives and engineering leaders must establish a verification system that validates whether the ‘reasoning process’ the model uses to reach an answer is valid, rather than relying on simple accuracy. This research provides a standard for enterprises to objectively measure AI performance by presenting an evolutionary path of evaluation from recognition to reasoning, and finally to execution.

Core Analysis

This research defines the history of multimodal AI evaluation not as a mere list of datasets, but as a process of increasingly rigorous ‘cognitive examinations.’ It analyzes the progression from the ‘Recognition Era (Level I),’ centered on early ImageNet and COCO, through the ‘Dawn of Reasoning (Level II)’ measuring logic and comprehension via VQA, GQA, and VCR, to the current stage of ‘Expert-Level Integrated Evaluation (Level III)’ for MLLMs.

The researchers categorize evaluation into four levels. Level III measures high-level domain knowledge and cross-modal reasoning through benchmarks such as MMMU, MMBench, and SEED-Bench. Specifically, it presents the performance of latest models, noting Gemini 1.5 Pro’s 81.6% accuracy on Video-MME (including captions) and GPT-4o’s approximately 84.8% accuracy on MMBench.

Furthermore, it explores Level IV—still an uncharted territory—where simulators such as VirtualHome, ALFRED, and EmbodiedBench are used to evaluate grounded planning, social intelligence, and creativity. In this stage, behavior-based metrics such as Success Rate (SR) and Success weighted by Path Length (SPL) become more critical than binary correct/incorrect answers.

Finally, the paper points out the problem of ‘data contamination,’ where static benchmarks are included in model training data, distorting performance. To resolve this, it concludes by proposing a transition toward ‘Living Benchmarks,’ where humans intervene to continuously generate new cases, and adversarial evaluation systems.

Deep Dive into Key Failure Factors

First is the issue of Shortcut Learning and spurious correlations. Instead of learning the intrinsic features of an object, models tend to use statistical biases present in the dataset to find the correct answer. For example, a model recognizing a cow might judge it as such based on the ‘green grass’ in the background rather than the cow itself. This causes vulnerability where performance drops sharply when the background changes, despite high benchmark scores.

Second is the Binding Problem and the failure of compositional generalization. This is a phenomenon where a model recognizes individual objects (red, cube, blue, sphere) but struggles to combine them into a relational structure like ‘a red cube and a blue sphere.’ The fact that latest models show performance near random guess levels on benchmarks like Winoground demonstrates that models fail to accurately map linguistic structures to visual constraints.

Third is the saturation of static evaluation and Data Contamination. Latest MLLMs trained at web-scale are highly likely to have encountered evaluation datasets during the training process. Consequently, models output memorized content rather than deriving answers through reasoning, resulting in a significant drop in the ability to handle new situations in actual deployment environments.

Source: Photo by Markus Spiske on Unsplash

Multimodal Measurement Framework: Analysis of 6 Orthogonal Axes

To precisely measure the performance of multimodal AI, a measurement framework based on six orthogonal axes—which act independently—is required, moving beyond simple accuracy metrics. This framework analyzes benchmark quality and actual model capability through six dimensions: Skill, Reliability, Robustness, Hygiene, Cost, and Fairness. Here, Skill refers to the cognitive abilities being measured, from pattern matching to complex reasoning; Reliability measures the statistical consistency of metrics such as Top-1 Accuracy or Elo Rating. Robustness evaluates the presence of mechanisms to prevent Shortcut Learning, and Hygiene assesses the level of data contamination prevention and validity assurance. Additionally, Cost measures the computational resources required from static inference to simulation, and Fairness measures bias detection and representation capabilities across diverse groups.

These six axes form distinct trade-off relationships as the cognitive stage (Level I~IV) increases. Most notably, as the depth of Skill increases, Cost rises exponentially. In the Level I recognition stage, evaluation is possible through simple discriminative classification, optimizing cost-efficiency. However, upon entering the Level IV grounded intelligence stage, agent behavior must be measured in simulation environments containing physics engines, making computational costs extremely high. In other words, attempts to measure higher-dimensional intelligence inevitably require massive computational resource investment.

