According to arxiv.org, researchers have introduced Solvita, an agentic evolution framework designed to enhance large language models for competitive programming tasks through continuous learning that does not require weight updates to the underlying model.
The paper, authored by Han Li, Jinyu Tian, Rili Feng, and colleagues, addresses a key limitation in current multi-agent frameworks: they “remain fundamentally stateless” and “discard the valuable problem-solving and debugging experience gained from previous tasks,” according to the abstract posted on arxiv.org.
Solvita reorganizes problem-solving into what the researchers describe as “a closed-loop system of strategy selection, program synthesis, certified supervision, and targeted hacking.” According to arxiv.org, this system is executed by four specialized agents: Planner, Solver, Oracle, and Hacker. Each agent is paired with “a trainable, graph-structured knowledge network” that captures experience over time.
As the system operates, “outcome signals, such as pass/fail verdicts” enable the knowledge networks to evolve, according to the abstract. This approach allows the framework to accumulate problem-solving experience while keeping the base LLM weights frozen.
The paper was published on May 18, 2026, and represents an attempt to address what the authors characterize as LLMs’ ongoing struggles with “the rigorous reasoning demands of hard competitive programming.”