awm-qwen3.6-27b-rl-d498-ep3

Qwen3.6-27B trained with GRPO (Dr.GRPO variant) to act as an orchestrator in a two-agent system: it decomposes a task, delegates subtasks to a second model, and integrates the results. Only orchestrator tokens receive gradient (train_roles: ["orch"]) — the sub-agent is never trained.

Training

Base model	Qwen/Qwen3.6-27B
Algorithm	GRPO / Dr.GRPO (--disable-grpo-std-normalization)
Training set	498 tasks
Epochs	3
Samples per prompt	8, DAPO dynamic sampling (over-sample 64 / batch 32)
Parallelism	8× p5e nodes, tp×cp = 16, dp = 4
Stack	slime + Megatron-LM + SGLang

Evaluation

196 held-out tasks × 3 seeds = 588 runs per setting. All rows share identical settings (n=3, base_seed=42, max_turns=200, orch_style=pure); the comparison against the base model is a paired t-test over per-task means.

The two "fixed sub" rows hold the sub-agent constant, so any difference is attributable to orchestration alone rather than to the executor improving.

setting	base Qwen3.6-27B	this model	paired diff
single agent (no delegation)	0.3808	0.4016 (n=584/588)	+0.0208 (t=+2.54) ✅
self-paired (orch = sub = this model)	0.3499	0.4023	+0.0524 (t=+5.18) ✅
fixed weak sub (Qwen3.6-35B-A3B)	0.3110	0.3464	+0.0354 (t=+2.96) ✅
fixed strong sub (DeepSeek-V4-Flash)	0.4153	0.4428	+0.0275 (t=+2.72) ✅

Reference: DeepSeek-V4-Flash orchestrating itself scores 0.4726 in the strong-sub setting. The 1001-task checkpoints match that number exactly; the 498-task ones remain significantly below it.

Usage

sglang serve --model-path VanWang/awm-qwen3.6-27b-rl-d498-ep3 --tp 8

The served model name must match what your client sends. The model expects the orchestration prompt format used during training.