EmberJin/Minecraft-Agent-CoT-Qwen3.5-9B-Control

Minecraft Agent CoT Control — no thoughts (Qwen3.5-9B, checkpoint 525 / epoch 5)

The attribution control: identical continue-SFT data and hyperparameters as CoT v1, but with all 18,495 thought annotations stripped before training. The only difference vs v1 is the presence/absence of thoughts — this model isolates the data-exposure effect from the thought-supervision effect.

This is also the current best model of the whole project (31.5%).

Training

• Base: EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 (24.9% easy-ng, h29)
• Data: the exact same 6,661 rows as v1 (5,329 CoT-sourced + 1,332 replay), with every Thought: line removed — pure action supervision
• LR 1e-6, 5 epochs, epoch-5 checkpoint (525 steps)

Results (easy-ng, h29, 3 rollouts, seed 42)

	Overall	Embodied	Combat
Base (no continue-SFT)	24.9%	24.8%	25.2%
v1 (raw thoughts, e4)	25.3%	27.5%	18.4%
v2 (cleaned thoughts, e4)	29.2%	31.2%	23.1%
This (no thoughts, e5)	31.5%	32.0%	29.9%

Epoch curve is monotonically increasing (22.6 → 22.3 → 24.6 → 28.5 → 31.5) with no overfitting at 5 epochs — unlike v1 which collapsed at e5.

What this model is for (QC / analysis)

1. The cleanest thought-effect comparator: run identical tasks on this vs v1/v2; any behavioral difference is attributable to thought supervision (Combat degradation, hallucination loops — all absent here by construction; this model never emits Thought: lines).
2. The "data exposure" anchor: +6.6pp over base proves the original single-epoch stage-3 training was undertrained — low-LR continue-SFT on a 3.3% data subset alone recovers most of the CoT experiments' gains.
3. For rollout QC teams: use this as the reference trajectory quality baseline — v1/v2 rollouts should be compared against this model's behavior on the same seeds.

Serving

vllm serve . --served-name cot-ctrl-e5 \
  --max-model-len 32768 --limit-mm-per-prompt image=30

Output format: plain Action: move(dx, dy) and press(keys) [and click(left|right)] — no thoughts, ever.