When Disagreements Elicit Robustness: Investigating Self-Repair Capabilities under LLM Multi-Agent Disagreements

TL;DR

This paper explores how disagreements among large language model agents can enhance robustness and self-repair in collaborative tasks, revealing that disagreement type influences success and solution flexibility.

Contribution

It demonstrates that general disagreements promote exploration and robustness, while task-critical disagreements can hinder reasoning but have limited impact on programming tasks.

Findings

General disagreements improve success through exploration.

Task-critical disagreements reduce success in reasoning tasks.

Agents often bypass edited facts in programming, enabling self-repair.

Abstract

Recent advances in Large Language Models (LLMs) have upgraded them from sophisticated text generators to autonomous agents capable of cooperation and tool use in multi-agent systems (MAS). However, it remains unclear how disagreements shape collective decision-making. In this paper, we revisit the role of disagreement and argue that general, partially overlapping disagreements prevent premature consensus and expand the explored solution space, while disagreements on task-critical steps can derail collaboration depending on the topology of solution paths. We investigate two collaborative settings with distinct path structures: collaborative reasoning (CounterFact, MQuAKE-cf), which typically follows a single evidential chain, whereas collaborative programming (HumanEval, GAIA) often adopts multiple valid implementations. Disagreements are instantiated as general heterogeneity among agents and as task-critical counterfactual knowledge edits injected into context or parameters. Experiments reveal that general disagreements consistently improve success by encouraging complementary exploration. By contrast, task-critical disagreements substantially reduce success on single-path reasoning, yet have a limited impact on programming, where agents can choose alternative solutions. Trace analyses show that MAS frequently bypasses the edited facts in programming but rarely does so in reasoning, revealing an emergent self-repair capability that depends on solution-path rather than scale alone. Our code is available at https://github.com/wbw625/MultiAgentRobustness.