Junyao Yang

About Me

Hi there, this is Junyao. I am a first-year graduate student at the School of Computing, National University of Singapore (NUS), where I am pursuing a specialization in Artificial Intelligence. I am currently at Tencent Hy, working on Agentic RL Stability and Long-Horizon Terminus Agent. My research interests lie in Agentic AI, Large Language Models, Reinforcement Learning and Explainable Artificial Intelligence.
My Chinese name is 杨竣尧 (/jɑːŋ dʒuːn jaʊ/).
CVs: EN, 中文

News

2026.07 New paper: Stale but Stable! Staleness-adaptive trust regions stabilize asynchronous RL by tightening high-mismatch updates while preserving ordinary-token behavior.
2026.07 New paper: Harness Handbook! A behavior-centric representation and BGPD framework that help agents localize implementation sites and plan edits for evolving agent harnesses.
2026.07 New paper: Long-Horizon-Terminal-Bench! A long-horizon terminal benchmark of 46 tasks with dense subtask-level rewards for partial-credit evaluation of frontier agents.
2026.05 First-Author paper: Entropy-Gradient Inversion! Leverage entropy gradient inversion to enhance reinforcement learning for reasoning models.
2026.04 Joined Tencent Hy LLM Team, working on Agentic RL Stability and Long-Horizon Terminus Agent.
2026.04 First-Author paper ReasonAny accepted to ACL 2026 Main.
2026.02 Blog post: The Entropy-Gradient Inversion. R1/o1-like reasoning models exhibit significant negative correlations between gradient strength and token entropy, emerging rapidly within the first 200 steps of SFT.
2026.01 Tech report: AgentDoG! State-of-the-art diagnostic guardrail framework with an Agentic XAI attribution module.
2026.01 Paper: Agentic Attribution! A hierarchical framework to unveil internal factors driving LLM-based agent actions.
2026.01 Attending AAAI 2026 at Singapore during Jan 20-27, 2026.
2026.01 Paper: ReasonAny! Contrastive gradient identification to resolve destructive performance collapse in model merging.
2025.11 First-Author paper RCP-Merging accepted to AAAI 2026 Main Track.
2025.08 RewardDS accepted to EMNLP 2025 Main.
2025.08 Joined Shanghai AI Lab as a Research Intern, advised by Dongrui Liu.
2025.08 New work: RCP-Merging! Integrating long CoT capability into domain-specific LLMs.
2025.05 Passed undergraduate thesis defense.
2025.05 Co-First-Author paper PrivacyRestore accepted to ACL 2025 Main.
2025.02 New papers: RewardDS and PrivacyRestore.
2024.07 Joined ZeroNLP as a Research Assistant, advised by Prof. Ziqian Zeng.
2024.07 Completed internship at Tencent as a machine learning intern.
2024.07 Contextless CS reached 20,000 DAU.
2024.04 Joined Tencent as a machine learning intern.
2024.03 Completed internship at ShenZhen Stock Exchange as a machine learning intern.

Selected Publications

ACL 2026 ReasonAny: Incorporating Reasoning Capability to Any Model via Simple and Effective Model Merging

Junyao Yang, Chen Qian, Dongrui Liu^†, Wen Shen, Yong Liu^†, Jing Shao^†

TL;DR: Merging robust chain-of-thought capabilities into domain-specific models (Safety, Biomedicine) using Contrastive Gradient Identification.

Paper Code

AAAI 2026 RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior

Junyao Yang, Jianwei Wang, Huiping Zhuang, Cen Chen, Ziqian Zeng*^†

TL;DR: Enhancing domain performance while preserving chain-of-thought reasoning abilities by treating reasoning as a prior.

Paper Code

arXiv Preprint Stale but Stable: Staleness-Adaptive Trust Regions for Stabilizing Asynchronous Reinforcement Learning

Junyao Yang, Yucheng Shi, Zongxia Li, Zhongzhi Li, Ruhan Wang, Xiangxin Zhou, Kishan Panaganti, Haitao Mi, Leowei Liang

TL;DR: Stabilizing asynchronous RL by adapting trust regions to rollout staleness, tightening high-mismatch updates while preserving ordinary-token behavior.

Paper

ACL 2025 PrivacyRestore: Privacy-Preserving Inference in Large Language Models via Privacy Removal and Restoration

Ziqian Zeng*^†, Jianwei Wang*, Junyao Yang*, Zhengdong Lu, Haoran Li, Huiping Zhuang, Cen Chen

TL;DR: Protecting privacy via activation steering using a protected meta-vector without retraining.

Paper Code

Tech Report

AgentDoG: A Diagnostic Guardrail Framework for AI Agent Safety and Security

Shanghai Artificial Intelligence Laboratory (Contributor)

TL;DR: A state-of-the-art diagnostic guardrail framework utilizing a unified three-dimensional taxonomy to provide fine-grained monitoring and root-cause analysis of AI agent safety risks.

Paper Code Model #1 Paper of the day 机器之心

arXiv Preprint Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading

Zongxia Li, Zhongzhi Li, Yucheng Shi, Ruhan Wang, Junyao Yang, Zhichao Liu, Xiyang Wu, Anhao Li, Yue Yu, Ninghao Liu, Lichao Sun, Haotao Mi, Leowei Liang

TL;DR: A dense-reward terminal benchmark for measuring how far agents can progress on long, complex terminal workflows.

Paper Code

arXiv Preprint Harness Handbook: Making Evolving Agent Harnesses Readable, Navigable, and Editable

Ruhan Wang, Yucheng Shi, Zongxia Li, Zhongzhi Li, Yue Yu, Junyao Yang, Kishan Panaganti, Haitao Mi, Dongruo Zhou, Leoweiliang

TL;DR: A behavior-centric map and BGPD framework that help agents find code and plan harness edits.

