Overview

Training data attribution studies how particular examples or datasets affect a model’s behavior, helping researchers debug errors, audit provenance, select useful data, and assess safety. Common approaches include influence functions and gradient similarity, Shapley-style valuation, counterfactual retraining, and membership or provenance tests; newer methods use activations, training trajectories, and compressed gradient representations to scale attribution to large models. Evaluations span language-model knowledge and alignment, image and music generation, classification, and data selection, often comparing scores with retraining or known-inclusion benchmarks. A central challenge is that attribution depends on the counterfactual being asked about, while estimates can be costly, sensitive to interactions and evaluation choices, or confounded by similarity that does not establish causal influence.

Latest papers

92 papers10 papers in the last four weeks, down 52% on the four weeks before. 0.1% of all new papers.

Nov 15, 2026

Paying Creators for Training Data Can Increase AI Displacement

Generative AI increasingly competes with creators while using their content as training data. In response, policymakers have called for platforms to compensate creators for training data use and allow opting out. We study how such policies affect creator displacement in a model of a platform where creators choose whether to contribute training data and consumers choose between purchasing from creators and prompting generative AI trained on their content. When compensation is required, the low-quality creator chooses to remain solely as a training data supplier while AI fully displaces her demand, despite AI quality and the price the platform charges for AI-generated content being identical with and without compensation. Without compensation, the risk of losing creators and their training data constrains the platform to set the price of creators' content to preserve their demand. This constraint also prevents further AI displacement as AI improves, with the platform raising AI prices while keeping creators' prices and demand unchanged to prevent exit. Compensation increases creators' payoffs and total surplus but reduces consumer surplus. Our findings highlight how compensation policies for training data can shape creators' economic roles in a future with more capable generative AI.
Nov 15, 2026

Data Attribution Should Steer Pretraining

Pretraining data curation decides what a model learns from, yet what makes a document worth learning from depends on what the model already knows. Influence-driven selectors already rank documents to improve one run; we argue that training-data attribution should steer pretraining by identifying testable data properties that become reusable curation criteria, evaluated on new data against cheaper selectors at matched total cost, discovery included. A document ranking is consumed by the run it was computed for, whereas an explained data property can be tested across corpora and checkpoints, especially where familiar quality signals diverge from measured usefulness. Because historical contribution, an understandable explanation, and future data value require distinct evidence, we specify a feedback loop that interprets attributed sources, tests candidate data changes, and states for five claim types what evidence each requires and what none establishes. We close with a falsifiable comparison against perplexity, classifier, and mixing baselines at equal total compute.
Nov 15, 2026

MathaMAGICal: Validating Trajectory Data Attribution During Capability Acquisition

Influence-function-based methods like MAGIC provide exact first-order scores for trajectory data attribution, yet only retraining indicates whether those scores accurately predict data removal. We evaluate MAGIC on 8686M- and 152152M-parameter GPT-2-style models learning multistep arithmetic from scratch on filtered web text and four synthetic math families. Full-run attribution at 8686M suffers from numerical overflow, and even with metagradient clipping, scores correlate near chance with retraining (linear datamodeling score, LDS, 0.170.17--0.220.22). In contrast, short checkpoint windows yield stable, accurate predictions at both scales (LDS 0.860.86--0.990.99 across four 100100-step windows at 152152M), whereas a single-checkpoint gradient-dot baseline shows near-zero correlation. Rankings of individual training blocks predict retraining only within that single run and do not transfer across seeds or hyperparameters (ρ≤0.09ρ\le 0.09), whereas format-level patterns replicate across runs. Two data families, worked examples and primitive facts, occupy 2.12.1--2.6×2.6\times their corpus share in both the helpful and harmful 5%5\% score tails, reflecting strong bidirectional local sensitivity distinct from net benefit under multi-epoch family deletion. These results show that trajectory attribution is reliable within short, validated windows and yields family-level insights that generalize across runs.
Nov 14, 2026

Efficient and Accurate Causal Validation of Training Data Attribution

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%3.2\% relative RMSE with as few as 44 subsets, and 3232 subsets match the accuracy of ordinary Spearman using 9696 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 22--5×5\times fewer retrainings.
Nov 14, 2026

