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..
Language-guided Skill Learning with Temporal Variational Inference
Authors: Haotian Fu, Pratyusha Sharma, Elias Stengel-Eskin, George Konidaris, Nicolas Le Roux, Marc-Alexandre Côté, Xingdi Yuan
ICML 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our results demonstrate that agents equipped with our method are able to discover skills that help accelerate learning and outperform baseline skill learning approaches on new long-horizon tasks in Baby AI, a grid world navigation environment, as well as ALFRED, a household simulation environment. |
| Researcher Affiliation | Collaboration | 1Brown University 2MIT 3University of North Carolina, Chapel Hill 4Mila 5Microsoft Research. |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and videos can be found at https://language-skill-discovery.github.io/. |
| Open Datasets | Yes | Baby AI (Chevalier-Boisvert et al., 2019) is an environment where an agent navigates and interacts in a grid world to achieve a goal described in language... The Baby AI dataset contains expert demonstrations collected from 40 different task types of varying levels of difficulty. ALFRED (Shridhar et al., 2020a) is a complex environment based on the AI2-THOR (Kolve et al., 2017) simulator... For the ALFRED dataset, we follow the settings in Pashevich et al. (2021) where the training dataset consists of more than 20000 trajectories. |
| Dataset Splits | Yes | For ALFRED, we follow the settings in (Pashevich et al., 2021). We train our algorithm on the training dataset with cross validation and test on the valid dataset. |
| Hardware Specification | No | The paper does not specify the exact hardware (e.g., CPU, GPU models, or memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions software like GPT-4, T5 encoder, and Faster R-CNN encoder but does not provide specific version numbers for these or other software dependencies required for replication. |
| Experiment Setup | Yes | Table 2: Hyperparameters of LAST. Hyperparameters Value: learning rate 3e-4, batch size 16, Size of skill library 100, weight of KL loss 0.0001, λ 1, 0.1, 0.01, γ 0.99, temperature (SAC) 1, α1 0.01, α2 1, training epochs 80,140. Also mentions: 'The transformers have 3 layers and 8 attention heads.' and 'we take the weighted sum of them as the total loss (weight 1 for the action type and weight 0.1 for the object type).' |