Training data attribution (TDA) should enable generative model interpretability and foster a variety of related downstream tasks. Nonetheless, current TDA approaches lack reliability and robustness, preventing their adoption in real-world setups. In this paper, we take a decisive step towards more reliable and robust TDA for diffusion models. We propose to perform TDA with mirrored unlearning and noise-consistent skew (MUCS). The idea is to fine-tune a second model with bounded mirrored gradient ascent, and to measure the normalized skew of this model with respect to the original one using consistent noise samples. We show that, while being conceptually simple and generic, MUCS systematically outperforms existing methods on three different datasets by a large margin. We additionally study the effect that core design choices have on final performance, and analyze novel aspects regarding the overlap of influential instances across generated items and the potential of ensembling TDA approaches. We believe that our findings may have broader implications for more general unlearning setups, as well as for tasks requiring the comparison of diffusion losses.

Data attribution aims to explain model predictions by estimating how they would change if certain training points were removed, and is used in a wide range of applications, from interpretability and credit assignment to unlearning and privacy. Even in the relatively simple case of linear regressions, existing mathematical analyses of leading data attribution methods such as Influence Functions (IF) and single Newton Step (NS) remain limited in two key ways. First, they rely on global strong convexity assumptions which are often not satisfied in practice. Second, the resulting bounds scale very poorly with the number of parameters ($d$) and the number of samples removed ($k$). As a result, these analyses are not tight enough to answer fundamental questions such as "what is the asymptotic scaling of the errors of each method?" or "which of these methods is more accurate for a given dataset?" In this paper, we introduce a new analysis of the NS and IF data attribution methods for convex learning problems. To the best of our knowledge, this is the first analysis of these questions that does not assume global strong convexity and also the first explanation of [KATL19] and [RH25a]'s observation that NS data attribution is often more accurate than IF. We prove that for sufficiently well-behaved logistic regression, our bounds are asymptotically tight up to poly-logarithmic factors, yielding scaling laws for the errors in the average-case sample removals. \[ \mathbb{E}_{T \subseteq [n],\, |T| = k} \bigl[ \|\hatθ_T - \hatθ_T^{\mathrm{NS}}\|_2 \bigr] = \widetildeΘ\!\left(\frac{k d}{n^2}\right), \qquad \mathbb{E}_{T \subseteq [n],\, |T| = k} \bigl[ \|\hatθ_T^{\mathrm{NS}} - \hatθ_T^{\mathrm{IF}}\|_2 \bigr] = \widetildeΘ\!\left( \frac{(k + d)\sqrt{k d}}{n^2} \right). \]

Training data attribution (TDA) methods aim to measure how training data impacts a model's predictions. While gradient-based attribution methods, such as influence functions, offer theoretical grounding, their computational costs make them impractical for large-scale applications. Representation-based approaches are far more scalable, but typically rely on heuristic embeddings that are not optimized for attribution, limiting their fidelity. To address these challenges, we propose AirRep, a scalable, representation-based approach that closes this gap by learning task-specific and model-aligned representations optimized explicitly for TDA. AirRep introduces two key innovations: a trainable encoder tuned for attribution quality, and an attention-based pooling mechanism that enables accurate estimation of group-wise influence. We train AirRep using a ranking objective over automatically constructed training subsets labeled by their empirical effect on target predictions. Experiments on instruction-tuned LLMs demonstrate that AirRep achieves performance on par with state-of-the-art gradient-based approaches while being nearly two orders of magnitude more efficient at inference time. Further analysis highlights its robustness and generalization across tasks and models. Our code is available at https://github.com/sunnweiwei/AirRep

Unlearning in large language models is becoming increasingly important due to regulatory compliance, copyright protection, and privacy concerns. However, a key challenge in LLM unlearning is unintended forgetting, where the removal of specific data inadvertently impairs the utility of the model and its retention of valuable, desired information. While prior work has primarily focused on architectural innovations, the influence of data-level factors on unlearning performance remains underexplored. As a result, existing methods often suffer from degraded retention when forgetting high-impact data. To address this problem, we propose GUARD, a novel framework for Guided Unlearning And Retention via Data attribution. At its core, GUARD introduces a lightweight proxy data attribution metric tailored for LLM unlearning, which quantifies the alignment between the Forget and Retain sets while remaining computationally efficient. Building on this, we design a novel unlearning objective that assigns adaptive, nonuniform unlearning weights to samples, inversely proportional to their proxy attribution scores. Through such a reallocation of unlearning power, GUARD mitigates unintended retention loss. We also provide rigorous theoretical guarantees that GUARD significantly improves retention while maintaining forgetting metrics comparable to prior methods. Extensive experiments on the TOFU and MUSE benchmarks across multiple LLM architectures demonstrate that GUARD reduces utility sacrifice on the TOFU Retain Set by up to 194.92 percent in terms of Truth Ratio when forgetting 10 percent of the training data, and improves knowledge retention on the MUSE NEWS Retain Set by 16.20 percent, with comparable or very moderate increases in privacy loss compared to state-of-the-art methods.

Data attribution for generative models seeks to quantify the influence of individual training examples on model outputs. Existing methods for diffusion models typically require access to model gradients or retraining, limiting their applicability in proprietary or large-scale settings. We propose a nonparametric attribution method that operates entirely on data, measuring influence via patch-level similarity between generated and training images. Our approach is grounded in the analytical form of the optimal score function and naturally extends to multiscale representations, while remaining computationally efficient through convolution-based acceleration. In addition to producing spatially interpretable attributions, our framework uncovers patterns that reflect intrinsic relationships between training data and outputs, independent of any specific model. Experiments demonstrate that our method achieves strong attribution performance, closely matching gradient-based approaches and substantially outperforming existing nonparametric baselines. Code is available at https://github.com/sail-sg/NDA.

