Causal validation of training data attribution (TDA) methods is costly, since evaluation involves retraining a model on many different subsets of the original training set. For example, the Linear Datamodeling Score (LDS), which has become a standard metric for evaluating TDA methods, measures the correlation between TDA predictions and expected retrained model outputs. In this paper, we reduce the amount of computation needed to estimate the LDS along three axes: the number of sampled training subsets, the number of retraining seeds per subset, and the number of models in the attribution ensemble. Each of these factors induces a form of bias in the LDS estimation, and removing these sources of bias significantly reduces the number of retrainings needed to match the same accuracy in LDS estimation. On ResNet-9/CIFAR-10, finite-subset debiasing achieves 3.2% relative RMSE with as few as 4 subsets, and 32 subsets match the accuracy of ordinary Spearman using 96 subsets, measured against a 300-subset reference. On ResNet-9/CIFAR-10, BERT/QNLI and GPT-2/WikiText-2, estimating the training-noise variance allows us to reproduce the standard five-seed LDS at the same accuracy with 2--5× fewer retrainings.
Figures & tables
Figure 1: Three ways to reduce LDS evaluation cost on CIFAR-10/ResNet-9. (a) Correcting for training variance approaches the Spearman reference computed from held-out retraining seeds with fewer retrainings, while naive seed averaging plateaus (EK-FAC). Curves show mean estimates ±1 SD over random draws, see Appendix D for details. (b) The circled small attribution ensembles predict the LDS of larger ensembles (dashed lines). (c) Correcting finite-subset Spearman bias reduces RMSE against the 300-subset reference (TRAK).
Figure 2: Noise correction reduces the cost of LDS evaluation. Rows: CIFAR-10/ResNet-9, QNLI/BERT, WikiText-2/GPT-2. (a, d, g) RMSE to the noise-free Spearman LDS versus total retrainings, for naive R -seed averaging and the noise correction (EK-FAC); dotted: predicted naive plateaus. (b, e, h) Reconstructing the five-seed Spearman target ρ5 from two seeds matches naive averaging with 2.1 – 5.2× fewer retrainings. (c, f, i) Measured noise-corrected Spearman LDS and predictions from the circled small ensembles. The first two columns use E=1 and disjoint retraining seeds for estimation and reference.
Figure 3: Spearman versus Pearson LDS. (a) Spearman minus Pearson LDS; curve: Gaussian prediction. (b) Residual from the finite- M Gaussian expectation. Large markers: full ensemble, all seeds.
Figure 4: Bias (squared) and variance of estimating TRAK LDS on CIFAR-10 using sample Spearman ρ^s and the unbiased estimator ρ^s,unb . Using the unbiased estimator removes the large bias in estimating LDS.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: The arcsin bridge in detail. Top: CIFAR-10 / ResNet-9; middle: QNLI / BERT; bottom: WikiText-2 / GPT-2; single attribution members ( E=1 ). (a, d, g) Bias of the reconstructed estimate (filled squares, 95% CI), its conversion component (open squares), and the corresponding finite-mask Gaussian baseline (crosses); grey: ±1 standard deviation across two-seed reconstructions. (b, e, h) Q–Q deviation of D2 for EK-FAC, with the readout averaged over R retrains per mask; 0 denotes the Gaussian reference. (c, f, i) Excess kurtosis of the readout relative to the mean of its rotation-based reference distribution, against R . Grey in (b, c, e, f, h, i): pointwise 90% rotation-based reference ranges.
Figure 6: RMSE of corrected Spearman LDS. Error to references computed from held-out retraining seeds over 200 draws, with E=10 . Columns: CIFAR-10, QNLI, WikiText-2. Rows: EK-FAC, TRAK, SOURCE/TracIn. Dashed: naive Spearman; solid: combined training-noise and finite-subset correction; dotted: predicted naive RMSE floors.
Figure 7: Spearman LDS estimates from a single draw. The first of the 200 draws in Fig. 6 , with the same E=10 ensembles and estimators. Black lines: reference computed from the first split’s held-out retraining seeds; dotted lines: predicted naive plateaus from that split.
Figure 8: Comparing the bias (squared) and variance of estimating ρs using the finite-sample Spearman correlation estimator ρ^s , across different numbers of query points Q .
Figure 9: Comparing RMSE in estimating the full-bank LDS using finite-sample Spearman and the finite-bank correction ( 18 ). All experiments use R=5 retraining seeds.
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×) faster than previous art. We further validate its practical utility through downstream applications including data selection, data contamination, and qualitative analysis.
Rishit Dagli, Abir Harrasse, Luke Zhang +4
Jinesis AI Lab, University of Toronto & Vector Institute · Max Planck Institute for Intelligent Systems, Tübingen, Germany · Thoughtworks +3
Training Data Attribution (TDA) seeks to trace model predictions back to influential training examples, enhancing interpretability and safety. We formulate TDA as a Bayesian information-theoretic problem: subsets are scored by the information loss they induce - the entropy increase at a query when removed. This criterion credits examples for resolving predictive uncertainty rather than label noise. To scale to modern networks, we approximate information loss using a Gaussian Process surrogate built from tangent features. We show this aligns with classical influence scores for single-example attribution while promoting diversity for subsets. For even larger-scale retrieval, we relax to an information-gain objective and add a variance correction for scalable attribution in vector databases. Experiments show competitive performance on counterfactual sensitivity, ground-truth retrieval and coreset selection, showing that our method scales to modern architectures while bridging principled measures with practice.
Data attribution methods for Language Models (LMs) are typically evaluated using downstream tasks like fact retrieval. However, existing evaluations vary multiple dataset properties simultaneously, making it impossible to isolate exactly why an attribution algorithm fails. To diagnose these failures, we explore the design space of LLM data attribution, with a primary focus on evaluation. We introduce the Controlled Attribution Benchmark (CATT) for fact tracing, in which we fine-tune LLMs on entirely fictitious knowledge, providing unambiguous ground truth for attribution. This synthetic setup allows us to systematically alter one dataset configuration at a time, such as document redundancy and lexical distractors. With this configurable evaluation, we demonstrate that redundancy and distractors consistently degrade attribution performance. Rather than proposing a new estimator, we examine two design choices that are underexplored: the loss used to compute gradients and how per-token attribution scores are aggregated into a document score. We find that the true causal evidence within a training document is highly localized, but standard methods dilute this evidence into noise by averaging attribution scores across the entire document. We find that replacing standard cross-entropy loss gradients with the negative log-odds gradients surfaces the critical tokens from other uninformative ones. We can then aggregate only the top token scores to improve ground truth recall by 35 to 72 percentage points across multiple baselines on our diagnostic benchmark. The negative log-odds correction also generalizes to FTRACE-TREX, improving counterfactual retraining performance for every method we test, by up to 4.9x.