EvalTree: Profiling Language Model Weaknesses via Hierarchical Capability Trees

We are grateful for the highest recognition of the motivation behind EvalTree, the clarity and effectiveness of our methodology, and the extensiveness and depth of our experimental analysis (including some auxiliary experiments like the Chatbot Arena study). We thank Reviewer pMeM for the strongest support and extremely positive assessment of our work!

We sincerely thank Reviewer n857 for their thoughtful and constructive review. We are grateful for their recognition of EvalTree’s novelty, its comprehensive and interpretable approach to diagnosing LM weaknesses, the comprehensiveness of our experimental validation, and the practical value demonstrated through weakness-guided data collection and interactive interface. We also appreciate the helpful pointers to related work, and completely agree that better contextualization within the broader literature will further strengthen our contribution.

Below, we address Reviewer n857’s concerns and suggestions in detail.

• Reviewer n857 raises the question of how each design choice within EvalTree contributes to its overall effectiveness (e.g., if the hierarchical structure matters).
  • We appreciate this question and agree that understanding the impact of different design choices is important. Many of EvalTree’s design choices follow naturally from its overarching goal; as such, most are relatively direct or standard (e.g., embedding instances before clustering). We have therefore focused our ablation efforts on the design choices that may require empirical validation.
  • Regarding the importance of hierarchy versus flat categorization: We view this as a central axis of EvalTree’s design and thus a key factor of EvalTree’s success. A natural way to ablate this dimension is by comparing against existing methods that use a flat (single-level) categorization structure. Among such works, QualEval [1] and GoalEx [2] are the most relevant; in our experiments, we chose QualEval as a baseline for this comparison, as GoalEx involves extremely extensive LLM querying that becomes computationally infeasible at the benchmark scales we target. This comparison can be interpreted as an ablation of EvalTree’s hierarchical structure. Furthermore, in Appendix E.3.2, we provide qualitative analysis of this distinction: flat categorization can only operate at a fixed granularity, and thus fails to sufficiently model the intricate and interrelated structure of capabilities tested within a benchmark; in contrast, a hierarchical structure successfully models the complexity of capabilities tested within a benchmark, which lets us analyze capabilities at varying granularities flexibly, from broad categories to specific skills.
  • As for how the hierarchical structure is constructed, EvalTree uses a recursive K-Means-based clustering algorithm by default. We ablate this choice in Appendix G by replacing it with a standard hierarchical clustering method. The results show that our default choice is more effective.
  • We also conducted new ablation experiments to study how variations in capability annotations (from Stage 1) and capability embeddings (from Stage 2) might affect the resulting capability trees. Please see our overall response for further details.
  • We appreciate the reviewer’s interest in understanding the contribution of individual components. We believe our analyses provide meaningful insights into the core design choices of EvalTree, and we thank the reviewer for the opportunity to clarify—we will make them more prominent in the paper.
• Reviewer n857 expresses concern that the ground-truth weakness assessment relies on manually curated weaknesses rather than weaknesses derived from real-world model behavior.
  • The reviewer is absolutely right to highlight the value of validating against real-world weaknesses. However, we argue that knowing such “ground-truth” weaknesses from real models is currently NOT feasible; and, in fact, this limitation is what motivates the need for formulating the research problem of weakness profiling and proposing methods like EvalTree in the first place. To reliably derive a ground-truth set of weaknesses from real models, one must already possess an accurate and comprehensive understanding of where and why the model fails, which is exactly the open research challenge that we hope to address.
  • That said, we agree it is a limitation that the weaknesses used in our Ground-Truth Assessment are not “real” ones. We address this by relying on our full suite of weakness profiling method assessments.
    • First, our Low-Performance Identification Assessment directly assesses how precisely each method can identify weaknesses using real model evaluation results.
    • Second, to compare comprehensiveness with real model evaluation results, we rely on Extrinsic Assessment to provide an indirect but reliable signal, which examines how effectively the identified weaknesses guide training data collection. The intuition is as follows: if one method identifies a broader set of true model weaknesses, then collecting data to target those weaknesses should yield greater model performance gains.
    • Ground-Truth Assessment provides a direct signal of how precisely and comprehensively each method can identify weaknesses (using synthetic model evaluation results).

