Actionability of Interpretation for LLM Safety

Ascertain the actionability of interpretation methods for autoregressive Transformer-based generative large language models by formally defining what constitutes actionable interpretation outputs, determining evaluation criteria across diverse stakeholder groups, and establishing procedures that operationalize these outputs to support concrete safety decisions.

Background

The survey emphasizes practical tools that help users understand and apply interpretation results to improve LLM safety, noting that stakeholders (e.g., developers, auditors, end-users) may differ in how they define practicality and usefulness. Despite a growing ecosystem of libraries and visual tools, the authors point out that it remains unresolved how interpretation outputs should be evaluated for their ability to lead to concrete, safety-relevant actions across contexts and user types.

This unresolved issue arises in the broader effort to bridge interpretation methods and safety enhancements. Clear, formal criteria and processes for assessing and operationalizing interpretation outputs are needed to ensure that insights derived from model internals, token attributions, or self-reasoning reliably translate into safer behavior in real-world deployments.