Inference of Human-derived Specifications of Object Placement via Demonstration

Authors: Alex Cuellar, Ho Chit Siu, Julie A Shah

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we present the results from a human study, which demonstrate our framework s ability to capture a human s intended specification and the benefits of learning from demonstration approaches over human-provided specifications.
Researcher Affiliation Academia 1Massachusetts Institute of Technology 2MIT Lincoln Laboratory EMAIL, julie a EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Candidate Disjunctive Formulas; Algorithm 2: Inferring Intended Formulas
Open Source Code Yes 1For code and datasets: https://github.com/Alex Cuellar/PARCC
Open Datasets Yes 1For code and datasets: https://github.com/Alex Cuellar/PARCC
Dataset Splits No The paper describes how demonstrations were collected and used for inference and human studies, but does not provide specific train/test/validation dataset splits with percentages or sample counts for reproducibility of a model's performance evaluation.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, or cloud computing specifications used for running the experiments or generating results.
Software Dependencies No The paper does not provide specific software dependency details such as library names with version numbers for reproducibility.
Experiment Setup Yes For each group, inference used 100 non-specification demonstrations. The algorithm then loops over every disjunctive formula ϕ C, calculating the probability that all human demonstrations unintentionally satisfied ϕ using Equation 11 (lines 4-5). Next, the algorithm checks whether this probability is under the cutoff probability pc (i.e., whether we are confident that ϕ was not randomly satisfied, we use pc = .05), and adds it to C (lines 6-7) if so.