The authors recast diagnostic evaluation as behavior-driven decision-tree induction.
Tree-based method reveals fine-grained LLM capabilities hidden by aggregate scores
TIDE builds a hierarchical capability tree from benchmark samples and model scores, using an LLM judge to propose splits and information gain to select them, exposing contrasts like Claude Opus 4.6's perfect score on microeconomic reasoning where GPT-5.4 scores 83.3%.
Chinese Tech
Jiayi Chen · Hui Li · Yujiong Shen · Qiyuan Peng · Yanxin Li · Chenhao Huang · +9 more
Fudan University · Tencent
Research Digest··2 min read
The authors propose TIDE, a method that induces a decision tree from benchmark samples and model scores to diagnose fine-grained capabilities.
§