The numbers on a paper are never the full story, but when an ICLR submission lands with ratings of 3, 3, and 2, and one of those weak accepts comes from a reviewer who admits low confidence, the math gets complicated fast. The honest answer for the author facing this split is that the average rating is a signal, but it is not the decisive one. Confidence scores exist precisely to help Area Chairs interpret that signal, and a 1-confidence review, even a positive one, carries less weight in practice than a 3-confidence review of the same score. That does not mean the low-confidence review is discarded; it means the AC will scrutinize its substance more closely.
The real question, then, is what the AC is weighing after the rebuttal and discussion. This is where the process reveals its human center. A reviewer who gives a 3 with confidence 3 has likely done a thorough job and stands by their assessment. A reviewer who gives a 3 with confidence 1 may have identified a genuine strength in the work but is signaling that they are not sure enough to advocate for it. In a split decision, the AC is not counting votes; they are evaluating whether the low-confidence reviewer's identified issue was actually resolved in the rebuttal. If the paper's authors addressed that key concern clearly, the weak accept may hold more weight than its confidence suggests. If the rebuttal glossed over it, the weak reject from the confident reviewer will likely dominate the decision.
This tension between numerical signals and human judgment is a recurring theme in peer review, and it connects directly to a challenge we explored in our piece on Navigating AC-Only Comments at ICLR 2027. In that discussion, the ability for an AC to leave private comments created a space for candid evaluation that went beyond the scores. Similarly, here the AC is acting as a judge of both the reviews and the rebuttal, not just a calculator. The split decision is the AC's moment to decide which reviewer's reasoning is more convincing, and that judgment often comes down to whether the paper's contribution is clear enough to survive one confident skeptic.
For authors in this situation, the takeaway is direct and practical: do not assume the low-confidence positive review is a gift. Treat it as a vulnerability. That reviewer likely pointed to a specific weakness, and if the rebuttal did not fully address it, the confident negative review will echo that point louder. The best strategy is to engage that low-confidence reviewer's concern head-on in the rebuttal, because the AC will be reading for that exact resolution. This mirrors a broader lesson from Anthropic's research influence outpaces OpenAI's in citation rates, where sustained impact came from addressing foundational questions rather than chasing volume. A paper that survives a split decision by resolving its core critique is one that earns its place.
The specific detail to watch in this case is the substance of the rebuttal. If the authors can show that the low-confidence reviewer's key issue was either a misunderstanding or a solvable limitation, the AC may elevate that 3. If not, the confident 2 will likely tip the balance. The process is not a formula, but it is predictable in one way: confidence amplifies conviction, and conviction shapes the final call.