ONTO: A Token-Efficient Columnar Notation for LLM Input Optimization
The paper introduces ONTO, a novel columnar serialization format that significantly reduces LLM input token consumption and inference latency by eliminating repetitive field names while preserving human readability and hierarchical structure, without compromising task accuracy.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to explain a massive list of 1,000 grocery orders to a very smart, but very hungry, robot chef.
The robot (an AI model) has a limit on how much it can "read" at once before it gets tired or confused. This limit is called a token budget. Every word, number, and punctuation mark costs the robot "energy" (tokens) to process.
The Problem: The "Over-Explaining" Robot
Currently, when we send data to these robots, we use a format called JSON. Think of JSON like a very polite, but overly verbose, butler who insists on introducing every single item with its full name, every time.
If you have 1,000 orders for "Apples," the JSON butler says:
"Here is an order: Apples. Here is another order: Apples. Here is a third order: Apples..."
He repeats the word "Apples" 1,000 times, along with all the fancy brackets {}, quotes ", and colons :.
- The Result: The robot spends 80% of its energy just reading the labels ("Apples", "Price", "Quantity") and the fancy punctuation, leaving very little energy to actually think about the numbers and the logic. It's like paying a taxi driver to drive you around the block 1,000 times just to say "Taxi" out loud every time you get in.
The Solution: Onto (The "Efficient Chef")
The author of this paper, Harshavardhanan Deekeswar, invented a new way to write these lists called Onto.
Think of Onto as a smart spreadsheet or a train schedule. Instead of repeating the column headers for every single row, you write the headers once at the top, and then just list the data below them.
How it looks:
Instead of:
{ "item": "Apple", "price": 1.00 },
{ "item": "Apple", "price": 1.00 },
{ "item": "Apple", "price": 1.00 }
Onto writes:
item: Apple|Apple|Apple
price: 1.00|1.00|1.00
The Magic Analogy:
Imagine you are packing a suitcase for a trip.
- JSON is like packing 1,000 identical t-shirts, but wrapping each one in a separate box, labeling the box "T-Shirt," and putting it in a bigger box labeled "Clothing."
- Onto is like folding all 1,000 t-shirts neatly into one stack, writing "T-Shirts" on the top of the stack, and putting them in the suitcase.
What Did They Find?
The researchers tested this new method against the old way (JSON) and some other formats. Here are the results, translated into plain English:
- Huge Savings: Onto cut the amount of text the robot had to read by nearly half (46–51%).
- Analogy: If the robot's brain was a 100-page book, JSON made it read 100 pages of text. Onto let it read the same information in just 50 pages.
- Faster Thinking: Because the robot had less "noise" to read, it started answering questions 5–10% faster.
- Analogy: It's like the difference between reading a novel with footnotes on every line versus reading a clean novel. You get to the story much quicker.
- Just as Smart: The robot didn't get confused. It could still find specific items, count things, and do math just as well as it did with the old, bloated format.
- The "Indentation" Cost: The only time Onto gets slightly less efficient is when the data is very complex (like a box inside a box inside a box). In those cases, Onto uses spaces to show the "nesting," which adds a tiny bit of extra text. But even then, it's still much better than the old way.
Why Should You Care?
Right now, using AI for business (like analyzing thousands of sensor readings from a factory or logs from a website) is getting expensive because you are paying for all that wasted "label" text.
Onto is a tool that strips away the fluff. It tells the AI: "Don't waste your brainpower reading the labels; I've already told you what the labels are. Just look at the numbers and solve the problem."
The Bottom Line
This paper introduces a new "language" for talking to AI that is leaner, faster, and cheaper, without making the AI dumber. It's a simple switch from "repeating everything" to "listing everything once," which is a game-changer for anyone trying to use AI on large amounts of data.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.