Solvita: Agentic Evolution Framework Boosts LLMs for Competitive Programming
A new framework called Solvita enhances large language models (LLMs) for competitive programming without weight updates. It uses four specialized agents—Planner, Solver, Oracle, and Hacker—in a closed-loop system with trainable graph-structured knowledge networks. The system learns from pass/fail verdicts, test certification quality, and adversarial vulnerabilities, enabling continuous improvement on hard programming tasks.
Key facts
- Solvita is an agentic evolution framework for LLMs in competitive programming.
- It does not require weight updates to the underlying LLM.
- The framework uses four agents: Planner, Solver, Oracle, and Hacker.
- Each agent is paired with a trainable, graph-structured knowledge network.
- The system learns from outcome signals like pass/fail verdicts and adversarial vulnerabilities.
- It addresses the stateless nature of current multi-agent frameworks.
- The framework enables continuous learning from previous problem-solving experience.
- Solvita reorganizes problem-solving into strategy selection, program synthesis, certified supervision, and targeted hacking.
Entities
—