Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Predictive Preference Learning from Human Interventions

Authors: Haoyuan Cai, Zhenghao (Mark) Peng, Bolei Zhou

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach with experiments on both autonomous driving and robotic manipulation benchmarks and demonstrate its efficiency and generality. Our theoretical analysis further shows that selecting an appropriate preference horizon L balances coverage of risky states with label correctness, thereby bounding the algorithmic optimality gap.
Researcher Affiliation Academia Haoyuan Cai, Zhenghao Peng, Bolei Zhou Department of Computer Science, University of California, Los Angeles
Pseudocode Yes We summarize our method PPL in Alg. 1. Algorithm 1 Predictive Preference Learning from Human Interventions (PPL)
Open Source Code Yes Demo and code are available at: https://metadriverse.github.io/ppl.
Open Datasets Yes We evaluate our algorithm on the Meta Drive [16] and Robosuite [49] benchmarks, using both neural experts and real human participants, showing that PPL requires fewer expert monitoring efforts and demonstrations to achieve near-optimal policies.
Dataset Splits No The paper mentions evaluating
Hardware Specification Yes The whole experiment of PPL takes only 12 minutes on a desktop computer with an Nvidia Ge Force RTX 4080 GPU.
Software Dependencies No The paper does not explicitly list software dependencies with specific version numbers.
Experiment Setup Yes We fix H = 10 for all the interactive imitation learning baselines. In PPL, we fix β = 0.1, choose L = 4 for the Meta Drive benchmark and Table Wiping task, and set L = 6 for the Nut Assembly task.