Thank you for the detailed and thoughtful rebuttal. I'm glad to see that the authors have clearly addressed my concerns, particularly regarding the design choices in EvalTree and the applicability of related work. I appreciate the clarifications and the new ablation results. Given these responses, I am satisfied with the changes and would like to raise my score.

We sincerely thank Reviewer SQGa for their thoughtful and constructive review. We appreciate the recognition of EvalTree’s novelty, effectiveness, and efficiency, its strengths as an interpretable diagnostic framework for language model evaluation, its practical value in guiding data collection, and the contribution of the interactive demo to real-world applicability.

Below, we address Reviewer SQGa’s concerns and suggestions in detail.

• Reviewer SQGa raises a point regarding the consistency and reproducibility of EvalTree, as some pipeline components rely on LLM-generated outputs that can vary even with fixed inputs.
  • We agree this is a reasonable concern. In our experiments, we used GPT-4o-mini and a temperature of 0.0 consistently across all components to ensure internal consistency; however, we acknowledge that this may lead to variability when people want to reproduce our experiments by themselves. To address this, EvalTree is model-agnostic by design, so its components can be replaced with open-weight LLMs to reduce variance and improve reproducibility.
  • To this end, we have conducted new experiments using open-weight LLMs within the EvalTree pipeline. Please see our overall response for further details.
• Reviewer SQGa asks about how variations in capability annotations and capability embeddings might affect the resulting capability trees.
  • We agree this is a thoughtful concern given EvalTree’s reliance on LLM-generated annotations and embeddings. To study this, we conducted new ablation experiments. Please see our overall response for further details.
• Reviewer SQGa suggests that running EvalTree only on failed instances may be more efficient under compute or budget constraints.
  • We agree that EvalTree can be applied to any subset of instances with a common known prior property, for example, the subset of instances unsolved or solved by an LLM. The capability tree generated from such a subset can indeed provide rich insights into those instances. In practical deployments where resources are constrained, using EvalTree on a filtered subset (e.g., only failures) should be a useful strategy.
  • However, for the research problem of weakness profiling, we intentionally include all instances for theoretical rigor. We identify weak capabilities by assessing the proportion of errors under a capability rather than the absolute count of errors. This ratio-based consideration is important: suppose a capability contains a disproportionately large number of instances, and the model answers most of them correctly; if we analyzed only the failed instances, this capability could appear to be a major weakness simply because of the volume of examples it contributes, while in fact the model performs well on it; by considering all instances, we avoid such statistical biases for more principled identification of real weaknesses.
• Reviewer SQGa suggests validating the capability tree against known categorizations, such as comparing second-layer nodes with MMLU subcategories.
  • We appreciate this suggestion and agree that grounding EvalTree’s structure against a known taxonomy can provide useful insights. To this end, we conducted a new experiment using the MMLU benchmark, which provides 57 human-defined subcategories (e.g., abstract algebra, high school psychology, etc). We compared this human-defined partitioning with a partitioning of instances based on EvalTree; specifically, we focused on the second-layer nodes of the tree, which yielded 46 capability clusters, and each instance was assigned to the second-layer node it descends from. We then computed the Adjusted Mutual Information (AMI) between the EvalTree and MMLU partitionings, obtaining a score of 0.503.
  • An AMI of 0.503 suggests a moderate-to-strong alignment between the EvalTree-generated partitioning and the MMLU-defined partitioning. This validates that EvalTree captures meaningful, structured distinctions that correspond in part to the topic-based groupings defined by humans.
  • Furthermore, EvalTree nodes often reflect more than just topical alignment, for instance, goals and reasoning styles, which go beyond what MMLU categories capture; this is why the alignment is not perfect (i.e., close to 1.0). For example, one single capability is described as “analyzing and synthesizing complex biological, biochemical, and physiological systems and processes to evaluate health impacts and functional interrelationships”, which contains instances from multiple MMLU categories such as “nutrition,” “clinical_knowledge,” and “college_medicine.”

We appreciate the very high recognition of EvalTree’s motivation, its improvements over prior work, and the clarity of our experimental setup. We sincerely thank Reviewer DECS for the thoughtful and very positive review!

For the release of code and data, please see our overall response for further details.