Fastest
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →https://www.patsnap.com/resources/blog/rd-blog/embedding-model-training-patent-landscape/ · Patsnap · data cut-off 2026-07-31 · downloaded from the live page
Patent Landscape · Vector Search & RAGEmbedding Model Training Patents: Filing Trends and Where the Field Is Still Open
- Filings nearly tripled in three years. 2021 to 2024 growth of +179%, from 130 to 363 records, before the most recent two years understate true activity due to publication lag.
- No single filer dominates. The leader holds 72 records out of 1,562, and the top 5 combined account for just 16.9% of all records in scope — concentration is modest, not a chokehold.
- Recent momentum is cooling across the largest filers. Several of the most active assignees show sharp year-on-year drops in their latest reported year, consistent with the lag in published data rather than a real slowdown.
1,562
Published Records
17%
Top-5 Share of All Records
+179%
Filing Growth 2021→2024
US
Leading Jurisdiction
Filing growth compares 2021 (130 records) with 2024 (363) — a three-year span. 2024 is the most recent year we treat as complete: publication lags filing by roughly 18 months, so 2025 onwards are still filling in and any growth rate that ends there would understate the field. Top-5 share is the combined record count of the five largest assignees divided by all 1,562 records in scope (CR5), not by the ranked leaders only.
Published byPatsnap Research··7 min readSourced from Patsnap Eureka
Overview
What the embedding model training patent record actually shows
Embedding model training sits at the intersection of representation learning and retrieval: contrastive learning, hard negative mining and domain adaptation are the recurring technical vocabulary across this corpus. The 1,562 records in scope span 2015 through mid-2026, capturing the period in which contrastive objectives and in-batch negative sampling moved from research papers into filed claims. Filing activity is not dominated by one lab or vendor — the leader holds 72 records, and even the top 10 combined reach only 25.9% of all records in scope.
The technology composition leans heavily on general digital data processing and AI model classes, with image and healthcare-specific applications forming a meaningful but secondary layer. That spread suggests embedding training claims are being filed as infrastructure — usable across many downstream tasks — rather than locked to a single vertical.
Filing activity and technology composition, 2015-2026
Source: Patsnap Eureka. Assignee ranking and totals. Derived from a Patsnap search on Embedding Model Training covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Let an AI agent run this analysis on your own technology
Pick a task. Every answer cites the patents behind it.
10,000 free credits to start
The Numbers
Filing trend and technology composition
Two views of the same 1,562-record corpus: filing activity over time, and where those records sit across IPC subclasses. Both use the full record set as the denominator.
Filings climbed sharply through 2024
Filings rose from 12 in 2017 to a peak of 423 in 2025, with 2021-2024 alone showing +179% growth (130 to 363 records). Treat 2025 and 2026 as undercounts: publication lags filing by roughly 18 months, so the most recent years will keep filling in as more applications publish.
G06F and G06N dominate, but overlap runs deep
G06F (electric digital data processing) appears in 47.1% of records and G06N (AI-based computing) in 40.3%, confirming embedding training is filed primarily as general-purpose computing infrastructure. Image recognition (G06V, 19.5%) and image data processing (G06T, 13.9%) form a secondary cluster, with healthcare informatics (G16H, 9.0%) and business data processing (G06Q, 8.6%) trailing as applied verticals. Because records can carry multiple classes, these shares sum past 100% and should be read as overlap, not as a partition of the field.
Shares are the percentage of the 1,562 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Source: Patsnap Eureka. Filing trend and technology composition. Derived from a Patsnap search on Embedding Model Training covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.
Go deeper on Embedding Model Training with Eureka
This page is one run against one query. Ask Eureka your own question about embedding model training and every answer comes back with the patent numbers behind it.
Try EurekaKey Patents
A representative claim and the most-cited prior art
Representative Filing
Source-free domain adaptation for a target classifier via an adaptive adversarial neural network (Adobe, 2023)
ADOBE INC.
The application describes generating a target classifier for a target domain via domain adaptation using a source classifier learned on a source domain. An embedding model, a target classifier and a source classifier analyse sets of target samples and generate classification probabilities based on that analysis; the classification probabilities are then used to modify the target classifier's parameters through adaptive adversarial inference.Filed by Adobe Inc., published 2023-03-30 as US20230094415A1.
