arXiv:2606. 11660v1 Announce Type: new Abstract: Data attribution is a promising field in interpretability that aims to explain model behavior through the influence of its training data, with applications including debugging undesirable model behavior and training dataset curation.
arXiv:2606. 04928v1 Announce Type: new Abstract: Large Language Models (LLMs) are increasingly deployed across diverse applications, raising critical questions for governance, accountability, and data provenance.
arXiv:2603. 21014v2 Announce Type: replace Abstract: Mechanistic interpretability seeks to understand how Large Language Models (LLMs) represent and process information.
TrueMuse is a new benchmark designed to evaluate data attribution in text-to-music models. It consists of a controlled dataset created by fine‑tuning three diffusion‑based models on curated attribution samples, providing known attribution targets. The benchmark covers four settings—melodic structure, timbral characteristics, artist‑level style, and genre‑level patterns—across 133 attributes, 648 models, and 95,456 generated samples, and is used to assess existing black‑box attribution methods along several dimensions.
Data Attribution at Scale via Influence Matrix Estimation proposes a scalable approach to quantify how individual training examples influence a model’s predictions. The authors introduce two algorithms, MAGE and SPELL, that reconstruct an influence matrix from a limited number of measurements without extra computational cost, improving over existing baselines across various training scales and budgets.
arXiv:2609.14248v1 Announce Type: cross Abstract: Faithful citation attribution begins with identifying the intended source for a scientific claim. We study this source-identification capability thro...
arXiv:2603. 02184v2 Announce Type: replace-cross Abstract: Multi-attribution learning (MAL), which enhances model performance by learning from conversion labels yielded by multiple attribution mechanisms, has emerged as a promising learning paradigm for conversion rate (CVR) prediction.
arXiv:2608.24524v1 Announce Type: cross Abstract: Feature attribution is a central tool of model interpretability, yet the software through which it is applied remains fragmented: individual tools sp...
arXiv:2606. 07996v1 Announce Type: cross Abstract: Pretraining is fundamental to the development of Large Language Models (LLMs), yet the opacity of pretraining data complicates model analysis and raises ethical, legal, and fairness concerns.
arXiv:2608. 02879v1 Announce Type: new Abstract: The widespread adoption of proprietary Large Language Models (LLMs) accessed strictly through closed APIs has created a critical challenge for responsible deployment: a fundamental lack of interpretability.
The paper introduces CAMS, a Claim‑Anchored Multi‑Document Summarization framework that decomposes source documents into atomic claims, resolves provenance deterministically from verbatim quotes to token spans, clusters equivalent claims across documents, and rewrites summaries so each sentence ends with claim identifiers linking back to source spans. CAMS separates provenance (an invariant for each emitted sentence) from faithfulness (an objective encouraged by selection, rewriting, and verification). Evaluations on MultiNews, DiverseSumm, and zero‑shot WCEP show that CAMS matches strong baselines in summary quality while improving faithfulness and citation precision, raising attribution accuracy from 38% to 64% and reducing human verification time per claim by 3.4×.
The paper introduces dattri-LLM, a library designed to make training data attribution (TDA) practical for large language models. It achieves efficiency by using compact gradient representations and a cost‑based routing system, while maintaining compatibility by capturing per‑example gradients from existing training loops without modifications, even in distributed settings. The library also offers extensibility through reusable gradient operations and callbacks, supporting various attribution methods and applications such as online data selection, and demonstrates significant performance gains and scalability up to 110B‑parameter models.
The paper introduces two algorithms, MAGE and SPELL, that enable efficient data attribution in neural networks by estimating a large influence matrix from a limited number of measurements. These methods leverage existing metagradient techniques without additional computational overhead, addressing the challenge of predicting the impact of removing training data in non‑convex models. Experiments show that MAGE and SPELL outperform current baselines across various training scales and measurement budgets.