Provsformer

Provsformer is a proposed neural architecture that blends data provenance tracking with transformer models to support auditable AI processes. The name signals its aim to make the transformation of information over neural networks transparent, with a formal record of how inputs are transformed into outputs. In this concept, runs of Provsformer generate a provenance graph that encodes the lineage of tokens, decisions, and intermediate representations, along with a compact digest of model configuration and data.

The architecture builds on standard transformer components but incorporates a provenance module. As data flows through layers, the module records origin references for tokens, edits applied by attention, and any augmentations, producing a directed acyclic graph. This provenance graph can be queried to answer questions such as which input features influenced a particular decision, or how a specific output was derived in a given run. Provsformer is designed to be compatible with existing transformer backbones and can operate in supervised or self-supervised settings.

Applications include regulated industries like finance and healthcare, scientific data pipelines, and any AI workflow where reproducibility and accountability are required. The concept emphasizes transparency, reproducibility, and governance, not performance alone.

Challenges include added computational overhead, complexity of provenance schemas, privacy concerns, and the need for standards to exchange provenance information. Research in this area remains exploratory, with ongoing work on efficient provenance encoding, secure auditing, and evaluation metrics.