Conflict aware abstention does not improve selective accuracy over confidence ranking in FEVER claim verification
This paper demonstrates that while conflict-aware abstention in FEVER claim verification provides an interpretable rationale for withholding answers, it fails to improve selective accuracy over standard confidence-based ranking at matched coverage levels and performs poorly when applied to scientific evidence.
Original paper licensed under CC BY 4.0 (https://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
In the world of artificial intelligence, there is a growing expectation that computer systems should not just answer questions, but know when to stay silent. Imagine a student taking a test who is confident enough to write an answer only when they are sure they are right, and who leaves the page blank when the question is too tricky or the information is missing. This idea, known as selective prediction, is central to making AI reliable. For systems that check facts, the goal is to verify a claim by finding evidence in a large library of text. If the evidence supports the claim, the system says "true"; if it contradicts the claim, it says "false"; and if the evidence is missing or confusing, the system should admit it does not know. The challenge lies in teaching the machine to recognize that confusion. A simple way to do this is to look at how "sure" the computer feels about its answer. A more complex way is to have the computer actually read the evidence, sort it into piles of support and contradiction, and decide to stay silent only if the piles are mixed up or empty. The big question for researchers is whether this more complicated, evidence-aware approach actually leads to better results than simply trusting the computer's confidence score.
A recent study set out to answer this question by putting two different decision-making strategies to the test. The researchers wanted to see if a system that explicitly checks for conflicting evidence performs better than a system that simply abstains when its internal confidence is low. They ran their experiments on two different sets of data. The first was a large, balanced collection of 1,200 claims drawn from general knowledge, where the answers were already known and verified. The second was a smaller, more difficult set of 209 claims taken from scientific literature, where the language is specialized and the reasoning is often subtle. The goal was to see if the "evidence-aware" method could find the right answers more often than the "confidence-based" method, specifically when both methods were forced to answer the same number of questions.
The results were clear and somewhat surprising. When the researchers compared the two methods on the general knowledge set, they found that the evidence-aware system did not outperform the simpler confidence-based system. Both methods chose to answer roughly the same number of questions—about 54 percent of the total. Among the questions they did answer, both methods were correct about 90 percent of the time. The evidence-aware system was correct 90.05 percent of the time, while the confidence-based system was correct 90.36 percent of the time. This tiny difference means that the extra effort of sorting evidence into piles of support and conflict did not buy any measurable improvement in accuracy. In fact, the two systems ended up picking almost the exact same set of questions to answer, agreeing on 97 percent of their choices. The study suggests that for general fact-checking, the complex mechanism of analyzing evidence structure offers no real advantage over simply trusting the model's confidence score.
The story changed slightly when the researchers looked at the scientific literature. Here, the task proved much harder, and the evidence-aware system struggled significantly. It managed to answer only about half of the scientific claims, and when it did, it was correct about 84 percent of the time. The main problem was that the system often refused to answer claims that were actually true. A detailed manual review of 81 of these missed opportunities revealed why the system was failing. It was not because the evidence was contradictory, but because the system could not recognize the support when it was hidden behind scientific jargon, indirect reasoning, or careful qualifications. For example, a claim might be supported by a sentence that uses different words or requires connecting several facts together, and the system simply labeled this as "not enough information" instead of seeing the support. The study found that the system was missing these connections, leading it to stay silent when it should have spoken up.
The researchers also looked at whether the system was just guessing based on patterns in the data rather than actually reasoning. They found that the scientific dataset had a strong bias toward true claims, and that simple features like the length of the text or the specific words used could predict the answer with surprising accuracy. This means that the low performance numbers on the scientific set might be partly due to the difficulty of the task, but also because the system was not yet sophisticated enough to handle the nuances of scientific writing. The study concludes that while the evidence-aware approach provides a nice explanation for why a system might stay silent—pointing to a mix of support and conflict rather than just a low confidence score—it does not actually make the system smarter or more accurate. The extra complexity does not translate into better results, especially when moving from general knowledge to the specialized world of science.
Ultimately, this work serves as a careful check on the field of artificial intelligence. It shows that adding more layers of reasoning does not automatically lead to better performance. The researchers found that the simpler method of relying on confidence scores was just as effective as the more complex evidence-sorting method for general facts. For scientific claims, the problem was not that the system was too cautious, but that it lacked the ability to understand the subtle ways scientific evidence supports a claim. The study suggests that future improvements will likely come from better training on scientific language and reasoning, rather than from adding more rules about how to sort evidence. The most important takeaway is that for now, the ability to explain why a system is unsure is valuable, but it does not yet guarantee that the system will be right more often.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.