Furthermore, the advancement of Skill tends to lead to a decrease in Reliability. In Levels I and II, reliability scores appear very high due to the use of deterministic metrics such as Top-1 Accuracy or mAP. However, in Level III expert-level integrated evaluation, models enter a stochastic judgment domain where they generate free-form text. At this stage, as subjective judgment metrics like LLM-as-judge are introduced, maintaining reproducibility and consistency becomes much harder, resulting in lower reliability metrics.

Changes in the Hygiene axis clearly illustrate the shift in the evaluation paradigm. In Levels I and II, hygiene management was sufficient through ‘structural isolation,’ separating fixed training/validation/test sets. However, in the era of Level III MLLMs training on opaque web-scale data, this has shifted toward estimating contamination through post-hoc diagnostic audits. The final solution proposed for this is the Level IV ‘dynamic prevention’ system. Efforts are being made to raise the hygiene level by having humans intervene to continuously generate new cases or by using data produced after the model’s training cutoff.

Regarding Robustness and Fairness, critical flaws in early stages became the drivers for subsequent stages. Recognition-era benchmarks in Level I had high reliability scores but very low robustness and fairness scores due to texture bias and Western-centric data distributions. These vulnerabilities led to models reaching correct answers using statistical shortcuts rather than the essence of objects. To resolve this, Level II introduced diagnostic stress tests such as VQA-CP, which intentionally shuffles distributions, and CLEVR, which requires compositional reasoning, achieving a dramatic increase in the robustness axis.

Consequently, the multimodal measurement framework suggests that AI evaluation is not a simple competition for scores, but an optimization process of balancing these axes. The recognition stage secured low cost and high reliability but lacked robustness; the reasoning stage attempted to secure robustness but faced hygiene and reliability issues. AI evaluation is now evolving away from static metrics toward measuring actual intelligence in dynamic environments, even while accepting the risks of increased cost and decreased reliability.

Evolution of Evaluation by Cognitive Stage: From Level I to Level IV

The evolution of multimodal AI evaluation can be analyzed not as a simple expansion of datasets, but as a gradual increase in the difficulty of the ‘exams’ measuring the model’s cognitive abilities. The core of Level I (Recognition Era) was a basic knowledge test to measure whether a machine could see the world like a human. ImageNet set the standard for large-scale visual recognition with over 20,000 categories and 14 million images, while PASCAL VOC established rules for object detection and segmentation. Later, COCO expanded evaluation to understanding multiple objects and context within complex everyday scenes beyond simple classification. However, models at this stage revealed vulnerabilities to ‘Shortcut Learning,’ relying on statistical characteristics of the background or texture rather than the essence of objects.

Level II (Logic and Understanding Era) emerged to diagnose these vulnerabilities. Benchmarks at this stage evaluate not just ‘what is there,’ but how the model integrates visual information and linguistic cues to reason. VQA laid the foundation for vision-language integration through open-ended questions, and VQA-CP introduced stress tests using linguistic biases in training data to verify if models answer based on actual images. GQA required compositional reasoning based on scene graphs, attempting to overcome the limits of simple pattern matching by measuring whether a model reaches an answer following a step-by-step logical chain.

Deepening logical reasoning led to benchmarks like CLEVR and NLVR2. CLEVR precisely diagnosed the ability to bind attributes and relations through synthetic images, and NLVR2 tested logical composition by judging truth values using real photos. Notably, Winoground exposed the ‘Binding Problem,’ where models fail to accurately map words to visual objects, proving that even latest models perform at random-guess levels in relation-reversal scenarios. Furthermore, OK-VQA and A-OKVQA evaluated the ability to integrate world knowledge external to the image, and VCR raised cognitive evaluation to higher-order thinking by requiring a justification process explaining the reason for the answer.

