Simulation of Graph Algorithms with Looped Transformers

Authors: Artur Back De Luca, Kimon Fountoulakis

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

Reproducibility Variable Result LLM Response
Research Type Experimental We prove by construction that this architecture can simulate individual algorithms... We provide empirical validation of the results in Appendix B.3.
Researcher Affiliation Academia 1David R. Cheriton School of Computer Science, University of Waterloo, Waterloo, Ontario, Canada.
Pseudocode Yes Algorithm 1 Looped Transformer (Giannou et al., 2023)
Open Source Code No No explicit statement about providing open-source code for the described methodology or a link to a code repository was found.
Open Datasets Yes We validate our theoretical results using graphs in the CLRS Algorithmic Reasoning Benchmark (Veliˇckovi c et al., 2022).
Dataset Splits Yes For graph-related tasks, the training and validation sets include graphs with 16 nodes, while the test set contains graphs with 64 nodes.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) used for running experiments were mentioned in the paper.
Software Dependencies No The paper mentions software components like ReLU and softmax, but does not provide specific version numbers for any libraries, frameworks, or programming languages used in the implementation.
Experiment Setup Yes In our empirical verifications, we consistently apply specific parameters: an angular increment δ set at 10 2, a maximum value Ωof 105, and a temperature parameter T at 10 7.