On Representational and Functional Alignment of Training Data Attribution Methods for Language Models

Existing approaches to training data attribution derive influence estimates from different sources of information, including gradients, model-internal states, and surface-level similarity. However, comparisons between attribution methods remain incomplete, often relying on indirect evaluations through downstream task performance. In this work, we introduce a framework for directly comparing methods along three dimensions: functional alignment, which measures correlation between full attribution vectors; representational alignment, which captures structural similarity of the representations underlying attribution; and cross-method predictability, which evaluates agreement on local pairwise ranking decisions. We observe strong representational and moderate functional alignment within paradigms. Alignment across paradigms is low and task-dependent, yet stable across models. Our results suggest that the similar downstream performance observed across paradigms in prior work can be attributed to agreement on the relative importance of documents, rather than to agreement on rankings of the full training data.
Nov 6, 2026

Functional Dependency beyond Similarity: Hierarchical Decomposition of Training Data Attribution in Music Generative Models

Copyright analysis focuses on the similarity between two works. However, generative AI models learn from training data through functional relations beyond similarity. Training Data Attribution (TDA) estimates the influence of each training instance on a generated output and may serve as a basis for remuneration. However, prior studies find global influence, where a few training instances rank as the most influential for almost every generation. Used directly, TDA scores would distribute part of every remuneration to these instances regardless of what was generated. To address this, we propose a new approach in which TDA scores are decomposed into three levels of influence, namely global indicating the influence of a training data to a trained model, group for its influence on a group of generated outputs (that could be clustered together), and specific for its influence on a single generated output. Our study finds that, across several music generative models and TDA methods, global- and group-level influence is non-negligible in every case. Based on these observations, we propose the term functional dependency for the relation of training data to a trained model or to its outputs. This notion both encompasses and distinguishes dependencies of different scopes, allowing TDA scores to be interpreted as distinct forms of contribution in legal, economic and ethical discussions on remuneration.
Oct 29, 2026

Elucidating the Design Space of LM Data Attribution

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.
Oct 27, 2026

Robust LLM Data Selection via Content-Invariant Gradient Attribution

Gradient-based data attribution has become an important tool for high-quality data selection in the instruction fine-tuning of large language models (LLMs), as it estimates the contribution of training examples through their influence on model optimization. However, we find that existing global gradient attribution methods suffer from surface-form bias: minor changes to semantically irrelevant details, such as wording or formatting, can produce highly inconsistent attribution scores even when the underlying content remains unchanged. To address this issue, we propose Content-Invariant Gradient Attribution (CIGA), a robust framework for instruction data selection. CIGA first identifies layers whose gradients are both influential for attribution and stable under semantics-preserving surface-form perturbations, thereby reducing formatting noise in gradient features. It then dynamically fuses these attribution signals with semantic representations of the input, combining optimization-level influence with content-level information. Extensive experiments across multiple benchmarks show that CIGA produces more stable rankings under surface-form perturbations while preserving downstream capabilities and improving average performance. CIGA also reduces attribution cost, offering a robust and efficient approach to data selection for LLM instruction tuning.
Oct 20, 2026

RL-IF: Tracking Non-local Training Data Influence for Online Reinforcement Learning

Online reinforcement learning (RL) has achieved significant success in decision-making tasks, but modern online RL remains highly sensitive to the quality of training experience. Understanding the contribution of individual training instances is therefore central to data-centric RL, with applications ranging from experience selection to policy interpretation. Existing data attribution methods for online RL primarily focus on local influence within a single training round, overlooking the non-local data influence across multiple training rounds. In this work, we formalize data attribution in online RL through a trajectory-level leave-one-out quantity, RL-LOO, which measures how one training sample influences a downstream target after subsequent online updates have taken place. We then derive RL-IF, a first-order estimator that propagates data influence through both optimization effects and policy-induced sampling effects, and provide theoretical guarantees on its approximation accuracy under smoothness assumptions. Our analysis further disentangles these two effects and shows both theoretically and empirically that the full RL-IF estimator can often be well approximated using the optimization effect alone. Building on this observation, we develop RINSE, a practical data filtering method for online RL. Experiments on standard control benchmarks and an RLHF-style toxicity-mitigation setting show that RINSE improves training efficiency and final performance over standard PPO and local attribution baselines. These results suggest that RL-IF provides a practical and principled framework for understanding and improving online RL.
Oct 20, 2026

