Zhiyuan Feng | 冯志远 | Tsinghua University

Selected Papers

	HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs Zhuoshi Pan, Zhiyuan Feng, Yukuo Cen, Honglin Lin, Junru Lu, Zhongxia Chen, Shuo Yang, H. Vicky Zhao^✉, Jie Tang^✉ * Equal contribution. Paper / Code arXiv 2026 We propose HarnessPro, a label-free framework that recursively improves agent harnesses through trajectory diagnosis, candidate editing, and Pareto-aware performance-cost optimization. It produces lite, medium, and heavy variants for different deployment budgets. On Terminal-Bench 2, SWE-Bench Pro, and ProgramBench with GPT-5.5/Codex, its heavy mode improves average pass rate by 2.0 percentage points while reducing per-task cost by 26.1% relative to RHO.
	EnactOpt: Diagnosis-Guided Harness Evolution for Embodied Agents Zhiyuan Feng, Zhuoshi Pan, Zhaolu Kang, Zhongxia Chen^‡, Shuo Yang^‡, Sen Ye, Heng Wang, Han Li Equal contribution. ‡ Project leads. Paper arXiv 2026 We propose EnactOpt, a diagnosis-guided framework that jointly evolves textual skills and context-construction code from limited execution feedback, without changing the underlying models or runtime. By distinguishing missing evidence from evidence that was not effectively used, it turns rollout diagnoses into reusable harness improvements. Across VLABench, ALFWorld, and EAI, EnactOpt improves task success by approximately 10-27 percentage points over initial harnesses and generalizes to unseen task conditions.
	TaskGround: Ground-Infer-Execute for Full-Scene Household Reasoning Zhiyuan Feng, Yu Deng, Ruichuan An, Zhenhua Liu, Qixiu Li, Keming Wu, Zhiying Du, Weijie Wang, Haoxiao Wang, Shuang Chen, Sicheng Xu, Yaobo Liang, Jiaolong Yang^✉, Baining Guo Paper / Webpage NeurIPS 2026 🔥🔥🔥 We formalize full-scene household reasoning, where an agent must infer executable task structure from a complete household scene and a situated request. We propose TaskGround, a training-free Ground-Infer-Execute framework that grounds the scene into compact task-relevant slices before producing skill-level action sequences. On our new FullHome benchmark (400 human-validated household tasks), TaskGround substantially boosts success rates across proprietary and open-weight models, making Qwen3.5-9B competitive with GPT-5 while cutting input-token cost by up to 18×.
	Seeing across views: Benchmarking spatial reasoning of vision-language models in robotic scenes Zhiyuan Feng, Zhaolu Kang, Qijie Wang, Zhiying Du, Jiongrui Yan, Shubin Shi, Chengbo Yuan, Huizhi Liang, Yu Deng, Qixiu Li, Rushuai Yang, Arctanx An, Leqi Zheng, Weijie Wang, Shawn Chen, Sicheng Xu, Yaobo Liang, Jiaolong Yang^✉, Baining Guo Paper / Webpage / Code / Data / SlideLive ICLR 2026 🔥🔥🔥 Cited by DreamZero 🌟🌟🌟 We propose MV-RoboBench, a benchmark for evaluating the multi-view spatial reasoning capabilities of VLMs in robotic manipulation.
	From Human Videos to Robot Manipulation: A Survey on Scalable Vision-Language-Action Learning with Human-Centric Data Zhiyuan Feng, Qixiu Li, Huizhi Liang, Rushuai Yang, Yichao Shen, Zhiying Du, Zhaowei Zhang, Yu Deng^✉, Li Zhao, Hao Zhao, Zongqing Lu, Oier Mees, Marc Pollefeys, Jiaolong Yang^✉, Baining Guo Paper / Webpage IJCAI 2026 Survey Track 🔥🔥🔥 We survey how to turn abundant human videos into actionable supervision for VLA models, organizing methods into four categories and outlining key challenges for reliable real-world transfer.
	AD-MIR: Bridging the Gap from Perception to Persuasion in Advertising Video Understanding via Structured Reasoning Binxiao Xu, Junyu Feng, Xiaopeng Lin, Haodong Li, Zhiyuan Feng, Bohan Zeng, Shaolin Lu, Ming Lu, Qi She, Wentao Zhang Paper ICML 2026 🔥🔥🔥 We propose AD-MIR, a two-stage framework that decodes advertising intent by first constructing a structured multimodal database via semantic retrieval and keyword matching, then reasoning over it with an iterative, evidence-grounded inquiry loop that mimics marketing expertise, achieving state-of-the-art on AdsQA.
	Discover, Learn, and Reinforce: Scaling Vision-Language-Action Pretraining with Diverse RL-Generated Trajectories Rushuai Yang, Zhiyuan Feng, Tianxiang Zhang, Kaixin Wang, Chuheng Zhang, Li Zhao, Xiu Su, Yi Chen, Jiang Bian Paper arXiv 2025 We propose DLR, an information-theoretic multi-pattern RL framework that generates diverse, high-success manipulation trajectories for scalable VLA pretraining, improving transfer and showing better data-scaling than standard single-pattern RL.
	WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors Keming Wu, Yijing Cui, Wenhan Xue, Qijie Wang, Xuan Luo, Zhiyuan Feng, Zuhao Yang, Sudong Wang, Sicong Jiang, Haowei Zhu, Zihan Wang, Ping Nie, Wenhu Chen, Bin Wang Paper / Dataset NeurIPS 2026 🔥🔥🔥 We introduce WorldReasonBench, a human-aligned benchmark that stress-tests video generators as future world-state predictors across four dimensions—World Knowledge, Human Centric, Logic Reasoning, and Information-Based reasoning—covering diverse tasks from physical dynamics and cultural understanding to spatial geometry and quantitative math.
