Prototype Language Models Offer Interpretable, Faster Attribution

Understanding why a language model produces a specific output is crucial for auditing, debugging, and improving its behavior. Current large language models (LLMs) distribute the influence of training data across numerous parameters, making it difficult and computationally expensive to trace back an output to its originating examples. This paper introduces PRISM (Prototypes for Interpretable Sequence Modeling), a novel prototype language model architecture designed to address this challenge. PRISM generates each prediction using a sparse, non-negative combination of learned prototypes. These prototypes are trained with clustering objectives, effectively anchoring each one to coherent groups of training examples. Across various model sizes and training data volumes, PRISM maintains performance comparable to traditional dense LLMs, often within 2.5 percentage points of accuracy. A key advantage is its sparse prototype structure, which localizes curvature in the loss landscape, leading to a more manageable Hessian. This enables training data attribution that is approximately 500 times faster than existing post-hoc methods, while consuming equivalent memory. Furthermore, PRISM allows for targeted corrections, such as improving downstream accuracy by calibrating linear prototype controllers or removing undesirable model behaviors through prototype suppression without extensive fine-tuning or loss of generation quality.