Influence Is the Target-Response Jacobian Times Residuals

We introduce the target-response Jacobian J=[∂y^i/∂yj]J = [\partial \hat y_i/\partial y_j], which measures how model predictions respond to changes in individual training targets. JJ complements training-data attribution (TDA) methods, which characterize sensitivity to the weights placed on training observations in the loss rather than to their targets, and can be computed matrix-free at the same cost as influence-based TDA methods. We show that sample-weight influence I\mathcal{I} and JJ follow from the same implicit differentiation argument and differ only by a per-example multiplier: for pointwise losses, I=Jdiag⁡(m)\mathcal{I} = J\operatorname{diag}(m) exactly at any stationary point with an invertible Hessian. For canonical-link losses (squared error, logistic, Poisson) the multiplier is the training residual, I=Jdiag⁡(r)\mathcal{I} = J\operatorname{diag}(r), and softmax cross-entropy obeys a multiclass version. The identity predicts three ways in which influence fails when used in place of JJ: small residuals shrink its predicted shift towards zero (interpolation collapse), residuals of mixed sign wash out its direction (direction washout), and under softmax cross-entropy it collapses a training point's per-class sensitivities towards its own true class (true-class collapse). We score both operators against refitting with a per-prediction R2R^2 across target perturbations, with no rescaling. On tanh networks trained with squared error (Friedman #1) and logistic loss (MAGIC, HIGGS), with NN from 2,000 to 10610^6, JJ reaches R2R^2 between 0.99 and 1.00, whereas influence's R2R^2 is negative on every dataset. On an MNIST softmax classifier, influence's R2R^2 is again negative, while JJ reaches 0.92. On HIGGS, one row of JJ, holding all 10610^6 sensitivities of one prediction, takes under 0.2 seconds on an RTX 3090 after a one-time setup of under five minutes, whereas refitting would take 10610^6 refits.
Oct 17, 2026

Tensor Network Datamodels for Training Data Attribution

Datamodels predict how a model's behavior changes when different parts of its training data are included or removed. However, linear datamodels assign fixed contributions to training units and cannot represent non-additive joint effects. We introduce Tensor Network Datamodels (TNDMs), which use low-rank tensor networks to model dependencies among training units while keeping the representation compact. TNDMs learn a map from subsets of the training data to the outcome of training on those subsets, allowing them to represent interactions among several training units. We prove that complementary and redundant training groups can induce high-degree subset functions with compact low-rank representations. A single fitted TNDM can then be used for several training data questions, including predicting the effect of removing data, measuring the importance of individual training units, identifying interactions between training units, and selecting useful subsets of the training data. We evaluate TNDMs on controlled tasks designed to contain strong interactions, as well as standard fine-tuning settings for classification and reasoning. On a controlled reasoning task, TNDMs outperform every baseline while observing only half as many training subsets, and TNDM-TT reduces prediction NMSE by 89%89\% relative to the linear datamodel after observing approximately 0.0061%0.0061\% of all possible training subsets. In factual-association fine-tuning, a single TNDM-CP raises held-out R2\mathrm{R}^2 from 0.5790.579 to 0.6680.668 over the linear datamodel. These results demonstrate that compact low-rank models improve prediction when training data effects depend on combinations of training groups.
Oct 16, 2026

Dataset Inference for Data Provenance and Privacy Auditing in Tabular Foundation Models

Tabular foundation models (TFMs) are increasingly deployed through black-box APIs and trained on large collections of real-world tabular datasets. As this training paradigm becomes more widely adopted, private or proprietary datasets may be incorporated into pre-training corpora, yet there are currently no dedicated methods for determining whether a given tabular dataset was used to train a TFM. As a solution, we introduce the first dataset inference method for TFMs, aiming to infer whether a suspect dataset was part of a model's pre-training data. We systematically analyze a broad collection of candidate signals that can be observed from a black-box TFM via input manipulation and find that individual signals provide only weak evidence of training data membership. However, by combining multiple complementary signals, we can infer dataset membership for several state-of-the-art TFMs trained on real tabular data, achieving up to 0.99 ROCAUC dependent on the TFM. We then study factors that influence dataset identification, including pre-training data composition, model capacity, and the use of real vs. synthetic data. Our results show that our dataset inference method is a practical auditing tool for detecting privacy leakage and the use of proprietary datasets in TFMs.
Oct 10, 2026

