Reasoning or Pattern Exploitation? Mechanistic Insights from RL-Trained Language Models
RL training raises Tic-Tac-Toe accuracy and robustness in a small model. Sparse autoencoders show heavier reliance on prompt facts than on strategy.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv