Motif: Intrinsic Motivation from Artificial Intelligence Feedback

Authors: Martin Klissarov, Pierluca D'Oro, Shagun Sodhani, Roberta Raileanu, Pierre-Luc Bacon, Pascal Vincent, Amy Zhang, Mikael Henaff

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Motif s performance and behavior on the challenging, open-ended and procedurally-generated Net Hack game.
Researcher Affiliation Collaboration 1 Mila, 2 FAIR at Meta, 3 UT Austin, 4 Universit e de Montr eal, 5 Mc Gill University
Pseudocode No The paper describes its method (Motif) and mentions the use of an RL algorithm (PPO), but it does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at: https://github.com/facebookresearch/motif
Open Datasets Yes To further encourage reproducibility and scientific discoveries, we also release our complete Llama 2 annotations for all experiments.
Dataset Splits Yes We split the dataset of annotation into a training set containing 80% of the datapoints and a validation set containing 20%.
Hardware Specification Yes Sample Factory includes a an extremely fast implementation of PPO (Schulman et al., 2017) which runs at about 20K frames-per-second using 20 computer cores and one V 100 GPU. If annotation is done on A100s GPUs the compute costs can be cut approximately in half.
Software Dependencies No The paper mentions software like "Sample Factory", "Llama 2", and the "vLLM Python module", but it does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We provide all hyperparemeters in Table 2.