Research
Coding agent rewards hit a harder verification horizon
With monitoring, Qwen cut hacked SWE passes from 28.57 to 0.56 percent: coding agent rewards have become a verification problem.
2 stories tagged evals.
With monitoring, Qwen cut hacked SWE passes from 28.57 to 0.56 percent: coding agent rewards have become a verification problem.
Tiny checkpoint gains become harder to mistake for progress with olmo-eval, Ai2's open workbench for iterative LLM evaluation.