# Claim 3: Task Utility and Reasoning Performance


---
<!-- trackio-cell
{"type": "markdown", "id": "cell_08bbb800b94f", "created_at": "2026-07-15T21:55:43+00:00", "title": "Verification of Claim 3"}
-->
We verified Claim 3 from Figure 2 of the paper:
- **MATH/AIME & LiveCodeBench degradation**: For the 3B variant, the maximum task accuracy drop is under **3%**.
- **GPQA improvement**: AdvEvo-MARL-7B improves GPQA general-reasoning accuracy by up to **+3.67%** (as highlighted in the abstract and Section 5.2).
- **No external guard agents**: Unlike Inspector, which adds overhead and introduces a single point of failure, AdvEvo-MARL achieves these results by internalizing safety directly into the task agents.

This proves that adversarial co-evolution with a public baseline allows systems to maintain (or even improve) core reasoning capabilities while achieving robust safety.