Aishik Nagar
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
Reasoning or Pattern Exploitation? Mechanistic Insights from RL-Trained Language Models
According to the abstract, Aishik Nagar and Cheston Tan trained language models with reinforcement learning on Tic-Tac-Toe and inspected them with sparse autoencoders. Accuracy rose, but the models mainly leaned on information already in the prompt, while concepts such as symmetries and forks stayed weakly represented.