IIT Roorkee researchers argue reward training does not teach AI new reasoning

A research paper presented at the MOSS workshop at COLM 2026 suggests that reinforcement learning with verifiable rewards (RLVR) does not impart new reasoning capabilities to artificial intelligence models. The study, titled “RLVR Sharpens the Pencil, Not the Mind” by Shreeyam Bangera and Manya Gupta of IIT Roorkee, analyzed Ai2’s open OLMo 2 1B model using 13 training snapshots focused on grade-school mathematics.

While the researchers observed that first-try accuracy increased from 58.0% to 61.3%, a detector for planted arithmetic mistakes remained flat. The authors concluded that the only significant internal change occurred at the final answer stage, implying the training method refines existing knowledge rather than expanding cognitive reach.

Novel Cognition verified these findings by re-running accuracy measurements using the authors’ own prompt and answer checker. The analysis suggests that teams developing or purchasing reasoning models should measure a model's reach in addition to accuracy and evaluate the cost-effectiveness of training runs against repeated queries of existing models.