TIDE (Tree‑Induced Diagnostic Evaluation) reframes diagnostic evaluation as a behavior‑driven decision‑tree induction task, enabling fine‑grained capability diagnosis for large language models (LLMs). The method recursively builds a hierarchical capability tree: at each node an LLM judge proposes several candidate splits and classifies benchmark samples into groups; then a behavior‑based information gain criterion selects the split that best separates the samples. This recursion continues until a full tree covering all samples is formed, with each leaf representing a specific capability subset.
We applied TIDE to five models across four public benchmarks, including Claude Opus 4.6 and GPT‑5.4. The results reveal capability contrasts invisible to aggregate scores. For instance, Claude Opus achieves 100% accuracy on micro‑economic market reasoning while GPT‑5.4 scores 83.3%, yet Claude Opus trails GPT‑5.4 on contract‑law reasoning (77.8% vs. 88.9%). Such nuances are completely masked by traditional overall accuracy or pass‑rate metrics.
Key advantages of TIDE are:
- Directly leverages model‑generated behavior signals without extra annotation;
- Automatically discovers the most discriminative capability dimensions via information gain;
- Produces a capability tree that can be visualized to pinpoint model strengths and weaknesses.
Future work may extend TIDE to legacy benchmarks and emerging evaluation tasks, fully exploiting the rich behavioral signals in evaluation results to guide model development and safety assessment.
Review