Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Sciworld

Benchmarks

Task NameDataset NameSOTA ResultTrend
Scientific ReasoningSciWorld
Accuracy95.9
164
Science SimulationSciWorld
Accuracy94.6
41
Embodied AgenticSciWorld
Accuracy88.9
21
One-step next-observation predictionSciWorld (test)
Token F196
16
Next-state predictionSciWorld (SW)
EM Accuracy98.64
16
Task successSciWorld
Real68.21
14
Scientific ReasoningSciWorld
Success Rate (SR)59.48
14
Agent TaskSciWorld
Success Rate @198.4
13
Text-based embodied taskSciworld
Success Rate77.77
13
Embodied science experimentsSciWorld Seen
Average Accumulated Reward76.3
12
Interactive Decision MakingSciWorld
Score67.3
12
Science simulationSciworld
Progress Rate82.6
12
Scientific SimulationSciWorld
Measure Score55.7
10
Lifelong Agent InteractionSciWorld
Success Rate (SR)73.5
10
Interactive AgentSciWorld
Pass@129.45
10
Interactive LLM Agent Task SuccessSciWorld Unseen
Success Rate82.4
7
Interactive LLM Agent Task SuccessSciWorld (Seen)
Success Rate86.4
7
Autonomous ExplorationSciWorld
Steps7.4
7
Agentic Action PredictionSCIWorld (out domain)
Success Rate44
4
Agentic Action PredictionSCIWorld (in domain)
Success Rate61
4
Science Experiment SimulationSciWorld
Average Reward69.1
3
Showing 21 of 21 rows