This repository is the official implementation of the NeurIPS Spotlight paper Benchmarking Egocentric Multimodal Goal Inference for Assistive Wearable Agents.

This repository contains the code to evaluate the Qwen2.5-VL-72B model on the benchmark, for both the Multi-Choice-Question (MCQ) and Generative-LLM-Judge settings. Also, we provide the raw predictions for all models used in the paper, and the analysis script to help you reproduce the all the results.

This repository requires python3.9.

To install requirements:

pip install -r requirements.txt

You may download the dataset here: https://dl.fbaipublicfiles.com/wagibench/ob2.tar.gz and unpack it locally.

You can download the Qwen2.5-VL-72B-Instruct model from HuggingFace: https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct

And the LLM Judge model from HuggingFace: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B

With both dataset and model downloaded, you should have a directory structure like this:

ROOT_DIR
├── ob2
│   ├── digital_state_v2.4.4.json
│   ├── ob2_v2.4.4_1_mcq.json
│   ├── structured_goals.json
│   ├── ob2_transcriptions_250501.json
│   ├── structured_goals_longitudinal_history_v244.json
│   ├── clipped_videos
│   │   ├── 001a1239-f4f6-4a93-8e1b-533f0901b7f3.mp4
│   │   ├── ...
├── models
│   ├── Qwen2.5-VL-72B-Instruct
│   │── DeepSeek-R1-Distill-Llama-70B
├── assets
│   ├── raw_predictions
│   │   ├── generative_internvl_2b_V.json
│   │   ├── ...

First, edit the ROOT_DIR variable in evaluation scripts eval_qwen_mcq.sh and eval_qwen_generative.sh to point to the correct path.

You need at least 4 GPUs with 80GB memory (e.g. A100-80G, H100) to run the evaluation.

To evaluate Qwen2.5-VL-72B-Instruct model on the Multi-Choice-Question task of the benchmark, run:

CUDA_VISIBLE_DEVICES=0,1,2,3 bash eval_qwen_mcq.sh

To evaluate Qwen2.5-VL-72B-Instruct model on the generative task of the benchmark, run:

CUDA_VISIBLE_DEVICES=0,1,2,3 bash eval_qwen_generative.sh

This will evaluate Qwen2.5-VL-72B-Instruct model using DeepSeek-R1-Distill-Llama-70B as the LLM Judge.

If the benchmark is properly setup, you should see similar results for both runs:

MCQ: 0.871299
Generative: 0.507109

The analysis script consolidates results from different models, creates multiple data subsets, and evaluates results on them using varying combinations of input modalities.

You need two specific OB2 input files under ob2 and all the raw prediction files to be present under assets/raw_predictions to run subset analysis:

ROOT_DIR
├── ob2
│   ├── digital_state_v2.4.4.json
│   ├── structured_goals.json
├── assets
│   ├── raw_predictions
│   │   ├── generative_internvl_2b_V.json
│   │   ├── generative_internvl_2b_VA.json
│   │   ├── ...

Run the command:

python subset_analysis.py -o ~/temp/

This will generate a temp directory in your home folder and save all the results and plots in it. If run successfully, you should see results similar to the ones below around script's termination output on the terminal:

#### MODEL RESULTS on generative task ####
internvl_2b: 0.2187 (0.2078, 0.2297)
qwen_3b: 0.2534 (0.2418, 0.2649)
llama3v: 0.3235 (0.3119, 0.3351)
internvl_8b: 0.3571 (0.3442, 0.3691)
qwen_7b: 0.4076 (0.3952, 0.42)
llama4v: 0.4484 (0.4346, 0.46)
internvl_78b: 0.4956 (0.4821, 0.5077)
qwen_72b: 0.4988 (0.4851, 0.5115)

This project is licensed under the Apache 2.0 license.

See contributing and Code of Conduct.

If you use WAGIBench, please feel free to use the following BibTex entry.

@inproceedings{
veerabadran2025benchmarking,
title={Benchmarking Egocentric Multimodal Goal Inference for Assistive Wearable Agents},
author={Vijay Veerabadran and Fanyi Xiao and Nitin Kamra and Pedro Matias and Joy Chen and Caley Drooff and Brett D Roads and Riley Williams and Ethan Henderson and Xuanyi Zhao and Kevin Carlberg and Joseph Tighe and Karl Ridgeway},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
year={2025},
url={https://openreview.net/forum?id=REG4cJItSZ}
}