RouteWeaver teaches an LLM router to pick its own mode, reaching 61.4% accuracy

A team of researchers has proposed RouteWeaver, a unified framework for routing between language models. It chooses not only which model to call but also how to structure the process itself: a single request, a series of rounds, or an agentic loop. On benchmarks across five domains, the system reaches an average accuracy of 61.4%. That is 11.03% higher in relative terms than the strongest of the compared methods, and RouteWeaver beats every baseline in each of the five domains. Working backward, the best competitor lands at roughly 55.3%.

The problem: a router locked into one scheme

LLM routing has long been a working tool: requests are distributed across heterogeneous models so that an expensive model isn't spent on simple tasks and hard tasks aren't handed to a weak one. But almost all existing approaches are built around a fixed execution paradigm. One router works in single-round mode: it picks a model and gets an answer. Another does multi-round, where models are called in sequence and refine each other's output. A third is agentic, with planning and tools.

This rigidity creates an obvious mismatch. A factual question doesn't need a multi-step dialogue between models, while a hard math problem may not be solved in a single call. When the way calls are organized is baked into the architecture, the router cannot adapt to the needs of a specific query.

RouteWeaver removes that constraint. All three modes live inside a single LLM policy, which jointly learns two things: which mode to apply and how to select models within it.

COMET-GRPO: separate signals for two levels of decisions

Training two levels of decisions together runs into a credit assignment problem. If the entire trajectory receives one shared advantage signal, it is unclear what worked: a good choice of mode or good execution within it. The authors address this with the COMET-GRPO algorithm (COordinated Mode-Execution Training). It decomposes the trajectory-level advantage into two separate signals: one for mode selection and one for execution within the mode. Each decision gets its own, more precise feedback.

The second element is progressive exploration. Training starts with modes that are assigned externally and gradually shifts toward autonomous selection. The logic is simple: the router first learns how to execute each mode, and only then decides when to apply which.

Ablations show that without this step, everything breaks down. If progressive exploration is removed, the policy collapses entirely into single-round: the router picks the simplest path and stops using the others. Both components matter, COMET-GRPO and the progressive scheme.

What is known from the evaluation

• Framework: a single LLM policy that unifies single-round, multi-round and agentic routing.
• Training algorithm: COMET-GRPO with separate advantage signals for mode selection and execution.
• Training scheme: progressive exploration, moving from assigned modes to independent selection.
• Evaluation domains: math, coding, knowledge, logical reasoning, factual recall.
• Average accuracy: 61.4%.
• Relative gain over the strongest baseline: 11.03%.
• Per-domain result: better than all compared methods in each of the five.

Practical takeaway for teams building routing

For teams building multi-model systems, the main signal is this: the choice of execution mode should be treated as part of the routing problem. If you currently have one router for models and separate logic for when to switch on chains or agents, RouteWeaver suggests merging them under a single trainable policy.

The second practical detail concerns the training itself. Collapse into single-round without progressive exploration is a typical trap for RL approaches: the policy finds the cheapest, most stable path and gets stuck in it. Teaching the router to execute modes first and only then choose between them is a technique that could carry over to other hierarchical decisions in agentic systems.

What remains to be seen is how such a router behaves on tasks outside the five tested domains, and how to account for the cost and latency of multi-round and agentic modes. But the direction is clear: routing is shifting from choosing a model to choosing the entire strategy for working with it.