slee-persis/GVS5H

What it solves

It addresses the high cost and size of proprietary frontier models required for high-end coding performance. It provides a training-free method to achieve frontier-level accuracy on hard coding problems using smaller, open-weight models or cheaper proprietary ones.

How it works

The project implements "ledger-based zero-shot self-orchestration." This is a multi-agent system where fresh instances of the same model decompose complex problems and coordinate through a shared filesystem. This shared filesystem acts as a ledger containing a plan, notes, and the current solution, allowing the model to organize its reasoning process without additional training.

Who it’s for

Developers and researchers interested in improving the LLM coding performance of open-weight models or reducing the cost of using high-end proprietary models for complex software engineering tasks.

Highlights

• Frontier-level performance: Qwen3.8-27B can reach 92.4% pass@1 on LiveCodeBench Hard, exceeding some proprietary models.
• Training-free: The method requires no fine-tuning or training, relying entirely on inference-time organization.
• Ledge-based coordination: Uses a shared filesystem for persistent context and problem decomposition.
• Cost-efficient: Can achieve high accuracy at a fraction of the cost of the largest proprietary models.