ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment

Haosen Sun^*, Hongjue Zhao^*, Jiangtao Kong, Xiaochang Li, Qineng Wang, Liwei Jiang, Qi Zhu, Tarek F. Abdelzaher, Yejin Choi, Manling Li^†, Huajie Shao^†

International Conference on Learning Representations (ICLR), 2026

[Project Page] [Paper] [Code]

• A unified ODE-based framework for multi-step and adaptive activation steering guided by barrier functions.
• Consistent gains on TruthfulQA (+5.7%), RealToxicityPrompts (+2.4%), UltraFeedback (+2.5%).

ProgressLM: Towards Progress Reasoning in Vision-Language Models

Jianshu Zhang^*, Chengxuan Qian^*, Haosen Sun, Haoran Lu, Dingcheng Wang, Letian Xue, Han Liu

Preprint, 2026; ICLR 2026 Workshop on World Models

[Project Page] [Paper] [Code]

• PROGRESS-BENCH: a benchmark for long-horizon progress reasoning in VLMs, with controlled modality, viewpoint, and answerability.
• Shows that progress reasoning is unstable in current VLMs, and becomes more robust with explicitly trained coarse-to-fine models (ProgressLM-3B).

T*: Re-thinking Temporal Search for Long-Form Video Understanding

Jinhui Ye^*, Zihan Wang^*, Haosen Sun, Keshigeyan Chandrasegaran, Zane Durante, Cristobal Eyzaguirre, Yonatan Bisk, Juan Carlos Niebles, Ehsan Adeli, Fei-Fei Li, Jiajun Wu, Manling Li

Conference on Computer Vision and Pattern Recognition (CVPR), 2025; Oral @ ICCV 2025 LongVid-Foundations, Featured by Stanford AI Blog

[Project Page] [Paper] [Code]

• We introduce LongVideoHaystack (LV-Haystack), a 480-hour dataset for keyframe search in long videos, with 15,092 human-annotated instances (SOTA scores 2.1% Temporal F₁).
• Our framework T* reframed temporal search as spatial search with adaptive zooming, boosting GPT-4o from 50.5% to 53.1% and LLaVA-OV from 56.5% to 62.4% on LongVideoBench XL.

Auto-DAS: Automated Proxy Discovery for Training-free Distillation-aware Architecture Search

Haosen Sun, Lujun Li^†, Peijie Dong, Zimian Wei, Shitong Shao

European Conference on Computer Vision (ECCV), 2024

[Paper] [Code]

• We present Auto-DAS, an automatic proxy discovery framework using an Evolutionary Algorithm (EA) for training-free Distillation-aware Architecture Search (DAS).
• Auto-DAS generalizes well to various architectures and search spaces (e.g. ResNet, ViT, NAS-Bench-101, and NAS-Bench-201), achieving state-of-the-art results in both ranking correlation and final searched accuracy.

Auto-GAS: Automated Proxy Discovery for Training-free Generative Architecture Search

Lujun Li, Haosen Sun, Shiwen Li, Peijie Dong, Wenhan Luo, Wei Xue, Qifeng Liu^†, Yike Guo^†

European Conference on Computer Vision (ECCV), 2024

[Paper] [Code]

• We introduce Auto-GAS, the first training-free Generation Architecture Search (GAS) framework enabled by an auto-discovered proxy, which achieves competitive scores with 110× faster search than GAN Compression.

Inpaint4DNeRF: Promptable Spatio-Temporal NeRF Inpainting with Generative Diffusion Models

Han Jiang^*, Haosen Sun^*, Ruoxuan Li^*, Yu-Wing Tai, Chi-Keung Tang

Arxiv, Dec 2023

[Project Page] [Paper] [Code]

• Inpaint4DNeRF can generate prompt-based objects guided by the seed images and their 3D proxies while preserving multiview consistency. Our generative baseline framework is general which can be readily extended to 4D dynamic NeRFs.

ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment

Haosen Sun^*, Hongjue Zhao^*, Jiangtao Kong, Xiaochang Li, Qineng Wang, Liwei Jiang, Qi Zhu, Tarek F. Abdelzaher, Yejin Choi, Manling Li^†, Huajie Shao^†

International Conference on Learning Representations (ICLR), 2026

[Project Page] [Paper] [Code]

• A unified ODE-based framework for multi-step and adaptive activation steering guided by barrier functions.
• Consistent gains on TruthfulQA (+5.7%), RealToxicityPrompts (+2.4%), UltraFeedback (+2.5%).

