Opening up the Pandora’s box: An Overview on Debugging Large Language Models | David Plaskowski
Opening up the Pandora’s box: An Overview on Debugging Large Language Models
David Plaskowski, RISS 2023 Cohort
Carnegie Mellon University
Robotics Institute Summer Scholars: https://riss.ri.cmu.edu/
Language models are often treated as black boxes, obscuring our understanding of their internal mechanisms. Mechanistic interpretability is a new field aimed at reverse engineering internal mechanisms through decomposition of neural networks into circuits that we can analyze, thus providing new directions for interpretability of otherwise black box models. However, despite the advances in this field, the current algorithms face significant limitations. This overview aims to navigate these challenges by exploring the largely unexplored territory of various mechanistic interpretability methods and how they can be used to understand internal representations of large language models.
*****
The Robotics Institute Summer Scholars (RISS) Program at Carnegie Mellon University's School of Computer Science
Carnegie Mellon University’s Robotics Institute is committed to opening doors and creating opportunities for future leaders in robotics. Carnegie Mellon University is home to the top-ranked School of Computer Science, the world’s first university robotics department, the world’s first Ph.D. in robotics, and the largest university-affiliated robotics research group. Launched in 2006, CMU’s Robotics Institute Summer Scholars (RISS) program (http://riss.ri.cmu.edu/)comprises a ten-week summer undergraduate research program that immerses a diverse cohort of scholars in cutting-edge robotics and extensive post-program mentoring. The program provides opportunities for students from across the country and the world to conduct research with leaders in the field. The program aspires to foster a diverse and inclusive working and learning environment where all students enjoy the educational benefits of diversity and are actively welcomed, included, and supported by the community. The quality and breadth of research, high level of institute and university engagement, and powerful professional development programming, graduate school application counseling, and alumni network create transformative experiences and remarkable post-program trajectories.
RESEARCH RESULTS:
Explore our research projects and results at:
Videos & Posters at https://riss.ri.cmu.edu/research_show...
Working Papers Journal at https://riss.ri.cmu.edu/research_show...
APPLY: Starting November 1st at https://riss.ri.cmu.edu/
SCHOLAR EXPERIENCE: Scholars contribute, communicate, & connect.
Contribute: Scholars contribute to robotics research projects through a guided research experience with multiple layers of mentorship.
Communicate: Scholars learn how to effectively communicate research ideas to various audiences (e.g., sponsors, academic audience, novice audiences) and in various formats (e.g., elevator pitches, short talks, research papers, and poster presentations).
Connect: Scholars forge long-lasting connections to Carnegie Mellon University researchers and partners.
PROFESSIONAL SKILLS DEVELOPMENT: The RISS communications workshop series includes workshops and one-on-one tutoring on writing research & technical papers, designing graphics, and presenting posters. Their impressive results are a reflection of the deep partnerships and commitment of partners’ teaching contributions.
Robotics Workshops & Talks: The technical professional development series exposes scholars to a wide range of robotics applications and projects through weekly robotics talks, visits to labs, and hands-on workshops.
COMMUNITY: We foster the creation of a supportive learning community through intentional messaging, welcoming events, and effective programming. Onboarding includes virtual orientation sessions and office hours, a cohort Slack group, on-site orientation, and weekly office hours. Programming is structured to engage students in deep interactions – from workshop teams to peer reviews for papers and posters. Projects require students to go beyond their current skill set and to learn from others. Over 75 individuals participate as mentors, presenters, or programming partners annually.