STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations
2026-06-03 • Machine Learning
Machine LearningComputation and Language
AI summaryⓘ
The authors propose a new method called STRIDE to figure out which training data influenced a large language model's predictions without expensive retraining. Instead of tracking tiny changes in the model’s parameters, they look at how the model’s internal activations change when influenced by small 'steering operators.' These operators help identify the effect of individual training examples more efficiently, making the process much faster and still accurate. The authors demonstrate that STRIDE improves speed and quality in tasks like selecting training data and detecting data contamination.
Training Data AttributionLarge Language ModelsParameter SpaceActivation SpaceSteering OperatorsSparse RecoveryCompressive SensingModel RetrainingData ContaminationData Selection
Authors
Rishit Dagli, Abir Harrasse, Luke Zhang, Florent Draye, Amirali Abdullah, Bernhard Schölkopf, Zhijing Jin
Abstract
Training Data Attribution (TDA) seeks to trace a model's predictions back to its training data. The gold standard for TDA relies on causal interventions, observing how a model changes when data is added or removed, but repeated retraining is computationally challenging for Large Language Models (LLMs). Consequently, most approaches approximate this effect in the parameter space using gradients. However, tracking gradients across billions of parameters is not only prohibitively expensive but relies on local approximations. In this work, we propose a shift: rather than estimating parameter changes, we model the functional effect of training data in the activation space. We introduce STRIDE (Steering-based Training Data Influence Decomposition), a framework that formulates TDA as a sparse recovery problem in the spirit of compressive sensing. STRIDE learns lightweight "steering operators" that mimic the behavioral shift caused by training on data subsets. By measuring how these operators perturb test predictions, we recover individual training example influences via sparse linear decomposition. STRIDE achieves state-of-the-art for LLM pre-training attribution while being an order of magnitude ($13\times$) faster than previous art. We further validate its practical utility through downstream applications including data selection, data contamination, and qualitative analysis.