PaCoRe (Parallel Collaborative Reasoning) - PDC Project

A competition-grade implementation of Parallel Cortex Reasoning for enhanced LLM inference.

This project demonstrates how parallel reasoning branches, multi-round refinement, and self-consistency voting can dramatically improve reasoning accuracy on a 3B parameter model.

Key Features

Advanced Reasoning Capabilities

• Multi-branch parallel reasoning: Run multiple independent reasoning paths simultaneously
• Multi-round refinement: Iteratively improve answers with reflection on prior attempts
• Self-consistency voting: Run k independent attempts and vote on the final answer
• Chain-of-thought prompting: Enhanced prompts with step-by-step reasoning instructions
• Answer verification: Built-in verification step to validate reasoning

Comprehensive Benchmarking

• GSM8K-style math problem evaluation
• Multi-metric scoring (accuracy, latency, consistency)
• Configuration comparison framework
• Detailed per-problem analysis

What's Implemented

Core Components

Module	Description
src/consensus.py	5 consensus algorithms (majority, weighted, ensemble, ranking, borda)
src/pacore_pipeline.py	Main inference pipeline with multi-round refinement
src/pacore_prompts.py	Chain-of-thought prompt templates
src/benchmark.py	Evaluation framework with metrics
src/pacore_trainer.py	PPO training skeleton (TRL-based)

Pipeline Features

• Parallel branches with temperature-based diversity
• Smart compaction with tagged answer preservation
• Synthesis with consensus fallback
• Multi-round refinement (num_rounds parameter)
• Self-consistency voting (self_consistency_k parameter)

Setup (Windows)

python -m venv .venv
.\.venv\Scripts\activate
python -m pip install -r requirements.txt

Usage Examples

1) Consensus Demo (no LLM required)

python examples/simple_usage.py

2) PaCoRe Inference Demo

python examples/run_pacore_pipeline.py "Solve: 23*17" --model meta-llama/Llama-3.2-3B-Instruct --branches 1 --branch_tokens 64

3) Run Benchmarks

# Quick validation
python examples/run_benchmark.py --quick

# Full benchmark with comparison
python examples/run_benchmark.py --samples 10 --compare --verbose

# Generate report
python examples/run_benchmark.py --samples 20 --report benchmark_results.json

4) Advanced Usage (Python)

from src.pacore_pipeline import PaCoRePipeline

# Initialize with defaults
pipeline = PaCoRePipeline()

# Simple inference
result = pipeline.run("What is 23 * 17?")
print(result["final_answer"])  # 391

# Multi-round refinement (more accurate)
result = pipeline.run("Complex math problem...", num_rounds=2)

# Self-consistency voting (most robust)
result = pipeline.run("Tricky problem...", self_consistency_k=5)
print(f"Answer: {result['final_answer']}, Confidence: {result['vote_confidence']}")

Hugging Face Access (Llama Models)

Some Meta Llama repositories are gated on the Hugging Face Hub.

1. Request access on the model page
2. Authenticate: huggingface-cli login
3. Or set environment variable: HF_TOKEN or HUGGINGFACE_HUB_TOKEN

Tests

python -m pytest -v

Repository Layout

PDC-Project/
    src/
        __init__.py
        consensus.py          # Consensus algorithms
        pacore_pipeline.py    # Main inference pipeline
        pacore_prompts.py     # Prompt templates
        pacore_trainer.py     # PPO trainer skeleton
        benchmark.py          # Evaluation framework
    examples/
        simple_usage.py       # Consensus demo
        run_pacore_pipeline.py # Inference demo
        run_benchmark.py      # Benchmark runner
    tests/
        test_consensus.py     # Unit tests
    data/
        math_train.jsonl      # Training data sample
    config.yaml
    requirements.txt
    pytest.ini
    README.md

Performance Notes

The PaCoRe approach can significantly improve reasoning accuracy:

• Baseline (1 round, k=1): Standard single-pass inference
• Multi-round (2+ rounds): Each round reflects on prior answer
• Self-consistency (k>1): Multiple attempts with majority voting
• Combined: Best accuracy but higher latency

For competition-grade results, use num_rounds=2, self_consistency_k=3 or higher.

License

This project is for educational/research purposes as part of a PDC course submission.