Frugal LMs Trained to Invoke Symbolic Solvers Achieve Parameter-Efficient Arithmetic Reasoning

Authors: Subhabrata Dutta, Ishan Pandey, Joykirat Singh, Sunny Manchanda, Soumen Chakrabarti, Tanmoy Chakraborty

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

Reproducibility Variable Result LLM Response
Research Type Experimental SYRELM shows massive improvements (e.g., +30.65 absolute point improvement in accuracy on the SVAMP dataset using GPT-J 6B model) over base LMs, while keeping our testbed easy to diagnose, interpret and within reach of most researchers.
Researcher Affiliation Collaboration 1IIT Delhi, India 2IIIT Delhi, India 3DYSL-AI, India 4IIT Bombay, India
Pseudocode Yes Figure 2: Example of parsable pseudocode generation from LMs.
Open Source Code Yes We release the code and data for reproducibility 2. 2https://github.com/joykirat18/SYRELM
Open Datasets Yes We built upon three existing mathword problem datasets: ASDiv (Miao, Liang, and Su 2020b), MAWPS (Miao, Liang, and Su 2020b), and Math23k (Shen et al. 2021).
Dataset Splits No The paper mentions 'training data' and 'testbed' but does not provide specific dataset split information (exact percentages, sample counts, or detailed splitting methodology) for training, validation, and testing.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. It only mentions 'modest hardware' in a general sense.
Software Dependencies No The paper mentions using 'Sym Py' as a symbolic solver and that Python code is generated, but it does not provide specific version numbers for Sym Py, Python, or any other relevant software libraries.
Experiment Setup Yes Details of the hyperparameters used in SYRELM are described in Appendix B.