A reinforcement learning method rewards guessing as much as genuine reasoning.

When artificial intelligence learns by trial and error, it tests multiple tries and scores them against each other using an approach called GRPOGRPOA reinforcement learning method that uses group comparisons to assign rewards based on relative performance.. But there is a problem. The system cannot tell if a correct answer came from true reasoning or just a lucky guess. Both get big positive scores, which the researchers call spurious advantagespurious advantageA false reward given to actions that succeed by chance rather than through skill or reasoning.. This happens in bounded tasks with few choices, open tasks with bounded subsets, and search tasks with many paths. Because of this, the AI learns to guess. To fix it, the researchers created SIGNBALANCE. It keeps the plus or minus sign from the verifier, applies a global scale, and balances the values. It matched GRPO on open-answer math and beat it on bounded-answer math and search agents.

Why it matters

Developers training AI for math or search tasks might get systems that guess correctly but don't understand, leading to unreliable outputs. This affects anyone relying on such AI for accurate reasoning.

Questions answered