LLM Agents Learn to Self-Escalate Tasks to Stronger Models

Current large language model (LLM) agent systems typically decide on task delegation either before starting reasoning, by routing to a specific model, or after a full response, by verifying and potentially retrying. This paper explores a novel approach where an agent identifies its own likelihood of failure during its reasoning process and proactively escalates the task to a more powerful LLM. The core of this method is a Bayesian optimal-stopping framework, which uses a learned "competence posterior" to estimate the agent's eventual success. This allows for dynamic, real-time decision-making on when to transfer control. The research provides theoretical proofs for the optimal policy and demonstrates its effectiveness in a controlled simulation and a real-world code generation cascade, showing superior performance compared to post-hoc routing.