The current frontier, Level III (Expert-Level Integration), takes the form of a ‘comprehensive exam’ measuring the general capabilities of MLLMs. MMMU requires college-level knowledge in STEM and professional fields to evaluate domain expertise beyond simple common sense. MMBench profiles 20 distinct ability dimensions and increases reliability through the CircularEval technique, which removes bias based on option order. SEED-Bench integrates spatial perception and temporal reasoning to analyze performance differences between image and video modalities. While models at this stage show high accuracy, they face new flaws: data contamination from web-scale training and ‘outcome-centric evaluation’ where they get the answer right but the process wrong.

Recently, tools to measure Process Fidelity have been introduced to overcome such outcome-centric evaluations. MathVista precisely diagnoses visual mathematical reasoning via charts and diagrams, and MM-Vet presents integrated tasks that require chaining multiple abilities such as OCR, math, and spatial perception. Specifically, GeoChain organizes geographical location reasoning into a step-by-step Chain-of-Thought (CoT) to identify the exact point where cognitive errors occur, and VCR-Bench distinguishes between ‘perception’ and ‘reasoning’ stages in video understanding to verify if the reason for a correct answer is based on correct logic.

Finally, Level IV (Abstract and Creative Intelligence) is an uncharted territory that evaluates dynamic interaction and execution capabilities beyond static Q&A. VirtualHome measures causal state tracking by executing household activities as programs, and ALFRED evaluates long-term planning abilities for performing complex instructions from an egocentric perspective. MUEP and EmbodiedBench comprehensively verify the planning and low-level control capabilities of MLLMs in simulated environments. In this stage, behavioral metrics like Success Rate (SR) or Success weighted by Path Length (SPL) become central, expanding the horizon of evaluation into subjective areas without single correct answers, such as social intelligence and creativity.

Deep-Dive Diagnostic Tool Analysis for Latest MLLMs

Recent multimodal large language model (MLLM) evaluation systems are evolving beyond simple performance measurement toward ‘deep-dive’ tools that precisely diagnose specific cognitive abilities. In particular, Video-MME evaluates videos ranging from 11 seconds to 1 hour by integrating visual information with captions and audio, measuring long-horizon reasoning capabilities. This tool verifies whether a model actually understands temporal compositionality rather than simple single-frame hints by analyzing the contribution of frames, captions, and audio individually via a modality toggle function.

MathVista, which precisely measures quantitative reasoning, diagnoses mathematical reasoning abilities by integrating disparate visual sources such as charts, diagrams, and formulas. It reveals that while general MLLMs are proficient in natural scene understanding, they remain vulnerable in geometric parsing and algorithmic reasoning. Even leading models like GPT-4V record approximately 50% accuracy, falling short of the 60%+ performance of human experts, demonstrating limits in precise numerical reasoning.

MM-Vet, which evaluates integrated performance rather than individual abilities, presents open-ended problems that require chaining six core abilities: recognition, OCR, knowledge, language generation, spatial perception, and math. Using the LLM-as-judge method, it ensures flexibility in correct answers while analyzing bottlenecks models encounter in tasks combining different abilities, such as ‘OCR + Math.’ This focuses on measuring how models compose skills to solve real problems rather than whether they possess a single skill.

Hallucinations, a key factor in model reliability, are systematically diagnosed through HallusionBench. This tool presents plausible but false claims along with sophisticatedly edited images to measure whether the model can reject them. The case where an early GPT-4V version scored a low 32% in question-pair accuracy proves that latest models, while appearing accurate, are highly vulnerable to knowledge-based hallucinations or visual illusions unrelated to visual evidence.

Notably, recent tools are moving away from outcome-centric evaluation toward verifying the ‘process’ of reaching an answer. GeoChain implements a 21-step hierarchical geographical reasoning chain (CoT) that narrows down from continent to country, city, and landmark from street-view images. By scoring the visual cues extracted and hypotheses formed at each step, it precisely diagnoses whether cognitive errors occurred in the intermediate reasoning process, even if the final location prediction was correct.

