AI & softwareResearch findings
A Sketching Method for Context Attribution in Language Models
Original published 2026.11.13 03:57 KSTAdded to Blackdeck 2026.11.14 18:17 KST
What has changed?
A sketching method reduces evaluation calls for context attribution by leveraging stability assumptions and backpropagation circuit differentiation
Key points
- Context attribution analyzes how removing prompt parts affects language model outputs.
- Traditional leave-one-out methods require n model evaluations, which is computationally expensive.
- Proposed sketching method reduces evaluation calls to O(√(log(1/ε)/ε)) under stability assumption
Connect it to your work
The method achieves error proportional to ε with O(√(log(1/ε)/ε)) model evaluations under stability assumption
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.