We automate deep step-by step reasoning in an LLM dialog thread by recursively exploring alternatives (OR-nodes) and expanding details (AND-nodes) up to a given depth. Starting from a single succinct task-specific initiator we steer the automated dialog thread to stay focussed on the task by synthesizing a prompt that summarizes the depth-first steps taken so far.

Our algorithm is derived from a simple recursive descent implementation of a Horn Clause interpreter, except that we accommodate our logic engine to fit the natural language reasoning patterns LLMs have been trained on. Semantic similarity to ground-truth facts or oracle advice from another LLM instance is used to restrict the search space and validate the traces of justification steps returned as answers. At the end, the unique minimal model of a generated Horn Clause program collects the results of the reasoning process.

As applications, we sketch implementations of consequence predictions, causal explanations, recommendation systems and topic-focussed exploration of scientific literature.

To run the code locally, put the OpenAI key in your Linux or OS X shell environment with:

export OPENAI_API_KEY=<your_key>

Clone from github with with:

git clone [email protected]:ptarau/recursors.git

If you have cloned this repo, you can install the package deepllm by typing in folder recursors

pip3 install -e .

You can also install it from pypi with

pip3 install deepllm

The DeepLLM API exposes its high-level functions ready to embed in your application with something as simple as (assuming the your OPENAI_KEY is exported by your environment):

for result in run_recursor(initiator='Using tactical nukes', prompter=conseq_prompter, lim=2):
    print(result)

Also, you can explore questions with less gruesome results like in:

for result in run_rater(initiator='Artificial General Intelligence', prompter=sci_prompter, lim=2, threshold=0.5):
    print(result)

• Take a look at folder deepllm/tests for typical uses.

• There are more extensive demos in folder deepllm/demos .

• There is a streamlit app showing typical use cases, also deployed on the cloud at https://deepllm.streamlit.app/ .

• If you install fastchat, there are examples of using Vicuna models with it in folder deepllm/local_llms.

After installing streamlit, try it with:

streamlit run deepllm/apps/app.py

If you find this software useful please cite it as:

@ARTICLE{tarau2023automation,
       author = {{Tarau}, Paul},
        title = "{Full Automation of Goal-driven LLM Dialog Threads with And-Or Recursors and Refiner Oracles}",
      journal = {arXiv e-prints},
     keywords = {Computer Science - Artificial Intelligence, Computer Science - Logic in Computer Science},
         year = 2023,
        month = jun,
          eid = {arXiv:2306.14077},
        pages = {arXiv:2306.14077},
          doi = {10.48550/arXiv.2306.14077},
archivePrefix = {arXiv},
       eprint = {2306.14077},
 primaryClass = {cs.AI},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2023arXiv230614077T},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}

You can also find the paper (and future related work) in folder docs.

Enjoy,

Paul Tarau