Cold-start SFT from Qwen3.5-4B for GUI next-action prediction. This checkpoint = the retrieved trajectory memory (main result) arm of a 3-arm A/B.

Action accuracy (n=498 test, AgentNetBench score_pair): 0.556 — +19.0pp vs basecur, +8.6pp vs basefull; usage-gap +11.4pp (memory is genuinely used)

arm	action acc (n=498)
basecur (current only)	0.366
basefull (full history)	0.470
mem (retrieved memory)	0.556

Status: v1, single-seed (positive; 3-seed confirmation pending). See the collection for the other arms.

Load

python

from transformers import AutoProcessor
from qwen_cua.modeling_qwen35_vl_latent import Qwen35VLLatentForConditionalGeneration as M
proc  = AutoProcessor.from_pretrained("hyunseoki/memrag-mem", max_pixels=1_000_000)
model = M.from_pretrained("hyunseoki/memrag-mem", torch_dtype="bfloat16", attn_implementation="flash_attention_2")

Plain Qwen3.5-VL arch (wm.enabled=false) — also loadable with the standard class.