Groups Don't Move as One: A Bregman Clustering View of Linearized Group Influence

Influence functions provide an efficient alternative to retraining for assessing the effects of removing or reweighting training data. For groups, however, the standard first-order estimate reduces individual contributions to a sum that can conceal substantial differences among members, including opposing effects that cancel. We establish a Bregman centroid formulation of the linearized group influence objective, interpreting this aggregate as a single representative and using Bregman information (BI) to quantify the variation it omits. Extending this formulation to multiple clusters reveals distinct contributions whose sum recovers the original estimate. We further propose an estimator that jointly optimizes the weights of these cluster contributions, rather than simply summing them. Experiments uncover cluster-specific influence patterns and a positive association between target-specific BI and approximation error. The proposed estimator achieves higher R2R^2 than standard group influence in predicting target changes under the nonlinear proximal Bregman response function. Simplified clustering and correction variants reduce computational cost. The centroid view thus explains what additive group influence leaves out and provides a basis for using that information to improve estimation.
Oct 9, 2026

Cross-Modal Latent Influence: Patch-Level Training Sample Attribution for Vision-Language Models

Explaining a vision--language model (VLM) prediction requires tracing training-example influence to the multimodal evidence that mediates it, yet existing attribution methods largely operate only at the sample level. Recent latent-mediated influence methods provide feature-level attribution in language models, but lack the multimodal and spatial structure needed to ground influence in VLMs. We introduce MOSAIC (Multimodal Orthogonalized Sparse Autoencoders for Influence Characterization). It learns a shared sparse image--text latent space whose concepts are 1) sparse within each modality, 2) supported on contiguous image regions, 3) aligned to the phrases that describe them, and 4) approximately decorrelated. It then applies EK-FAC to decompose training-example influence through these concepts and localize their contributions to specific text phrases and image regions. We evaluate MOSAIC on Qwen and Gemma across captioning, chest X-ray understanding, and visual question answering. The learned features are more localized and coherent without degrading task behavior, and ablating influence-ranked latents changes model outputs more than ablating latents ranked by activation or frequency. We find that visual evidence can substantially reorganize training-data attribution while preserving more of the shared latent structure. MOSAIC thus connects where model behavior comes from with what multimodal evidence carries that influence.
Oct 8, 2026

Attributing HOW, Not Just WHICH: Counterfactual Response Trajectories for Diffusion Models

Diffusion models have achieved remarkable success in image generation, yet tracing their outputs to individual training examples remains challenging. Existing attribution methods often compress factor-specific effects into scalar responses, making distinct internal changes indistinguishable. This is particularly limiting for diffusion models, where semantic factors emerge through evolving representation dynamics during denoising. We therefore reformulate diffusion data attribution as attributing factor-induced internal response trajectories. In this paper, we propose a novel Concept Attribution method through Dynamic Trajectories(CADT). We argue that attribution should therefore ask not only which examples matter, but also how their influence unfolds during generation. Specifically, we construct matched counterfactual pairs at identical noisy states to isolate factor-specific representation displacements, and model their directional and magnitude evolution across denoising as dynamic attribution signatures. For each training example and generated query, CADT extracts stage-wise feature vectors and integrates them along the denoising process to form a trajectory descriptor. Applying the same construction across the training set yields a bank of factor-specific trajectory descriptors. The covariance statistics of this bank are then used to construct . CADT uses this covariance-aware positive-semidefinite kernel to calibrate the query and training representations, and compares the calibrated query trajectory with each training trajectory to produce the final training-sample attribution scores. Experiments on multiple public datasets show consistent improvements over existing diffusion attribution baselines across hierarchical, compositional, and style attribution.
Oct 7, 2026

Which Rollout Taught It That? BehaviorTrace and the Limits of Training-Data Attribution in Online RL

