What’s Left? Concept Grounding with Logic-Enhanced Foundation Models

Authors: Joy Hsu, Jiayuan Mao, Josh Tenenbaum, Jiajun Wu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate LEFT on four domains and seven tasks, and show its effectiveness in multiple settings.
Researcher Affiliation Academia Joy Hsu Stanford University Jiayuan Mao MIT Joshua B. Tenenbaum MIT Jiajun Wu Stanford University
Pseudocode No The paper describes the model's components and execution strategy in prose, but does not include formal pseudocode or algorithm blocks.
Open Source Code Yes We publicly release our code. See our project website for additional details.
Open Datasets Yes We use the official data splits released for each dataset, CLEVR [Johnson et al., 2017a], Refer It3D [Achlioptas et al., 2020], Human Motion QA [Endo et al., 2023], and Cliport [Shridhar et al., 2022].
Dataset Splits Yes We use the official data splits released for each dataset, CLEVR [Johnson et al., 2017a], Refer It3D [Achlioptas et al., 2020], Human Motion QA [Endo et al., 2023], and Cliport [Shridhar et al., 2022].
Hardware Specification Yes We trained with 1 NVIDIA Titan RTX per experiment for all datasets, from an internal cluster.
Software Dependencies No The paper mentions using GPT-3.5 and LLAMA but does not provide specific version numbers for other key software dependencies or libraries (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes The core hyperparameters were set as 128 for concept embedding dimensions, and learning rate taken from prior neuro-symbolic concept learning repositories that we use as baselines.