ProgressLM: Towards Progress Reasoning in Vision-Language Models

Jianshu Zhang^*, Chengxuan Qian^*, Haosen Sun, Haoran Lu, Dingcheng Wang, Letian Xue, Han Liu

Preprint, 2026; ICLR 2026 Workshop on World Models

[Project Page] [Paper] [Code]

• PROGRESS-BENCH: a benchmark for long-horizon progress reasoning in VLMs, with controlled modality, viewpoint, and answerability.
• Shows that progress reasoning is unstable in current VLMs, and becomes more robust with explicitly trained coarse-to-fine models (ProgressLM-3B).

Flow-Based Knowledge Transfer for Efficient Large Model Distillation

Xinye Yang^*, Junhao Wang^*, Haosen Sun^‡, Xuesheng Zhang, Zebang Liu, Rui Li, Gaochao Xu, Yiwei Chen^†

Association for the Advancement of Artificial Intelligence (AAAI), 2026

[Paper] [Code]

• FlowDistill leverages invertible normalizing flows to achieve lossless teacher-to-student knowledge transfer.
• It outperforms diffusion-based distillation (+2.1%) while cutting inference 3.5×, advancing SOTA across tasks.

T*: Re-thinking Temporal Search for Long-Form Video Understanding

Jinhui Ye^*, Zihan Wang^*, Haosen Sun, Keshigeyan Chandrasegaran, Zane Durante, Cristobal Eyzaguirre, Yonatan Bisk, Juan Carlos Niebles, Ehsan Adeli, Fei-Fei Li, Jiajun Wu, Manling Li

Conference on Computer Vision and Pattern Recognition (CVPR), 2025; Oral @ ICCV 2025 LongVid-Foundations, Featured by Stanford AI Blog

[Project Page] [Paper] [Code]

• We introduce LongVideoHaystack (LV-Haystack), a 480-hour dataset for keyframe search in long videos, with 15,092 human-annotated instances (SOTA scores 2.1% Temporal F₁).
• Our framework T* reframed temporal search as spatial search with adaptive zooming, boosting GPT-4o from 50.5% to 53.1% and LLaVA-OV from 56.5% to 62.4% on LongVideoBench XL.

Auto-DAS: Automated Proxy Discovery for Training-free Distillation-aware Architecture Search

Haosen Sun, Lujun Li^†, Peijie Dong, Zimian Wei, Shitong Shao

European Conference on Computer Vision (ECCV), 2024

[Paper] [Code]

• We present Auto-DAS, an automatic proxy discovery framework using an Evolutionary Algorithm (EA) for training-free Distillation-aware Architecture Search (DAS).
• Auto-DAS generalizes well to various architectures and search spaces (e.g. ResNet, ViT, NAS-Bench-101, and NAS-Bench-201), achieving state-of-the-art results in both ranking correlation and final searched accuracy.

Auto-GAS: Automated Proxy Discovery for Training-free Generative Architecture Search

Lujun Li, Haosen Sun, Shiwen Li, Peijie Dong, Wenhan Luo, Wei Xue, Qifeng Liu^†, Yike Guo^†

European Conference on Computer Vision (ECCV), 2024

[Paper] [Code]

• We introduce Auto-GAS, the first training-free Generation Architecture Search (GAS) framework enabled by an auto-discovered proxy, which achieves competitive scores with 110× faster search than GAN Compression.

Inpaint4DNeRF: Promptable Spatio-Temporal NeRF Inpainting with Generative Diffusion Models

Han Jiang^*, Haosen Sun^*, Ruoxuan Li^*, Yu-Wing Tai, Chi-Keung Tang

Arxiv, Dec 2023

[Project Page] [Paper] [Code]

• Inpaint4DNeRF can generate prompt-based objects guided by the seed images and their 3D proxies while preserving multiview consistency. Our generative baseline framework is general which can be readily extended to 4D dynamic NeRFs.

Registering Neural Radiance Fields as 3D Density Images

Han Jiang^*, Ruoxuan Li^*, Haosen Sun, Yu-Wing Tai, Chi-Keung Tang

Arxiv, May 2023

[Paper]

• We proposes a method to align and merge pre-trained NeRF models of partially overlapping 3D scenes using a generalized registration pipeline, incorporating key point detection, point set registration, and universal pre-trained descriptor networks with contrastive learning strategy.