When reinforcement learning teaches a language model a new behavior, can we find the training rollouts that taught it? And when an attribution method says it can, how do we know the answer is real? We study both questions on online RL fine-tuning with GRPO, using a planted behavior with a known cause. We release BehaviorTrace, an open evaluation harness that combines full-gradient sketching, the planted-behavior setup, and controls for gradient magnitude, fluency, headroom, and variation across seeds and generation draws. Across three seeds on Qwen2.5-1.5B, much of the apparent attribution signal comes from confounds. A control that ranks training steps by gradient size alone, with no behavior target, reaches 4.2 to 4.5 times chance and matches or beats the best targeted estimator on two of three seeds. At saturated checkpoints, model fluency predicts the behavior label at least as well as every gradient method we compared it with. Once fluency is controlled, the per-rollout results change from seed to seed and from one generation draw to the next, so a single run cannot settle the question. One signal does hold on all three seeds. The gradient of the trigger tokens aligns with a target built where the behavior actually occurs. We turn these findings into a checklist for evaluating attribution in RL. We test existing estimators, including GAS (renormalized TracInCP) and a TRAK-style estimator, and do not propose a new one.
Oct 6, 2026

TICDA: Tabular In-Context Data Attribution

Tabular foundation models (TFMs) achieve strong predictive performance by conditioning on labeled demonstrations provided in context, without any parameter update. Yet how individual demonstrations shape a given prediction remains poorly understood. This gap matters in practice: the context is often assembled from whatever labeled data is available, potentially leading to the inclusion of mislabeled, redundant, or low-quality examples that degrade performance. Standard data attribution methods do not transfer to the TFM setting: resampling-based approaches such as DemoShapley require a combinatorial number of forward passes, and gradient-based estimators such as influence functions require computing training point's effect on the model parameters, which in-context learning never updates. We introduce TICDA, a method that measures the influence of every demonstration in the context directly from linear surrogates trained on TFM latent embeddings, in a single forward pass and at negligible cost. We show that TICDA offers the best compromise against competitors across four tasks: detecting labeling errors, curating context to preserve predictive accuracy while lowering inference cost, producing attribution scores that transfer across TFMs, and supporting an acquisition strategy for efficient active learning.
Oct 2, 2026

When Does Synthetic Relational Data Teach Models to Use Relations? Tracing Predictive Structure from Pretraining Data to Model Behavior

Relational foundation models are increasingly pretrained on synthetic databases, yet downstream benchmarks reveal little about why one synthetic corpus produces a better model than another. In particular, strong performance may arise from realistic row-level statistics without the model ever learning to use relational structure. We study this as a data-attribution problem: which property of synthetic pretraining data induces relational computation? Using four Relational Transformer checkpoints trained with the same architecture, initialization, objective, and compute budget on corpora produced by four relational data generators, we trace a measurable property of the data to learned computation and downstream behavior. We hypothesize that relational mechanisms emerge when cross-table information is predictively necessary for the masked-cell pretraining objective. RelDiff exhibits by far the largest predictive gain from foreign-key-linked parents, and its corresponding model is uniquely sensitive to foreign-key interventions on unseen databases. This dependence survives a random-initialization control, grows monotonically with the fraction of corrupted links, and localizes to a serial cross-table pathway. Finally, disrupting the same mechanism during downstream inference removes RelDiff's advantage on relational tasks while leaving structure-insensitive models nearly unchanged. These results connect a property of synthetic training data to a learned mechanism and, through intervention, to downstream behavior.
Oct 2, 2026

LUMOS: Tracing Parametric Knowledge from Training Data to Behavioral Outputs in LLMs

Current analyses of LLMs' parametric knowledge are largely output-centric, drawing conclusions about what a model knows without verifying what it was actually trained on. This leaves fundamental questions, such as whether a correct response reflects genuine generalization or rote memorization, grounded in speculation rather than evidence. To resolve these ambiguities, we introduce LUMOS, a diagnostic framework that traces knowledge across training-data exposure, internal representations, and behavioral outputs, leveraging OLMo 2 with its fully transparent training corpus. By grounding analysis in verified exposure, we reveal that models internally encode rare facts with high separability (84%) yet fail to express them behaviorally (54%), though this retrieval gap narrows with scale. Furthermore, when models are asked to self-reflect on their own answers, they perform reliably on trained content (83%) but drop to random-baseline levels (49%) on unseen content. This collapse persists even under chain-of-thought prompting, which inflates confidence signals rather than improving calibration. Collectively, these findings demonstrate that incorporating the training-data axis into LLM evaluation transforms speculative diagnoses into verifiable claims, and we advocate that this axis should be a standard component of knowledge assessment in LLMs.
Oct 1, 2026

