AI Reviewer Precision Doesn't Guarantee Critique Uptake in Math Agents

Many AI agent systems designed for math and science problems employ hierarchical structures with specialized reviewer roles, assuming that a dedicated review stage will improve accuracy by correcting errors. This research investigates that assumption using 4,181 Omni-MATH problems and matched GPT-OSS-120B actors. The findings indicate that while collaboration offers minimal gains on easier problems, its benefits sharply increase from tier 4 onwards. Interestingly, a broadcast-style peer discussion approach achieved higher final accuracy than a traditional planner-executor-reviewer (PER) pipeline in these harder regimes. The study further explored why this gap exists, finding that it's not solely due to reviewer quality. Although the PER reviewer was more precise in identifying errors, its useful critiques were significantly less likely to influence the next candidate answer, resulting in poorer reviewer-guided repair. Forcing explicit acknowledgment of critiques in PER actually lowered accuracy, while embedding guidance directly into the solver's context offered only partial improvement. This suggests that the ability to detect errors (precision) and the ability to act on those critiques (uptake) are distinct challenges in multi-agent systems.