Inside LinkedIn’s AI overhaul: Job search powered by LLM distillation

The emergence of research in the natural language encouraged people to change how to search for information, Winkedid, which works with many artificial intelligence models over the past year, hoping this transformation will extend to the search for a job.

The search for jobs on behalf of LinkedIn, now available to all LinkedIn users, is used as distilled models on the basis that have been trained on the basis of knowledge of the professional social media platform to narrow possible job opportunities on the basis of natural language.

“This new search experience allows members to describe their goals with their own words and get results that really reflect what they are looking for,” Iran, Berger, Vinturebeat, Venturebeat, told Venturebeat in an email. “This is the first step on a larger trip to make the search for work more easy, comprehensive and empowering for everyone.”

LinkedIn previously said in a blog post that there is a big problem that users faced when searching for jobs on the platform was excessive dependence on microscopic keywords. Often, users write more general job ownership and get functions that do not fully match. From the personal experience, if you write a “correspondent” on LinkedIn, get search results for the correspondent jobs in media publications, along with the correspondent of the court, which is a set of completely different skills.

Vice president LinkedIn Engineering told Venturebeat in a separate interview that they saw the need to improve how jobs are fully suitable, and began to understand better what they were looking for.

“In the past, when we use keywords, we mainly look at a major word and try to find accurate matching. Sometimes in a job description, the correspondent job description may say, but they are not a truly correspondent; we still recover this information, and it is not perfect for the candidate,” said Zhang.

LinkedIn has improved her understanding of user information and now allows people to use more than just major words. Instead of searching for a “software engineer”, they can ask, “finding software engineering functions in the Silicon Valley were recently published.”

How did they build that?

One of the first things that LinkedIn had to do is to fix the ability of the search job to understand.

“The first stage is when you write a query, we need to be able to understand the query, then the next step is that you need to recover the appropriate type of information from our job library. Then the last step is that you now have like a hundred final candidates, and how you do the arrangement so that the most relevant function appears on top,” Zang said.

LinkedIn relied on the classified styles based on classification, arrangement models, and the oldest LLMS, which said “lacked the ability to deep semantic understanding.” After that, the company turned into a more modern and ablution large language (LLMS) to help enhance the capabilities of natural language processing (NLP).

But LLMS also comes with expensive account costs. Therefore, LinkedIn turns into distillation methods to reduce the cost of using expensive graphics processing units. They divided LLM into two steps: one to work on retrieving data and information and the other to classify the results. Using a teacher model to classify query and function, LinkedIn said he is able to align all of the recovery and classification models.

The method also allowed LinkedIn to reduce the stages used by a job search system. At one point, “there were nine different stages that form the pipeline to search for a job and match”, which are often repeated.

“To do this, we use a common technique for multi -targets.

LinkedIn also developed a query engine that generates dedicated suggestions for users.

More based on artificial intelligence

LinkedIn is not alone in seeing the possibility of searching for the LLM foundation. Google claims that 2025 will be the year when Search Enterprise will become more powerful, thanks to advanced models.

Models such as Cohere’s Rrank 3.5 help break the language silos within institutions. Various “deep research” products from Openai, Google and Anthropology indicate that there is an increased regulatory demand for agents who reach and analyze internal data sources.

LinkedIn has been offered many features of artificial intelligence last year. In October, Amnesty International’s assistant to help employment has launched the best candidates.

Deepak Agharwal, chief artificial intelligence employee at LinkedIn, will discuss the company’s artificial intelligence initiatives, including how to expand the employment assistant from the initial model to productionDuring converting VB in San Francisco this month. Register now to attend.