From Behavior to Provenance: Attributing Tabular Foundation Models to Synthetic Pretraining Data

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±\pm0.004 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
Oct 1, 2026

Generation Provenance Before Behavior Attribution: Auditing Synthetic Speech Research Objects

Attributing model behavior to synthetic training data requires knowing what produced each training item before estimating what that item caused. A waveform-label pair does not preserve this knowledge. We propose a generation-provenance substrate in which a synthetic research object binds source specification, generated content, waveform, target, fact requirements, quality signals, review lineage, and immutable manifest identity. Producer and selection mechanism determine evidentiary meaning; storage location and variable name do not. We audit this substrate in a private Japanese care-handoff pipeline. A 113-asset review population contains 1.552 hours of synthetic speech across six scenario families; all items have linked audio, transcripts, candidate notes, and fact checklists, but human evidence is selective and source-specific. Two faithful-only manifests are scenario-seed-disjoint and immutably versioned, while exact upstream attribution remains blocked by floating generator aliases, missing per-clip TTS and code stamps, and an unversioned checking prompt. We argue that generation provenance is necessary but not sufficient for behavior attribution: it defines the candidate causal graph and audit units, whereas contributive attribution still requires frozen training runs and intervention or influence evidence. The paper contributes a compact provenance contract, an audit protocol, and a bounded case study for synthetic-data attribution; controlled research access may be offered, but we do not claim causal training-data attribution, clinical validity, or unrestricted public release.
Sep 30, 2026

TrueMuse: A Benchmark for Data Attribution in Text-to-Music Models

Text-to-music generation models are trained on massive music collections, creating a growing need for data attribution methods that can quantify the contribution of individual training samples. However, existing attribution methods are difficult to rigorously evaluate due to the lack of reliable ground truth, making it challenging to reliably assess their actual effectiveness. To address this gap, we introduce TrueMuse, a controlled dataset and benchmark for text-to-music data attribution. TrueMuse is constructed by fine-tuning three diffusion-based text-to-music models on carefully curated attribution samples, whose known inclusion in fine-tuning provides controlled attribution targets for evaluation. The benchmark covers four attribution settings, spanning melodic structure, timbral characteristics, artist-level stylistic signatures, and genre-level shared patterns, and includes 133 attributes, 648 fine-tuned models, and 95,456 generated samples across two prompt types. Using TrueMuse, we systematically evaluate existing black-box attribution methods along four dimensions: fine-tuning improvement, prompt-type difficulty, multi-task training, and fine-tuning data size. Our results show that attribution remains challenging, with existing methods exhibiting substantial variation across evaluation settings, highlighting the need for more reliable and generalizable attribution methods for text-to-music generation.
Sep 30, 2026

Source Identification Is Not Fitness Testing: Measuring the Limits of Synthetic-Data Attribution

Repeated training on model-generated data can degrade later models. One possible response is to use provenance when deciding which generated examples to reuse. We test both how reliably that provenance can be recovered and whether it helps identify better training data. Using financial-risk text, we first identify the source of generated passages and then repeat the test after rewriting them. Generator attribution is 98.7% accurate on the original passages but falls to 53.1% after paraphrasing and 29.0% after style rewriting. Generated-versus-human detection remains close to perfect against the tested human comparison set. We then compare two ways of selecting generated examples over three rounds of generation and retraining. One uses source information. The other uses a score from a separate reference model. The two rules select different examples, but the planned comparison does not detect a stable difference in the degradation of the resulting models. The results show that identifying where data came from and identifying which data are useful for training are separate problems. The experiment therefore separates source identity, criterion-facing selection, and recursive training outcome: neither the provenance score nor the tested criterion-facing proxy is established as sufficient for future recursive behaviour.
Sep 30, 2026

