LOCA: Logical Chain Augmentation for Olympiad-Level Physics Reasoning

Overview

LOCA (Logical Chain Augmentation) is a novel multi-agent framework for complex physics reasoning, capable of solving Olympiad-level physics problems (IPhO/CPhO). LOCA leverages iterative review and augmentation to enhance LLM reasoning quality on challenging scientific problems.

Installation

Prerequisites

• Python 3.12+
• LLM API key
• uv (recommended)

Setup

1. Clone the repository

git clone xxx
cd LOCA

2. Install dependencies

Using uv (recommended):

uv sync

Project Structure

LOCA/
├── src/                         # Source code
│   ├── loca/                    # LOCA implementation (Solver, Reviewers, Augmentation)
│   ├── vanilla_review/          # Vanilla review implementation (Baseline)
│   ├── single_llm.py            # Single LLM solver (Baseline)
│   ├── api/                     # LLM API interfaces (OpenAI, Anthropic, LiteLLM)
│   ├── utils.py                 # Common utilities
│   └── ...
├── configs/                     # Configuration files
├── problem_set/                 # Test datasets
├── test_results/                # Evaluation results
└── scripts/                     # Analysis scripts

Usage

Running LOCA

uv run main.py --config PATH_TO_YAML --config-name CONFIG_NAME_IN_YAML