SciGroveBeta
Machine Learning

STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations

Rishit Dagli, Abir Harrasse, Luke Zhang, Florent Draye, Amirali Abdullah, Bernhard Schölkopf, Zhijing Jin

Featured June 5, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

STRIDE identifies which training data influenced a model's answer by learning how to steer the model's internal signals, making it much faster and more accurate than traditional gradient-based methods.

In depth
The paper introduces STRIDE, a framework that reformulates Training Data Attribution (TDA) as a sparse recovery problem in the activation space. Instead of computing expensive parameter-space gradients, the authors learn low-rank steering operators that simulate the functional impact of training on specific data subsets. By applying these operators to test queries, the method generates perturbation responses that are then decomposed into individual example influences using compressive sensing techniques.

Key Takeaways

  • 1
    STRIDE shifts the attribution paradigm from parameter-space approximations to activation-space steering, significantly reducing computational overhead.
  • 2
    The framework recovers per-example influence scores by treating the attribution problem as a sparse linear decomposition task.
  • 3
    The method achieves state-of-the-art Linear Datamodeling Score (LDS) performance while being over 12 times faster than existing gradient-based baselines.

Conceptual Flow

HIGH LEVEL
1
Methodology

The model learns how to mimic the effect of training on data subsets by adjusting internal signals, then uses math to figure out which specific examples caused those changes.

Training Data Subsets
Learn Steering Operators and Solve Sparse Recovery
Individual Example Influence
2
Results

The new method is much faster and more accurate at tracing model answers back to the original training data compared to older, slower techniques.

Old Slow Method

Faster and More Accurate Attribution

New Efficient Method