VCR-Bench, evaluating process fidelity in video understanding, goes a step further by assigning ‘perception’ and ‘reasoning’ tags to each reasoning step. It calculates step-by-step precision and recall by comparing the rationale generated by the model with human-authored gold standards. Analysis reveals that even powerful models like GPT-4o or Gemini 1.5 Pro tend to score lower in the perception stage than in the reasoning stage. This suggests that the primary bottleneck for MLLMs lies in the perception stage of extracting accurate visual evidence rather than in logical reasoning capability itself.

Such deep-dive tools reveal detailed cognitive flaws that broad expert exams like MMMU often miss. While MMMU acts as a ‘qualification exam’ measuring broad college-level domain knowledge, the diagnostic tools mentioned above act as ‘precision check-ups’ determining the failure of specific cognitive circuits. Consequently, a system is being established to verify if a model is ‘right for the right reasons’ through a combination of these tools.

Living Benchmarks: Construction and Operational Strategy

To overcome the chronic problems of saturation and data contamination in static benchmarks, Living Benchmarks are proposed as dynamic evaluation systems centered on continuous human intervention and adversarial updates. Unlike traditional methods providing fixed datasets, they aim to measure whether a model exhibits actual intelligence rather than simple memorization by updating evaluation items in real-time to match the speed of model development. For example, Dynabench builds more discriminative datasets through adversarial rounds involving humans and models, and RealTimeQA fundamentally blocks the possibility of data contamination through weekly evaluations covering latest events after the training cutoff.

One of the core design patterns of living benchmarks is Adversarial Data Collection. This is a method where skilled annotators or other models intentionally design valid cases that cause even latest SOTA models to fail. As seen in cases like AdVQA, humans create new failure cases using subtle visual cues or complex common-sense reasoning that models overlook, continuously pushing the limits of the model. Additionally, human preference arenas like LMSYS Chatbot Arena introduce random pairwise comparisons and Elo-based aggregation to measure alignment from a user perspective, such as instruction following and fluency, which static metrics often miss.

Maintenance playbooks for effective operation emphasize strict governance and version control. Living benchmarks are managed using a Semantic Versioning (SemVer) system beyond simple data updates. A MAJOR version is increased when the test definition itself changes, a MINOR version when new items or splits are added, and a PATCH version for simple error corrections. This maintains a consistent comparison standard between evaluators and evaluatees and increases benchmark reliability by transparently disclosing change history.

Secure evaluation protocols are also essential to ensure evaluation integrity. To prevent models from using test sets as training data, API-based server evaluation or code-upload methods are adopted to run models on invisible test sets instead of direct dataset distribution. Specifically, in embodied AI evaluations like the Habitat Challenge, rather than just receiving predictions, agent code is submitted and executed in unseen environments to verify actual problem-solving capability rather than simple pattern matching.

For data quality management, living benchmarks undergo a multi-stage quality control process. Collected adversarial cases go through double-review and arbitration, and difficulty and discriminative power are measured through item-response analysis. Subsequently, deduplication and contamination audits are performed to ensure no overlap with training data before promotion to the final test set. This series of processes allows the benchmark to function as a ‘diagnostic tool’ that continuously discovers model vulnerabilities rather than a set of ‘solvable problems.’

Finally, living benchmarks manage their lifecycle through a Sunset Policy. When a particular benchmark’s performance exceeds 90% and variability between models decreases, reaching saturation, the dataset is demoted from evaluation to training or development (dev) use. It is then replaced by successor versions applying stricter constraints or new stress factors, such as MMLU-Pro or MMLU-CF, to maintain discriminative power. This dynamic circular structure suggests that AI evaluation is not a history of records, but a process of continuous adversarial evolution to create more intelligent systems.

Multimodal AI Cognitive Ability Verification Checklist

To prevent such system failures in practical engineering pipelines, the following core defense systems must be enforced in the operational stage:

• Reasoning Process Verification: Do you verify if the step-by-step logic reaching the answer is valid via Chain-of-Thought (CoT), rather than just the final answer?
• Data Contamination Check: Do you perform post-hoc audits to diagnose the possibility that the evaluation dataset was included in the model’s pre-training data?
• Relational Binding Test: Do you test whether the model accurately binds spatial/causal relationships between objects, rather than just individual object recognition?
• Adversarial Case Input: Do you include modified adversarial samples that induce the model to answer incorrectly, rather than general cases with high correct-answer probability?
• Domain-Specific Evaluation: Do you possess ‘expert-level’ evaluation sets requiring professional knowledge of core business domains, beyond general benchmarks?
• Time-Series Consistency Measurement: When processing video or audio, do you measure whether logical consistency is maintained over time?
• Hallucination Rejection Test: Does the model accurately answer ‘I don’t know’ or ‘False’ to inputs containing questions without answers or false claims?
• Dynamic Evaluation System: Instead of fixing a once-built evaluation set, have you established a living benchmark system that updates continuously in line with model performance improvements?

II. Multimodal Measurement Framework

To systematically understand the evolution of multimodal AI, a measurement framework that strictly defines each evaluation stage is required, moving beyond simple historical narration. This research presents six orthogonal axes that act independently to analyze benchmark quality and model capability from multiple angles. This framework serves as a tool to precisely diagnose the characteristics of performance metrics that change as cognitive levels rise and the underlying trade-off relationships.

The first axis, Skill, refers to the specific cognitive ability being tested. This starts with simple pattern matching in early stages and expands to complex logical reasoning and high-level integrated abilities. The second axis, Reliability, evaluates the statistical consistency of measurement metrics. This includes everything from deterministic metrics with clear results, such as Top-1 Accuracy, to relative evaluation methods like Elo Rating, measuring how consistently the measured value represents model performance.

The third axis, Robustness, evaluates the presence of mechanisms designed to prevent ‘Shortcut Learning,’ where models exploit statistical loopholes in datasets. For example, it verifies whether a model has learned intrinsic causal relationships rather than superficial features of training data through Out-of-Distribution (OOD) datasets. The fourth axis, Hygiene, measures measures to ensure data validity and prevent contamination, where training and test data overlap.

The fifth axis, Cost, refers to the computational resources required for evaluation. Unlike early stages where simple static inference was sufficient, cost rises exponentially in advanced stages requiring complex simulation or interaction. The final sixth axis, Fairness, measures the ability to detect bias against various demographic groups and maintain balanced representation without leaning toward specific groups.

These six axes form very distinct correlations and trade-offs as the AI cognitive level (Level I~IV) increases. The most prominent feature is that as the depth of Skill increases, Cost increases sharply. In recognition-centered Level I, cost-efficiency was maximized as evaluation was possible through simple discriminative classification, but in the Level IV grounded intelligence stage, physics-engine-based simulation environments are essential, requiring massive computational resources.

Furthermore, the advancement of Skill tends to cause a decline in Reliability. In Levels I and II, reliability scores remained very high using deterministic metrics like Top-1 Accuracy or mAP. However, in the professional reasoning stages of Level III and above, the form of the answer changes to open-ended generation, and the use of stochastic judgment tools like LLM-as-judge leads to lower reproducibility and consistency of metrics.

Consequently, the transition from Level I to Level IV can be viewed as a process of gaining technical depth at the expense of cost and reliability. Therefore, evaluators should not simply look at the final benchmark score, but comprehensively analyze which orthogonal axis the benchmark prioritized and what trade-offs are occurring at the current model level to judge the actual advancement of intelligence.

III. Level I: Basic Knowledge (Recognition Era)

The Recognition Era (approx. 2009-2015), which formed the foundation of modern AI, focused on securing core cognitive abilities to identify objects through large-scale standardized tests. Benchmarks of this period served as ‘basic knowledge tests’ to measure whether AI could see the world like humans, setting common tasks, metrics, and goals that led the deep learning revolution.

