Ruixiang Tang
-
Why Language Models Love the Em Dash: How Post-Training Recruits a Sparse Writing Circuit
August 12 2026
Why do some language models overuse the em dash while others almost never do? We trace this writing habit to a sparse set of late-layer neurons, study how post-training amplifies an internal mechanism already present in the base model, and examine why explicit instructions sometimes fail to suppress it.
-
When a Linear Probe Reads Your Data Pipeline: Why a Near-Perfect AUC Can Say Nothing About the Model
August 16 2026
A linear probe is the standard tool for asking whether a concept or a state is represented inside a language model. Replicating two papers from top conferences, we find a flaw that is easy to walk into. When the two groups of examples are assembled by different procedures, the probe can learn the procedure instead of the concept, and report a near-perfect score that says nothing about what is inside the model.