Salesforce Patents a Way to Trim AI Models by Measuring What Each Layer Actually Does
Most AI models carry dead weight: layers that do almost nothing but still consume time and computing power. Salesforce has filed a patent for a method that measures each layer's contribution and removes the freeloaders automatically.
How Salesforce's entropy pruning cuts AI bloat
Large AI models today are built from dozens (sometimes hundreds) of stacked processing stages. The problem is that many of those stages add very little to the final answer, yet every one of them still burns computing resources every time you ask the AI a question.
Salesforce's approach is to measure how much each stage actually changes the information flowing through the model. Think of it like measuring water pressure before and after each section of pipe: if the pressure barely changes, that section probably isn't doing much useful work. Stages that score near zero on that measurement get removed, and the hardware assigned to the job gets scaled down to match.
The result is a leaner model that needs fewer processors to run, which can translate into faster responses and lower costs for anyone using an AI service built on this technology.
… computing a respective change in entropy for each block based on a difference between an entropy estimate value for the respective output intermediate value and an entropy estimate value for the respective input intermediate value; …
Translation: It measures how much data complexity changes as it passes through each section of the AI.
How entropy scores decide which layers get cut
The patent describes a technique called entropy-based pruning. Entropy, in information theory, is a measure of how much variety or uncertainty exists in a set of data. A block (one processing layer in the AI model) that dramatically reshapes the data passing through it produces a big change in entropy; a block that leaves the data nearly unchanged produces almost no change.
Here is how the method works step by step:
- Forward pass measurement: The model processes a set of input data normally, and the system records the data values both entering and leaving each individual block.
- Entropy delta calculation: For each block, the system computes an entropy estimate (a single number summarizing information richness) for the input values and another for the output values, then finds the difference.
- Pruning decision: Blocks whose entropy change falls below a threshold are flagged as low-contribution and removed from the model entirely.
- Resource reallocation: Once blocks are removed, the hardware budget (number of processors, memory allocation) is reduced to match the now-smaller model.
This makes the approach dynamic: the same technique can be applied at deployment time rather than only during the original training phase, meaning a model can be trimmed after the fact without retraining it from scratch.
Blocks are pruned based on their respective changes in entropy, and hardware resources allocated to the pruned model are adjusted accordingly.
Translation: Useless parts of the AI are removed to free up computer processing power.
What this means for AI running costs and speed
For people using AI-powered tools at work, the practical effect would be responses that arrive faster or cost the company running the service less to deliver. Bloated AI models are one reason enterprise AI deployments are expensive; cutting redundant layers without measurably hurting output quality addresses a real operational pain point.
The filing is also part of a broader pattern of AI companies trying to make large models run on less hardware rather than always scaling up. Salesforce's approach of using information-theoretic scoring (rather than rule-of-thumb size limits) to decide what to cut is a methodological bet that precision beats blunt trimming. Coverage of this area and similar AI-efficiency approaches appears regularly among the latest Big Tech patents, reflecting how central the cost-of-inference problem has become across the industry.
If you use an AI assistant inside a Salesforce product, the difference this could make is straightforward: the same quality of answer, delivered faster, from a model that costs less to run. The entropy measurement step is what makes this more than routine model compression, because it gives engineers an objective signal about which layers matter rather than guessing by position or size. The concrete payoff for end users depends on how aggressively Salesforce applies the pruning threshold in practice, but the method at least gives them a principled way to find out.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
15 drawing sheets from US 2026/0244926 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →