What this article covers
- Explains the LLM-Based Essentiality Score used in the Memory Technology Package and how its 1–100% score and score category indicate likely patent essentiality.
- Describes how patent claims are compared with technical standards using vector-based retrieval, LLM claim analysis, and continuous expert validation.
- Explains how to interpret High, Medium, and Low score categories when assessing a patent's alignment with mandatory or SEP-like requirements.
- Clarifies how the LLM based Essentiality Score for Memory, differs from the Semantic Essentiality Score for all other technology packages, and the classifier-based declared/undeclared SEP methodology used elsewhere in IPlytics.
- Defines the limitations of the LLM-based Essentiality Score, including why the score supports portfolio prioritization but does not replace legal or technical claim-charting review.
Note: It is important to note that both the Semantic Essentiality Score methodology, and LLM Essentiality Score for Memory (both listed under the Essentiality Score, or "ES" in the product) estimate the likelihood a patent is essential to a given technical standard, proving a 1-100 score and score category. The difference between them is explained below.
Understanding the LLM-Based Essentiality Score
How to Interpret the score and score category rating
How the LLM-Based Essentiality Methodology Works
How It Differs from Other Scoring Methodologies in IPLytics
Overview
Determining whether a patent is essential to a technical standard is one of the most fundamental and challenging tasks in standard-essential patent (SEP) analysis. Traditional approaches rely on manual claim charting which is accurate, but time-consuming, expensive, and difficult to scale across large portfolios.
To address this, in addition to the Semantic Essentiality Score, LexisNexis IPlytics has developed an LLM-assisted essentiality methodology that also estimates the likelihood a patent is essential to a given technical standard. The DRAM Memory patent database covering HBM, GDDR, LPDDR, and DDR in the Memory Technology Package is the first to use this LLM-based approach.
How to Interpret the score and score category rating
For each patent, the methodology produces:
- An essentiality score from 1–100%.
- A score category of High, Medium, or Low.
Band |
Meaning (short)
|
Example next actions
|
🟩 High LLM-ES |
High likelihood the patent aligns with mandatory or SEP-like technical requirements. | Queue for expert confirmation/claim charting; include in licensing binders; prioritize for FRAND/licensing prep. |
🟨 Medium LLM-ES |
Clearly related to the standard, but mapping is incomplete, optional, or only partially aligned. | Route to analyst review; check which claim elements are supported vs. missing; escalate or defer based on business context. |
🟥 Low LLM-ES |
Only a remote or generic relationship to the standard, based on broader technical similarity rather than mandatory functionality. |
Spot-check a small sample; monitor only for strategic owners/technologies. |
How the LLM-Based Essentiality Score Works
The methodology compares each independent patent claim against all relevant sections of the applicable standard, using a multi-layered framework:
- Vector-based retrieval first identifies the most relevant sections of the standard, reducing hallucination risk and improving retrieval quality before LLM evaluation begins.
- LLM claim analysis, guided by carefully engineered prompting, identifies individual claim elements and evaluates whether corresponding technical features exist within the standard, producing transparent reasoning for why each claim element is or isn't supported.
- Continuous validation, including rigorous error analysis, subject-matter-expert (SME) review, patent owner feedback, error correction, bias reduction, and iterative model refinement.
This combination is what distinguishes the approach from a purely generative AI method. The LLM's output is checked, corrected, and refined rather than taken at face value.
How It Differs from the Semantic Essentiality Score in IPLytics
Both the Semantic Essentiality Score and LLM-Based Essentiality Score provide a score and score category, estimating the likelihood a patent is essential to a given technical standard. They just work differently 'under the hood' and both are noted under Essentiality Score or "ES" in the product.
They are different methodologies as follows:
| Semantic Essentiality Score | LLM-Based Essentiality Score | |
| Underlying method | Word2Vec, TF-IDF, and SVD (context-aware vector modelling) | LLM claim-by-claim reasoning + vector retrieval |
| Output | 1–100% score + High/Medium/Low confidence, across most packages | 1–100% score + High/Medium/Low confidence, across Memory package only. |
| Where it's used | Video, Audio, Cellular, Wi-Fi, Wireless Charging, Cellular IoT | Memory (DRAM) package only, currently |
It also is not a replacement for the classifier-based declared/undeclared methodology described in "Technology Packages In IPLytics." That process builds a landscape by classifying patents into declared and undeclared (predicted) SEP layers.
The LLM-based approach doesn't produce a declared/undeclared split at all. Instead of classifying each patent into one of two categories, it assigns every patent a continuous essentiality score (1–100%) with a score category, giving a graduated view of likely essentiality rather than a binary classification.
Finally, this is not a substitute for legal or technical claim-charting review. It's designed to help prioritize large portfolios efficiently, not to serve as a legal conclusion on essentiality, validity, infringement, or licensing obligations.