Few-shot Language Coordination by Modeling Theory of Mind

Authors: Hao Zhu, Graham Neubig, Yonatan Bisk

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We examine our hypothesis that the instructions generated with To M modeling yield better communication performance in both a referential game and a language navigation task. Positive results from our experiments hint at the importance of explicitly modeling communication as a socio-pragmatic progress. Code can be found at https://github.com/CLAW-Lab/To M.
Researcher Affiliation Academia Hao Zhu 1 Graham Neubig 1 Yonatan Bisk 1 1Language Technologies Institute, Carnegie Mellon University. Correspondence to: Hao Zhu <zhuhao@cmu.edu>.
Pseudocode Yes Procedure 1. General Theory-of-Mind (To M) model training procedure. Algorithm 1 Evaluate To M Model
Open Source Code Yes Code can be found at https://github.com/CLAW-Lab/To M.
Open Datasets Yes Following Lazaridou et al. (2016); Lowe et al. (2019a), we use 30k image-caption pairs from MSCOCO dataset (Lin et al., 2014).
Dataset Splits Yes These listeners are randomly divided into training, validation, and testing listeners (80/20/20). These listeners are randomly divided into training, validation, and testing listeners (30/10/10).
Hardware Specification No No specific hardware details (like GPU models, CPU types, or cloud instance specifications) used for running experiments were mentioned in the paper.
Software Dependencies No The paper mentions software components like LSTMs, ResNet, and Gumbel-softmax, but does not provide specific version numbers for any libraries, frameworks, or programming languages used in the implementation or experiments.
Experiment Setup Yes The MAML hyper-parameters are = 0.01, Ninner = 5, outer = 0.0001, Nouter = 500, and batch size is 2. Within one session, the maximum number of interactions between speaker and listener is K = 100, and maximum number of interactions in a game is 20.