Skip to main content

​
Overview

​
Created as part of Eleuther SOAR by Nadya Devani, Pouya Sadeghi, Purva Kandalgaonkar, Suparnojit Sarkar

The Pareto Merging method is an advanced technique for combining multiple specialized AI models into a single, highly efficient model. Unlike traditional methods that create one static merged model, Pareto Merging produces a flexible framework that can generate an infinite number of model variations tailored to specific needs without retraining. This method merges two types of models: a concise, non-reasoning Large Language Model (LLM) and a powerful Large Reasoning Model (LRM). The key innovation is a two-part architecture:
  1. Preference-Independent Base Model: A single, fixed base model combines the general knowledge of the LLM with the reasoning capabilities of the LRM.
  2. Preference-Dependent Tensor: A small, trainable “control knob” that is adjusted on-the-fly based on a predicted question difficulty score.
This allows the system to be dynamic: for easy questions, it leans towards the concise base model to save costs, and for difficult questions, it engages more reasoning power to ensure high accuracy.

​
Architecture

pareto

​
Pareto Merging API

​
Overview

The ScaleDown Pareto Merging API provides dynamic, cost-effective reasoning. It merges a concise LLM with a powerful LRM, allowing you to balance response accuracy and token cost on a per-query basis. Instead of using a static model, our system adjusts its reasoning depth based on the predicted difficulty of your prompt, reducing costs by up to 30% while maintaining high accuracy.

​
Quick Start

Get started in minutes by using our hosted API endpoint or by running the model locally with our Python package.

​
Installation

# Placeholder for package installation
pip install scaledown-paretomerge