Symbolic Regression with a Learned Concept Library

Authors: Arya Grayeli, Atharva Sehgal, Omar Costilla Reyes, Miles Cranmer, Swarat Chaudhuri

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally compare LASR on Feynman Equations... We validate LASR on the Feynman equations... On these benchmarks, LASR substantially outperforms a variety of state-of-the-art SR approaches based on deep learning and evolutionary algorithms.
Researcher Affiliation Collaboration Arya Grayeli UT Austin, Foundry Technologies Atharva Sehgal UT Austin Omar Costilla-Reyes MIT Miles Cranmer University of Cambridge Swarat Chaudhuri UT Austin
Pseudocode Yes The full LASR algorithm is presented in Algorithm 1 and visualized in Figure 2.
Open Source Code Yes Artifacts available at https://trishullab.github.io/lasr-web (Footnote 2) and 'We provide a link to an open source implementation of LASR in Julia.' (Checklist item 4).
Open Datasets Yes We evaluate LASR on the Feynman Equations dataset... and Big Bench dataset [16].
Dataset Splits Yes We fit the free parameters of each equation on the training set (43, 049 samples) and measure the MSE loss between the actual grade and the predicted grade on the validation set (10, 763 samples).
Hardware Specification Yes We run all experiments on a server node with 8x A100 GPUs with 80 GB of VRAM each.
Software Dependencies Yes We instantiate LASR using gpt-3.5-turbo-0125 [4] as the backbone LLM... and llama3-8b [17]... We chose to run llama3-8b using v LLM [25].
Experiment Setup Yes We instantiate LASR using gpt-3.5-turbo-0125 [4] as the backbone LLM and calling it with p = 0.01 for 40 iterations. (Section 4.1 Setup)... Figure 7 showcases the hyperparameters used for all our experiments.