	VolSplat: Rethinking Feed-Forward 3D Gaussian Splatting with Voxel-Aligned Prediction Weijie Wang, Yeqing Chen, Zeyu Zhang, Hengyu Liu, Haoxiao Wang, Zhiyuan Feng, Wenkang Qin, Feng Chen, Jia-Wang Bian, Zheng Zhu, Donny Y. Chen, Bohan Zhuang Paper / Code ECCV 2026 We propose VolSplat, a voxel-aligned feed-forward 3D Gaussian Splatting framework that predicts Gaussians from sparse 3D features, improving multi-view consistency and adapting Gaussian density to scene complexity for novel-view synthesis.
	HiSpatial: Taming Hierarchical 3D Spatial Understanding in Vision-Language Models Huizhi Liang, Yichao Shen, Yu Deng, Sicheng Xu, Zhiyuan Feng, Tong Zhang, Yaobo Liang, Jiaolong Yang Paper / Webpage / Code CVPR 2026 🔥🔥🔥 We propose a hierarchical framework that decomposes 3D spatial understanding in VLMs into four progressive levels, and build an automated pipeline over ~5M images to generate 3D spatial VQA data for fine-tuning, achieving state-of-the-art on multiple spatial reasoning benchmarks.
	VITRA: Scalable vision-language-action model pretraining for robotic manipulation with real-life human activity videos Qixiu Li, Yu Deng, Yaobo Liang, Lin Luo, Lei Zhou, Chengtang Yao, Lingqi Zeng, Zhiyuan Feng, Huizhi Liang, Sicheng Xu, Yizhong Zhang, Xi Chen, Hao Chen, Lily Sun, Dong Chen, Jiaolong Yang, Baining Guo Paper / Webpage / Code ICRA 2026 🔥🔥🔥 We propose VITRA, a novel approach for pretraining Vision-Language-Action (VLA) models for robotic manipulation using large-scale, unscripted, real-world videos of human hand activities. My contribution: I focus on the data processing pipeline, leveraging HaWoR to reconstruct 3D hand poses from human-centric videos.
	Beyond human demonstrations: Diffusion-based reinforcement learning to generate data for vla training Rushuai Yang, Hangxing Wei, Ran Zhang, Zhiyuan Feng, Xiaoyu Chen, Tong Li, Chuheng Zhang, Li Zhao, Jiang Bian, Xiu Su, Yi Chen Paper arXiv 2025 We propose a diffusion policy optimization method that uses RL to autonomously generate smooth, low-variance long-horizon manipulation trajectories, enabling VLA training that outperforms models trained on human or Gaussian RL demonstrations on LIBERO.
	HiMoE-VLA: Hierarchical Mixture-of-Experts for Generalist Vision-Language-Action Policies Zhiying Du, Bei Liu, Yaobo Liang, Yichao Shen, Haidong Cao, Xiangyu Zheng, Zhiyuan Feng, Zuxuan Wu, Jiaolong Yang, Yu-Gang Jiang Paper / Code arXiv 2025 We propose HiMoE-VLA, a vision-language-action framework that tackles cross-embodiment heterogeneity via a Hierarchical Mixture-of-Experts action module, adaptively handling diverse action spaces, sensor configurations, and control frequencies to deliver consistent gains and robust generalization across simulation benchmarks and real-world robotic platforms.
	HSSBench: Benchmarking Humanities and Social Sciences Ability for Multimodal Large Language Models Zhaolu Kang, Junhao Gong, Jiaxu Yan, Wanke Xia, Yian Wang, Ziwen Wang, Huaxuan Ding, Zhuo Cheng, Wenhao Cao, Zhiyuan Feng, Siqi He, Shannan Yan, Junzhe Chen, Xiaomin He, Chaoya Jiang, Wei Ye, Kaidong Yu, Xuelong Li Paper / Code ICLR 2026 HSSBench is a multilingual (UN six languages) benchmark of 13,000+ expert-curated samples designed to evaluate MLLMs’ cross-disciplinary, concept-to-vision reasoning in Humanities and Social Sciences, revealing clear gaps even in state-of-the-art models.
	MIRA: Medical Time Series Foundation Model for Real-World Health Data Hao Li, Bowen Deng, Chang Xu, Zhiyuan Feng, Viktor Schlegel, Yu-Hao Huang, Yizheng Sun, Jingyuan Sun, Kailai Yang, Yiyao Yu, Jiang Bian Paper NeurIPS 2025 AI4X-AC 2026 Oral We introduce MIRA, a medical time-series foundation model pretrained on 454B+ time points that handles irregular sampling and missingness to deliver stronger zero-shot and fine-tuned forecasting across datasets and tasks.
	TransDiff: Diffusion-Based Method for Manipulating Transparent Objects Using a Single RGB-D Image Haoxiao Wang, Kaichen Zhou, Binrui Gu, Zhiyuan Feng, Weijie Wang, Peilin Sun, Yicheng Xiao, Jianhua Zhang, Hao Dong Paper / Webpage ICRA 2025 We propose TransDiff, a diffusion-based single-view RGB-D depth completion method for accurate grasping of transparent objects.
	Deep evidential learning in diffusion convolutional recurrent neural network Zhiyuan Feng, Kai Qi, Bin Shi, Hao Mei, Qinghua Zheng, Hua Wei Paper CIKM 2022 DIT Workshop We integrate evidential deep learning into DCRNN to provide sampling-free uncertainty quantification for spatiotemporal traffic forecasting, achieving improved predictive intervals measured by MIS.