TLDR: A study on LLM agents in mathematical question answering found that collaboration among multiple agents reliably improves accuracy, with the largest gains seen up to five agents. However, this multi-agent setup does not close the adversarial robustness gap; LLMs remain highly vulnerable to noisy inputs, especially human-like typos (WikiTypo), regardless of the number of agents. Punctuation noise is more easily mitigated by agent collaboration, while challenging datasets amplify the impact of input perturbations.
Large Language Models (LLMs) have shown impressive capabilities in solving complex reasoning tasks, particularly in mathematics. A recent study delves into whether the collaborative power of multiple LLM agents can not only boost accuracy but also enhance their resilience against adversarial inputs, such as typos and punctuation noise.
The Power of Collaboration
The research, titled “More Agents Helps but Adversarial Robustness Gap Persists,” investigates the performance of LLM agents when tackling mathematical questions. It builds on the idea that having multiple LLM instances work together, often through simple strategies like sampling and voting, can significantly improve their problem-solving accuracy. This collaborative approach, termed ‘Agent Forest’ in the study, involves each agent generating a candidate solution, which are then aggregated by majority vote to produce a final answer.
The findings consistently show that increasing the number of collaborating agents reliably improves accuracy. The most substantial gains are observed when moving from a single agent to five agents, with diminishing returns typically seen beyond ten agents. This suggests that while teamwork is beneficial, there’s a sweet spot for the number of agents before the additional benefit becomes minimal.
The Persistent Vulnerability to Noise
However, the study also highlights a critical weakness: LLMs’ fragility when faced with perturbed inputs. Even minor surface-level changes, like random punctuation insertions or human-like typos, can drastically reduce a model’s accuracy. The researchers systematically introduced various types of noise into math questions across four established benchmarks: GSM8K, MATH, MMLU–Math, and MultiArith.
The noise types included:
- Synthetic Punctuation Noise: Random insertions of punctuation marks at varying intensities (10%, 30%, 50% of question length).
- Human-like Typos: Real-world spelling mistakes from Wikipedia (WikiTypo) and adversarial typographical errors (R2ATA) that mimic common keyboard usage mistakes.
The results indicate that while collaboration does improve overall accuracy, the fundamental “adversarial robustness gap” persists. This means that even with multiple agents, LLMs remain vulnerable to noisy inputs. The Attack Success Rate (ASR), which measures how often a correct prediction flips to incorrect due to noise, remains largely stable as the agent count increases, especially for human-like typos.
Noise Type Matters Most
A key takeaway is that the type of noise significantly impacts performance. Punctuation noise, while degrading accuracy in proportion to its intensity, is largely mitigated by the collaborative efforts of five to ten agents. In contrast, human-like typos, particularly those from WikiTypo, proved to be the most challenging. They consistently caused the largest drops in accuracy and the highest ASR, even when a large number of agents (up to 25) were involved. R2ATA, another form of human-like typo, fell somewhere in between.
The study suggests that human typos introduce heterogeneous and context-dependent distortions that simple majority voting among agents cannot fully absorb, making them a persistent source of errors.
Dataset Difficulty and Model Performance
The research also revealed differences in robustness across datasets. Easier datasets, such as GSM8K and MultiArith, were less affected by noise, and multi-agent collaboration more easily compensated for any performance degradation. Conversely, more challenging datasets like MATH and MMLU suffered greater absolute performance drops under noise, and even with more agents, the recovery remained incomplete. This highlights that task difficulty amplifies the impact of input perturbations.
Regarding the models tested (Qwen3-4B/14B, Llama3.1-8B, Mistral-7B, Gemma3-4B/12B), larger models generally achieved higher absolute accuracies. However, smaller models often showed steeper relative gains with increasing agent counts, indicating that collaboration can be particularly beneficial for them. Interestingly, robustness didn’t strictly correlate with parameter count across all models, though within the same model families (like Qwen and Gemma), larger versions tended to be more inherently robust.
Also Read:
- Unmasking Vulnerabilities: A New Benchmark for Multi-Agent LLM System Security
- MAESTRO: Orchestrating LLM Teams for Smarter Problem Solving
Looking Ahead
In conclusion, while multi-LLM-agent collaboration is a powerful strategy for enhancing accuracy in mathematical reasoning, it does not fully solve the problem of adversarial robustness. The vulnerability to human-like typos remains a significant challenge. Future research directions could explore noise-aware sampling and aggregation methods, the use of verifier- or tool-assisted agents, and training-time augmentation specifically targeting human typos. For more details, you can read the full paper here.
