New rule improves AI accuracy by verifying sources before voting

The problem with simple voting

When multiple AI models are asked the same question, researchers often look to a majority vote to determine the best answer. However, this method has a significant flaw: models can frequently share the same mistake. If several models provide the same incorrect information, a simple vote will not catch the error.

To solve this, researchers developed a method called "Verify, Then Vote." Instead of just counting votes, the system checks the sources each model cites. It then applies a fixed rule to pick the final answer based on whether the cited passages actually support the claims made by the models.

As blog.grainulation.org reported, the team tested this on 1,000 questions. They used five small local models, including Llama 3.2 3B and Phi-3 mini, running on a single machine. By separating the steps of answering, checking, and choosing, they created a system that is more robust than a simple majority count.

How the source check works

The process begins with five models generating search queries to find information from Wikipedia or research abstracts. Each model then produces an answer, provides a confidence score, and cites up to three passages. The system saves all of this data before any final decision is made to ensure the process remains modular.

To verify the information, the researchers used Claude Sonnet 5.5 as a separate checker. This model is shown only one sentence of the answer and one cited passage. It is tasked with stating whether the passage supports the answer, contradicts it, or does not contain enough information. Crucially, the checker never sees the correct answer during this step.

The rule follows a specific logic: it starts with the answer that received the most votes. It only switches to a different answer if the checker confirms that a cited passage supports that alternative. This ensures that a switch only happens when there is evidence for it, and it keeps a permanent record of every check for future auditing.

Testing the limits of the rule

The researchers first tested the rule on a practice set and found that it improved accuracy from 77.1% to 85.0%. However, they noticed a problem: the rule occasionally broke correct answers that the original vote had gotten right. They set a safety limit of 5% for how many correct answers the rule could break, but the initial version slightly exceeded that limit.

To fix this, they refined the rule to be more conservative. The new version states that if the starting answer is already backed up by a source, the system should keep it. When they ran this revised rule on a new set of models, it fixed 73 questions and only broke three, staying well under the 5% error limit.

The team also wanted to ensure the checker was actually reading the text. They ran tests where they swapped the cited text for unrelated content. They found that 76 of the 91 fixes disappeared when the text was removed, proving that the checker was indeed using the provided passages to make its decisions.

Efficiency and rerunning results

One common question is whether it would be cheaper to just let a large model like Claude choose the best answer directly. The researchers found that while Claude was slightly more accurate at picking the right answer, their "Verify, Then Vote" method was actually cheaper per question. Furthermore, their method provides a unique advantage by saving every intermediate check.

Because the answering, checking, and choosing steps are kept separate, the researchers can rerun decisions without asking the models for new information. They used a Python script to drop models one by one and a Rust program to re-evaluate the remaining saved answers. This process took only 23 seconds for 3,510 reruns without any new model calls.

This modularity means that researchers can change the rules for how answers are chosen without paying to collect new data. They can experiment with different logic, measure the effects, and see exactly why a decision was made. This provides a level of transparency and cost-efficiency that a single-pass large model cannot offer.

What comes next for the research

The researchers have already published their paper, code, and saved experiment records on GitHub. These resources include scripts that allow other people to reproduce their results or test their own custom rules on the same dataset. This open approach allows the community to see exactly how the rule behaves across different scenarios.

The next step for the team is to move beyond testing new rules on old questions. Because they have saved every run, they can now focus on generating new questions to see how the system performs on fresh data. They want to see how the rule holds up as the variety of queries increases.

The study also highlights the importance of mathematical proofs in AI safety. The team used Lean 4, a software for checking mathematical proofs, to prove that if the rule ever swaps a correct answer for a wrong one, a source check must have backed the wrong answer. This helps pinpoint exactly where a failure occurred in the chain of logic.