Grammar-Aligned Decoding

Authors: Kanghee Park, Jiayu Wang, Taylor Berg-Kirkpatrick, Nadia Polikarpova, Loris D'Antoni

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our evaluation on code generation and structured NLP tasks shows how ASAp often produces outputs with higher likelihood (according to the LLM s distribution) than existing GCD techniques, while still enforcing the desired grammatical constraints. We evaluate ASAp on two structured prediction tasks: formal program synthesis and constituency parsing.
Researcher Affiliation Academia University of Wisconsin-Madison University of California San Diego
Pseudocode Yes Algorithm 1 ASAp algorithm
Open Source Code Yes Our code, datasets, and checkpoints are available at: https://github.com/ebmoon/transformers-GAD.
Open Datasets Yes Our code, datasets, and checkpoints are available at: https://github.com/ebmoon/transformers-GAD.
Dataset Splits Yes We employ a standard train-validation-test split of 70-10-20%.
Hardware Specification Yes Our experiments are conducted on 4 NVIDIA RTX A6000 GPUs and 4 NVIDIA A100 GPUs.
Software Dependencies Yes Our implementation is based on Python 3.10 and Py Torch 2.1.2.
Experiment Setup Yes Key hyperparameters include a learning rate of 2e-4, a warmup ratio of 0.03, a maximum sequence length of 2048, Lo RA alpha of 32, Lo RA dropout of 0.05, and Lo RA rank of 64.