Web-Shepherd: Advancing PRMs for Reinforcing Web Agents

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Chae, Hyungjoo, Kim, Sunghwan, Cho, Junhee, Kim, Seungone, Moon, Seungjun, Hwangbo, Gyeom, Lim, Dongha, Kim, Minjin, Hwang, Yeonjun, Gwak, Minju, Choi, Dongwook, Kang, Minseok, Im, Gwanhoon, Cho, ByeongUng, Kim, Hyojun, Han, Jun Hee, Kwon, Taeyoon, Kim, Minju, Kwak, Beong-woo, Kang, Dongjin, Yeo, Jinyoung
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908676246208512
author Chae, Hyungjoo
Kim, Sunghwan
Cho, Junhee
Kim, Seungone
Moon, Seungjun
Hwangbo, Gyeom
Lim, Dongha
Kim, Minjin
Hwang, Yeonjun
Gwak, Minju
Choi, Dongwook
Kang, Minseok
Im, Gwanhoon
Cho, ByeongUng
Kim, Hyojun
Han, Jun Hee
Kwon, Taeyoon
Kim, Minju
Kwak, Beong-woo
Kang, Dongjin
Yeo, Jinyoung
author_facet Chae, Hyungjoo
Kim, Sunghwan
Cho, Junhee
Kim, Seungone
Moon, Seungjun
Hwangbo, Gyeom
Lim, Dongha
Kim, Minjin
Hwang, Yeonjun
Gwak, Minju
Choi, Dongwook
Kang, Minseok
Im, Gwanhoon
Cho, ByeongUng
Kim, Hyojun
Han, Jun Hee
Kwon, Taeyoon
Kim, Minju
Kwak, Beong-woo
Kang, Dongjin
Yeo, Jinyoung
contents Web navigation is a unique domain that can automate many repetitive real-life tasks and is challenging as it requires long-horizon sequential decision making beyond typical multimodal large language model (MLLM) tasks. Yet, specialized reward models for web navigation that can be utilized during both training and test-time have been absent until now. Despite the importance of speed and cost-effectiveness, prior works have utilized MLLMs as reward models, which poses significant constraints for real-world deployment. To address this, in this work, we propose the first process reward model (PRM) called Web-Shepherd which could assess web navigation trajectories in a step-level. To achieve this, we first construct the WebPRM Collection, a large-scale dataset with 40K step-level preference pairs and annotated checklists spanning diverse domains and difficulty levels. Next, we also introduce the WebRewardBench, the first meta-evaluation benchmark for evaluating PRMs. In our experiments, we observe that our Web-Shepherd achieves about 30 points better accuracy compared to using GPT-4o on WebRewardBench. Furthermore, when testing on WebArena-lite by using GPT-4o-mini as the policy and Web-Shepherd as the verifier, we achieve 10.9 points better performance, in 10 less cost compared to using GPT-4o-mini as the verifier. Our model, dataset, and code are publicly available at LINK.
format Preprint
id arxiv_https___arxiv_org_abs_2505_15277
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Web-Shepherd: Advancing PRMs for Reinforcing Web Agents
Chae, Hyungjoo
Kim, Sunghwan
Cho, Junhee
Kim, Seungone
Moon, Seungjun
Hwangbo, Gyeom
Lim, Dongha
Kim, Minjin
Hwang, Yeonjun
Gwak, Minju
Choi, Dongwook
Kang, Minseok
Im, Gwanhoon
Cho, ByeongUng
Kim, Hyojun
Han, Jun Hee
Kwon, Taeyoon
Kim, Minju
Kwak, Beong-woo
Kang, Dongjin
Yeo, Jinyoung
Computation and Language
Web navigation is a unique domain that can automate many repetitive real-life tasks and is challenging as it requires long-horizon sequential decision making beyond typical multimodal large language model (MLLM) tasks. Yet, specialized reward models for web navigation that can be utilized during both training and test-time have been absent until now. Despite the importance of speed and cost-effectiveness, prior works have utilized MLLMs as reward models, which poses significant constraints for real-world deployment. To address this, in this work, we propose the first process reward model (PRM) called Web-Shepherd which could assess web navigation trajectories in a step-level. To achieve this, we first construct the WebPRM Collection, a large-scale dataset with 40K step-level preference pairs and annotated checklists spanning diverse domains and difficulty levels. Next, we also introduce the WebRewardBench, the first meta-evaluation benchmark for evaluating PRMs. In our experiments, we observe that our Web-Shepherd achieves about 30 points better accuracy compared to using GPT-4o on WebRewardBench. Furthermore, when testing on WebArena-lite by using GPT-4o-mini as the policy and Web-Shepherd as the verifier, we achieve 10.9 points better performance, in 10 less cost compared to using GPT-4o-mini as the verifier. Our model, dataset, and code are publicly available at LINK.
title Web-Shepherd: Advancing PRMs for Reinforcing Web Agents
topic Computation and Language
url https://arxiv.org/abs/2505.15277