A Jury, Not an Electorate: Label-Free Reliability Estimation for Ensembles of Language Models

In the rapidly evolving world of artificial intelligence, researchers often turn to a strategy called "ensemble learning" to solve difficult problems. Instead of relying on a single, massive computer program to find the right answer, they gather a group of different programs, each trained on different data or designed with different strengths, and ask them all to work on the same task. The hope is that while one program might stumble, another will succeed, and by combining their efforts, the group can achieve a level of accuracy that no single member could reach alone. However, a persistent challenge has been how to decide which member of the group to trust when they disagree. Traditional methods often act like an election, where every model gets a vote, and the answer chosen is simply the one that receives the most support. This approach assumes that the group's collective wisdom is found in the majority, but it ignores the possibility that the most popular answer might be wrong, or that a quieter, more reliable expert is being drowned out by a louder, less accurate one.

A new study proposes a different way to think about this problem, suggesting that a group of artificial intelligence models should be treated not as an electorate, but as a jury. In a courtroom, a jury does not simply count how many people agree on a verdict; instead, they listen to the testimony of each witness and try to judge who is credible based on what they say and how their stories align. The researchers behind this work asked whether a group of language models could do the same thing: infer the reliability of each member solely from the pattern of their answers, without needing to know the correct answers in advance. They built a system where the models act as witnesses, and the system learns to trust the ones whose testimony consistently aligns with the truth, even though the system never sees the truth itself.

The researchers tested this idea on a benchmark of 1,560 questions covering a wide range of topics, from math and science to law and general reasoning. They gathered nine different language models, varying in size and training, to serve as the experts. In a standard setup, these models would be given a set of questions, and their answers would be combined using a simple majority vote or a weighted average where the weights are determined by how well the models performed on a separate set of questions with known answers. The new method, however, skips the step of looking at the correct answers to determine trust. Instead, it uses a mathematical process to analyze the answers the models gave to the questions they were asked. By looking at which models agreed with each other and which ones disagreed, the system estimates how reliable each model is. It treats the correct answer as a hidden fact that the models are trying to report, and it calculates the likelihood that each model is telling the truth based on the configuration of the group's responses.

The results of this experiment were striking. The jury-based system, which learned to trust its members without ever seeing a single correct answer, achieved an accuracy of 76.9 percent on the full benchmark. This performance was significantly better than a simple majority vote, which scored 63.0 percent, and also outperformed the best single model in the group, which reached 75.0 percent. Perhaps most importantly, the jury performed just as well as a system that used the same combination rules but was trained with the correct answers to determine the weights. This finding suggests that the complex process of labeling data to teach a system which models to trust is not strictly necessary; the system can learn to trust the right witnesses just by listening to what they say. The study also found that keeping the full range of possible answers a model might give, rather than just its top choice, significantly improved the results. When the system considered the probability distribution of answers, it gained an additional 8.4 percentage points over a simple vote, and weighting the experts correctly added another 5.5 points.

The researchers were careful to rule out the idea that this success was due to simply having more data or better sampling techniques. They compared their method against several other advanced techniques that also try to combine models without labels, and the jury consistently outperformed them. For instance, it beat the strongest alternative label-free method by 5.5 percentage points. The study also examined whether the system was simply lucky or if it was genuinely learning the right patterns. They tested the jury on different subsets of the models and found that it consistently outperformed both the equal-weight voting and the best single expert across almost all combinations. The only time the jury did not improve on the best single model was when the group consisted of very weak experts, a scenario where no combination method could realistically help.

One of the most significant aspects of this work is that it challenges the assumption that supervision is always required to get the best results. In many fields of artificial intelligence, the standard practice is to train a system on a large dataset where the correct answers are known, so the system can learn which models are good. This new approach shows that for combining models, the system can infer reliability directly from the agreement and disagreement among the models themselves. The researchers found that the difference between the jury and the supervised system was so small that it was not statistically significant when accounting for the natural variation in how the models were fitted. This implies that the extra effort of labeling data to train the combination rule might not be worth it in many cases.

The study also looked at how the system handles uncertainty. When the models are asked to generate answers, they often produce a range of possibilities rather than a single definitive statement. The jury method uses these full distributions of answers, which allows it to capture more nuance than a simple vote. For example, if one model is very confident in a wrong answer while another is slightly less confident in the right answer, the system can weigh the evidence appropriately. The researchers found that this approach was particularly effective on math problems, where the best single model in the group struggled, but the jury was able to leverage the strengths of other models to reach the correct solution.

While the method is powerful, the researchers noted some limitations. The system requires that every model in the group provides an answer, and for generative tasks, it needs to run multiple samples to build a reliable distribution of answers. This increases the computational cost compared to simply asking one model for an answer. However, the study suggests that this cost is justified by the significant gain in accuracy. The researchers also observed that the system's confidence in its answers was often higher than its actual accuracy, meaning it sometimes felt very sure of a wrong answer. This is a common issue in artificial intelligence, and the researchers noted that the system would need recalibration before its confidence scores could be used as a reliable measure of truth.

Ultimately, this research offers a new perspective on how we can harness the collective intelligence of artificial intelligence. By treating models as a jury rather than an electorate, we can build systems that are smarter than the sum of their parts without needing to know the answers in advance. The ability to learn trust from testimony alone opens the door to more efficient and adaptable ways of using artificial intelligence, where the system can continuously refine its understanding of which experts to rely on as it encounters new problems. The findings suggest that the key to unlocking the full potential of these groups lies not in forcing them to agree, but in teaching them how to listen to each other and judge the credibility of their own voices.