Guaranteed Generation from Large Language Models

Authors: Minbeom Kim, Thibaut Thonet, Jos Rozen, Hwaran Lee, Kyomin Jung, Marc Dymetman

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To validate these theoretical concepts, we conduct extensive experiments on two text generation settings with hard-to-satisfy constraints: a lexical constraint scenario and a sentiment reversal scenario. These experiments show that GUARD achieves perfect constraint satisfaction while almost preserving the ideal distribution with highly improved inference efficiency.
Researcher Affiliation Collaboration 1Seoul National University 2NAVER Labs Europe 3NAVER AI Lab 4Sogang University 5Independent Researcher
Pseudocode Yes Algorithm 1 GUARD sampler
Open Source Code Yes 1The code is available at https://github.com/naver/guard.
Open Datasets Yes We then selected our set of openings X by collecting negative story openings from the ROCStories test set (Mostafazadeh et al., 2016)
Dataset Splits No The paper uses the ROCStories test set as a source for story openings, but it does not specify explicit training/test/validation splits for its own experimental methodology beyond referring to pre-existing dataset divisions.
Hardware Specification No The paper does not provide specific details regarding the hardware used to run its experiments, such as GPU/CPU models or memory specifications.
Software Dependencies No The paper mentions the 'disco toolkit' but does not provide specific version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes Detailed hyperparameters are provided in Table 3, and the list of constraint-aware prompts used in the experiments can be found in Table 4.