The tool integration problem that’s holding back enterprise AI (and how CoTools solves it)

Researchers University of Soochow China has introduced a Roman frame (COTOOLS) designed to use large language models (LLS) external means. Cotools aims to provide a more efficient and flexible approach compared to existing methods. This will allow you to use extensive tools, including wide range of LLS tools, including those not taught.

For businesses that want to create advanced AI agents, this ability can unlock more powerful and adaptable applications without typical shortcomings of integrated techniques in integration techniques.

When the model LLMS is a database or applications for many tasks or applications, you need to interact with foreign sources and means for many tasks. To supply llms with external instruments-Third is very important to extend their own capabilities to practical, realistic world applications of the APIS or functions.

However, the current methods to use the tool use significant trade. Covers a general approach Fine arrangement llm in instrument use patterns. Although this is skillful to call the special tools taken during the training, it is only limiting only the model of these tools. In addition, the subtle regulation process itself sometimes reduces the main strengths of the LLM’s overall justification skills, such as capital (cot) and the basic model.

An alternative approach is trusting Learning in context (ICL), where the LLM is provided with examples of existing vehicles and how to use them directly within the desire. This method offers flexibility that allows the model to use the means previously visible. However, it can be difficult to build these complex tips, and the efficiency of the model is significantly reduced because the number of existing means is less practical for the scenarios with large, dynamic tools.

As researchers noted paper Chain tools, a LLM agent “An LLM agent” is not possible to use a large number of vehicles effectively manage and not visible people, because many new vehicles can appear in real-world application scenarios. “

Cotools offer an attractive alternative to existing methods by combining the aspects of subtle regulation and semantic understanding of subtle regulation and semantic understanding to protect the original weight and strong reasoning opportunities of the original weight and strong substantiation. Instead of delicate adjusting the whole model, Cotools are a flight, specialized modules that work with LLM during the generation process.

“The researchers are writing to the semantic representation of the frozen basic models to make calls to call and call what tools to call and call,,” the researchers write.

In fact, Cotools, a rich concept within the rich anniversary of the LLM’s domestic representations, the “Secret States”, which is the text of the model, sets the text of the model and creates answers.

The Cotools Framework consists of three main components working consistently during the R. R. Ringing Process:

Tool Judge: LLM intends to answer by Token, the tool analyzes the hidden situation associated with the potential potential of the power and accepts whether it is suitable for a special point in the main point.

Tool recipient: If a judge determines the need for a vehicle, Retriever chooses the most suitable means for the task. The tool buyer was trained to compare a placement of the survey and comparing to existing means. This allows you to select the most semantically relevant tool from the existing tool pool, including “unprecedented” tools (ie part of training information for COTOOLS modules).

Tool call: Once the best tool is selected, Cotools uses a ICL offer that demonstrates the instrument’s settings based on the context of the tool. This target usage is afraid to add thousands of demonstrations to the initial tool option using ICL. After the selected vehicle was executed, its results were re-entered to the offspring of the LLM response.

Parameter Filling (calling with the ICL in the spotlight) Semantic understanding based on semantic understanding (referee) and options (retriever), protecting the main abilities of the LLM and achieves the effectiveness of mass tools that allow the flexible use of new tools. However, because the COTOOLS requires access to the model’s secret situations, it can only be applied to open weight models such as GPT-4O and Clams instead of special models such as GPT-4O and Claude.

Researchers evaluated Cotools along two different application scenarios: Numerical justification using arithmetic tools and knowledge based questions (KABA) obtained from arithmetic tools and knowledge bases.

GSM8K-XL (use of basic transactions) and funca (using more complex functions) (using more complex functions), applied cutools Llama2-7b Chatgpt in GSM8K-XL and a slightly superior or other tool learning method, a little superior or comparable performance in FunCA options. The results stressed that the capacity of the cotools is effectively increased the capacity of the basic model.

A simple symbol of the Camel database and a newly built symbol (s. Especially in the scenarios and unseen instruments that are in the scenario-numbers, the use of descendants. Despite the quality training information, the cotools also protected the strongest performance of Cotools.

Effects for the enterprise

Chain tools provide a promising direction for the establishment of agents equipped with more practical and powerful LLM in the company. This is especially useful as new standards Model context protocol (MCP) allows developers to easily connect foreign tools and resources to easily connect their application. Enterprises can potentially place agents adapted to new internal or foreign APIs and functions with minimal preparation.

In the positions of a semantic understanding of the framework of secret states, the positions that require interactions with various sources and systems, allow you to make more reliable AI assistants and a clear tool.

“Cotools explore the way to supply LLS in a simple way,” Mengsong Wu, Cotools in the University of SOOCHOW, the author of the learning author Venturebeat said.

However, WU noted that they have only been initial intelligence. “To apply it in a real world environment, you still need to find a balance between the effectiveness of the generalized tool with a subtle regulation cost,” Wu said Wu.

Researchers left the code to prepare the Magic and Retriever modules Entrusted.

“We believe that our main frame-based LLMS learning agency can be useful in real world applications based on frozen LLMS, and even in further developing tool learning,” researchers are writing.