Speakers
TBA
3rd Workshop on Attributing Model Behavior at Scale: Data Attribution and Provenance
Sydney, Australia
Contact info: attrib-neurips26 [at] googlegroups [dot] com
Submissions: OpenReview
How can we attribute model outputs and actions back to data? How can we design attributions that inform downstream uses like alignment, data use litigation, and regulatory audits?
As generative AI permeates rapidly across society, an increasingly relevant problem is attribution: how can we attribute model outputs and actions back to data? The appropriate notion of attribution depends on the specific applications. Contributive attribution aims to determine which training data causally influenced the generated output. A growing body of work has made this increasingly tractable, though challenges remain. In parallel, corroborative attribution methods leverage lexical techniques, conditional generation, and textual entailment to identify sources that textually entail, or otherwise semantically correspond to, a model output, i.e., citations.
However, contributive methods produce scores that are often difficult to interpret or verify and corroborative methods identify semantically supporting sources but make no claims about causal responsibility. In practice, however, stakeholders in law, journalism, and the arts need attribution that is interpretable, auditable, and grounded in identifiable sources, beyond computability. Meanwhile, the rapid growth of synthetic and model-generated data further complicates attribution by blurring the boundary between training data and model outputs. Bridging the gap between methods and applications will require not only technical innovation but also clearer problem formulations informed by the concrete constraints of real applications.
This workshop will bring together researchers from both the contributive and corroborative attribution communities with practitioners in law, music, journalism, and AI safety. Our goal is to identify where existing methods fall short of practical needs, surface shared technical challenges across application domains, and establish concrete research directions that can move attribution from a purely academic tool toward one that is useful in practice.
ATTRIB uses the same LLM usage policy as NeurIPS 2026. Please see here for details.
TBA
