MAMM-Refine: A Recipe for Improving Faithfulness in Generation with Multi-Agent Collaboration

TL;DR

This work addresses the challenge of faithfulness in long-form generation by introducing MAMM-Refine, a multi-agent, multi-model refinement recipe that decomposes refinement into Detect, Critique, and Refine subtasks. It systematically evaluates how multiple agents and model types influence each subtask, exploring Generate and Rerank framings and showing that a discriminative reranking approach often outperforms free-form generation in aggregation tasks. Across three summarization benchmarks and a long-form QA dataset, the authors demonstrate that the proposed recipe yields consistent improvements in faithfulness, with the best configurations achieving statistically significant gains. The work further shows that the approach generalizes beyond summarization to long-form QA, highlighting practical impact for safer and more reliable generative systems.

Abstract

Multi-agent collaboration among models has shown promise in reasoning tasks but is underexplored in long-form generation tasks like summarization and question-answering. We extend multi-agent multi-model reasoning to generation, specifically to improving faithfulness through refinement, i.e., revising model-generated outputs to remove factual inconsistencies. We investigate how iterative collaboration among multiple instances and types of large language models (LLMs) enhances subtasks in the refinement process, such as error detection, critiquing unfaithful sentences, and making corrections based on critiques. We design intrinsic evaluations for each subtask, with our findings indicating that both multi-agent (multiple instances) and multi-model (diverse LLM types) approaches benefit error detection and critiquing. Additionally, reframing critiquing and refinement as reranking rather than generation tasks improves multi-agent performance. We consolidate these insights into a final "recipe" called Multi-Agent Multi-Model Refinement (MAMM-Refine), where multi-agent and multi-model collaboration significantly boosts performance on three summarization datasets as well as on long-form question answering, demonstrating the effectiveness and generalizability of our recipe.

Figures (4)

• Figure 1: Illustration of the refinement pipeline (top-left) and how multi-agent debate is applied to different subtasks. In the Detect subtask (top-right), agents collectively choose among a discrete set of options, such as making yes/no decisions or selecting the most faithful candidate. For the Critique and Refine subtasks, we explore two approaches. In the bottom-left panel, we frame the task as generative (using Generate), where each agent updates its own critique or output based on other agents' responses. In the bottom-right panel, we frame it as a discriminative task using Rerank, where agents choose the best output from the candidates. While discriminative tasks converge to a single solution, generative tasks result in updated responses from each agent.
• Figure 2: Illustration of our setup for intrinsic evaluations for different subtasks. We convert TofuEval, a dataset of system summaries with human-annotated faithfulness labels and critiques, to tasks for evaluating the performance of different multi-agent setups for Detect, Critique, and Refine subtasks with Rerank and Generate.
• Figure 3: Detect and rerank multi-agent performance across multiple iterations.
• Figure 4: Error match rate for Critique and faithfulness score for Refine across multiple iterations.