The most representative, ImageNet/ILSVRC, set the standard for large-scale visual recognition by providing over 20,000 categories organized by WordNet and over 14 million manually annotated images. In particular, ILSVRC, utilizing a 1,000-class subset, popularized the ‘closed-book exam’ format of directly comparing model performance through Top-1 and Top-5 accuracy. The event in 2012 where AlexNet recorded a Top-5 error rate of 15.3%, showing an overwhelming improvement of 10.8 percentage points over existing models, proved the potential of Deep Convolutional Neural Networks (CNN) and acted as a catalyst for modern deep learning.

PASCAL VOC standardized the technology for finding object locations beyond simple classification. This challenge, running from 2005 to 2012, established evaluation protocols across 20 categories, including image classification, object detection, and semantic segmentation. Specifically, it introduced a method of calculating Average Precision (AP) and mean AP (mAP) based on an Intersection-over-Union (IoU) threshold of 0.5 to measure the overlap between predicted and ground-truth boxes, precisely measuring how accurately a model localizes objects beyond simple recognition.

Subsequently, COCO (Common Objects in Context) shifted the evaluation perspective from single-object-centered recognition to complex scene understanding where multiple objects interact in everyday environments. It expanded beyond object detection and instance segmentation across 80 categories to introduce the concept of ‘stuff’ (sky, grass, road, etc.), strengthening the importance of background and context. Furthermore, it enforced precise localization and instance distinction by applying a stricter metric that averages mAP in 0.05 increments from IoU 0.50 to 0.95.

However, it became evident that the high scores of these Level I benchmarks were ‘fragile metrics’ distant from real-world capabilities. In the case of ImageNet, the data collection process involved showing specific class images to workers and requesting confirmation, leading to a bias where typical (iconic) frontal views were over-represented. As a result, models relied on Shortcut Learning, using simple textures or statistical regularities rather than the intrinsic features of objects.

As a concrete example, when recognizing a ‘Pickelhaube (German helmet),’ it was found that models arrived at the correct answer through more prominent background information, such as the accompanying ‘military uniform,’ rather than the features of the helmet itself. Research showing that accuracy drops by about 10% in multi-object images compared to single-object images suggests that high scores at the time were at the level of solving statistical quizzes of the dataset rather than causal understanding.

Consequently, Recognition Era benchmarks achieved optimal results in terms of Reliability and Cost through deterministic metrics and fixed test sets. However, they showed serious flaws in the Robustness and Fairness axes due to the aforementioned texture bias and Western-centric data distribution. These cracks in the foundation created the necessity to diagnose the model’s reasoning process and logical binding capability rather than simply creating harder problems, which naturally drove the transition to Level II, the ‘Dawn of Reasoning.’

VI. Level IV: Abstract and Creative Integration (Uncharted Territory)

As models show strong performance in knowledge-based reasoning tasks with clear answers, the frontier of evaluation is now expanding into the ambiguous, subjective, and dynamic domains central to human intelligence. Level IV is a stage that measures planning in interactive environments, understanding complex social dynamics, and generating valuable new outcomes beyond simple static Q&A. In this stage, behavior-based evaluation centered on the process becomes key rather than the existence of a single correct answer, verifying whether AI has evolved from a passive observer into an active agent.

Various simulation environments have been introduced to evaluate Grounded Intelligence (Embodied AI). VirtualHome encodes everyday household activities into atomic behavior sequences in program form, measuring procedural induction abilities that map language or video to executable programs. Here, LCS (Longest Common Subsequence) similarity and execution success rate (SR) within the simulator are used as metrics rather than simple accuracy. Meanwhile, ALFRED evaluates the ability of agents to complete household tasks following high-level goals and step-by-step instructions in the AI2-THOR environment, verifying planning precision through Success Rate (SR) and SPL (Success weighted by Path Length).

