Albert No

Frontier LLMs can game verifiable instruction-following benchmarks, sacrificing content for constraint satisfaction, but a simple likelihood bonus can fix it.

RLVR agents are surprisingly sensitive to the very first token they generate, and diversifying only that token unlocks significant gains in exploration and reasoning performance.