AdaPlanner: Adaptive Planning from Feedback with Language Models

Authors: Haotian Sun, Yuchen Zhuang, Lingkai Kong, Bo Dai, Chao Zhang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments in the ALFWorld and Mini Wo B++ environments demonstrate that Ada Planner outperforms state-of-the-art baselines by 3.73% and 4.11% while utilizing 2x and 600x fewer samples, respectively.
Researcher Affiliation Academia Haotian Sun1 , Yuchen Zhuang1 , Lingkai Kong1, Bo Dai1, Chao Zhang1 1 Georgia Institute of Technology {haotian.sun, yczhuang, lkkong, chaozhang}@gatech.edu, bodai@cc.gatech.edu
Pseudocode Yes Figure 2: An illustrative example from ALFWorld to show the proposed adaptive closed-loop planning through code. The task is to put some clean lettuce on the diningtable... (a) Initial Plan (Extracted) def solution(agent, start_from=3):...
Open Source Code Yes The implementation of Ada Planner is available at https://github.com/haotiansun14/Ada Planner.
Open Datasets Yes We test Ada Planner on two text-based decision-making environments: 1) ALFWorld [20] is a text-based virtual household environment... 2) Mini Wo B++ [11] is a simulation environment...
Dataset Splits No The paper refers to
Hardware Specification No The paper mentions using specific LLM models (e.g., GPT-3.5, text-davinci-002) but does not provide details on the hardware (CPU, GPU, RAM) used to run the experiments or host these models.
Software Dependencies Yes RCI and Ada Planner harness GPT-3.5 (gpt-3.5-turbo and text-davinci-003) as backends.
Experiment Setup Yes The Setup details and prompts for Ada Planner are depicted in Appendix A and D.