Influence functions offer a principled way to trace model predictions back to training data, but their use in deep learning is hampered by the need to invert a large, ill-conditioned Hessian matrix. Approximations such as Generalised Gauss-Newton (GGN) and Kronecker-Factored Approximate Curvature (K-FAC) have been proposed to make influence computation tractable, yet it remains unclear how the departure from exactness impacts data attribution performance. Critically, given the restricted regime in which influence functions are derived, it is not necessarily clear better Hessian approximations should even lead to better data attribution performance. In this paper, we investigate the effect of Hessian approximation quality on influence-function attributions in a controlled classification setting. Our experiments show that better Hessian approximations consistently yield better influence score quality, offering justification for recent research efforts towards that end. We further decompose the approximation steps for recent Hessian approximation methods and evaluate each step's influence on attribution accuracy. Notably, the mismatch between K-FAC eigenvalues and GGN/EK-FAC eigenvalues accounts for the majority of the error and influence loss. These findings highlight which approximations are most critical, guiding future efforts to balance computational tractability and attribution accuracy.

Training data attribution (TDA) plays a critical role in understanding the influence of individual training data points on model predictions. Gradient-based TDA methods, popularized by \textit{influence function} for their superior performance, have been widely applied in data selection, data cleaning, data economics, and fact tracing. However, in real-world scenarios where commercial models are not publicly accessible and computational resources are limited, existing TDA methods are often constrained by their reliance on full model access and high computational costs. This poses significant challenges to the broader adoption of TDA in practical applications. In this work, we present a systematic study of TDA methods under various access and resource constraints. We investigate the feasibility of performing TDA under varying levels of access constraints by leveraging appropriately designed solutions such as proxy models. Besides, we demonstrate that attribution scores obtained from models without prior training on the target dataset remain informative across a range of tasks, which is useful for scenarios where computational resources are limited. Our findings provide practical guidance for deploying TDA in real-world environments, aiming to improve feasibility and efficiency under limited access.

Data attribution methods aim to answer useful counterfactual questions like "what would a ML model's prediction be if it were trained on a different dataset?" However, estimation of data attribution models through techniques like empirical influence or "datamodeling" remains very computationally expensive. This causes a critical trust issue: if only a few computationally rich parties can obtain data attributions, how can resource-constrained parties trust that the provided attributions are indeed "good," especially when they are used for important downstream applications (e.g., data pricing)? In this paper, we address this trust issue by proposing an interactive verification paradigm for data attribution. An untrusted and computationally powerful Prover learns data attributions, and then engages in an interactive proof with a resource-constrained Verifier. Our main result is a protocol that provides formal completeness, soundness, and efficiency guarantees in the sense of Probably-Approximately-Correct (PAC) verification. Specifically, if both Prover and Verifier follow the protocol, the Verifier accepts data attributions that are ε-close to the optimal data attributions (in terms of the Mean Squared Error) with probability 1-δ. Conversely, if the Prover arbitrarily deviates from the protocol, even with infinite compute, then this is detected (or it still yields data attributions to the Verifier) except with probability δ. Importantly, our protocol ensures the Verifier's workload, measured by the number of independent model retrainings it must perform, scales only as O(1/ε); i.e., independently of the dataset size. At a technical level, our results apply to efficiently verifying any linear function over the boolean hypercube computed by the Prover, making them broadly applicable to various attribution tasks.

The goal of data attribution is to trace the model's predictions through the learning algorithm and back to its training data. thereby identifying the most influential training samples and understanding how the model's behavior leads to particular predictions. Understanding how individual training examples influence a model's predictions is fundamental for machine learning interpretability, data debugging, and model accountability. Influence functions, originating from robust statistics, offer an efficient, first-order approximation to estimate the impact of marginally upweighting or removing a data point on a model's learned parameters and its subsequent predictions, without the need for expensive retraining. This paper comprehensively reviews the data attribution capability of influence functions in deep learning. We discuss their theoretical foundations, recent algorithmic advances for efficient inverse-Hessian-vector product estimation, and evaluate their effectiveness for data attribution and mislabel detection. Finally, highlighting current challenges and promising directions for unleashing the huge potential of influence functions in large-scale, real-world deep learning scenarios.

While diffusion models excel at image generation, their growing adoption raises critical concerns around copyright issues and model transparency. Existing attribution methods identify training examples influencing an entire image, but fall short in isolating contributions to specific elements, such as styles or objects, that matter most to stakeholders. To bridge this gap, we introduce \emph{concept-level attribution} via a novel method called \emph{Concept-TRAK}. Concept-TRAK extends influence functions with two key innovations: (1) a reformulated diffusion training loss based on diffusion posterior sampling, enabling robust, sample-specific attribution; and (2) a concept-aware reward function that emphasizes semantic relevance. We evaluate Concept-TRAK on the AbC benchmark, showing substantial improvements over prior methods. Through diverse case studies--ranging from identifying IP-protected and unsafe content to analyzing prompt engineering and compositional learning--we demonstrate how concept-level attribution yields actionable insights for responsible generative AI development and governance.