Solvita: An Agentic Evolution Framework Enhancing LLMs for Competitive Programming

Forum topic · 小凯 · 2026-05-19

Summary

Solvita (arXiv:2505.10883) is an agentic evolution framework that improves large language model performance on hard competitive programming without updating LLM weights. It reorganizes problem solving into a closed-loop system executed by four specialized agents: Planner (strategy selection), Solver (program synthesis), Oracle (certified supervision), and Hacker (targeted adversarial attacks). Each agent is paired with a trainable, graph-structured knowledge network. As the system runs, outcome signals such as pass/fail verdicts, test certification quality, and vulnerabilities discovered by the Hacker are converted into reinforcement learning updates to these network weights. This allows the agents to dynamically route future queries based on accumulated successes and failures, enabling continuous, transferable reasoning experience—unlike prior stateless multi-agent frameworks that rely on static retrieval. Evaluated on CodeContests, APPS, AetherCode, and live Codeforces rounds, Solvita establishes a new state of the art among code generation agents, outperforming multiple existing agentic pipelines and nearly doubling the accuracy of single-pass baselines.

Paper Overview

Field: Machine Learning
Authors: Han Li, Jinyu Tian, Rili Feng
Published: 2025-05-15
arXiv: 2505.10883

Introduction

Large language models (LLMs) still struggle with the rigorous reasoning demands of hard competitive programming. Recent multi-agent frameworks attempt to bridge this reliability gap, but they remain fundamentally stateless: they rely on static retrieval and discard the valuable problem-solving and debugging experience gained from previous tasks.

The Solvita Framework

Solvita is an agentic evolution framework that enables continuous learning without requiring weight updates to the underlying LLM. It reorganizes problem solving into a closed-loop system of:

1. Strategy selection 2. Program synthesis 3. Certified supervision 4. Targeted hacking

These stages are executed by four specialized agents:

Planner — strategy selection
Solver — program synthesis
Oracle — certified supervision
Hacker — targeted adversarial attacks

Crucially, each agent is paired with a trainable, graph-structured knowledge network. As the system operates, outcome signals—such as pass/fail verdicts, test certification quality, and adversarial vulnerabilities discovered by the Hacker—are converted back into reinforcement learning updates to these network weights. This allows agents to dynamically route future queries based on past successes and failures, effectively accumulating transferable reasoning experience.

Results

Evaluated on CodeContests, APPS, AetherCode, and live Codeforces rounds, Solvita:

Establishes a new state of the art among code generation agents
Outperforms multiple existing agentic pipelines
Nearly doubles the accuracy of single-pass baselines