RepBench: A Benchmark-Grounded Data Layer for Capability-Aligned Representation Probing in LLMs

A recent research paper presents RepBench, a data layer anchored in benchmarks aimed at aligning representation probing capabilities in large language models (LLMs). This study tackles a major challenge in representation engineering: existing evaluation methods often rely on synthetic data specific to certain papers, complicating comparisons and reproducibility, and may highlight superficial patterns instead of genuine capabilities. RepBench organizes 13,427 benchmark papers into 182 capability clusters across 13 families and aggregates 353 public benchmark datasets, resulting in 46,149 audited probe texts that cover 94 capabilities, each validated by at least two independent benchmarks. The findings indicate that raw per-text vectors lack natural cluster granularity, while benchmark-pooled capability vectors reveal an optimal clustering at a limited number of clusters across all 12 models assessed, showing minimal alignment with human taxonomy. This research, identified as arXiv 2607.28008v2, offers the AI community a stronger, reproducible basis for probing and guiding LLM capabilities, potentially enhancing model interpretability and control.