List your company

Long-Horizon User-Centric Interaction on τ2-Bench

46.9Telecom Success RateGemini-3-Flash
7.918.02528.1538.275Feb 28, 2026

Evaluation results

MethodLinks
Gemini-3-FlashType=Prompting, Backbone=Closed-Source Model2026.0246.961.955.8
GPT-4.1Type=Prompting, Backbone=Closed-Source Model2026.0238.854.440
InfoPO (Ours)Type=RL Training, Backbone=Qwen2.5-7B-Instruct2026.0218.118.816.3
RAGENType=RL Training, Backbone=Qwen2.5-7B-Instruct2026.0217.517.515
InfoPO w/o stdType=RL Training, Backbone=Qwen2.5-7B-Instruct2026.0216.916.212.5
Search-R1Type=RL Training, Backbone=Qwen2.5-7B-Instruct2026.0215.61510
InfoPO (Ours)Type=RL Training, Backbone=Qwen3-4B2026.0215.624.410
InfoPO w/o GateType=RL Training, Backbone=Qwen2.5-7B-Instruct2026.021515.610
Qwen2.5Type=Prompting, Backbone=Qwen2.5-7B-Instruct2026.0214.413.17.5
InfoPO w/o stdType=RL Training, Backbone=Qwen3-4B2026.0214.221.58.8
ReflexionType=Prompting, Backbone=Qwen2.5-7B-Instruct2026.0213.813.17.5
UserRLType=RL Training, Backbone=Qwen2.5-7B-Instruct2026.0213.816.97.5
RAGENType=RL Training, Backbone=Qwen3-4B2026.0213.723.18.8
ReActType=Prompting, Backbone=Qwen2.5-7B-Instruct2026.0213.113.810
InfoPO w/o GateType=RL Training, Backbone=Qwen3-4B2026.0213.119.87.5
Search-R1Type=RL Training, Backbone=Qwen3-4B2026.0211.823.48.8
GPT-4o-miniType=Prompting, Backbone=Closed-Source Model2026.0211.34017.5
InfoPO w/o RextType=RL Training, Backbone=Qwen2.5-7B-Instruct2026.0211.212.58.8
Qwen3Type=Prompting, Backbone=Qwen3-4B2026.0210.622.56.2
ReflexionType=Prompting, Backbone=Qwen3-4B2026.0210207.5
UserRLType=RL Training, Backbone=Qwen3-4B2026.021016.37.5
InfoPO w/o RextType=RL Training, Backbone=Qwen3-4B2026.029.515.25.5
ReActType=Prompting, Backbone=Qwen3-4B2026.029.422.58.8