Back to Arena

BrowserGym

by ServiceNow Research

System Card

OrganizationServiceNow Research
Released2024-02
Architectureagentic-workflow / Standardized Gym env for web agents (memory evaluation)
DetailsGym-style framework for web-agent research supporting MiniWoB, WebArena, WorkArena, VisualWebArena. Standard API for state/action/reward, memory across steps, and multi-benchmark integration.
Parameters—
Domainagent-memory
Open SourceYes
PaperView Paper
CodeRepository
servicenowgymwebarenabenchmark-harness

Capability Profile

Benchmark Scores

5 of 14 benchmarks
Long-Context Retrieval
0/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
no data
∞Bench
no data
Multi-Turn Recall
1/2
LoCoMo
80.388p
MemoryBank
no data
Cross-Session Memory
1/1
LongMemEval
80.784p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
7579p
HotpotQA
75.371p
Agent Task Memory
1/1
AgentBench-Mem
7226p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:BrowserGym paper (arXiv:2412.05467); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)BrowserGym paper (arXiv:2412.05467); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)BrowserGym paper (arXiv:2412.05467); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)BrowserGym paper (arXiv:2412.05467); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)BrowserGym paper (arXiv:2412.05467); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)