Distilling Diffusion Score Discrepancy for Efficient Training Data Attribution

Training data attribution for diffusion models aims to identify the training samples that influence a generated instance, but existing methods either require costly per-sample gradient computation or query-specific model optimization. Moreover, most methods attribute changes in a proxy loss rather than changes in the actual model's generative behavior. We address these limitations by formulating attribution directly with a local score discrepancy measure, which applies to any diffusion variant (including DDPM, EDM, and flow matching), and by showing that such measure can be estimated without retraining, as a preconditioned gradient similarity. We instantiate this estimator as Training-data Influence via score Discrepancy (TID), which uses Kronecker-factored curvature to avoid random projections and per-sample gradient storage. We then distill TID into TIDE, a forward-only student trained online to reproduce the teacher's rankings from the diffusion model's internal activations. Under counterfactual evaluation on CIFAR-10, ArtBench-10, and MS-COCO, TID matches or outperforms state-of-the-art approaches, while TIDE retains most of TID's accuracy at four to five orders of magnitude lower per-query cost, attributing generated samples in milliseconds and faster than the generation itself.
Sep 30, 2026

dattri-LLM: A Unified and Efficient Library for Training Data Attribution at LLM Scale

Training data attribution (TDA) estimates the contribution of individual training examples to model outputs. Most scalable TDA methods rely on per-example gradients, whose computation and use at LLM scale pose challenges in efficiency, compatibility, and extensibility. We introduce dattri-LLM, a TDA library that makes gradient-based attribution more practical at scale. For efficiency, dattri-LLM uses compact gradient representations and dynamically routes gradient operations based on a cost model. For compatibility, its capture mechanism collects per-example gradients from existing training loops that call backward(), without requiring changes to the loop or its configuration. This includes distributed training with DDP and FSDP and pipelines built with HuggingFace Transformers, TRL, and OLMo. For extensibility, dattri-LLM exposes reusable gradient operations and training-time callbacks for implementing attribution methods and applications. These interfaces support a variety of attribution methods, including gradient similarity, curvature-based influence, and trajectory-based methods, as well as applications that act on gradients during training, such as online data selection. On the same hardware and workload, dattri-LLM achieves 3.2x the throughput of the fastest competing library on average, scales multiple attribution methods to 110B-parameter models across four H200 GPUs, and offers superior attribution fidelity-cost trade-offs across a range of models with different model families and scales. The source code of dattri-LLM is available at https://github.com/TRAIS-Lab/dattri-llm.
Sep 29, 2026

The Unequal Influence of Bad Advice: Using Training Data Attribution to Modulate Emergent Misalignment

Fine-tuning large language models on narrow, misaligned tasks can undo their post-training alignment and induce novel misaligned behaviors -- a phenomenon known as emergent misalignment (EM). EM has been linked to persona-like representations, where fine-tuning might reduce loss by amplifying a harmful or 'evil' persona. It remains unclear which properties of the training data drive this effect: whether all harmful examples contribute approximately equally to misalignment and whether different models are equally affected by the same fine-tuning examples. In this work, we use training data attribution to quantitatively estimate how much each harmful example contributes to EM. We benchmark the quality of the attribution via retraining -- a sound attribution score should enable us to enhance or attenuate EM by filtering data on that score. Score-based filtering can substantially enhance or attenuate EM; we find that both data-attribution scores and a black-box harmfulness score can identify consequential examples. All models we test become misaligned when trained on the same dataset, and influence scores perform best when filtering data from the same model that computed them. We find cross-model generalization of influence scores from scores derived from the three model families we tested, but this generalization does not recover same model filtering performance.
Sep 29, 2026

Scaling Influence Functions in LLMs through Eigenbasis-Corrected One-Bit Gradient Projection

Influence functions estimate how individual training examples affect the behavior of large language models (LLMs). Analyzing how training data influence different behaviors of an LLM involves repeated influence computation. Reusing stored training gradients reduces the computational cost, but storing full gradients is prohibitively expensive at LLM scale. We study how to compress these gradients while preserving influence estimates for future queries that are unknown at storage time. Through a worst-case analysis, we characterize the optimal fixed-dimensional linear representation and propose eigenbasis-corrected one-bit gradient projection (EOGP) to approximate it at scale. Specifically, EOGP uses EK-FAC to reduce gradient dimensionality, then applies PCA within the retained subspace to learn compression directions from the training gradients. We then apply one-bit quantization to the resulting coordinates, allowing more coordinates to be retained within a fixed storage budget. On GPT-2, EOGP predicts retraining outcomes more accurately than the evaluated compression baselines while using one-sixteenth of their per-example storage. On OLMo 2 SFT models from 1B to 32B parameters, EOGP remains competitive with the baselines allocated over 100 times as much storage per example.
Sep 27, 2026

