agentification/rafa_code

Reason for Future, Act for Now: A Principled Framework for Autonomous LLM Agents with Provable Sample Efficiency

Code for Reason for Future, Act for Now: A Principled Framework for Autonomous LLM Agents with Provable Sample Efficiency, International Conference on Machine Learning (ICML), 2024.

Project page: https://agentification.github.io/RAFA.

Authors: Zhihan Liu*, Hao Hu*, Shenao Zhang*, Hongyi Guo, Shuqi Ke, Boyi Liu, Zhaoran Wang (* indicates equal contribution)

Please follow the instructions in the respective directories to reproduce our results in the four benchmarks:

• Game of 24
• ALFWorld
• BlocksWorld
• Tic-Tac-Toe

---

Game of 24

Environment setup

• Set OPENAI_API_KEY environment variable to your OpenAI API key:

export OPENAI_API_KEY=<your key>

Run the code

Experiment replication

python run.py --backend gpt-4 --task game24 --task_file_path 24.csv --task_start_index 900 --task_end_index 1000 --prompt_sample standard --n_generate_sample 10 --method_generate propose --method_evaluate value --method_select greedy --n_select_sample 1 --n_evaluate_sample 3 --feedback

Params for different method

• baseline ToT method (b=1, b=2)

python run.py --backend gpt-4 --task game24 --task_file_path 24.csv --task_start_index 900 --task_end_index 1000 --prompt_sample standard --n_generate_sample 10 --method_generate propose --method_evaluate value --method_select greedy --n_select_sample 1 --n_evaluate_sample 3 --planning tot

python run.py --backend gpt-4 --task game24 --task_file_path 24.csv --task_start_index 900 --task_end_index 1000 --prompt_sample standard --n_generate_sample 10 --method_generate propose --method_evaluate value --method_select greedy --n_select_sample 2 --n_evaluate_sample 3 --planning tot

• baseline Reflexion method

python run.py --backend gpt-4 --task game24 --task_file_path 24.csv --task_start_index 900 --task_end_index 1000 --prompt_sample standard --n_generate_sample 10 --method_generate propose --method_evaluate value --method_select greedy --n_select_sample 1 --n_evaluate_sample 3 --planning naive --feedback

• RAFA (b=1, b=2)

python run.py --backend gpt-4 --task game24 --task_file_path 24.csv --task_start_index 900 --task_end_index 1000 --prompt_sample standard --n_generate_sample 10 --method_generate propose --method_evaluate value --method_select greedy --n_select_sample 1 --n_evaluate_sample 3 --planning tot --feedback

python run.py --backend gpt-4 --task game24 --task_file_path 24.csv --task_start_index 900 --task_end_index 1000 --prompt_sample standard --n_generate_sample 10 --method_generate propose --method_evaluate value --method_select greedy --n_select_sample 2 --n_evaluate_sample 3 --planning tot --feedback

GPT 3.5

To run gpt-3.5-turbo, just replace --backend gpt-4 with --backend gpt-3.5-turbo. You can use --backend gpt-3.5-turbo-16k to avoid context length error if possible.

---

ALFWorld

Environment setup

• Install the required packages:

pip install -r requirements.txt

• Install the ALFWorld environment. Please refer to https://github.com/alfworld/alfworld.

• Set OPENAI_API_KEY environment variable to your OpenAI API key:

export OPENAI_API_KEY=<your key>

Run the code

./run_rafa.sh

---

BlocksWorld

Environment setup

• Our experiments are conducted with Vicuna-13B/33B (v1.3). The required packages can be installed by

  pip install -r requirements.txt
  

Run the code

• To run the RAP experiments, here is a shell script of the script

  CUDA_VISIBLE_DEVICES=0,1,2 nohup python -m torch.distributed.run --master_port 1034 --nproc_per_node 1 run_mcts.py --task mcts --model_name Vicuna --verbose False --data data/blocksworld/step_6.json --max_depth 6 --name m6ct_roll60 --rollouts 60 --model_path lmsys/vicuna-33b-v1.3 --num_gpus 3

• To run the RAFA experiments, here is a shell script example

  CUDA_VISIBLE_DEVICES=0,1,2 nohup python -m torch.distributed.run --master_port 36977 --nproc_per_node 1 run_rafa_mcts.py --model_name Vicuna --verbose False --data data/blocksworld/step_6.json --max_depth 6 --name rafm_step6_33b_try60 --rollouts 60 --model_path lmsys/vicuna-33b-v1.3 --num_gpus 3

• For details on the runtime arguments, one can use python run_rafa_mcts.py --help.

---

Tic-Tac-Toe

Environment setup

• Set OPENAI_API_KEY environment variable to your OpenAI API key:

export OPENAI_API_KEY=<your key>

Run the code

Experiment replication

python run.py --X gpt-4 --O gpt-4 --O_MPC 3 --num_train_epochs 12 --num_eval_epochs 10

Parameters

--X, --O: the backend model for X player and O player (default: gpt-3.5-turbo-16k)
--X_MPC, --O_MPC: how many actions to propose for X player and O player (default: 1, just base model without MPC)
--temperature: temperature for gpt (default: 0.2)
--eval_freq: evaluation frequency (default: 1)
--num_train_epochs: number of epochs for training (default: 1)
--num_eval_epochs: number of epochs for evaluating (default: 1)
--verbose: auxiliary outputs (default: 1)

---

Citation

@article{liu2023reason,
      title={Reason for Future, Act for Now: A Principled Framework for Autonomous LLM Agents with Provable Sample Efficiency},
      author={Liu, Zhihan and Hu, Hao and Zhang, Shenao and Guo, Hongyi and Ke, Shuqi and Liu, Boyi and Wang, Zhaoran},
      journal={arXiv preprint arXiv:2309.17382},
      year={2023}
}

rafa_code's Issues

[Question] Evaluation against web environment

Amazing paper and results.
Was the RAFA agent evaluated on a web-based environment such as WebArena or MiniWob?

Experimental Results on BlocksWorld (4-steps, Vicuna-13B-V1.3)

Hello authors, appreciate to your great work!

When I attempted to replicate the experimental results on Blocksworld (4-step setting, Vicuna-13B-V1.3), I found some discrepancies with the results stated in the paper. Could you please help me to check where the issue is?

Here are my results:

For RAP, the command I used was:

python -m torch.distributed.run --master_port 1034 --nproc_per_node 1 run_mcts.py --task mcts \
    --model_name Vicuna \
    --verbose False \
    --data data/blocksworld/step_4.json \
    --max_depth 4 \
    --name rap_roll60_13b_all \
    --rollouts 60 \
    --model_path lmsys/vicuna-13b-v1.3 \
    --num_gpus 1

For RAFA, the command was:

python -m torch.distributed.run --master_port 36977 --nproc_per_node 1 run_rafa_mcts.py \
    --model_name Vicuna \
    --verbose False \
    --data data/blocksworld/step_4.json \
    --max_depth 4 \
    --name rafa_roll60_13b_all \
    --rollouts 60 \
    --model_path lmsys/vicuna-13b-v1.3 \
    --num_gpus 1

Thank you very much!

Best,
Peng

ENV Package issue

Hi, I am interested in your masterpieces, and trying to reproduce the result, however, I am a little confused with Environment and DATA_PATH as attached, it seems like those variables are missing in the env package, how can I fix that issue, looking forward to your reply!