Jagged Frontier of LLM Capability

The observation that LLMs can be strong in some tasks while making errors in others

The Jagged Frontier is the observation that LLMs can be strong in some tasks while making severe errors in others DellAcqua et al. (2023)

A recent paper LLMs Corrupt Your Documents When You Delegate showed how through a multi-step backtranslation-inspired approach, they demonstrate that LLMs had severe issues with degradation across all document domains except Python.

References

Fabrizio Dell'Acqua, Edward McFowland, Ethan R. Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R. Lakhani. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. SSRN Electronic Journal, 2023. doi:10.2139/ssrn.4573321.