Qwen3-1.7B · Agentick gridworld · online ROSE 10+10 with a Qwen3-32B teacher — step 200

Student: Qwen/Qwen3-1.7B. Part of a study comparing teacher-trajectory SFT, online multi-turn ROSE and on-policy distillation for multi-turn agents.

Training

Online multi-turn ROSE: the student plays the first 10 turns in a live environment, the teacher continues the same environment for up to 10 turns, and only the teacher's turns are cross-entropy targets. 32 episodes per step, AdamW constant lr 1e-5, bf16, thinking off, 512 tokens per turn. Teacher: Qwen3-32B (bf16, local). Agentick ms60 training split (20000 tasks, 625 steps per epoch); the run was stopped after step 400, this is step 200. 256 tokens per turn.

Evaluation

Agentick official eval (245 tasks, max 60 steps, temperature 0.4, thinking off), one pass: success 11.4%. Reference from a collaborator's run on the same tasks: base Qwen3-1.7B 5.70%, Qwen3-32B 22.40%, teacher-SFT 25.30%.

Prompt format

ReAct, multi-turn chat: the environment's AgentGym instruction as the first user turn, a canned assistant acknowledgement, then one user turn per observation; the model replies Thought:\n...\n\nAction:\n<one action>. Render with Qwen3's chat template and enable_thinking=False.