Elucidating the Design Space of LM Data Attribution
Source check: Source checked. Not independent scientific confirmation.
Finding: A controlled attribution benchmark fine-tuned language models on fictitious knowledge, allowing ground-truth source tracing while varying redundancy and lexical distractors. Standard document-level averaging diluted localized attribution evidence; negative-log-odds gradients with top-token aggregation improved recall in the evaluated benchmarks.
Technology: Language models with token-level data-attribution scoring
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Token-level language-model attribution scores were evaluated under controlled fictitious-knowledge, redundancy and distractor settings.
- Test and result
- On CATT, top-token aggregation with negative-log-odds gradients improved ground-truth recall by 35–72 percentage points across several baselines; the correction also improved counterfactual retraining performance on FTRACE-TREX by up to 4.9-fold.
- Limitation / still unknown
- These results use controlled attribution benchmarks and do not establish reliable provenance for arbitrary language-model training data.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.40655
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.