Improved Generalized Planning with LLMs through Strategy Refinement and Reflection

Code and data for the paper Improved Generalized Planning with LLMs through Strategy Refinement and Reflection, accepted at ICAPS 2026 (ICAPS version will be uploaded to arxiv soon).

A previous version of the paper was presented at the Workshop on Planning in the Era of LLMs @ ICAPS25
The code for the previous paper version is availble as release v1.0

See also the project website for an overview.

Prerequisites

Current version of the code is set up for

• OpenAI models and tested with OpenAI non-reasoning models
• DeepSeek reasoning models
• Both reasoning and non-reasoning models where a local model is run as a server using sglang

Requires

• requirements specified in requirements.txt
• cuda
• sglang

Create a file 'set_env.py' in the utils folder

import os

def set_env_vars():
    os.environ['OPENAI_API_KEY'] = 
    os.environ['DEEPSEEK_API_KEY']

Code tested with Python 3.10 on Linux

Requires a compiled version of the plan validator VAL and fastdownward.
Set the FASTDOWNWARD and VAL path variables in the ./utils/paths.py script accordingly (line 7-8)

Data Availability

All data will be made available in a form that prevents crawling the data easily for training new LLMs. The exact way is not decided yet. To get access to the data in the meantime contact [email protected]

How to run

Generating Generalized Plans

python run_pipeline.py --env [dataset]-[domain] --config [config_path]

• env: dataset name (i.e. subfolder of the ./data directory) and name of the domain, separated by '-', e.g. 'silver-ferry'
• config_path: path to the configuration .json file

See also scripts in the sh_scripts_generation folder.
To generate additional sh_scripts run: python create_sh_scripts.py -d [dataset]-[domain] --conf [config_dir] --sh [sh_dir]
This will generate a file in the sh_dir directory with the commands for running the pipeline with all configuration files in the config_dir.

Configuration Files
Baseline

• baseline.json: our baseline

Full framework:

• full_3_6.json: F3-6
• full_5_3.json: F5-3

Ablations

• full_no_code_sr.json: no self-reflection during code debugging
• ful_no_multicode.json: only one initial program version
• full_no_strat_debug.json: no debugging on strategy

Evaluating Generalized Plans

python run_evaluation.py --env [dataset]-[domain] --eval_env [dataset_ev]-[domain_ev] --config [config_path] --out [output file]

• env: dataset name, needs to match the dataset name from generating the generalized plans
• eval_env: dataset name for the data to use for evaluation
• config_path: path to the evaluation configuration .json file
• output file: the name of the file where the results are saved; output directory is defined in the config file

Note on env vs. eval_env

• env determines the name of the subfolders in the output directory from which the generalized plans are read
• eval_env determines in which subfolders of the ./data directory the evaluation data can be found
• the domains should be the same but can come from different datasets, e.g. --env silver-ferry --eval-env additional-ferry

See also the scripts in the sh_scripts_evaluation folder

Configuration Files

See ./eval_configs/eval_config_all.json

• experiments_output_folder: folder with the output files of the program generation, i.e. with the programs to test
• experiments_results_folder: folder with the corresponding result files from running the pipeline
• experiments_names: names of the approaches, i.e. sub-folders in the outputs and results folder that should be considered
• split_file: name of the .json file in the data directory which specifies which tasks to run the evaluation on,
• eval_split_name: the key to extract from the json file
• val_split_file: the name of the .json file that specified the data splits for generating the programs

Use generalized plan to create a plan for a specific instance

python generate_plan_for_instance.py [-t timeout] [-p] domain_file instance_file code_file output_path log_path

• timeout: Set a specific timeout, default is 45 (given in seconds)
• -p, --print: Set this flag to print the log file contents to console
• domain_file: Path to the domain file for the relevant domain
• instance_file: Path to the problem instance file
• code_file: Path to the file with the generated python program
• output_path: Filepath to store the plan at
• log_path: Filepath to store the log notes at