More Agents Improve Math Problem Solving but Adversarial Robustness Gap Persists

Research reveals that while increasing the number of LLM agents improves mathematical problem-solving accuracy, these multi-agent systems remain vulnerable to adversarial attacks. The study found that human-like typos pose the greatest threat to robustness, and the adversarial vulnerability gap persists regardless of agent count.

• →Multi-agent LLM systems show improved math problem-solving accuracy, with largest gains occurring when scaling from 1 to 5 agents.
• →Human-like typos remain the most significant vulnerability for multi-agent systems, causing higher attack success rates than punctuation noise.
• →Adversarial robustness gaps persist regardless of the number of agents deployed in the system.
• →Diminishing returns in performance improvements occur beyond approximately 10 agents.
• →Punctuation noise damage scales directly with severity levels (10%, 30%, 50%).