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

τ2-bench

Benchmarks

Task NameDataset NameSOTA ResultTrend
Agentic Task Performanceτ2-Bench Airline 1.0 (test)
CAP96.4
48
Agentic Task Performanceτ2-Bench Retail 1.0 (test)
Completion Accuracy (CAP)91.4
48
Agentic Workflow Successτ2-bench
Airline Success Rate76.5
43
Agentτ2-Bench
Accuracy85.4
41
Uncertainty quantificationτ2-bench Retail
AUROC0.899
32
Uncertainty quantificationτ2-bench Airline
AUROC86.5
32
Agent Task Completionτ2-BENCH (test)
Average Task Reward0.921
27
Long-Horizon User-Centric Interactionτ2-Bench
Telecom Success Rate46.9
23
Agentic Task Completionτ2-bench Airline
Success Rate97
22
Agenticτ2-Bench
Score91.6
20
Agentic Task Completionτ2-Bench
Airline Success Rate84
19
Multi-turn tool callingτ2-bench
Airline Score38
19
Agentic taskτ2-Bench Telecom
Avg@2 Score45
8
Agentic taskτ2-Bench Airline
Avg@460
8
Agentic taskτ2-Bench Retail
Avg@469.7
8
Web-based Decision-makingτ2 Bench Retail, Telecom, Airline
Retail Score48.3
5
Agentic Task Completionτ2-bench Retail
Success Rate100
4
Tool-use Policy Auditingτ2-BENCH Telecom (test)
PASS^430
3
Tool-use Policy Auditingτ2-BENCH Telecom base
PASS^420.2
3
Tool-use Policy Auditingτ2-BENCH Retail (base)
PASS^459.6
3
Tool-use Agent Taskτ2-bench (full)
Retail Success Rate82.46
3
Memory-Poisoning Attackτ2-Bench
Attack Hit Rate (AHR)90.5
3
Agentic Task Completionτ2-bench Telecom
Success Rate100
3
Task Resolutionτ2-bench (test)
Success Rate (Airline)38.3
2
Showing 24 of 24 rows