Reproduce “ProbeLLM: Automating Principled Diagnosis of LLM Failures” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yue Huang, Zhengzhe Jiang, Yuchen Ma, Yu Jiang, Xiangqi Wang, Yujun Zhou, Yuexing Hao, Kehan Guo and 3 more
ProbeLLM introduces an automated, benchmark-agnostic framework for diagnosing large language model failures by discovering structured, interpretable 'failure modes' instead of isolated errors. It leverages a hierarchical Monte Carlo Tree Search for efficient test generation, employs tool-augmented generation for verifiable ground truth, and consolidates individual failures into actionable patterns.
Key insights
- The hierarchical MCTS effectively balances global exploration and local refinement, enabling efficient discovery of diverse failure types while also concentrating on specific error patterns.
- Tool-augmented generation is critical for producing high-quality, verifiable test cases with reliable ground truths, mitigating the noise and ambiguity common in purely model-generated evaluations.
- Synthesizing individual failure instances into structured 'failure modes' provides deeper, more diagnostic insights into systematic LLM weaknesses, moving beyond superficial error reporting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.