OSWorld

https://os-world.github.io/

HKU NLP + Salesforce's OSWorld — full-desktop agent benchmark with 369 tasks spanning real applications (LibreOffice, GIMP, Chrome, VS Code, Thunderbird, Files) across Ubuntu, Windows, and macOS. Tasks require multi-application workflows; evaluation via post-task state inspection (not just screenshot match). Released NeurIPS 2024.

At a glance

Type
Desktop OS agent benchmark — 369 tasks across Ubuntu / Windows / macOS
Tier
T3
Created
2024-04 (Xie et al. arXiv 2404.07972)
Latest release
not applicable — not OSS
License
Apache-2.0
Pricing
not applicable — open benchmark; free to use
Funding
not applicable — not commercial

Taxonomy

storage
n/a
retrieval
n/a
persistence
n/a
update
read-only
unit
n/a
governance
n/a
conflict
n/a

When to use

Optimised for: not applicable - eval dataset, not a system

Anti-fit: not applicable - eval dataset, not a system

Pros & cons

Pros

Multi-OS (Linux + Windows + macOS) coverage unique among OS benchmarks; real apps; state-based evaluation more robust than screenshot match.

Cons

VM setup is heavyweight (gigabytes); SOTA still ~22% — long horizon to human parity; OSWorld game tasks (per spec) are not the focus — productivity apps dominate.

Claims & capabilities

NeurIPS 2024; 369 real-OS tasks across 3 operating systems; SOTA (Claude 3.5 Sonnet + Anthropic Computer Use) ~22% success

Technical surface

API surface
not applicable — eval dataset, not a system
Backend storage
not applicable — eval dataset, not a system
Deployment
not applicable — run locally
Embedding model
not applicable — eval dataset, not a system
Multi-tenancy
not applicable — eval dataset, not a system
MCP
not applicable — benchmark / evaluation harness
A2A
not applicable — benchmark / evaluation harness
OpenTelemetry
not applicable — benchmark / evaluation harness

Similar systems

Other memory benchmarks & evaluation in the catalog, ranked by inbound references.

  • LoCoMo T3

    Conversations spanning up to 35 sessions / 300 turns / ~9K tokens, built from LLM-agent personas grounded on temporal event graphs and verified by human annotators. Tests QA (single-hop, multi-hop, temporal, commonsense, adversarial), event-graph summarisation, multimodal dialogue.

  • LongBench T3

    Tsinghua THUDM. 21 datasets across 6 categories (single-doc QA, multi-doc QA, summarisation, few-shot ICL, synthetic, code completion) in English + Chinese. ~6.7K words English / 13.4K chars Chinese average. v2 at ACL 2025 with harder, longer instances.

  • LongMemEval T3

    5 core memory abilities — info extraction, multi-session reasoning, temporal reasoning, knowledge updates, abstention — via 500 curated questions in synthetic chat histories. Two scales: LongMemEval-S (115K tokens), LongMemEval-M (up to 1.5M).

  • RULER T3

    NVIDIA. Extends NIAH with multi-needle retrieval, multi-hop entity tracing, aggregation, QA — configurable for arbitrary length / difficulty. Evaluated 17 long-context models on 13 tasks.

  • InfiniteBench (∞Bench) T3

    Tsinghua / OpenBMB. First benchmark with average input >100K tokens. 12 tasks across English + Chinese: long novel QA, code debugging, math in long context, summarisation, retrieval. Realistic documents (novels, code) rather than synthetic filler.

  • Atari 100k T3

    Sample-efficiency benchmark protocol on Arcade Learning Environment (ALE) — agents limited to 100k environment steps (~2 hours of game-play) before evaluation. Introduced by SimPLe (Kaiser et al. 2019) and adopted by EfficientZero, IRIS, BBF, DreamerV3, Storm. Tests world-model / model-based methods where memory-as-imagination matters more than huge replay buffers.

Row last verified 2026-05-14. Catalog data is CC-BY-4.0 — see how to read this.