Solve problems confidently and efficiently using a statistical approach with LLM

# This command installs the ConSol package from PyPI.
pip install consol

$ consol --prompt "1 + 1 = ?"
2

from consol import ConfidentSolver

# Initialize the ConfidentSolver with the following parameters:
# llm_model: The language model to use, e.g., "gpt-4o-mini", "o3-mini-low".
# confidence_model: The statistical model to determine confidence, e.g., "pvalue", "sprt", "bayesian".
# output_schema: The format of the output, e.g., "reasoned_float", "float".
consol = ConfidentSolver(
    llm_model="gpt-4o-mini",
    confidence_model="pvalue",
    output_schema="reasoned_float",
)
answer = consol.invoke("1 + 1 = ?")
print(answer)
# => 2

ConSol is a framework designed to solve various problems, primarily mathematical, by leveraging a statistical approach. This approach suppresses randomness and results in higher accuracy cost-efficiently.

• Higher Accuracy: ConSol improves OpenAI's GPT-o3-mini-medium performance on AIME24 Benchmark by 20 percentage points from 73% to 93%.
• Cost Efficiency: ConSol can reduce from 50% to 66% of output tokens of OpenAI's GPT-o3-mini-medium for AIME24 Benchmark. The number of output tokens is directly linked to the money cost. ConSol saves $50, $16, and $4 for o3-mini-high, o3-mini-medium, and o3-mini-low, respectively.

For the details, please refer to the publication.