Divergent Path Decoding

This repository contains the implementation of Divergent Path Decoding (DPD), an inference-time approach that modifies the autoregressive token generation process to improve output reliability in large language models.

Project Structure

├── charts/               # Visualization tools for analysis plots
├── gsm8k/               # Output directory for GSM8K evaluation results
├── mmlu/                # Output directory for MMLU evaluation results
├── scripts/             # Utility scripts for experiments
├── utils/               # Common utility functions
├── custom_greedy_counterfactual.py
├── environment.yaml     # Environment configuration
├── main.py             # Main execution script
├── multi_causal_counterfactual_analysis.py
├── selective_distributed_multi_mmlu.py
├── selective_distributed_random_mmlu.py
├── selective_greedy_counterfactual.py
└── selective_greedy_counterfactual_analysis.py

Overview

This project implements a novel decoding method that:

• Modifies the autoregressive token generation process at inference time
• Explores alternative generation paths through controlled token interventions
• Requires no model fine-tuning or architectural changes
• Improves output reliability while maintaining model capabilities

Requirements

The system was tested on a high-performance computing environment with:

• NVIDIA A100 GPUs (8 per node, 40GB VRAM each)
• Dual 64-core AMD EPYC processors
• 512GB system RAM
• SLURM workload manager
• PyTorch with DistributedDataParallel (DDP)
• Hugging Face Transformers library

Installation

# Clone the repository
git clone https://github.com/mmjerge/counterfactual_ensemble_decoding.git

# Create and configure environment from yaml
conda env create -f environment.yaml
conda activate divergent_decoding

Key Components

• main.py: Primary script for running experiments and generating results
• custom_greedy_counterfactual.py: Implementation of basic greedy counterfactual generation
• multi_causal_counterfactual_analysis.py: Multi-causal analysis framework
• selective_distributed_multi_mmlu.py: Distributed MMLU evaluation
• selective_distributed_random_mmlu.py: Random sampling for distributed evaluation
• selective_greedy_counterfactual.py: Selective greedy counterfactual generation
• selective_greedy_counterfactual_analysis.py: Analysis tools for selective approaches

Usage

Run experiments using the main.py script:

# Run evaluation on GSM8K dataset with 150 samples
python3 main.py --dataset gsm8k --num_samples 150

# Run evaluation on MMLU dataset with 150 samples
python3 main.py --dataset mmlu --num_samples 150

Results will be generated in the respective output directories:

• mmlu/: Contains MMLU evaluation results
• gsm8k/: Contains GSM8K evaluation results

Method

The implementation follows these key steps:

1. Generates multiple decoding paths for each input prompt
2. Introduces controlled deviations at specific token positions
3. Explores alternative reasoning trajectories
4. Aggregates results using majority voting

Citation

@article{jerge2024divergent,
  title={Divergent Path Decoding},
  author={Jerge, Michael},
  institution={University of Virginia},
  year={2024}
}

License

See LICENSE file for details.

Contact

• Author: Michael Jerge
• Institution: University of Virginia
• Repository: https://github.com/mmjerge/divergent_path_decoding