[NeurIPS 2023] Reflexion: Language Agents with Verbal Reinforcement Learning

This repo holds the code, demos, and log files for Reflexion: Language Agents with Verbal Reinforcement Learning by Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao.

We have released the LeetcodeHardGym here

To Run: reasoning (HotPotQA)

We have provided a set of notebooks to easily run, explore, and interact with the results of the reasoning experiments. Each experiment consists of a random sample of 100 questions from the HotPotQA distractor dataset. Each question in the sample is attempted by an agent with a specific type and reflexion strategy.

Setup

To get started:

1. Clone this repo and move to the HotPotQA directory:

git clone https://github.com/noahshinn/reflexion && cd ./hotpotqa_runs

2. Install the module dependencies into your environment:

pip install -r requirements.txt

3. Set OPENAI_API_KEY environment variable to your OpenAI API key:

export OPENAI_API_KEY=<your key>

Agent Types

Agent type is determined by the notebook you choose to run. The available agent types include:

• ReAct - ReAct Agent

• CoT_context - CoT Agent given supporting context about the question

• CoT_no_context - CoT Agent given no supporting context about the question

The notebook for each agent type is located in the ./hotpot_runs/notebooks directory.

Reflexion Strategies

Each notebook allows you to specify the reflexion strategy to be used by the agents. The available reflexion strategies, which are defined in an Enum, include:

• ReflexionStrategy.NONE - The agent is not given any information about its last attempt.

• ReflexionStrategy.LAST_ATTEMPT - The agent is given its reasoning trace from its last attempt on the question as context.

• ReflexionStrategy.REFLEXION - The agent is given its self-reflection on the last attempt as context.

• ReflexionStrategy.LAST_ATTEMPT_AND_REFLEXION - The agent is given both its reasoning trace and self-reflection on the last attempt as context.

To Run: decision-making (AlfWorld)

Clone this repo and move to the AlfWorld directory

git clone https://github.com/noahshinn/reflexion && cd ./alfworld_runs

Specify the run parameters in ./run_reflexion.sh. num_trials: number of iterative learning steps num_envs: number of task-environment pairs per trial run_name: the name for this run use_memory: use persisting memory to store self-reflections (turn off to run a baseline run) is_resume: use logging directory to resume a previous run resume_dir: the logging directory from which to resume the previous run start_trial_num: if resume run, then the trial number of which to start

Run the trial

./run_reflexion.sh

The logs will be sent to ./root/<run_name>.

Another Note

Due to the nature of these experiments, it may not be feasible for individual developers to rerun the results as GPT-4 has limited access and significant API charges. All runs from the paper and additional results are logged in ./alfworld_runs/root for decision-making, ./hotpotqa_runs/root for reasoning, and ./programming_runs/root for programming

Other Notes

Check out the code for the original code here

Read a blog post here

Check out an interesting type-prediction implementation here: OpenTau

For all questions, contact [email protected]

Cite

@misc{shinn2023reflexion,
      title={Reflexion: Language Agents with Verbal Reinforcement Learning}, 
      author={Noah Shinn and Federico Cassano and Edward Berman and Ashwin Gopinath and Karthik Narasimhan and Shunyu Yao},
      year={2023},
      eprint={2303.11366},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}