Open Problems in Mechanistic Interpretability

TMLR Paper4504 Authors

Abstract: Mechanistic interpretability aims to understand the computational mechanisms underlying neural networks' capabilities in order to accomplish concrete scientific and engineering goals. Progress in this field thus promises to provide greater assurance over AI system behavior and shed light on exciting scientific questions about the nature of intelligence. Despite recent progress toward these goals, there are many open problems in the field that require solutions before many scientific and practical benefits can be realized: Our methods require both conceptual and practical improvements to reveal deeper insights; we must figure out how best to apply our methods in pursuit of specific goals; and the field must grapple with socio-technical challenges that influence and are influenced by our work. This forward-facing review discusses the current frontier of mechanistic interpretability and the open problems that the field may benefit from prioritizing.

Submission Length: Long submission (more than 12 pages of main content)

Changes Since Last Submission:

In response to reviewer feedback, we have made a number of changes to improve the paper. These changes include a new figure and table to highlight the review's structure, updated figures to improve their informativeness, and various content changes and added references.

For full details of the changes, please see individual responses to reviewers.

Assigned Action Editor: ~Sarath_Chandar1

Submission Number: 4504