Most-cited records in this corpus
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20180165554A1 | Semisupervised autoencoder for sentiment analysis | 793 |
| 2 | US20240386015A1 | Composite symbolic and non-symbolic artificial intelligence system for advanced reasoning and semantic search | 370 |
| 3 | US20160247061A1 | Systems and Methods for Neural Language Modeling | 334 |
| 4 | US20240412720A1 | Real-time contextually aware artificial intelligence (AI) assistant system and a method for providing a conte… | 231 |
| 5 | US10782986B2 | Assisting users with personalized and contextual communication content | 185 |
| 6 | US20230072423A1 | Wearable electronic devices and extended reality systems including neuromuscular sensors | 166 |
| 7 | US20190251952A1 | Systems and methods for neural voice cloning with a few samples | 156 |
| 8 | US20180349477A1 | Tensor-Based Deep Relevance Model for Search on Online Social Networks | 150 |
| 9 | US9443320B1 | Multi-object tracking with generic object proposals | 119 |
| 10 | US20190130221A1 | Method and device for generative adversarial network training | 115 |
Citation counts are drawn from within this searched corpus and favour older filings that have had more time to accumulate citations — read them as a signal of influence, not of current technical importance.
Each row carries its publication number; clicking a row searches Eureka by that number.
Source: Patsnap Eureka. Citation counts and representative records. Derived from a Patsnap search on Embedding Model Training covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Run it yourself
Put your own technology through the same analysis
Where to run it
For builders
MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →Insights
What the data means for filing strategy
Three findings that change how a filing or freedom-to-operate strategy should be shaped in this space.
Concentration
16.9%
top 5 share of 1,562 records
The top of the field is thin, not thick
The leading assignee holds 72 records and the top 5 combined reach only 16.9% of all 1,562 records in scope. That is a much flatter distribution than in more mature AI hardware or chip-design fields, meaning no single portfolio forms an unavoidable blocking position across the whole space.
Read the full assignee ranking for the complete distribution.
Growth
+179%
filing growth, 2021 to 2024
Filing activity accelerated hard through 2024
Records rose from 130 in 2021 to 363 in 2024, a +179% increase over three years, with the field peaking at 423 published records in 2025. Because publication lags filing by roughly 18 months, the 2025-2026 figures are still incomplete and should not be read as a slowdown.
Filing counts by year are shown in the data section above.
Composition
47.1%
records tagged G06F
Claims skew toward general computing infrastructure
Nearly half of all records carry a G06F digital-data-processing class and 40.3% carry G06N, indicating embedding training is being claimed as reusable infrastructure rather than tied to one application. Vertical classes like healthcare informatics (9.0%) and business data processing (8.6%) are present but secondary.
Class shares overlap because records can carry multiple IPC codes.
Eureka AI Agent
Looking for what nobody has claimed yet?
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to embedding model training, with the prior art for and against each one.
Co-filing is rare in this corpus
| Assignee | Co-assignee | Shared families |
|---|---|---|
| International Business Machines Corporation | IBM (China) Co., Ltd. | 3 |
| Adobe Inc. | University of Surrey | 2 |
| International Business Machines Corporation | National Chung Hsing University | 2 |
| International Business Machines Corporation | IBM UNITED KINGDOM LTD INTELLECTUAL PROPERTY DEPARTMENT | 2 |
| International Business Machines Corporation | IBM ISRAEL SCI & TECH LTD | 1 |
Only five co-assignee pairs appear across the dataset, and the strongest of them involves related entities within the same corporate group rather than independent joint filers — cross-company co-filing is not a meaningful pattern here.
Source: Patsnap Eureka. Co-assignee relationships and derived observations. Derived from a Patsnap search on Embedding Model Training covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Players
Who is filing, and where activity is easing
The assignee ranking covers 100 companies counted by record count — not a top-50 or top-100 cut, but the full ranking the underlying data returns.
Leader
72
records
A clear but not dominant leader
The top-ranked assignee holds 72 records, well ahead of fifth place at 37 and tenth place at 21 — a real gap at the very top, but nowhere near a majority share of the field.
See the full ranking table for exact positions.
Momentum
-88% YoY
at two of the most active filers
Several leaders show sharp recent-year drops
Two of the largest historical filers show roughly -88% year-on-year in their latest reported year, and others drop to zero. Given the 18-month publication lag, this reads as incomplete recent data rather than firms exiting the space.
Momentum figures are drawn from the latest available filing year per assignee.
