Abstract
Zhao Xinjie, Fan Gao, Xingyu Song, Yingjian Chen, Rui Yang, Yanran Fu, Yuyang Wang, Yusuke Iwasawa, Yutaka Matsuo, Irene Li. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.
2025.
Introduction: The Challenge of Multi-Hop QA
Multi-hop Question Answering (QA) is a complex task requiring models to synthesize information from multiple sources. Traditional approaches, such as Chain-of-Thought (CoT) or standard Retrieval-Augmented Generation (RAG), typically follow a forward-only reasoning pipeline.
A significant drawback of these methods is "error propagation": a single mistake or incorrect retrieval in an early step can invalidate the entire reasoning chain. Because these systems lack a mechanism to re-evaluate or "change their minds" when encountering contradictions, they often produce incorrect answers with high confidence.
ReAgent: A Reversible Multi-Agent Architecture
To overcome the limitations of forward-only reasoning, the authors propose ReAgent, a novel multi-agent framework that introduces "undo" functionality through a hierarchical backtracking mechanism. ReAgent is structured into three distinct layers:
- Execution Layer: Comprises specialized agents for question decomposition ($A_Q$), evidence retrieval ($A_R$), fact verification ($A_V$), and answer assembly ($A_A$).
- Supervisor Layer: Coordinates the system. The Supervisor Agent ($A_S$) manages global rollbacks, while the Controller Agent ($A_C$) provides high-level strategic regulation.
- Interaction Layer: Provides a "Persistent Log" and "Temporal Tracker," acting as a save-state system that allows agents to return to previous valid states.
Hierarchical Backtracking
The core innovation of ReAgent is its two-level backtracking:
- Local Backtracking: Verifier agents can autonomously discard contradictory evidence and retry specific steps without halting the entire process.
- Global Backtracking: If a late-stage contradiction is detected, the Supervisor triggers a system-wide rollback to a known consistent state to explore alternative reasoning paths.
Key Insights
- Non-Monotonic Reasoning: ReAgent demonstrates that the ability to revoke or revise statements is as critical as generating them for complex reasoning tasks.
- Role Specialization: By separating retrieval from verification, the system creates a checks-and-balances environment that significantly reduces noise and error propagation.
- Interpretability: The explicit logging of backtracking steps makes the reasoning process more transparent, allowing researchers to trace how the model identifies and corrects its own errors.
Evaluation and Results
ReAgent was evaluated on three multi-hop QA benchmarks: HotpotQA, 2WikiMultiHopQA, and MuSiQue.
- Performance Gains: ReAgent achieved an average accuracy improvement of ~6% over strong forward-only baselines.
- Comparison with Large Models: On HotpotQA, ReAgent (using GPT-4o) achieved an F1 score of 0.795, notably outperforming specialized reasoning models like OpenAI’s O1 (0.661) and O3 (0.696).
- Efficiency: Despite the increased number of calls (averaging 29 per question), ReAgent proved more cost-effective than using O1 or O3, while delivering superior results.
- Ablation Study: Removing the backtracking mechanism resulted in a significant performance drop, confirming it as the primary driver of the system's success.
Limitations and Future Work
The primary trade-off for ReAgent's robustness is increased inference latency due to its iterative nature. The multi-agent architecture also introduces higher deployment complexity. Future research will focus on optimizing agent collaboration to reduce computational overhead and extending this reversible logic to other knowledge-intensive domains such as legal and medical reasoning.
Summary updated