IC-477Released LLMs achieve limited success rates as web agents on WebArena-Lite, with open-source models substantially below proprietary ones

Zehan Qi, Xiao Liu, Iat Long Iong, Hanyu Lai, Xueqiao Sun, Jiadai Sun, Xinyue Yang, Yu Yang, Shuntian Yao, Wei Xu, Jie Tang, Yuxiao Dong

SourceWebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning

The paper evaluates several released LLMs on the WebArena-Lite benchmark (165 test cases across five websites) using a simple prompt that defines feasible actions and provides examples. GPT-4-turbo achieves 17.6% average success rate, GPT-4o achieves 13.9%, LLaMA-3.1-70B-Instruct achieves 12.7%, LLaMA-3.1-8B-Instruct achieves 4.8%, and GLM-4-chat achieves 6.1%. The previously released AutoWebGLM agent achieves 18.2%. These results establish a significant gap between proprietary and open-source LLMs as web navigation agents.

Evidence
correlational
Key metric
GPT-4-turbo 17.6%, GPT-4o 13.9%, LLaMA-3.1-8B-Instruct 4.8%, LLaMA-3.1-70B-Instruct 12.7%, GLM-4-chat 6.1%, AutoWebGLM 18.2% average SR on WebArena-Lite
Caveat
Results use a single simple prompt; the paper notes that high-performing agents rely on 'meticulously crafted prompts' and proprietary APIs, so the simple-prompt setting may understate model capability. AWM+GPT-4-0613 (35.5%) and WebPilot+GPT-4o (37.2%) from prior work show that method scaffolding substantially changes the picture.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report GPT-4 Turbo, GPT-4o, Llama 3.1, GLM-4 GLM-4-chat, AutoWebGLM
Datasets
WebArena-Lite [eval]
Related work
AWM [compared-to], WebPilot [compared-to]
Related findings
IC-478
Extraction
automatic-extraction