Controlled benchmark clarifies failures in language-model data attribution

A research paper introduces CATT, a controlled fact-tracing benchmark that fine-tunes language models on entirely fictitious knowledge with unambiguous ground truth to isolate why data-attribution methods fail. The work argues that existing evaluations vary several dataset properties at once, confounding results, and suggests attribution research adopt controlled synthetic benchmarks before claiming progress on real training data.

The paper studies why language-model data-attribution methods fail, noting evaluations often vary several dataset properties at once. It introduces CATT, a controlled fact-tracing benchmark using fine-tuning on entirely fictitious knowledge with unambiguous ground truth. Why it matters Reliable attribution is needed to trace model outputs to training data, audit sources, and diagnose failures. Analysis Attribution research may adopt controlled synthetic benchmarks before claiming progress on real, confounded training data. Source: arXiv https://arxiv.org/abs/2610.40655v1 AI disclosure: this article was produced with AI assistance and reviewed for accuracy.