acceptodds
Withdrawn from ICLR 2027

In-Run Data Shapley for Adam Optimizer

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reliable data attribution is essential for understanding, debugging, and curating modern machine learning training data, with the Shapley value serving as a principled framework for data valuation. Recent In-Run Data Shapley methods avoid the prohibitive cost of retraining by decomposing data value into step-wise contributions along a single realized training trajectory. However, existing scalable in-run estimators are derived from the linear update structure of stochastic gradient descent (SGD), while modern deep learning pipelines are widely trained with adaptive optimizers such as Adam and AdamW. In this work, we show that data attribution is optimizer-dependent: SGD-induced and Adam-induced data values exhibit extremely low agreement in both magnitude and ranking, indicating that SGD-based attribution can lead to misleading rank-based decisions under Adam training. To address this mismatch, we propose Adam-Aware In-Run Data Shapley, an optimizer-consistent attribution framework for Adam training. We define a fixed-state Adam local utility and derive a first-order approximation that explicitly accounts for momentum and coordinate-wise variance normalization. To make the estimator scalable, we introduce a Linearized Ghost Approximation, which restores an additive dot-product form without materializing per-sample gradients. Experiments show that our method achieves near-perfect fidelity to exact local Shapley values under the fixed-state Adam utility (), substantially improves efficiency over direct per-sample Adam-aware computation, and improves downstream rank-based attribution tasks including semantic source identification and data pruning.

Primary area:
interpretability and explainable AI
Keywords:
Shapley Value

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_T5NhOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends the In-Run Data Shapley attribution method from SGD to Adam. Under SGD, In-Run Data Shapley credits a training example, at each step it appears in, with the dot product between its gradient and the validation loss gradient. To apply the same per-step Shapley decomposition under Adam, the authors linearize the Adam update at the batch gradient with the previous step's moment estimates held fixed. This yields an approximate Adam Shapley value that, up to a momentum-dependent correction, is the same dot product as under SGD with the validation loss gradient divided coordinate-wise by Adam's running gradient scale. It can therefore still be obtained without explicitly computing per-example gradients. Experimentally, the authors report that the SGD score is a poor proxy on Adam runs, and that their score gives better attributions.

Critical strengths

The approach is a sound way to obtain an approximate in-run Shapley value under Adam. The problem is relevant given that Adam is a common optimizer choice in deep learning practice and the SGD-based score is not designed for it.

Critical weaknesses

The operation the method arrives at is what LESS [1, Section 3.1] and TrackStar [2, Section 3] already do: dividing gradients coordinate-wise by the square root of the optimizer’s second-moment estimate before the dot product. Neither work is cited. That an Adam-aware score matches or beats the SGD score on Adam-trained models was also shown by LESS [1, Table 9]. The momentum-dependent term the submission adds on top of LESS’s operation is never ablated.

Contribution 1 is overstated. It is unsurprising that SGD-based attributions are not good for Adam-trained models: attribution is a property of the training algorithm used. Also, the experiment supporting this claim (Figure 1) lacks the basic control of comparing attributions across seeds under the same optimizer. Those can be weakly correlated too, so the near-zero correlation reported may just be noise.

Contribution 4 is overstated too. The near-perfect fidelity only shows that the proposed linearization is faithful to the per-step Adam Shapley values computed without it. It does not show that the method produces good attributions, which the paper never evaluates the standard way, against retraining-based attributions, except on a toy task.

What seems novel is the derivation through a linearization, which also lets the score be obtained without explicitly computing per-example gradients.

[1] Xia et al. LESS: Selecting Influential Data for Targeted Instruction Tuning. ICML, 2024.

[2] Chang et al. Scalable Influence and Fact Tracing for Large Language Model Pretraining. ICLR, 2025.

Questions for authors
  1. Which values of Adam's beta_1 and beta_2 were used in each experiment?
  2. Figure 1 shows about thirty points, while Appendix B says 10,000 samples were tracked to produce the figure. What does each point represent?
Reviewer_r336OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors in this paper investigated the reliable data attribution problem. They show that data attribution is optimizer-dependent: SGD-induced and Adam-induced data values exhibit fundamental mismatch. They propose Adam-Aware In-Run Data Shapley to address the mismatch.

Critical strengths

The paper investigated an important and interesting problem. The contribution and problem setting are stated clearly in this paper.

Critical weaknesses

This paper only compared two algorithms, SGD and Adam. From my point, it is more persuable to include more benchmark optimizers, like AdaGrad and Muon.

Questions for authors

Can you provide more benchmark optimizers, like AdaGrad and Muon?

Reviewer_7sPDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper is motivated by a practical scenario in which existing In-Run Shapley value (SV) methods rely on SGD-based parameter updates, while many deep learning models are trained using Adam-based optimizers. This mismatch may lead to inaccurate attribution results. The concern itself is interesting but the experimental design and results do not provide sufficient evidence to support the claim that the proposed approximation is necessary or effective. As a result, substantial additional work and empirical evidence are required to improve the quality of the paper. See the points below.

Critical strengths

The motivation of this paper arises from a practical concern that existing SV methods are based on SGD updates, while many deep learning models are trained using Adam-based optimizers. Whether this mismatch could lead to misleading or inaccurate attribution is an interesting question worth investigating.

Critical weaknesses

1.Section 4.1 attempts to verify that the motivation is indeed a concern, but what is actually reported is the Pearson and Spearman correlations between SGD-based and Adam-based Shapley values obtained by training the same model architecture on the same dataset using SGD and Adam, respectively. However, phi(U_{SGD}) and phi(U_{Adam}) can both be valid Shapley value vectors, and there is no reason to expect phi(U_{SGD})=phi(U_{Adam}), since they correspond to different training procedures and potentially different trained models. Therefore, this experiment cannot establish that SGD-based attribution is unreliable when applied to Adam-trained models, and the motivation is not adequately supported by this experiment.

  1. In summary, the model training optimizer could be either SGD or Adam, and the attribution method could be either SGD-based or Adam-based Shapley estimation, resulting in a total of 4 scenarios. The experiments should be systematically designed to evaluate these four scenarios and compare their performance against the corresponding optimizer-specific ground-truth SV. However, the current experiments do not provide such a systematic comparison. Therefore, the experimental evidence is insufficient to fully support the paper's claims about optimizer-consistent attribution.
Questions for authors

See above weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.