Generalizing Experience for Language Agents with Hierarchical MetaFlows

Shengda Fan, Xin Cong, Zhong Zhang, Yuepeng Fu, Yesai Wu, Hao Wang, Xinyu Zhang, Enrui Hu, Yankai Lin

• Proposed the framework MetaFlowLLM, a framework that builds a hierarchical exgitperience tree of previously completed tasks for language-agent workflows; it employs a hierarchical MetaFlow merging algorithm to enable experience reuse, leading to a higher success rate on AppWorld and WorkBench

WORKFLOWLLM: Enhancing Workflow Orchestration Capability of Large Language Models

Shengda Fan, Xin Cong, Yuepeng Fu, Zhong Zhang, Shuyan Zhang, Yuanwei Liu, Yesai Wu, Yankai Lin†, Zhiyuan Liu, Maosong Sun

Code

• Introduced WorkflowLLM, a data‑centric framework to improve the orchestration capabilities of large language models (LLMs) by constructing a large dataset (WorkflowBench) and fine‑tuning an LLM to yield strong performance on previously unseen APIs and instructions.

LogicST: A Logical Self‑Training Framework for Document‑Level Relation Extraction with Incomplete Annotations

Shengda Fan, Yanting Wang, Shasha Mo, Jianwei Niu

Code

• Introduced LogicST, a neural-logical self-training framework that diagnoses pseudo-label conflicts through logical rules, improving the performance of document-level relation extraction tasks.

Boosting Document‑Level Relation Extraction by Mining and Injecting Logical Rules
Shengda Fan, Shasha Mo, Jianwei Niu
Code

• Proposed MILR, a framework that mines logical rules from annotations and injects them into the training and inference process to improve consistency and F1 scores in Document Relation Extraction.