A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever
This paper proposes a system where a frozen language model achieves 100% accuracy and zero token generation on verified problem families by relying on a persistent, bit-exact memory store of independently verified solutions, thereby decoupling execution-bound capability from parameter scaling and enabling massive context windows on consumer hardware.
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 solve a math problem. In the world of modern artificial intelligence, the usual way to get an answer is to ask a super-smart robot to think about it from scratch every single time you ask. This robot is like a brilliant but very expensive student who has to re-read the textbook, re-derive the formula, and re-write the solution for every single question, even if you ask the exact same question a thousand times. It takes a lot of time, uses a massive amount of electricity, and sometimes, if you ask the same question twice, the robot might give you two slightly different answers because it's a bit unpredictable. This is how most "frontier" AI models work today: they are powerful, but they are also slow, costly, and a bit messy when you need perfect consistency.
But what if, instead of making the robot think harder, we just gave it a perfect, verified notebook? Imagine if we could solve a problem once, check the answer with a strict teacher to make sure it's 100% correct, and then write that solution down in a special book. From then on, whenever the same type of problem comes up, the robot doesn't need to think at all. It just flips to the right page in the book, reads the answer, and writes it down instantly. It would cost almost nothing, happen in a blink of an eye, and the answer would be exactly the same every single time. This is the core idea behind a new approach called "verified reuse," where the intelligence isn't in the robot's brain getting bigger, but in a growing library of perfect solutions that the robot can access without doing any heavy lifting.
The Frozen Brain and the Magic Library
Meet Galahad, a new system that changes the rules of the game. The creators took a standard AI model (a 12-billion-parameter "brain" called Gemma-4-12B) and did something radical: they froze it. They turned off its ability to learn or change. It's like taking a brilliant chef and telling them, "You can't invent new recipes anymore; you can only cook from this specific menu."
Next to this frozen chef, they built Merlin, a magical, super-fast library. Here is how the magic works:
- The Deposit: When a new problem comes in, the system solves it once. But before it puts the answer in the library, a strict "gatekeeper" checks it. This gatekeeper doesn't look at the answer key; it uses math and logic to prove the answer is correct. If it passes, the solution is stored.
- The Zero-Cost Reuse: Now, imagine a million people ask the same type of question. Instead of the chef cooking from scratch every time, the system looks up the verified recipe in the library. It tells the frozen chef, "Just copy this." The chef does it instantly.
- The Result: The system answers 180 brand-new variations of these problems with 100% accuracy. It uses zero "generation tokens" (the digital currency AI usually charges for thinking). It takes only 6 to 23 milliseconds (faster than a blink) and uses about 36 milliwatt-hours of energy—that's roughly the energy an LED lightbulb uses for 13 seconds.
Why the Old Way is Wasteful
The paper argues that the current way of using AI is like paying a master architect to design a house from scratch every time you want to build a shed. The "frontier" models (the most powerful, expensive AI available) are great at solving new problems they've never seen. But for problems that have already been solved and verified, asking them to think again is a waste.
The authors measured this waste. They found that if you use a standard AI API to answer a solved problem, you pay for a full "thinking pass" every single time. It's non-deterministic (the answer might change slightly), it's slow, and it costs money. In contrast, Galahad's frozen model, once it has the verified solution in its library, answers the same problem forever for free. The "break-even point"—where the system saves more money than it cost to build the library—happens after just 17 to 34 questions. After that, every single answer is pure savings.
The "Memory" is the Real Star
One of the coolest parts of this experiment is proving that the "brain" doesn't matter as much as the "book." The researchers tested this with four different AI models from four different companies (including Gemma, Qwen, DeepSeek, and Phi). They gave all of them the exact same frozen weights and the exact same library of verified solutions.
The result? All four models got 180 out of 180 correct.
It didn't matter if the model was "dense" or a "mixture of experts." The capability came entirely from the library, not the brain. When they emptied the library, the models couldn't solve anything. When they filled it, they were perfect. This proves that for verified, repeatable tasks, you don't need a bigger, smarter, more expensive brain; you just need a better, verified memory.
The Problem with "Guessing" (Vector Search)
You might be thinking, "Can't we just use a normal search engine to find the answer?" The paper says no, and here is why. Most AI systems use something called "vector retrieval," which is like finding a book by looking for one that looks similar to the one you want. It's fuzzy.
The researchers tested this on a library of 4,500 verified items. When they used the fuzzy search, it picked the wrong item 94.3% of the time. That's a disaster. If you are looking for a specific medical formula or a legal clause, getting the "almost right" answer is useless. Galahad uses exact addressing, meaning it finds the exact item with 0% errors. It's the difference between guessing the name of a song by humming a few notes and knowing the exact barcode to scan. For things that need to be perfect, fuzzy matching fails catastrophically.
A Window into the Infinite
Another wild discovery is about how much information the system can hold in its "working memory" (the space it can think about at once).
- Standard AI engines (like vLLM and SGLang) crash or silently forget things when you try to feed them more than about 32,000 tokens (a token is a piece of a word).
- Galahad managed to hold a 6,000,000-token window on a single graphics card.
- It could reach back to the very beginning of that 6-million-token history just as fast as it could reach the end.
- It did this without using up extra computer memory. It's like having a library with infinite shelves, but you can only look at one book at a time, and you can swap that book for any other book in the library instantly.
What This Means for the Future
The paper is very careful to say what it doesn't do. It doesn't claim this system can solve brand-new, never-before-seen problems better than the biggest AI models. If you ask it a fresh, creative question, the frozen brain might struggle. But for any task that is verifiable (like math, coding, or logic) and recurring (happens over and over), this system is a game-changer.
It turns the AI industry on its head:
- Old Way: Pay to think every time.
- New Way: Pay once to verify, then execute forever for free.
The authors have even set up a public testbench where anyone can see this happen in real-time. They challenge anyone to find an AI that can answer a solved problem with zero cost, perfect accuracy, and instant speed. So far, no one has. The message is clear: for the work we do every day, we don't need to build bigger brains; we just need to build better, verified libraries.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.