Tag
This paper investigates methods to steer Arabic LLMs toward dialect-specific generation by identifying sparse neuron populations and extracting dialect activation directions, enabling dialect control at inference time without fine-tuning.
This paper demonstrates that transformers trained on Sudoku solving traces build structured world models organized by domain constraints, and identifies a sparse, monosemantic circuit responsible for the naked-single decision rule. The work provides a fully interpretable algorithmic account of transformer reasoning on a combinatorial task.