When the number of in-context examples D increases, the prediction loss for both single-head and multi-head attention in transformers is in O(1/D), but the prediction loss for multi-head attention has a smaller multiplicative constant.

Authors

• Person: Samuel Tesfazgi, Leonhard Sprandl, Sandra Hirche Organization: AISTATS
  Track: Poster Session 3 - aistats 2026

Sources

• Track: Poster Session 3 - aistats 2026 virtual.aistats.org Samuel Tesfazgi, Leonhard Sprandl, Sandra Hirche · AISTATS via serper

Referenced by nodes (2)

• Transformers concept
• In-Context Learning concept