List your company

Tool Use on Tool-use multi-turn (test)

76.8AccuracyDiverge (ExIt ablation)
59.84864.24968.6573.051Sep 4, 2025

Evaluation results

MethodLinks
Diverge (ExIt ablation)Backbone=Qwen2.5-7B-Instruct, Improvement steps (K)=4, Number of training runs=32025.0976.80.6
Full ExItBackbone=Qwen2.5-7B-Instruct, Improvement steps (K)=4, Number of training runs=32025.0976.41.2
Improve (ExIt ablation)Backbone=Qwen2.5-7B-Instruct, Improvement steps (K)=4, Number of training runs=32025.0975.53.4
GRPO + curriculumBackbone=Qwen2.5-7B-Instruct, Improvement steps (K)=4, Number of training runs=32025.0975.41
GRPOBackbone=Qwen2.5-7B-Instruct, Improvement steps (K)=4, Number of training runs=32025.0973.5-0.5
Base modelBackbone=Qwen2.5-7B-Instruct, Improvement steps (K)=4, Number of training runs=32025.0960.50.1