Paper Project

EMNLP 2025 RewardDS: Privacy-Preserving Fine-Tuning for Large Language Models via Reward Driven Data Synthesis

Jianwei Wang, Chengming Shi, Junyao Yang, Haoran Li, Qianli Ma, Huiping Zhuang, Cen Chen, Ziqian Zeng^†

TL;DR: Using client-side reward models to filter synthetic data, mitigating noise while protecting privacy.

Paper Code

ACL 2026 ReasonAny: Incorporating Reasoning Capability to Any Model via Simple and Effective Model Merging

Junyao Yang, Chen Qian, Dongrui Liu^†, Wen Shen, Yong Liu^†, Jing Shao^†

TL;DR: Merging robust chain-of-thought capabilities into domain-specific models (Safety, Biomedicine) using Contrastive Gradient Identification.

Paper Code

AAAI 2026 RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior

Junyao Yang, Jianwei Wang, Huiping Zhuang, Cen Chen, Ziqian Zeng*^†

TL;DR: Enhancing domain performance while preserving chain-of-thought reasoning abilities by treating reasoning as a prior.

Paper Code

arXiv Preprint Stale but Stable: Staleness-Adaptive Trust Regions for Stabilizing Asynchronous Reinforcement Learning

Junyao Yang, Yucheng Shi, Zongxia Li, Zhongzhi Li, Ruhan Wang, Xiangxin Zhou, Kishan Panaganti, Haitao Mi, Leowei Liang

TL;DR: Stabilizing asynchronous RL by adapting trust regions to rollout staleness, tightening high-mismatch updates while preserving ordinary-token behavior.

Paper

ACL 2025 PrivacyRestore: Privacy-Preserving Inference in Large Language Models via Privacy Removal and Restoration

Ziqian Zeng*^†, Jianwei Wang*, Junyao Yang*, Zhengdong Lu, Haoran Li, Huiping Zhuang, Cen Chen

TL;DR: Protecting privacy via activation steering using a protected meta-vector without retraining.

Paper Code

arXiv Preprint Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading

Zongxia Li, Zhongzhi Li, Yucheng Shi, Ruhan Wang, Junyao Yang, Zhichao Liu, Xiyang Wu, Anhao Li, Yue Yu, Ninghao Liu, Lichao Sun, Haotao Mi, Leowei Liang

TL;DR: A dense-reward terminal benchmark for measuring how far agents can progress on long, complex terminal workflows.

Paper Code

arXiv Preprint Harness Handbook: Making Evolving Agent Harnesses Readable, Navigable, and Editable

Ruhan Wang, Yucheng Shi, Zongxia Li, Zhongzhi Li, Yue Yu, Junyao Yang, Kishan Panaganti, Haitao Mi, Dongruo Zhou, Leoweiliang

TL;DR: A behavior-centric map and BGPD framework that help agents find code and plan harness edits.

Paper Project

Tech Report

AgentDoG: A Diagnostic Guardrail Framework for AI Agent Safety and Security

Shanghai Artificial Intelligence Laboratory (Contributor)

TL;DR: A state-of-the-art diagnostic guardrail framework utilizing a unified three-dimensional taxonomy to provide fine-grained monitoring and root-cause analysis of AI agent safety risks.

Paper Code Model #1 Paper of the day 机器之心

arXiv Preprint Entropy-Gradient Inversion: Moving Toward Internal Mechanism of Large Reasoning Models

Junyao Yang, Chen Qian, Kun Wang, Linfeng Zhang, Quanshi Zhang, Yong Liu, Dongrui Liu^†

TL;DR: Identifying Entropy-Gradient Inversion—a robust negative correlation between token entropy and logit gradients—as a geometric fingerprint of LRM reasoning capability, and proposing CorR-PO that embeds this signature into RL reward regularization to stabilize reasoning optimization without costly external verifiers.

Paper

arXiv Preprint The Why Behind the Action: Unveiling Internal Drivers via Agentic Attribution

Chen Qian, Peng Wang, Dongrui Liu^†, Junyao Yang, Dadi Guo, Ling Tang, Jilin Mei, Qihan Ren, Shuai Shao, Yong Liu, Jie Fu, Jing Shao, Xia Hu

TL;DR: A hierarchical framework for agentic attribution, using temporal likelihood and perturbation-based analysis to unveil internal factors driving LLM-based agent actions.

Paper Code

EMNLP 2025 RewardDS: Privacy-Preserving Fine-Tuning for Large Language Models via Reward Driven Data Synthesis

Jianwei Wang, Chengming Shi, Junyao Yang, Haoran Li, Qianli Ma, Huiping Zhuang, Cen Chen, Ziqian Zeng^†

TL;DR: Using client-side reward models to filter synthetic data, mitigating noise while protecting privacy.

Paper Code

Blogs

Long-Horizon Terminal-Bench: Measuring the Progress Agents Can Sustain, Not Just What They Can Finish

TL;DR: LHTB evaluates agent progress on long terminal tasks in Docker with hidden dense-reward grading, not just final completion.

Harness Handbook: Making Agent Harnesses Understandable, Auditable, and Editable

TL;DR: Harness Handbook maps behavior to implementation, making agent harnesses easier to inspect, navigate, and edit.

The Entropy-Gradient Inversion: A New Perspective on LLM Reasoning Capabilities

TL;DR: We discover that reasoning models exhibit a unique "fingerprint": a significant negative correlation between gradient strength and token entropy, which contradicts traditional base models. This capability emerges rapidly within the first 200 steps of SFT.