Скачать или смотреть Web-Shepherd: First Web Navigation PRM

In this AI Research Roundup episode, Alex discusses the paper:
'Web-Shepherd: Advancing PRMs for Reinforcing Web Agents'
Web-Shepherd introduces the first process reward model (PRM) specifically for web navigation, tackling the challenge of creating reliable and affordable reward signals for web agents. Current multimodal large language models (MLLMs) are often slow, expensive, and inaccurate when used as reward models for these long-horizon tasks. To address this, Web-Shepherd proposes a novel step-level reward assessment methodology. This research also contributes the WebPRM Collection, a large dataset of step-level preference pairs, and WebRewardBench, the first benchmark for evaluating PRMs in web navigation, aiming to improve how web agents learn complex online tasks.
Paper URL: https://huggingface.co/papers/2505.15277

#AI #MachineLearning #DeepLearning #WebNavigation #RewardModel #WebAgents #ReinforcementLearning #LanguageModels