Adversarial Multi-Agent Consensus for Fact Verification
Replace self-reflection loops with a multi-agent adversarial debate protocol to verify generated claims and prevent unprovable hallucinations.
Source paper
Chen Hu, Haikuo Du, Heng Wang, Lin Lin, Mingrui Chen, Peng Liu, Ruihang Miao, Tianchi Yue, Wang You, Wei Ji, Wei Yuan, Wenjin Deng, Xiaojian Yuan, Xiaoyun Zhang, Xiangyu Liu, Xikai Liu, Yanming Xu, Yicheng Cao, Yifei Zhang, Yongyao Wang, Yubo Shu, Yurong Zhang, Yuxiang Zhang, Zheng Gong, Zhichao Chang, Binyan Li, Dan Ma, Furong Jia, Hongyuan Wang, Jiayu Liu, Jing Bai, Junlan Liu, Manjiao Liu, Na Wang, Qiuping Wu, Qinxin Du, Shiwei Li, Wen Sun, Yifeng Gong, Yonglin Chen, Yuling Zhao, Yuxuan Lin, Ziqi Ren, Zixuan Wang, Aihu Zhang, Brian Li, Buyun Ma, Kang An, Li Xie, Mingliang Li, Pan Li, Shidong Yang, Xi Chen, Xiaojia Liu, Yuchu Luo, Yuan Song, YuanHao Ding, Yuanwei Liang, Zexi Li, Zhaoning Zhang, Zixin Zhang, Binxing Jiao, Daxin Jiang, Jiansheng Chen, Jing Li, Xiangyu Zhang, Yibo Zhu
Generated by AI. The paper’s authors did not write these.
Replace self-reflection loops with a multi-agent adversarial debate protocol to verify generated claims and prevent unprovable hallucinations.
Train agents with tool domain randomization and meta-learning objectives to build resilience against arbitrary API interface changes.
Train a Process Reward Model specifically for STEM domains to evaluate step-by-step mathematical derivations rather than solely optimizing for final output quality.
This paper introduces Step-DeepResearch, a cost-effective 32B-parameter agent designed to perform open-ended, long-horizon research tasks by internalizing atomic capabilities like planning, retrieval, and reporting. The authors propose a progressive training pipeline—moving from agentic mid-training with synthesized data to Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) with a checklist-style reward model—and establish a new benchmark, ADR-Bench, to evaluate performance against commercial systems.
The model utilizes a single-agent ReAct architecture trained via a three-stage pipeline: (1) Mid-training on synthesized 'atomic capability' data (planning, reflection, etc.) derived from reverse-engineering expert reports; (2) SFT on curated deep search/research trajectories; and (3) PPO-based Reinforcement Learning using a 'Rubrics Judge' model that provides binary reward signals based on fine-grained evaluation criteria.