LearnLM
Last reviewed
Jun 3, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,446 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,446 words
Add missing citations, update stale details, or suggest a clearer explanation.
LearnLM is a family of generative AI models from Google that were fine-tuned for teaching and learning, grounded in principles drawn from learning science. Built on top of the company's Gemini models and developed jointly by Google DeepMind, Google Research, and Google product teams, LearnLM was designed to do more than answer questions: it aimed to act like a good tutor, encouraging students to work through problems, adapting to their level, and explaining how an answer is reached rather than simply stating it. Google first announced LearnLM at its I/O developer conference on 14 May 2024 [1][2]. Over the following year the line evolved from a separate experimental model into a set of capabilities that Google folded directly into its mainline Gemini 2.5 models [3][4].
Google framed LearnLM as an attempt to close a gap between what large language models do by default and what good teaching requires. A general-purpose chatbot tends to hand over a finished answer, which is often the opposite of what helps someone learn. To build models that behave more like tutors, Google said it worked with learning scientists, cognitive psychologists, and pedagogy experts, and tested the approach with educators and learners [2][5].
The technical lineage runs through several Gemini generations. An early research version known as LearnLM-Tutor was fine-tuned over Gemini 1.0 and described in a July 2024 paper, "Towards Responsible Development of Generative AI for Education," which introduced a set of pedagogical benchmarks and fine-tuning datasets and reported that the tutor model was consistently preferred over a prompt-tuned base model by educators and learners [5]. A later iteration described in December 2024 was built on Gemini 1.5 Pro [6]. Google has described LearnLM as a family of models and capabilities rather than a single product, and the work was a collaboration across DeepMind, Google Research, and teams behind Search, YouTube, and Classroom [1][2].
The defining idea behind LearnLM is that the model's behavior should be shaped by established findings about how people learn. Google organized this around five principles, summarized below [2].
| Principle | What it means in practice |
|---|---|
| Active learning | Allow for practice and healthy struggle with timely feedback rather than handing over the answer |
| Managing cognitive load | Present relevant, well-structured information in multiple modalities |
| Adaptivity | Dynamically adjust to a learner's goals and needs, grounded in relevant material |
| Stimulating curiosity | Inspire engagement and provide motivation through the learning journey |
| Deepening metacognition | Help learners plan, monitor, and reflect on their own progress |
In the December 2024 work, Google described its training method as "pedagogical instruction following." Rather than baking one fixed definition of good teaching into the model, the team conditioned it on system-level instructions that describe the teaching behaviors desired in a given conversation. That left developers free to specify the pedagogy they wanted instead of locking everyone into a single style [6].
From the start Google positioned LearnLM less as a standalone chatbot and more as an ingredient inside products people already used. At launch and in the months after, it described several integrations [1][2][7]:
Google also released LearnLM to developers. In December 2024 it made an experimental model available in Google AI Studio, accompanied by documentation, example use cases, and suggested prompts [6][8]. An updated developer model, LearnLM 2.0 Flash Experimental, followed in 2025 [9].
To test the approach beyond its own products, Google said it worked with outside institutions including MIT RAISE, Columbia Teachers College, Arizona State University, NYU Tisch, and Khan Academy [2].
Google's claims about LearnLM rest on human evaluation by people with teaching expertise rather than only automatic benchmarks. The December 2024 study used a three-stage process: pedagogy experts designed dozens of learning scenarios across academic subjects, experts role-played as learners to collect thousands of conversations, and a separate set of experts scored those conversations against a detailed rubric [6].
In that study, expert raters preferred LearnLM over contemporaneous flagship models when judging pedagogical quality. Google reported the margins as follows [6]:
| Comparison | Expert preference for LearnLM |
|---|---|
| vs. GPT-4o | +31% |
| vs. Claude 3.5 Sonnet | +11% |
| vs. Gemini 1.5 Pro | +13% |
The paper was candid about limitations. It noted that the model could struggle with maintaining an encouraging tone, was better at spotting student errors than confirming correct answers, and remained susceptible to hallucination [6].
At Google I/O 2025, announced on 20 May 2025, Google said it was bringing LearnLM's pedagogical strengths into its general-purpose flagship rather than maintaining a separate education model. "We're infusing LearnLM directly into Gemini 2.5," the company wrote, calling the result the world's leading model for learning [3]. Google reported that Gemini 2.5 Pro "outperformed top models on every one of the five principles of learning science used to build AI systems for learning," and that educators preferred it across a range of scenarios [3]. DeepMind's I/O 2025 post echoed the point: "Since incorporating LearnLM, our family of models built with educational experts, 2.5 Pro is also now the leading model for learning" [4].
A separate study released in May 2025, "Evaluating Gemini in an arena for learning," set up blind comparisons judged by educators across leading models. It reported that experts preferred Gemini 2.5 Pro in 73.2% of match-ups, ranking it first overall, against Claude 3.7 Sonnet, GPT-4o, and OpenAI o3 [10].
As LearnLM moved into Gemini 2.5, the standalone developer model was wound down. Google's documentation states that LearnLM is no longer a separate listing in AI Studio and that its capabilities have been integrated into Gemini starting with the 2.5 series [8]. The earlier experimental model, learnlm-2.0-flash-experimental, was slated for retirement, with developers directed to the newer Gemini models [9].
Coverage of the 2024 launch treated LearnLM as Google's bid to build models tuned specifically for education at a time when general chatbots were already being used in classrooms. TechCrunch described it as "Google's new family of AI models for education" and walked through the product integrations and outside partnerships [1]. Trade outlets covering schools, including THE Journal and Campus Technology, framed the announcement alongside Gemini's wider rollout into Google Workspace for Education and noted data-protection commitments for school accounts, such as student interactions not being used to train Google's models [11][7].
Education-technology analysts also picked up on a tension that ran through the launch: Google was selling LearnLM as a way to make learning more active and effortful, while many of its consumer features still optimized for fast answers. The December 2024 paper's own caveats about hallucination and tone gave outlets a concrete basis for caution even as they reported the favorable expert-preference results [6][1].
By the 2025 I/O cycle, the framing in the press shifted from a distinct model toward Gemini 2.5 as an education tool, in line with Google's own decision to fold LearnLM into its flagship and to ship learning-focused features such as guided, step-by-step study modes in the Gemini app [3][4].