CADAQUES: A Cost-Aware Dual Architecture for Query-Efficient Autonomous Discovery

Imagine you are a detective trying to solve a mystery in a giant, foggy city. You have a limited amount of money and time to ask questions to find the culprit. In the world of science, this "detective work" is called autonomous discovery. Scientists use smart computer programs to run experiments, like testing how a new material behaves or finding the perfect temperature for a chemical reaction. These programs act like a loop: they ask a question (run an experiment), get an answer, and then decide what to ask next.

Usually, scientists treat every question as if it costs the same amount of time and money. But in reality, some questions are cheap and fast (like a quick guess), while others are expensive and slow (like a deep, detailed analysis). Furthermore, the "detective" itself—the smart computer program deciding what to ask next—can also get tired and spend resources just by thinking! If you don't track these costs carefully, you might run out of money before you find the answer, or worse, you might spend all your cash on expensive questions that don't actually help you solve the mystery. This paper tackles the problem of how to manage these costs so scientists can get the best results without going broke.

---

The Detective's New Ledger: Meet CADAQUES

The paper introduces a new open-source tool called CADAQUES (which stands for Cost-Aware Dual Architecture for QUery-Efficient autonomous diScovery). Think of CADAQUES as a super-smart, strict accountant for scientific experiments. Its main job is to make sure that every single step of the discovery process—both the questions asked and the thinking done to decide those questions—is tracked against a shared budget.

Before CADAQUES, most software frameworks for these experiments were like a game where you just count how many turns you took. If you had 100 turns, you played 100 times, regardless of whether Turn 1 took one second or Turn 100 took one hour. CADAQUES changes the rules. Instead of counting turns, it counts resources. It treats time, computer power, money, and even "tokens" (the currency used by AI language models) as a single, shared budget.

The Two-Player Game: The Oracle and the Driver

CADAQUES splits the scientific loop into two distinct characters, each with a specific job:

1. The Oracle (The Answerer): This is the part that actually does the work. It could be a computer simulator, a robot in a lab, or a mathematical model. When asked a question, the Oracle says, "I think this will cost me 5 minutes and $2 to answer," and then it does the work.
2. The Driver (The Questioner): This is the smart brain that decides what to ask next. It looks at the answers it's already gotten and the money it has left, then proposes a new question.

The magic of CADAQUES is that it charges both characters against the same wallet. If the Driver spends 10 seconds just thinking about which question to ask, that 10 seconds is deducted from the total budget, just like the time the Oracle spends answering.

The "Declared" vs. "Settled" Bill

One of the coolest features is how it handles the bill.

• Declared Cost: Before the Oracle starts working, it gives an estimate: "This will take 5 minutes." This is the "declared" cost. The system checks if you can afford it. If you can't, it stops you before you waste time.
• Settled Cost: After the Oracle finishes, it sends the real bill: "Actually, it took 7 minutes." This is the "settled" cost.

CADAQUES records both numbers. If the Oracle underestimated the time, the system notes the difference. This creates a perfect history, or "ledger," of exactly what happened, how much it really cost, and how accurate the predictions were. This helps scientists learn to budget better next time.

What Did They Find?

To test if this system works, the authors ran a simulation based on a famous physics problem: finding the "critical temperature" of a 2D Ising model (a mathematical model of magnets). They set up a noisy, tricky environment where the answers weren't perfect and could be misleading.

They tested three different "Drivers" (strategies for asking questions) with the same 10-second budget:

1. Random Search: Just asking questions randomly.
2. Annealed Local Search: A strategy that starts broad and gets more focused as it spends money.
3. Gaussian Process (GP-EI): A very sophisticated, expensive AI planner that tries to predict the best answer.

The Results:

• The "Smart" Planner Didn't Win: Surprisingly, the fancy, expensive AI planner (GP-EI) did not perform better than the simple random search in this specific noisy setting. In fact, it spent about 3.2% of its total budget just on the "thinking" part (deliberation) before it even asked a single question!
• Noise Traps: The smart planners got tricked by "noise"—random spikes in the data that looked like great answers but were actually fake. They wasted their remaining budget investigating these fake peaks. The random search, which didn't get stuck on one spot, actually found the answer just as well (or better) in many cases.
• The Power of Mixing Costs: The most successful strategy was a "two-phase" plan. The scientists used cheap, low-quality questions to explore the whole city quickly, and then switched to expensive, high-quality questions only in the most promising areas. This "explore cheap, refine expensive" approach found the answer with less error than using expensive questions the whole time.

The Cost of the System Itself

Finally, the authors checked how much time CADAQUES itself took to do the accounting. They found that the system added only about 25 microseconds (0.000025 seconds) of overhead per question. This is roughly 1,000 times faster than the cheapest experiment they ran. So, the tool is incredibly lightweight and doesn't slow down the science.

Why This Matters

The paper suggests that for autonomous discovery to work well, we need to stop treating "thinking" and "doing" as free. We need to track every penny and every second. The study shows that being "smart" with a complex algorithm doesn't always beat being simple and thorough, especially when the data is noisy. It also proves that mixing cheap and expensive experiments is a winning strategy.

Most importantly, CADAQUES provides a way to see the real cost of scientific discovery. It lets researchers look at their "ledger" and say, "We spent 10 seconds, but 3 of those seconds were just the computer thinking." This transparency helps scientists build better, more efficient experiments for the future, ensuring they don't run out of resources before they find the truth.