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

SkillsBench

Benchmarks

Task NameDataset NameSOTA ResultTrend
Downstream skill executionSkillsBench (test)
Reward48.9
30
Skill executionSkillsBench
Overall Success Rate (avg@5)56.9
26
Agent task executionSkillsBench 11 domains
Overall Score44.8
16
Skill-Augmented Agent Performance EvaluationSkillsBench 86 runnable tasks (frozen-policy transfer)
Valid Task Count79
15
Downstream task performanceSkillsBench
Pass Rate51.1
12
Prefix-risk rankingSkillsBench (held-out)
AUPRC53.3
11
Skill predictionSkillsBench Real-task holdout n=65
Set F174.2
10
Skill predictionSkillsBench Synthetic n=494 (test)
Set F173.9
10
Agent Task CompletionSkillsBench
Pass Rate22.6
9
Skill SelectionSkillsBench 1000-skill scale
Reward (%)36.8
8
Agent task executionSkillsBench 1.0 (test)
Pass Rate (With Skills)71.1
8
Poisoning Injection AttackSkillsBench n=81
Verification Rate25.9
7
Skill-based Task ExecutionSkillsBench
Accuracy68.4
6
Downstream task executionSkillsBench
Reward Mean (%)29.26
6
Skill-Augmented Agent Performance EvaluationSkillsBench 20-task balanced panel (held-out evaluation)
Valid Task Count19
5
Skill-augmented Task CompletionSkillsBench 77 tasks (test)
Mean Reward0.46
4
Task completion pass rateSkillsBench 77 tasks
Coverage72.7
4
Video IndexSkillsBench
Solve Tokens105,000
4
PPTXSkillsBench
Tokens Used (Total)39,000
4
Offer LetterSkillsBench
Tokens Used41,000
4
Mars CloudsSkillsBench
Solve Tokens (K)52
4
JAXSkillsBench
Solve Tokens30,000
4
CitationSkillsBench
Solve Tokens Count66
4
3D ScanSkillsBench
Solve Tokens (K)15
4
Skill GenerationSkillsBench
Baseline Success Rate22
4
Showing 25 of 32 rows