Prototype-based interpretability for open-ended language generation

Develop prototype-based interpretability methods for open-ended text generation in generative language modeling, where the output space has vocabulary-scale cardinality, enabling faithful exemplar-based explanations that scale to open-ended outputs.

Background

The paper introduces an automated, geometry-aware prototype selection technique for interpretable reinforcement learning, removing reliance on human-curated prototypes and showing competitive performance with black-box agents. While this approach works across several RL environments, it assumes a manageable number of discrete action classes for prototype discovery and explanation.

In the context of generative language modeling, outputs are open-ended and the action space can expand to vocabulary scale, complicating direct adoption of prototype-based explanations. Although Proto-LM provides initial progress for sentence classification, the authors explicitly note that extending prototype-based interpretability to open-ended generation remains unresolved. Addressing this gap would broaden prototype-based interpretability to LLMs and other generative systems.