Add new paper:

[wyzh0912]

Title

Do Attention Heads Compete or Cooperate during Counting?

Published Date

2025-02-10

Source

arXiv

Head Name

Mixed Head

Summary

• Innovation: The paper investigates the behavior of attention heads in transformers when performing a counting task, revealing that attention heads act as a pseudo-ensemble, each solving the same subtask (i.e., counting), but requiring non-uniform aggregation to conform to task syntax.

• Tasks: The study involves training small transformers on the Count01 language, a simple counting task where the model determines if a string has more '1's than '0's, with '2's as noise. The analysis focuses on understanding whether attention heads compete or cooperate during this task.

• Significant Result: The research shows that while attention heads individually learn to solve the counting task, they do so as a pseudo-ensemble. The output layer must aggregate their outputs non-uniformly to address both the main task (counting) and an auxiliary syntactic task (ending the sentence).