Paper Overview
Field: NLP Authors: Steven Yu, Sheng-Yu Wang, Kimia Hamidieh, Antonio Torralba, David Bau Published: 2026-10-29 arXiv: 2610.40655
Summary
Data attribution methods for language models (LMs) are usually evaluated with downstream tasks such as fact retrieval. However, existing evaluations vary multiple dataset properties simultaneously, making it impossible to pinpoint exactly why an attribution algorithm fails. To diagnose these failures, the authors explore the design space of LLM data attribution, with a primary focus on evaluation methodology.
They introduce the Controlled Attribution Benchmark (CATT) for fact tracing, in which LLMs are fine-tuned on entirely fictitious knowledge, providing unambiguous ground truth for attribution. This synthetic setup allows systematically altering one dataset configuration at a time, such as document redundancy and lexical distractors. Using this configurable evaluation, they demonstrate that redundancy and distractors consistently degrade attribution performance.
Rather than proposing a new estimator, the paper examines two underexplored design choices: the loss function used to compute gradients, and how per-token attribution scores are aggregated into document scores. They find that the true causal evidence in a training document is highly localized, but standard methods dilute this evidence into noise by averaging attribution scores across the entire document.
Replacing the standard cross-entropy loss gradient with negative log-odds gradients highlights key tokens over uninformative ones. Aggregating only the top token scores then improves ground-truth recall by 35 to 72 percentage points across multiple baselines. The negative log-odds correction also generalizes to FTRACE-TREX, improving counterfactual retraining performance by up to 4.9x across every method tested.
Key Findings
- CATT benchmark: fine-tuning LLMs on fully fictitious knowledge yields unambiguous ground truth, enabling one-variable-at-a-time attribution evaluation.
- Redundancy and distractors systematically reduce attribution performance.
- Localized evidence: causal evidence in training documents is concentrated in a few tokens, and whole-document averaging dilutes it.
- Negative log-odds gradients surface key tokens; top-token-only aggregation improves recall by 35–72 percentage points.
- Generalization: the correction improves counterfactual retraining results on FTRACE-TREX by up to 4.9x.