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..

Offline Goal-conditioned Reinforcement Learning with Quasimetric Representations

Authors: Vivek Myers, Bill Zheng, Benjamin Eysenbach, Sergey Levine

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On existing offline GCRL benchmarks, our representation learning objective improves performance on stitching tasks where methods based on contrastive learning struggle, and on noisy, high-dimensional environments where methods based on quasimetric networks struggle. 5 Experiments In our experiments, we evaluate the performance of TMD on tasks from the OGBench benchmark [43].
Researcher Affiliation Academia Vivek Myers Bill Chunyuan Zheng Benjamin Eysenbach Sergey Levine UC Berkeley Princeton University
Pseudocode Yes Algorithm 1: Temporal Metric Distillation (TMD)
Open Source Code Yes Website and code: https://tmd-website.github.io/ The TMD agent is implemented in https: //github.com/vivekmyers/tmd-release/blob/master/impls/agents/tmd.py.
Open Datasets Yes We evaluate TMD across evaluation tasks in OGBench for the environments and datasets listed in Table 1. datasets used from Park et al. [43] are openly available.
Dataset Splits Yes We evaluate TMD across evaluation tasks in OGBench for the environments and datasets listed in Table 1. The evaluation and base agent structure follows the OGBench codebase [43].
Hardware Specification Yes Experiments were run using NVIDIA A6000 GPUs with 48GB of memory, and 4 CPU cores and 1 GPU per experiment.
Software Dependencies No We implemented TMD using JAX [47] within the OGBench [43] framework.
Experiment Setup Yes General hyperparameters are provided in Table 2. batch size 256 learning rate 3 10 4 discount factor 0.995 invariance weight ΞΆ 0.01 in medium locomotion environments, 0.1 otherwise