Multi-agent simulation boosts robustness of conversational agents
MADS: Multi-Agent Dialog Simulation for Robust Conversational Agents
Computation and Language
Summary
Conversational AI systems often struggle when users act unexpectedly or unfriendly. The authors present MADS, a method to create many simulated dialogues that include challenging conversations where things often go wrong. They built a large set of these simulated dialogues called MADS-Bench to better train conversational agents to handle tough situations. Their training combines usual supervised learning with reinforcement learning based on clear rewards to improve agent reliability.
What this means in practice
- •For chatbot developers: Train and evaluate chatbots with diverse difficult user interactions to improve robustness under real-world conditions.
- •For customer service teams: Develop conversational agents that better handle unhappy or adversarial customer dialogues for improved service quality.$Commercial implications: Enables building more reliable customer support AI products that handle difficult dialogues, enhancing customer satisfaction and reducing human intervention.
Authors
Arkajyoti Chakraborty, Adam Zernik, Ishika Kulkarni, Mudit Arora, Hugo Lin, Justin Chiu, Alessandro Di Bari, Neha Gupta, Andreas Stolcke
Abstract
Task-oriented conversational agents remain fragile under realistic conditions, especially when users exhibit non-cooperative behavior. Existing function-calling benchmarks underrepresent adversarial conversation trajectories, thereby limiting the training resources available for developing robust agents. We present MADS (Multi-Agent Dialog Simulation), a synthetic data generation pipeline that produces structured, diverse dialogue trajectories. We inject adversarial conversation, producing validated conversation trajectories that capture realistic, unhappy paths. Using MADS, we construct MADS-Bench: a benchmark spanning six domains, 55 personas, and 1095 conversations, with approximately 66% of conversations containing failure-prone conversation scenarios. We evaluate the robustness of the trained models using a two-stage training approach that combines supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR).