Modelpublic

JamesK2W/viewagent-ai2thor-qwen25vl7b-ivp

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes6downloads

Model Card

ViewAgent — Qwen2.5-VL-7B trained on AI2-THOR Interactive View Planning (IVP)

Qwen2.5-VL-7B-Instruct trained with the GraphRL pipeline (verl RL ↔ view-graph SFT distillation) on the AI2-THOR port of the ViewSuite proxy-task suite.

Test results (AI2-THOR test split, n=252)

Model	P2V	V2P	IVP	Overall
Qwen2.5-VL-7B (base)	40.9	27.4	4.0	24.1
This model (trained)	25.0	41.3	61.1	42.5
GPT-5.4 (zero-shot)	73.8	69.4	36.1	59.8

IVP (multi-turn active exploration): 61.1% — +57 over base, +25 over the best frontier zero-shot model (GPT-5.4). IVP-only RL trades some single-turn MCQ accuracy (P2V 40.9→25.0). Checkpoint = iter3 RL globalstep460.

—Code: https://github.com/JamesKrW/ViewAgent (branch feat/ai2thor-env)
—Dataset: https://huggingface.co/datasets/JamesK2W/viewagent-ai2thor