Publications
Selected work on embodied agents, reinforcement learning, self-evolving data systems, and multimodal reasoning. * denotes equal contribution.
Preprints & Technical Reports
Self-Evolving Embodied Agents via Skill-Harness Evolution
Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Xiaocui Yang, Shi Feng, Yuqing Yang, Dongsheng Li
SHAPER · arXiv, 2026 · Under review
Embodied Agents · VLA · Skill-Harness Evolution
SHAPER adapts embodied agents through target-environment rollouts that evolve reusable skills and an execution harness while keeping model parameters and low-level executors frozen. VLABench success improves from 28.25% to 34.50%, and ESI-Bench accuracy from 32.5% to 49.8%. The implementation is available as an official recipe in Microsoft’s Agent Lightning repository.
[Paper] [Agent Lightning recipe]
AI4AI at Scale: A Full-Pipeline System for Enhancing LLM Agentic Capabilities
Technical report, 2026 · Co-author
Foundation Model Post-Training · Agentic System Optimization
I contributed as an author during my foundation-model internship at XYZ AI Lab. The report describes bounded exploration and verification-gated agentic system optimization; the team released XYZ-Aquila-mini/pro Deep Search Agents together with training, evaluation, and replay records.
What Are Key Factors for Updates in RL for LLM Reasoning?
ACPO · Under review, 2026
LLM RL · Off-Policy Optimization · Theoretical Analysis
Starting from conflicting observations about token-level updates, we analyze how off-policy degree, token probability, and importance-ratio clipping affect gradient magnitude and direction. The analysis motivates ACPO, which groups tokens by behavior-policy probability and adapts clipping bounds using within-group importance-ratio dispersion. Experiments cover 3B/7B models on mathematics, table question answering, and logical reasoning under near- and off-policy settings.
Published Papers
SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning
Peidong Wang*, Zhiming Ma*, Xin Dai*, Yongkang Liu, Shi Feng, Xiaocui Yang, Wenxing Hu, Zhihao Wang, Mingjun Pan, Li Yuan, Daling Wang
ACL 2026
Multimodal RL · Audio Reasoning · GRPO
An end-to-end audio-text fraud detection framework with rule-based reasoning rewards and efficient, chunk-based risk assessment. Deployed in China Mobile’s anti-fraud business pipeline, it processes over 70,000 calls daily. SAFE-QAQ achieves 87.49 average classification F1 on TeleAntiFraud-Bench and reduces reasoning length by 48.87%; its real-time variant detects risk in 8.98 seconds on average.
Language Models as Continuous Self-Evolving Data Engineers
Peidong Wang, Ming Wang, Zhiming Ma, Xiaocui Yang, Shi Feng, Daling Wang, Yifei Zhang, Kaisong Song
EMNLP 2025
Recursive Self-Improvement · Data Agents
LANCE enables language models to generate, clean, review, and annotate their own training data with preference information. Iterative self-training connects data construction with model improvement, reducing reliance on external supervision for post-training data.
TeleAntiFraud-28k: An Audio-Text Slow-Thinking Dataset for Telecom Fraud Detection
Zhiming Ma*, Peidong Wang*, Minhua Huang, Jingpeng Wang, Kai Wu, Xiangzhao Lv, Yachun Pang, Yin Yang, Wenjie Tang, Yuchen Kang
ACM Multimedia 2025
Audio-Text Dataset · Fraud Detection Benchmark
A multi-task audio-text dataset for telecom fraud detection, built through real-call transcription, LLM simulation, and multi-agent generation. The accompanying TeleAntiFraud-Bench evaluates fraud detection and slow-thinking audio-language reasoning.
STICKERCONV: Generating Multimodal Empathetic Responses from Scratch
Yiqun Zhang*, Fanheng Kong*, Peidong Wang*, Shuang Sun, Lingshuai Wang, Shi Feng, Daling Wang, Yifei Zhang, Kaisong Song
ACL 2024
Multimodal LLMs · Multi-Agent Systems · Empathetic Dialogue
Agent4SC uses collaborative agents to construct a multimodal empathetic dialogue dataset. The PEGS framework generates contextually relevant text and sticker responses, supported by LLM-based evaluation of empathy and consistency.
[Paper] [Project] [Code] [Dataset]
TIGER: A Unified Generative Model Framework for Multimodal Dialogue Response Generation
Fanheng Kong, Peidong Wang, Shi Feng, Daling Wang, Yifei Zhang
LREC-COLING 2024
Multimodal Dialogue · Response Generation
A unified generative framework for dialogue responses containing both text and images, evaluated through automatic metrics and human judgments.