For AI agents, the verifier may matter more than the model
Across about 15 studies on , the strongest pattern is that successful systems have a strict verifier that checks the work. ComPilot connected a standard to a compiler, so the system could check whether code was legal and whether it actually ran faster, then try again when needed. It achieved a 2.66x speedup in one run and a 3.54x speedup when choosing the best result out of five, without .
AlphaCodium repeatedly tested generated code, raising GPT-4’s CodeContests result from 19% to 44%. DeepSeek-R1 trained with rewards that are easy to check in math and code; R1-Zero rose from 15.6% to 71.0% on AIME during training, and reached 86.7% with . o3 reached 87.5% on ARC-AGI, but only in a high-compute setting that cost roughly hundreds of thousands of dollars for the run.
Failed agent systems usually lacked a verifier, or used a check that the model could exploit.
Key points
- Successful usually include a strict verifier.
- ComPilot used a compiler to check code and measure real speed gains without .
- AlphaCodium improved coding results by repeatedly running generated code against tests.
- DeepSeek-R1 benefited from rewards that are easy to verify in math and code.
- o3’s strong ARC-AGI result came with a very high compute bill.