New 1.5B router model achieves 93% accuracy without costly retraining

The Katanemo Labs researchers introduced Arch-Router, a new guidance model and a work frame designed to set the user’s intelligence intelligence for the most suitable Grand Language model (LLM).

For multi-LLMS building products, Arch-Router aims to solve the main challenge: How to direct queries to the best job model without relying on strict logic or re-training assigned each time something changes.

LLM guidance challenges

With the growth of the LLMS number, developers move from one -mode settings to multiple models that use unique strengths for each model for specific tasks (for example, generating the code, summarizing the text or editing photos).

LLM direction has emerged as a major method of building and publishing these systems, as it acts as a traffic controller that directs every user’s inquiry to the most appropriate model.

Current guidance methods are generally divided into two categories: “task -based guidance”, where quotes are directed based on pre -determined tasks, and “performance -based guidance”, which is looking for an ideal balance between cost and performance.

However, the task -based guidance is struggling with the uninterrupted or changing user intentions, especially in multiple conversations. On the other hand, the performance -based guidance gives priority strictly standard degrees, often neglects the user’s preferences in the real world and badly adapts to new models unless it is subject to costly manufacturing.

More importantly, as researchers note at Katanemo Labs in their paper, “current guidance curricula have restrictions on use in the real world. They usually improve standard performance while neglecting human preferences that are driven by self -evaluation criteria.”

The researchers highlight the need for router systems that “are in line with self -preferences, provide more transparency, and they remain easily adaptive to the development of models and cases of use.”

A new framework for preference alignment

To address these restrictions, the researchers propose a “preference alignment” framework that matches the quotes with guidance policies based on specific preferences by the user.

In this context, users define their guidance policies in the natural language using the “field movement classification”. This is a hierarchical sequence on two levels that reflects how people describe the tasks naturally, starting with the general topic (field, such as “legal” or “financing”) and narrows a specific task (the procedure, such as “summarizing” or “generating software instructions”).

Each of these policies is then linked to a preferred model, allowing developers to take guidance decisions based on the needs of the real world instead of standard degrees. The paper also states, “This classification is a mental model to help users identify clear and structural guidance policies.”

The guidance process occurs in two phases. First, the virtue router model takes the user’s inquiry and the full set of policies and chooses the most appropriate policy. Second, the appointment function that the policy has set to the specified LLM links.

Since the logic of choosing the model is separated from politics, models can be added, removed or replaced simply by editing guidance policies, without any need to re -train or modify the router. This chapter provides the flexibility required for practical publication, as models and cases of use are constantly evolving.

Policy selection is run by bow, which is a 1.5B parametering language model that is well set for preference. Arch-Router receives the user’s inquiry and the full set of policy descriptions within his demands. Then he generates the best matching policy.

Since policies are part of the inputs, the system can adapt to new or modified methods at the time of reasoning through learning within context and without re -training. This obstetric approach allows the arc to use its pre -trained knowledge to understand the connotations of both inquiries and policies, and to address the entire conversation date at the same time.

There is a common concern with the inclusion of extensive policies in a wave is the possibility of increased cumin. However, researchers designed the bow arc to be very effective. “While the length of guidance policies can go away for a long time, we can easily increase the window of the bow context with the minimum impact on cumin,” explains Salman Parasha, co -author of the paper and founder/CEO of Katanemo Labs. It indicates that the cumin is mainly driven by the length of the output, and for the bow, the output is simply the short name of the guidance policy, such as “Image_eding” or “Dockument_creation”.

Sagittarius at work

To create the bow, the researchers set the QWEN 2.5 1.5B parameter on a coordinated data collection of 43,000 examples. Then they tested their performance against modern ownership models from Openai, Anthropic and Google on four general data sets designed to assess artificial intelligence systems for conversation.

The results showed that the arc achieves the highest level of 93.17 % total guidance, bypassing all other models, including the best ownership, at a rate of 7.71 % on average. The model feature has grown through longer conversations, indicating its strong ability to track the context on multiple turns.

In practice, this approach is already applied in many scenarios, according to Parsha. For example, in open source coding tools, the developers use the bow to direct the different stages of their workflow, such as “software design”, “understanding of software instructions”, and “code generation”, to the most appropriate LLMS for each task. Likewise, institutions can direct applications for documents to a model like Claude 3.7 Sonnet while sending photo editing tasks to Gemini 2.5 Pro.

Parasha said the system is also ideal for “personal assistants in various fields, where users have a variety of tasks from summarizing the text to realistic queries,” adding that “in these cases, developers can help developers to unify and improve the total user experience.”

This framework was combined with Arch, the original agent of Katanemo Labs for agents, which allows developers to implement advanced rules for forming traffic. For example, when combining a new LLM, the team can send a small portion of traffic to a specific guidance policy to the new model, check its performance with internal standards, then transfer the traffic completely with confidence. The company is also working to integrate its tools with evaluation platforms to simplify this process for institutional developers.

In the end, the goal is to bypass artificial intelligence applications. “Sagittarius and Sagittarius developers-from developers and institutions from divided LLM applications to a unified system based on politics, says.” “In the scenarios where the user’s tasks are varied, our framework helps to convert this task and break up LLM into a unified experience, making the final product smoothly for the final user.”