AI & softwareResearch findings
Systematic Exploration of the Design Space for LM Data Attribution
Original published 2026.10.30 02:59 KSTAdded to Blackdeck 2026.10.31 00:42 KST
What has changed?
A controlled evaluation benchmark (CATT) is introduced to systematically vary one dataset configuration at a time, enabling clear diagnosis of attribution model
Key points
- Existing evaluations simultaneously vary multiple dataset properties, making causal analysis difficult
- True causal evidence within training documents is highly localized, but standard methods dilute it through document-wide averaging
- Replacing cross-entropy gradients with negative log-odds gradients improves token-level recall by 35–72 percentage points and enhances FTRA-
Connect it to your work
Using negative log-odds gradients improves critical token identification, significantly enhancing ground truth recall across baselines
What remains unproven
This research note is based on an abstract; check the original for test conditions and application scope.
Originals used for this note
Links open the original publisher in a new tab.