From 58c2412ced028e8f2a13fcb5f6d65b6f3ad5d9fc Mon Sep 17 00:00:00 2001 From: genitrix Date: Tue, 28 Jul 2026 17:13:31 +0800 Subject: [PATCH] Add ClawBench web-agent benchmark --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 2616d5d..03fc094 100644 --- a/README.md +++ b/README.md @@ -56,6 +56,7 @@ Save hours of research time—get straight to what matters. | CompWoB | https://github.com/google-research/google-research/tree/master/compositional_rl/compwob | 2023 | Tests agents on sequential, compositional tasks that require state management across multiple steps, simulating real-world automation scenarios. | 50 compositional tasks | Task Success Rate | Element Match | Web | | TURKING BENCH | https://turkingbench.github.io | 2024 | Uses natural HTML tasks from crowdsourcing to assess interaction skills with real-world web layouts and elements. | 32.2K instances | Task Success Rate | Text Match, Element Match, Image Match | Web | | VisualWebBench | https://visualwebbench.github.io | 2024 | Provides a fine-grained assessment of multimodal large language models (MLLMs) on web-specific tasks. | 1,534 instances from 139 real websites across 87 sub-domains | Task Success Rate, Turn Success Rate, Efficiency Score, Metrics | Text Match, Image Match, Element Match, Action Match | Web | +| ClawBench | https://claw-bench.com | 2026 | Evaluates AI agents on everyday online tasks in isolated live-web environments with submission interception, execution traces, and judge-assisted outcome assessment. | 283 tasks across 163 live websites in V1+V2 | Task Success Rate, trajectory evidence, judge assessment | HTTP/request interception, browser actions, screenshots, recordings, agent messages | Web | | WONDERBREAD | https://github.com/HazyResearch/wonderbread | 2024 | Focuses on business process management (BPM) tasks like documentation, knowledge transfer, and process improvement. | 2,928 human demonstrations across 598 distinct workflows | Task Success Rate, Step Success Rate, Efficiency Score, Completion under Policy | Text Match, Action Match, State Information | Web | | WebOlympus | / | 2024 | An open platform for web agents that simplifies running demos, evaluations, and data collection for web agents on live websites. | 50 tasks | Task Success Rate, Step Success Rate | Action Match | Web | | AndroidEnv | https://github.com/google-deepmind/android_env | 2021 | Provides an open-source platform based on the Android ecosystem with over 100 tasks across approximately 30 apps, focusing on reinforcement learning for various Android interactions. | 100+ tasks | NA | NA | Android |