Paper says localized gradients improve language-model data attribution
Original title: Elucidating the Design Space of LM Data Attribution
On the synthetic CATT benchmark, top-token negative log-odds scoring lifts ground-truth recall by 35 to 72 points.
Researchers argue that fact-retrieval evaluations change several dataset properties at once, so attribution failures are hard to diagnose. They introduce the Controlled Attribution Benchmark, which fine-tunes language models on fictitious knowledge and varies one setting at a time, including redundancy and lexical distractors. Both factors consistently degrade attribution. They report that causal evidence is highly localized and that averaging token scores across a document dilutes it; negative log-odds gradients plus top-token aggregation also improve results on FTRACE-TREX by up to 4.9x.
- CATT uses entirely fictitious knowledge as unambiguous attribution ground truth.
- The paper does not propose a new attribution estimator.
- Redundancy and lexical distractors consistently reduce attribution performance.
- The log-odds correction improved every method tested on FTRACE-TREX.
From the original article
1 min read at arxiv.orgData attribution methods for Language Models (LMs) are typically evaluated using downstream tasks like fact retrieval. However, existing evaluations vary multiple dataset properties simultaneously, making it impossible to isolate exactly why an attribution algorithm fails. To diagnose these failures, we explore the design space of LLM data attribution, with a primary focus on evaluation.
By Steven Yu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.