More Agents Improve Math Problem Solving but Adversarial Robustness Gap Persists

TL;DR

This study shows that while increasing the number of collaborating LLM agents improves mathematical problem-solving accuracy, their vulnerability to adversarial inputs, especially human-like typos, remains largely unmitigated.

Contribution

It introduces a unified framework to evaluate the robustness of multi-agent LLM systems against adversarial perturbations in math questions.

Findings

Collaboration improves accuracy with more agents, especially from 1 to 5 agents.

Adversarial robustness gap persists regardless of the number of agents.

Human-like typos cause the largest accuracy drops and attack success rates.

Abstract

When LLM agents work together, they seem to be more powerful than a single LLM in mathematical question answering. However, are they also more robust to adversarial inputs? We investigate this question using adversarially perturbed math questions. These perturbations include punctuation noise with three intensities (10%, 30%, 50%), plus real-world and human-like typos (WikiTypo, R2ATA). Using a unified sampling-and-voting framework (Agent Forest), we evaluate six open-source models (Qwen3-4B/14B, Llama3.1-8B, Mistral-7B, Gemma3-4B/12B) across four benchmarks (GSM8K, MATH, MMLU-Math, MultiArith), with various numbers of agents n = {1,2,5,10,15,20,25}. Our findings show that 1) Noise type matters: punctuation noise harm scales with its severity, and the human typos remain the dominant bottleneck, yielding the largest gaps to Clean accuracy and the highest attack success rate (ASR) even with a large number of agents; 2) Collaboration reliably improves accuracy as the number of agents, n, increases, with the largest gains from n=1 to n=5 and diminishing returns beyond n$\approx$10. However, the adversarial robustness gap persists regardless of the agent count.