In the current era of artificial intelligence, we are obsessed with leaderboards. We crown a new "state-of-the-art" model because its accuracy is 1.2% higher than the previous champion. However, these aggregate metrics—like Accuracy, F1, or BLEU—are summative grades that provide a "pass/fail" result without explaining the underlying behavior. As Ian Porada and colleagues argue in Challenges to Evaluating the Generalization of Coreference Resolution Models: A Measurement Modeling Perspective, these single-number summaries often mask significant failures in specific linguistic sub-phenomena.
The research direction emerging across the field is a shift from summative ranking to fine-grained diagnostic evaluation. Instead of asking "How well does it perform?", researchers are asking "What specific skills has it mastered, and exactly where does it break?"
The Failure of Aggregate Metrics
The fundamental problem with aggregate metrics is that they are "lossy." They collapse complex multidimensional behaviors into a single scalar value. Tongshuang Wu demonstrates this in Beyond Accuracy: Behavioral Testing of NLP Models with Checklist, showing that models with high held-out accuracy can fail on trivial "sanity checks," such as handling simple negations or name changes.
Without granular analysis, developers cannot distinguish between different types of errors. For instance, in speech recognition, KEBAP: Korean Error Explainable Benchmark Dataset for ASR and Post-processing highlights that Word Error Rate (WER) fails to tell you if a model is struggling with acoustic noise or grammatical complexity. Similarly, in computer vision, Yukai Guo and team note in A Unified Interactive Model Evaluation for Classification, Object Detection, and Instance Segmentation in Computer Vision that standard metrics like mAP fail to reveal whether a model is failing due to poor localization, incorrect scaling, or simple misclassification.
Decomposing Intelligence into Capabilities
To solve this, researchers are decomposing "intelligence" into a hierarchy of specific capabilities. This approach, often called faceted evaluation, treats a model more like a student with a report card than a calculator with an output.
Hierarchical Capability Trees
Frameworks like SkillVerse : Assessing and Enhancing LLMs with Tree Evaluation by Yufei Tian and Jiao Sun organize model skills into dendrograms. By structuring evaluation this way, researchers can pinpoint that a model might be excellent at general reasoning but specifically fails at "instruction-following with negative constraints." Yizhong Wang furthered this with EvalTree: Profiling Language Model Weaknesses via Hierarchical Capability Trees, which generates human-readable "weakness profiles" to guide targeted data collection.
Dissociating Knowledge from Reasoning
A key insight in diagnostic evaluation is that a failure to answer a question might stem from two different sources: the model doesn't know the facts, or it knows the facts but can't reason with them. Xiaoqiang Wang and Lingfei Wu introduced the FAC²E: Better Understanding Large Language Model Capabilities by Dissociating Language and Cognition framework to separate these variables. They found that many models "know" the necessary information but require explicit reasoning steps (like chain-of-thought) to apply it correctly.
Borrowing from Psychometrics
If we treat AI models as "agents" or "students," it makes sense to use tools from human psychology. Rahul Ramachandran and colleagues applied Item Response Theory (IRT) in On Evaluation of Vision Datasets and Models using Human Competency Frameworks.
IRT allows researchers to separate the difficulty of a specific test item from the ability of the model. This is a crucial shift: instead of just seeing a "wrong" answer, IRT helps us understand if the model failed because the question was uniquely difficult (high IRT difficulty) or because the model lacks a specific latent trait (like spatial reasoning). Lucy Cheke expands on this in Inferring Capabilities from Task Performance with Bayesian Triangulation, using Bayesian methods to "triangulate" a model's cognitive profile across diverse tasks.
From Rankings to Actionable Feedback
The ultimate goal of this research direction is to move from rankings to insights. As Zongqi Wang and Yujiu Yang argue in From Rankings to Insights: Evaluation Should Shift Focus from Leaderboard to Feedback, evaluation should serve as a feedback loop for development.
When we use frameworks like QualEval: Qualitative Evaluation for Model Improvement by Ameet Deshpande and others, we gain the ability to automatically discover which sub-tasks (e.g., "logical deduction in medical contexts") are the primary bottlenecks. This allows for "precision engineering"—fixing specific model weaknesses through targeted data augmentation or architectural adjustments rather than just training on more of everything.
By moving beyond the single-number score, the field is finally beginning to look inside the "black box," transforming evaluation from a simple measurement of success into a roadmap for progress.