Co-filing
5
co-assignee pairs total
Independent joint filing is uncommon
Across 1,562 records, only five co-assignee pairs exist at all, and the strongest pairings sit within the same corporate family. Most activity in this field is filed by single entities working alone.
This favours solo freedom-to-operate analysis over partner-mapping.
Recent-year filing momentum by assignee
| Assignee | Recent year | YoY |
|---|---|---|
| Citibank, N.A. | 2 | -88% |
| Microsoft Technology Licensing, LLC | 1 | -88% |
| Oracle International Corporation | 1 | -83% |
| Insitro, Inc. | 0 | -100% |
| Google LLC | 0 | -100% |
| Adobe Inc. | 0 | -100% |
| International Business Machines Corporation | 0 | -100% |
| AIRIA LLC | 0 | -100% |
Source: Patsnap Eureka. Assignee-level momentum. Derived from a Patsnap search on Embedding Model Training covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
What's Next
Where to take this analysis
The dataset points to a field that is growing fast but still structurally open. These are the natural next steps for turning that into a filing or FTO decision.
Map claim language against the leader's portfolio
With the top assignee holding 72 of 1,562 records, a claim-by-claim comparison against that portfolio is the fastest way to spot genuinely open independent claim scope before drafting.
Explore assignee portfolios in EurekaTrack the under-claimed branches as they mature
Multilingual alignment and embedding drift correction show thinner claim density than the core G06F/G06N cluster today; monitoring filing velocity there over the next few publication cycles will show whether that gap is closing.
Set up monitoring in EurekaRe-run the trend once 2025-2026 fully publish
Because publication lags filing by about 18 months, the 2025 peak of 423 and the 2026 partial-year count will both revise upward; a follow-up pull in a year will give a truer read on whether growth held past 2024.
Revisit this landscape in EurekaSource: Patsnap Eureka. Forward-looking reading of the same dataset. Derived from a Patsnap search on Embedding Model Training covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
FAQ
Common questions about embedding model training patents
The leading assignee in this dataset holds 72 of the 1,562 records in scope, with fifth place at 37 and tenth place at 21. That gap between first and fifth place shows a real but modest lead rather than dominance, since the top 5 combined only reach 16.9% of all records. Anyone doing freedom-to-operate work here should treat the leader's portfolio as important but not as a single blocking gatekeeper across the whole field.
Yes, sharply: filings rose from 130 records in 2021 to 363 in 2024, a +179% increase over that three-year span, with a peak of 423 published records in 2025. Because patent publication typically lags filing by around 18 months, the 2025 and 2026 figures are still incomplete and will revise upward as more applications publish. The safest complete-year comparison currently available is 2021 to 2024.
General digital data processing (IPC class G06F) appears in 47.1% of the 1,562 records, and AI-based computing (G06N) in 40.3%, showing most claims are written as reusable computing infrastructure rather than tied to one application. Image and video recognition (G06V) and image data processing (G06T) form a secondary cluster at 19.5% and 13.9% respectively. Healthcare informatics and business data processing appear as applied verticals but at lower shares, around 9% and 8.6%.
Based on the composition and citation data, sub-areas like multilingual embedding alignment, embedding drift detection and cross-lingual hard negative mining show comparatively thin representation relative to the dominant G06F and G06N cluster. Co-filing is also rare across the dataset, with only five co-assignee pairs found among 1,562 records, suggesting collaborative filing strategies in these adjacent branches remain largely untried. These are reasonable areas to search more deeply before assuming claim space is occupied.
This is a data artefact, not a real slowdown: patent applications typically publish about 18 months after filing, so 2025 and 2026 records are still being added to public databases. Several of the largest historical filers show year-on-year drops of roughly -88% or more in their latest reported year purely because of this lag. The reliable growth comparison in this dataset is the complete 2021-2024 window, which shows +179% growth, not the partial recent years.
Answers are grounded in the same dataset. Derived from a Patsnap search on Embedding Model Training covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Research Embedding Model Training in depth with Eureka
Go past this page: query the whole embedding model training corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries inherent limitations: recent filings (typically the most recent 18–24 months) are under-counted due to standard publication lag; counts may be reported at either a patent-family or a patent-record basis and are not always directly comparable; classification, applicant-name, and citation data may contain errors, duplicates, or omissions; and the underlying search query defines and constrains the scope shown. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, or validity search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. Any patent, commercial, or strategic decision should be verified independently and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company's registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.