Scalable Attribution and Control of Model Behavior During Training

Attributing and controlling model behavior during training requires identifying each example's contribution quickly enough to act before the next update. However, examples in the same training batch can produce similar behavioral changes, making their individual contributions difficult to distinguish. We address this ambiguity through mutual information, accounting for interference within the batch by quantifying how much the combined behavioral change reveals about each example's contribution. We show that this mutual information is a logarithmic function of Behavioral Gradient Uniqueness (BGU). BGU gives the information measure its geometric interpretation. Our Batch-Space Ghost (BS-Ghost) algorithm makes these scores practical inside the training loop through shared computation in batch space, without storing model-sized example gradients. On a complete 1,000-example Qwen2.5-7B-Instruct workload, our BS-Ghost implementation adds 27 seconds (8.0%) to 5.5 minutes of ordinary training. Removal and retraining demonstrate that BGU identifies data that causally shapes final behavior. At each training step, signed information identifies which examples strengthen or weaken the target behavior, explaining how behavior develops during training. Signed information also enables cheap intervention during training: it predicts how changing example weights will affect behavior in the next update. We then use these predictions to choose weights that steer behavior toward a desired target. This makes our framework a practical foundation for scalable oversight and verification of training pipelines and processes, helping evaluators assess model alignment, understand how it develops during training, and guide interventions that shape ongoing learning.
Sep 27, 2026

Learning Dynamics of Continual Learning: A Unified View of Data Attribution, Forgetting, and Plasticity Loss

Modern language models are likely to be updated throughout their lifetime rather than trained once and frozen. Each update therefore participates in a recurring cycle: decide which experience to learn from, understand what that update changes, and remain capable of learning from what comes next. We show that these challenges are governed by the same evolving update--behavior interaction. We derive a token- and layer-wise decomposition of how learning from one token changes another prediction. By separating the softmax force, shared readout geometry, and residual connections, it exposes two interaction channels and yields a forward-computable approximation. Following this interaction through time reveals a unified picture of continual adaptation. Positive interaction identifies useful experience; negative interaction produces either concentrated collision or accumulated erosion; over longer horizons, updates reshape the shared geometry mediating future learning signals, reducing their transmission. These predictions lead to effective data selection, mechanism-specific controls for interference, and a readout-based diagnostic of future learnability whose degradation predicts the benefit of restoring the readout. Across models and training regimes, the same local interaction thus explains both what an update changes now and how learning today changes what can be learned tomorrow. This view connects data attribution, forgetting, and plasticity loss as distinct regimes of the same evolving learning dynamics.
Sep 26, 2026

Traceprop: Training Data Attribution in a Single Pass

Data attribution tools used in practice, TRAK, LoGRA/LogIX, and EK-FAC, run after training: they recompute per-sample gradients in a separate pass over the training set, and curvature-aware variants need a further pass to estimate covariance. Traceprop records projected per-sample gradients and K-FAC covariance statistics inside the training backward pass itself. On Pythia-1B LoRA fine-tunes on an A100, this adds 2.6% to training time, against 10.7% for LogIX's best inline configuration, random-init (4.1x lower, one-sided Mann-Whitney p = 9e-5, n = 10 per arm); at Pythia-6.9B the numbers are 11.3% and 27.0% (p = 0.004). LogIX's default configuration, PCA-init, is both slower and lower quality than random-init, so we compare against random-init throughout. On a small transformer where LDS is measurable, Traceprop matches LogIX's best configuration: pooled LDS difference +0.0022, 95% CI [-0.0038, +0.0075]. Two things do not work: at Pythia-160M/SST-2, every method is indistinguishable from a low noise ceiling, and on planted-backdoor and mislabel detection, gradient attribution does not beat gradient norm or representation similarity. The contribution here is systems, not a new estimator: LogIX's attribution quality, in one pass instead of two.