Source-linked AI summary
From Behavior to Provenance: Attributing Tabular Foundation Models to Synthetic Pretraining Data
Mohamed Bouadi, Nassim Bouarour, Shivam Dubey, Aditya Tanna, Vinay Kumar Sankarapu
TL;DR
Training-data attribution seeks to identify examples shaping model behavior, but its faithfulness is difficult to validate when causal contribution is unobserved. The paper uses O’PRIOR’s recorded synthetic-task lineage with counterfactual retraining and provenance-conditioned interventions. Influence shows the strongest removal faithfulness, while TracIn has greater shortcut enrichment despite a smaller within-provenance counterfactual gap.
Problem
Attribution aims to identify training examples that shaped behavior, but faithfulness is difficult to establish because example-level contribution is unobserved and provenance does not show behavior absent an example.
Source: §1 · p. 1Method
Results
Influence has the strongest counterfactual removal faithfulness, while TracIn has greater shortcut enrichment (1.81× vs. 1.60×) but a smaller within-provenance counterfactual gap (0.023 vs. 0.030).
Source: §4 · p. 8Takeaways & Limitations
Synthetic provenance is a structured reference for experimentally testing attribution claims when combined with counterfactual retraining, behavior-specific evaluation, and reliability analysis.
Source: §4 · p. 8The study uses a finite corpus of 5,000 O’PRIOR tasks and a NanoTabPFN-scale model, so the same relationships may change with larger models or corpora, or natural-data pretraining.
Source: §4 · p. 8
Abstract
from arXiv · showhide
Training-data attribution aims to identify which training examples shape model behavior, yet validating such claims is difficult because causal training influence is rarely observable. We argue that controlled synthetic pretraining makes attribution experimentally testable. Using O'PRIOR, a provenance-rich synthetic task generator for tabular foundation models, we construct a testbed in which every pretraining task carries explicit lineage over structural mechanisms, missingness, confounding, shortcuts, and distribution shift. We combine behavior-conditioned attribution with counterfactual retraining and provenance-aware interventions to test both task-level faithfulness and mechanism-level consistency. On held-out real tasks, removing the top-attributed 5% of synthetic tasks decreases mean ROC-AUC by 0.013, compared with 0.002$\pm$0.004 under random removal, while removing bottom-attributed tasks improves performance by 0.003. Within shortcut-provenance tasks, targeted removal yields an effect of 0.043 versus 0.016 for matched random removal. Provenance discrimination is more modest by ranking AUROC (0.55-0.62), despite substantial top-k enrichment, revealing that provenance association and interventional faithfulness need not coincide. Our results establish synthetic provenance as a controlled setting for verifiable contributive attribution
1 Introduction
The introduction argues that attribution is difficult to validate because training influence is unobserved, and proposes controlled synthetic pretraining as an experimental testbed. It uses O’PRIOR provenance and counterfactual retraining to distinguish attribution’s behavioral faithfulness from provenance association.
1 Introduction
- Attribution methods seek to identify training examples that shaped a behavior, but their faithfulness is hard to establish because example-level contribution is not directly observed.Provenance records data origins, not how model behavior would change if the data were absent.Source: §1 · p. 1
- The study uses O’PRIOR, where each synthetic task records structural mechanisms and interventions for missingness, confounding, spurious correlation, and distribution shift.The authors repurpose its controllable lineage as a provenance-rich testbed for attribution evaluation.Source: §1 · pp. 1–2
- Removing Influence’s top-attributed 5% decreases mean ROC-AUC by 0.013, versus 0.002 ± 0.004 under random removal; bottom removal slightly improves performance.Within shortcut provenance, Influence has the largest top-vs-random counterfactual gap (0.030).Source: §1 · p. 2
- Provenance discrimination remains modest at macro ROC-AUC 0.55–0.62, and TracIn’s shortcut Lift@5% is 1.81× versus Influence’s 1.60× despite weaker interventional faithfulness.The study compares methods across intervention-derived effects, provenance association, and reliability, which need not rank methods alike.Source: §1 · p. 2
- The framework links downstream behavior to influential pretraining tasks and then to those tasks’ known generative provenance.It treats task attribution, recorded provenance, and behavioral effects of removal and retraining as distinct quantities.Source: §1 · p. 2
- Counterfactual task removal and provenance-conditioned interventions test attribution against behavioral effects rather than plausibility alone.The study compares highly attributed task removals with equal-sized random or weakly attributed subsets.Source: §1 · p. 2
2 Provenance-Grounded Attribution
The framework separates recorded task provenance, behavior-conditioned attribution scores, and the effects of removing tasks and retraining. It uses this separation to test attribution faithfulness, provenance association, and reliability without treating provenance as causal ground truth.
