Jiayi Pan, Yichi Zhang, Nickolas Tomlin, Yifei Zhou, Sergey Levine, Alane Suhr. COLM 2024 / ⭐️ MAR Workshop @ CVPR 2024 Best Paper.

We design model-based evaluators to both evaluate and autonomously improve agents' performance. We show that these open-ended evaluators can significantly improve agents' performance, through either fine-tuning or inference-time guidance, without any extra supervision.

OpenHands Community. Preprint.

We introduce OpenHands, a platform for developing AI agents that interact with the digital world. OpenHands is a community project from over 180 contributors and 30K+ stars.

Hao Bai*, Yifei Zhou*, Mert Cemri, Jiayi Pan, Alane Suhr, Sergey Levine, Aviral Kumar. NIPS 2024.

We develop reinforcement learning techniques to post-train device-control language agents. Our 2B VLM, when post-trained with an autonomous evaluator, improves its success rate from 17% to 67% on Android device-control tasks.

Yuexiang Zhai, Hao Bai*, Zipeng Lin*, Jiayi Pan*, Shengbang Tong*, Yifei Zhou*, Alane Suhr, Saining Xie, Yann LeCun, Yi Ma, Sergey Levine. NIPS 2024.

We provide infrastructure and environment for training VLMs with RL on decision-making tasks. We show RL training enables our 7B model to outperform GPT-4V on these tasks. Additionally, we show the intriguing effectiveness of CoT reasoning for performance improvement

Yifei Zhou, Andrea Zanette, Jiayi Pan, Sergey Levine, Aviral Kumar. ICML 2024.

We present ArCHer, a new framework of multi-turn RL algorithms for training LM agents. It preserves the flexibility of mainstream single-turn LM RL methods like PPO, while effectively handling multiple turns, long horizons, and delayed rewards.

Sihan Xu*, Yidong Huang*, Jiayi Pan, Ziqiao Ma, Joyce Chai. CVPR 2024.

We present an inversion-free editing (InfEdit) method which enables consistent natural language guided image editing. InfEdit excels in complex editing tasks and is ~10X faster than prior methods.

Yichi Zhang, Jiayi Pan, Yuchen Zhou, Rui Pan, Joyce Chai. EMNLP 2023.

Do Vision-Language Models, an emergent human-computer interface, experience visual illusions similarly to humans, or do they accurately depict reality? We created GVIL dataset to study this. Among other findings, we discover that larger models align more closely with human perception.

Ziqiao Ma*, Jiayi Pan*, Joyce Chai. ⭐️ ACL 2023 Outstanding Paper.

We study grounding and bootstrapping in open-world language learning through Grounded Open Vocabulary Acquisition. Our visually-grounded language model, OctoBERT, excels in learning grounded words quickly and robustly.

Team SEAGULL at UMich. 🏆 1st Place in the inaugural Alexa Prize SimBot Challenge.

We introduce SEAGULL, an interactive embodied agent which completes complex tasks in the Arena simulation environment through dialog with users. SEAGULL is engineered to be efficient, user-centric, and continuously improving.

Jiayi Pan, Glen Chou, Dmitry Berenson. ICRA 2023.

We present a learning-based approach to translate from natural language commands to LTL specifications with only a handful of labeled data. It enables few-shot learning of LTL translators while achieving state-of-the-art performance.

Yichi Zhang, Jianing Yang, Jiayi Pan, Shane Storks, Nikhil Devraj, Ziqiao Ma, Keunwoo Peter Yu, Yuwei Bao, Joyce Chai. EMNLP 2022, Oral.

We introduce DANLI, a neural-symbolic agent that proactively reasons and plans according to its past experiences. DANLI achieves a 70% improvement on the challenging TEACh benchmark while improving transparency in its behaviors.