A new benchmark tries to fix a basic problem in AI interpretability research: nobody agrees on how to measure what is actually happening inside a model's head.
Researchers built RepBench by crawling 13,427 benchmark papers to construct a taxonomy of 182 capability clusters across 13 families, covering things like reasoning, coding, and safety behavior. From that taxonomy, they pulled from 353 public benchmark datasets to produce 46,149 audited probe texts spanning 94 distinct capabilities, each backed by at least two independent benchmarks. The goal is to give representation engineering researchers, who try to locate and steer the internal signals a model uses for a given skill, a shared, reproducible dataset instead of the one-off synthetic examples most papers currently rely on. The team then tested the corpus across twelve models using four different methods for extracting those internal signals.
The results are not reassuring. Raw probe vectors do not cluster into anything meaningful on their own, but once probes are grouped by benchmark, a small number of clusters does emerge, and it barely matches the human-made taxonomy the researchers started with. The two leading extraction methods also disagree on which one performs better: a simple difference-in-means approach wins on most models overall, while logistic regression wins more individual capability-model matchups. For a field that increasingly claims to identify and edit specific behaviors inside a model, that disagreement is a warning sign, not a footnote.
The pipeline and corpus are released as open tools, so the next paper claiming to have found a model's "honesty direction" can at least be checked against the same yardstick instead of a new one invented for the occasion.