Jurassic-X Bridges the Neuro-Symbolic Gap with the MRKL System

AI21 Labs presented Jurassic-X and the MRKL (Modular Reasoning, Knowledge and Language) system. The core idea is a structure that, instead of relying on a single large language model for everything, attaches external knowledge sources and symbolic reasoning experts so that different types of problems are handled by different components.

The company shared the response to its previously released Jurassic-1 (178B parameters), noting that over 10,000 developers have registered and hundreds of commercial applications are in development. At the same time, it points out the limitations of language models. It notes that they cannot access company databases, do not know up-to-date information, and fall short even of a 1970s HP calculator when it comes to complex arithmetic. The cost of updating them is also presented as a problem.

In Jurassic-X, a router breaks down a question and passes it to the appropriate expert. The ways of using it shown in the article are as follows.

• Searching and updating a database in natural language
• Connecting to up-to-date knowledge sources like Wikidata to generate timely sentences
• Calling a calculator to handle complex math problems written in natural language and showing the calculation process
• Breaking multi-step problems, like comparing Tel Aviv and Berlin, into sub-questions to solve them
• Fetching constantly changing information such as weather and exchange rates via external APIs to answer questions

The key point is not just accuracy but transparency. By showing which expert handled which sub-problem and what the intermediate calculations were, the direction is to build a system that is more trustworthy than a black-box language model.