Small Reasoning Models are Instruction Followers in Function Calling
This paper introduces Instruction-Followed Function Calling (IFFC), a framework that delegates function-calling logic to a smaller, instruction-following model to achieve superior accuracy and robustness under quantization compared to native or prompt-based approaches, particularly for reasoning-oriented LLMs.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the rapidly evolving world of artificial intelligence, a shift is occurring from computers that simply generate text to systems capable of taking action. These new systems, often called "agentic" models, are designed to solve problems by reaching out to the outside world. They do this by selecting tools, such as weather apps or calculators, and using them to gather information before answering a question. For a long time, the most powerful and expensive computer models were the only ones trusted to perform these tasks. They were built with a specific, rigid internal language to talk to these tools, a method known as native function calling. However, as the technology has matured, researchers have begun to look at much smaller, more efficient computer models. These compact models are easier to run on everyday devices like smartphones and laptops, offering better privacy and lower costs, but they have historically struggled with the strict, complex rules required to use external tools correctly.
A team of researchers from universities in Iran has challenged the assumption that these smaller models must follow the same rigid rules as the giant ones to succeed. In their recent work, they discovered that small models actually perform much better when they are asked to use tools through a simple conversation, rather than through a specialized, technical command structure. They found that when a small model is treated like a helpful assistant following a clear set of instructions in plain language, it makes far fewer mistakes than when it is forced to speak a specific machine code. This insight led them to create a new system that splits the work between two different models. One small, fast model acts as a decision-maker, listening to the user and figuring out if a tool is needed. If a tool is required, this small model handles the request using natural language instructions. The main, larger model then takes that information and formulates the final answer for the user.
The researchers tested this approach using a variety of small models, some with as few as one billion parameters, which is a measure of the model's size and complexity. They compared their new method against the standard ways these models are usually taught to use tools. The results were striking. The new system, which they named Instruction-Followed Function Calling, allowed larger compact models, such as the 4-billion-parameter Qwen-3, to outperform massive proprietary baselines like GPT-5.2 and Claude 4.5 Sonnet, which are typically considered the gold standard. For instance, the Qwen-3 4B model using their new method achieved an accuracy of 94.1% on a Non-Live evaluation subset of a difficult test of tool usage, significantly surpassing the performance of the same model using traditional methods. While the smallest models tested (0.6B) still trailed behind the very largest proprietary giants in certain categories, the framework demonstrated that specialized, reasoning-capable small models could match or exceed the performance of massive systems when the architectural burden of rigid tool definitions was removed.
A key part of their discovery was understanding how these models think. The researchers found that models designed to "think" through a problem before answering were far more reliable than those that just generated an immediate response. When these reasoning models were given the task of deciding which tool to use, they could correct their own mistakes and follow complex instructions with high precision. This was particularly true when the models were compressed to run on smaller devices. Usually, making a model smaller or running it on less powerful hardware causes its performance to drop sharply. However, the researchers found that their new method was incredibly robust. Even when the models were shrunk down to use very little memory, the ones that used the natural language instruction method maintained their high accuracy. This suggests that the ability to reason and follow a conversation is a more stable foundation for tool use than memorizing a rigid technical format.
The study also highlighted a flaw in the current way many systems are built. Often, a single large model is asked to do everything: understand the user, decide which tool to use, and write the final answer. The researchers showed that this approach often confuses the model, as the complex details of the tools get mixed up with the conversation. By separating the tasks, their new system keeps the conversation clean. The small model handles the heavy lifting of figuring out which tool to use, and the main model focuses solely on talking to the user. This separation prevents the system from getting overwhelmed by too much information at once. The researchers demonstrated that this method works well even in real-world scenarios where speed and privacy are critical, such as on a car's dashboard or a home assistant, where sending data to a distant server is not an option.
Ultimately, this work suggests a new path forward for artificial intelligence on everyday devices. It shows that we do not need to rely on massive, energy-hungry computers to perform complex tasks. Instead, by changing how we ask smaller models to work—treating them as intelligent assistants following instructions rather than machines executing code—we can achieve results that are both highly accurate and efficient. The findings indicate that the future of personal, private, and responsive artificial intelligence may not lie in building bigger models, but in smarter ways of using the smaller ones we already have.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.