Does Reward Training Teach AI New Reasoning? A New Paper Says No. Checked and Re-Run
0 View
Share this Video
Publish Date:
9 November, 2026
Category:
Sports
Video License
Standard License
Imported From:
Youtube
Tags
By Evan Vega
A MOSS @ COLM 2026 workshop paper (arXiv:2611.02574) argues that reinforcement learning with verifiable rewards only sharpens what a model already knows: tracking OLMo 2 1B across 13 training snapshots, first-try accuracy rose 58.0% to 61.3% while a detector for planted arithmetic mistakes stayed flat. Novel Cognition checked it against its own tables and code and re-ran its accuracy measurements, adding the model from before training that the paper leaves out.