2.1 Setup and behavior-conditioned attribution
- Each frozen synthetic task carries compositional provenance for mechanisms such as missingness, confounding, shortcuts, and support–query shift.O’PRIOR controls these properties independently, and provenance records active mechanisms and their severities.Source: §2.1 · p. 2
- Attribution is conditioned on a behavior measured on held-out real tasks, rather than assigned globally to the model.The framework compares Influence, TracIn, TRAK, and a representation-similarity baseline; attribution scores remain hypotheses about influence, not ground truth.Source: §2.1 · p. 3
2.2 Interventional validation and provenance analysis
- Removal faithfulness is tested by removing ranked tasks and retraining while holding the architecture, optimizer, and evaluation protocol fixed.The resulting effect is specific to this finite-corpus intervention under the training algorithm, not an intrinsic influence value independent of the learning procedure.Source: §2.2 · p. 3
- Because mechanisms co-occur, the framework tests within-provenance faithfulness by comparing top-attributed tasks with an equal number of random tasks inside the same mechanism stratum.This tests whether ranking identifies consequential instances beyond mechanism membership alone.Source: §2.2 · p. 3
2.3 Reliability
- Reliability analysis compares task-level ranking stability with stability after aggregating attribution by provenance across seeds or perturbations.The comparison tests whether mechanism-level aggregation yields a more reproducible signal than individual-task rankings.Source: §2.3 · p. 4
2 Provenance-Grounded Attribution
- The framework jointly evaluates task attribution, behavior change under counterfactual removal, and known provenance, keeping these three quantities distinct.It distinguishes removal faithfulness, provenance association, and attribution reliability, which need not coincide.Source: §2 · p. 2, §2.3 · p. 4
3 Experiments
The experiments use a frozen O’PRIOR corpus and counterfactual retraining to evaluate task-level faithfulness, provenance association, behavior specificity, and attribution reliability. The results show that these evaluation axes can differ, and that provenance aggregation modestly improves stability.
3.2 Task-level faithfulness and provenance association
- At r = 5%, removing top-attributed tasks decreases mean ROC-AUC by 0.013 versus 0.002±0.004 under random removal; bottom-task removal instead improves performance by 0.003.At r = 10%, the top-versus-random gap persists, and Influence has the largest gap among the compared estimators.Source: §3.2 · p. 5, Figure 1 · p. 5
- Behavior-matched provenance discrimination is modest at approximately 0.60 ROC-AUC, while matched Lift@5% is approximately 1.3× overall and shortcut Lift@5% reaches 1.60× under Influence.The full ranking is not interpreted as recovering the generative mechanism; provenance information is clearest among highly attributed tasks.Source: §3.2 · p. 6, Figure 2 · p. 6
- Within shortcut-generated tasks, Influence yields a top-versus-random removal gap of +0.030, showing ranking information beyond shortcut-provenance membership.Other behavior–mechanism pairs show heterogeneous or inconclusive separation, so within-provenance faithfulness is not universal.Source: Figure 1 · p. 5, §3.2 · p. 6
3.3 Behavior specificity and estimator disagreement
- Across graded missingness, covariate-shift, and shortcut interventions, behavior-matched provenance changes with severity, with ρsev ≥0.63 over at least four severity levels.The response varies by mechanism; some controlled behaviors produce weaker or control-indistinguishable shifts.Source: Figure 2 · p. 6, §3.3 · p. 7
- Shortcut provenance enrichment and counterfactual faithfulness rank estimators differently: TracIn has higher Lift@5% than Influence, while Influence has the larger within-shortcut faithfulness gap.The paper treats provenance association as evidence of where attribution concentrates and counterfactual retraining as evidence of behavioral consequence.Source: §3.3 · p. 7
3.4 Attribution reliability
- Provenance aggregation modestly improves mean pairwise Spearman stability for every estimator, including Influence from 0.85 at task level to 0.91 at provenance level.Stability under repeated estimation is distinct from agreement across attribution assumptions.Source: Figure 3 · p. 7, §3.4 · p. 8
4 Conclusion
The framework separates task attributions, recorded provenance, and removal effects to test contributive attribution through intervention. Results show that removal faithfulness and provenance association are distinct, while the study’s conclusions remain bounded by its controlled setting.
4 Conclusion
- The framework keeps task attribution, recorded generative provenance, and behavioral effects of removal and retraining separate rather than treating either scores or provenance as ground truth.Source: §4 · p. 8
- Influence shows the strongest counterfactual removal faithfulness, while TracIn has greater shortcut-provenance enrichment despite a smaller within-provenance counterfactual gap.TracIn’s shortcut-provenance enrichment is 1.81× versus 1.60× for Influence; their within-provenance counterfactual gaps are 0.023 and 0.030, respectively.Source: §4 · p. 8
- Synthetic provenance provides a structured reference for experimentally testing attribution claims, but records how tasks were generated rather than whether they caused learned behavior.Source: §4 · p. 8
- Conclusions are bounded by a finite corpus of 5,000 O’PRIOR tasks and a NanoTabPFN-scale model; relationships may change with larger models or corpora, or natural-data pretraining.Removal effects also depend on the optimizer, training budget, initialization, and retraining procedure, while task-set interactions mean subset effects need not sum across members.Source: §4 · p. 8
- Future work should extend evaluation to larger models, richer multi-generation provenance, and settings combining synthetic and natural pretraining data.Cheaper validation procedures that retain retraining-based evidential value are also identified as a direction for larger corpora.Source: §4 · p. 8