Recent tools like MUEP and EmbodiedBench diagnose these planning abilities in more detail. MUEP emphasizes multi-turn planning based on multimodal observations and measures replanning ability through metrics such as the Reasoning Direction Loss Index (RDI) and Language Compliance (LC). EmbodiedBench, presented at ICML 2025, integrates four environments including ALFRED and Habitat, dividing evaluation into six competency subsets such as common sense, complex instructions, and spatial perception. Notably, even latest models like GPT-4o show a low average success rate of about 29% in low-level manipulation tasks, revealing a significant gap between planning ability and actual physical execution capability.

Beyond physical execution, exploration into Social Intelligence, a higher-order cognitive domain of humans, is also taking place. Benchmarks like Social-IQ measure the ability to reason about emotion, intention, and social norms through real human interaction videos. However, unlike humans who make confident inferences through rich visual and auditory cues, AI models tend to evade ambiguous situations with ‘unable to answer.’ To resolve this, ‘Inverse Reasoning’ methods, which reverse-infer motivation through behavior, are being presented as new evaluation paths.

Creativity measurement is the most difficult area where objective metrics are impossible, and quantitative approaches borrowing frameworks from cognitive science are being attempted. A representative example is the automation of the Alternative Uses Test (AUT), measuring how many (fluency), in how many diverse categories (flexibility), and how rarely (originality) a model suggests new uses for common objects. Additionally, based on Margaret Boden’s theory, attempts are underway to quantify value and novelty by dividing creativity into three types: combinational, exploratory, and transformational.

Overall, Level IV completely departs from the existing ‘get the right answer’ paradigm, integrating successful task completion in dynamic environments with human-centered value judgments. In this process, the generation of infinite unknown environments through simulators becomes the key to solving the hygiene crisis (Data Contamination) that occurred in Level III. Although a trade-off occurs where computational cost rises extremely due to physics-engine-based simulations and the reliability of creativity evaluation decreases, this is a verification process that must be undergone for AI to move beyond a simple ‘test-taker’ and become a true ‘collaborative partner.’

Author’s Thoughts and Practical Recommendations

The most common mistake when adopting AI models in the field is believing, ‘Since GPT-4o is #1 on this benchmark, it will perform well in our business.’ However, as this paper points out, current evaluation systems focus on ‘what’ was answered correctly rather than ‘how’ the model thinks. From an engineering perspective, more granular diagnostic tools are needed than the single metric of accuracy.

The proposal for Embodied AI evaluation in Level IV is particularly impressive. Moving away from static Q&A to measuring actual task completion in simulated environments provides critical implications for companies seeking to utilize AI not as simple chatbots, but as ‘agents.’ We must move toward an era of measuring ‘AI that works well’ rather than ‘AI that speaks well.’

Additionally, the concept of ‘Living Benchmarks’ is very important from an MLOps perspective. Just as one detects model performance degradation or data drift, the evaluation set itself suffers from drift. The automation of an ‘evaluation pipeline’ that manages the lifespan of evaluation sets and continuously injects adversarial cases will be a core competitive advantage in the future.

Ultimately, the key is ‘trust.’ I believe the capability to distinguish whether an AI’s answer is a coincidence, a memorized response, or a result of genuine reasoning will become the core of AI governance.

Conclusion and Enterprise Governance Guide

For internal AI projects to settle into actual business value rather than remaining at the early demo stage of PoC, the following governance principles must be established from the planning phase:

1. From Outcome-Centric to Process-Centric: Set the precision of the reasoning stage (Rationale) as a core KPI, not just the model’s final output value.
2. From Static to Dynamic Evaluation: Reduce dependence on fixed benchmarks and mandate the operation of continuously updated adversarial evaluation sets (Living Benchmarks).
3. Data Contamination Prevention Principle: Before trusting external benchmark scores, always perform cross-verification through internal closed datasets.
4. Agentic Performance Definition: Use task success rate (Success Rate) through interaction with the environment as the measure of actual intelligence, rather than simple response capability.
5. Transparency of Evaluation Tools: When using LLM-as-a-judge, periodically audit the rubric (judgment criteria) and the bias of the judgment model.

References and Original Analysis Materials