Model AtlasA selection map of agent foundation models中文
Start researchModelsPapersGuideCompare中文

Search the atlas

Esc to close · Cmd/Ctrl + K to open

Why these models were selected

Roles and weight updates come from the paper case record. Selection reasons are written only when the paper or code states them; unknown is never inferred as fact.

Paper model roles and selection-evidence status
ModelRoleWeights updatedSelection basis
GPT-4policyNoNot recorded: verify in the paper/code
GPT-4VpolicyNoNot recorded: verify in the paper/code
Claude 3 OpusbaselineNoNot recorded: verify in the paper/code
Gemini 1.5 ProbaselineNoNot recorded: verify in the paper/code

OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments

2024-04-12 · Verified

Paper info

Evolution targets
Tool use, Multimodal policy, Actor policy
Benchmark
OSWorld
Category
Self-evolving agent, Tool use, Web agent
Model relations
4
Paper link ↗Code link ↗

Start research from this paper

Choose a research mode to carry this paper into the workspace context.

Strict reproductionMethod reproductionModern rerun

Reproducibility status

CodeNot verified
CheckpointNot verified
ConfigNot verified
EnvironmentNot verified

No structured reproducibility record has been added yet.

Model-role topology

This view only shows recorded model-role relations; it does not invent workflow edges.

Policy
GPT-4GPT-4V
Baseline
Claude 3 OpusGemini 1.5 Pro

This paper record has no workflow edges; the topology above only shows recorded roles and does not treat Actor, Environment, Critic, or Optimizer relations as facts.

Model roles

GPT-4

Policy · weights not updated · GPT-4 evaluated as an OSWorld computer-use agent.

GPT-4V

Policy · weights not updated · GPT-4V evaluated as a multimodal OSWorld agent.

Claude 3 Opus

Baseline · weights not updated · Claude 3 Opus benchmarked on OSWorld.

Gemini 1.5 Pro

Baseline · weights not updated · Gemini 1.5 Pro benchmarked on OSWorld.