Junyao Yang

RSR: Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite

Zongxia Li^‡, Yucheng Shi^‡, Zhongzhi Li^‡, Junyao Yang, Ruhan Wang, Chengsong Huang, Fuxiao Liu, Haitao Mi, Jordan Boyd-Graber, Leowei Liang

‡ Core contribution (marked † in the source paper)

TL;DR Recursive Self-Rewrite turns successful trajectories discovered under diverse harnesses into reusable capabilities under a general harness. It expands 2,001 trajectories to 11,094 and improves Qwen-3.8-27B pass@3 from 57.0% to 74.2% on Terminal-Bench 2.

SWR: Self-Supervised Scaling of Terminal Environments for Scientific Domains

Zhongzhi Li^*, Yucheng Shi^*‡, Zongxia Li^*, Junyao Yang, Ruhan Wang, Yu Wang, Jingyuan Huang, Jichao Yu, Ninghao Liu^†, Haitao Mi, Leowei Liang

* Equal contribution · † Corresponding author · ‡ Project lead; source role symbols normalized

TL;DR Software-in-the-loop reconstruction derives reference behavior and verifiers from scientific software workflows. Scaling to 500 workflows across 46 software families and six domains improves Qwen3.8-27B on Terminal-Bench 2 from 47.94% to 53.56%.

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, Chengsong Huang, Ergan Shang, Yue Yu, Ninghao Liu, Lichao Sun, Haitao Mi, Leowei Liang

‡ Core contribution (marked † in the source paper)

TL;DR A benchmark of 46 long-horizon terminal tasks evaluates sustained agent execution using dense subtask-level rewards. Partial-credit grading exposes intermediate progress that binary task-completion metrics miss.

RST: Recursive Synthesis for Long-Horizon Terminal Tasks

Zhongzhi Li^*, Yucheng Shi^*‡, Zongxia Li^*, Ruhan Wang, Anhao Li, Zixun Huang, Junyao Yang, Lei Ke, Ninghao Liu, Haitao Mi, Leowei Liang

* Equal contribution · ‡ Project lead (marked † in the source paper)

TL;DR Recursive verified synthesis expands long-horizon terminal tasks by extending successful solutions and realigning instructions and verifiers. It yields 37,484 tasks at approximately $0.05 each and improves Qwen3.5 by up to 10 points across terminal benchmarks.

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^†

† Corresponding authors · ‡ Project lead; work conducted during an internship at Shanghai AI Lab

TL;DR Contrastive Gradient Identification separates reasoning-related parameters from task-specific knowledge to prevent destructive collapse during model merging. The method transfers robust chain-of-thought capabilities into domain-specific models without additional training.

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^†

† Corresponding author

TL;DR Treating reasoning capability as a prior identifies and preserves reasoning-sensitive parameters during model merging. This enhances domain-specific performance while retaining long chain-of-thought abilities.

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

Shanghai Artificial Intelligence Laboratory

TL;DR A diagnostic guardrail framework uses a unified three-dimensional risk taxonomy to monitor agent trajectories. Fine-grained attribution identifies the causes of unsafe behavior and supports root-cause analysis.

Harness Self-Evolution for Long-Horizon Terminal Agents

Ruhan Wang, Yucheng Shi^‡, Zongxia Li, Junyao Yang, Zhongzhi Li, Yue Yu, Haitao Mi, Dongruo Zhou, Leowei Liang

‡ Internship mentor (marked † in the source paper) · Work conducted during an internship at Tencent

TL;DR RiseHarness iteratively improves agent harnesses while keeping model weights fixed. Successful and failed trajectories, internal execution traces, and a Harness Handbook guide behavior-level edits that are validated through repeated task evaluation.

Long-Horizon-Terminal-Bench 2.0: Automating Diverse Task Construction and Measuring Sustained Agent Progress

Zongxia Li^‡, Zhongzhi Li^‡, Yucheng Shi^‡, Ruhan Wang, Junyao Yang, Chengsong Huang, Zhichao Liu, Xiyang Wu, Anhao Li, Yue Yu, Ergan Shang, Fuxiao Liu, Tong Zheng, Ninghao Liu, Lichao Sun, Haitao Mi, Leowei Liang

‡ Core contribution (marked † in the source paper)

TL;DR LHTB 2.0 expands process-aware terminal evaluation to 77 tasks across 13 domains, including systematically reconstructed scientific and engineering workflows. Three-hour runs and verifier checkpoints measure sustained progress rather than only binary final outcomes.

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

† Corresponding author · ‡ Project lead; source role symbols normalized

TL;DR Staleness-adaptive trust regions stabilize asynchronous reinforcement learning by contracting high-mismatch outward updates. The method preserves ordinary-token behavior while controlling heterogeneous rollout staleness.

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, Leowei Liang

‡ Lead project collaborator (marked † in the source paper) · Ruhan Wang conducted this work during an internship at Tencent Seattle

TL;DR A behavior-centric representation maps agent harness responsibilities to their implementations. The BGPD framework uses this handbook to localize code and plan edits as harnesses evolve.

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^†

* Equal contribution · † Corresponding authors

TL;DR Entropy-Gradient Inversion identifies a robust negative correlation between token entropy and logit gradients as a geometric signature of reasoning capability. CorR-PO incorporates this signature into reward regularization to stabilize reasoning-model reinforcement learning.

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

† Corresponding author

TL;DR A hierarchical attribution framework explains the historical factors driving LLM-based agent actions, independent of task success or failure. Temporal likelihood identifies critical interaction components, and perturbation analysis localizes influential sentences.

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

* Equal contribution · † Corresponding author

TL;DR A plug-and-play privacy-preserving inference framework removes sensitive spans before sending inputs to an LLM service. Protected restoration vectors recover useful information through activation steering without retraining the model.

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^†

† Corresponding author

TL;DR Reward proxy models guide the filtering and self-optimizing refinement of synthetic data for privacy-preserving LLM fine-tuning. This reduces noise in generated supervision while protecting sensitive training data.