When an AI grades two nearly-identical answers, why does it become least reliable — and why won't the usual fixes help?

This explores why a judge is least trustworthy on close calls, where the verdict could go either way, and why those cases resist the usual fixes. The corpus has no paper on decision boundaries as such, so what follows is assembled from neighboring findings. They point to four separate obstacles.

The first is that the cheap fixes don't reach it. Telling a judge to be less biased does not reliably work, so the practical advice is to contain judge errors with structural checks rather than better instructions (Can prompting reduce bias in LLM judges reliably?). That matters most at the boundary, because when two answers are nearly equal, surface features like authority, verbosity, position and polish are what tip the scale. The one lever that does reduce those biases is training judges with reinforcement learning to reason through an evaluation before scoring (Can reasoning during evaluation reduce judgment bias in LLM judges?). That is a retraining project, not a prompt tweak.

The second obstacle is that the judge's output is often too coarse to express a close call. A judge that emits one discrete score token will tie two solutions that differ slightly. Averaging over the probabilities of the scoring tokens gives continuous scores and substantially fewer ties, without extra training (Can reading logit distributions break ties in LLM judging?). A judge can also drift toward the safe middle. One paper claims that deleting the scoring rubric improved a judge, but it leaves out the placeholder's error and the expert score distribution. That leaves open whether the gain was real or just a bias toward central scores, which look accurate on average while telling you nothing at the boundary (Did deleting the rubric actually improve the judge's performance?).

The third is that calibration isn't a single dial. Reasoning-trained models under-abstain and answer everything, because abstaining was never rewarded. Safety-trained models over-abstain and refuse benign questions. Each training objective leaves its own failure signature (Does training objective determine which direction models fail at abstention?). Pushing the boundary one way tends to worsen the opposite error. Binary right/wrong rewards make it worse, because they never penalize confident wrong answers. Adding a Brier score, which punishes confident mistakes, restores calibration without costing accuracy (Does binary reward training hurt model calibration?).

The fourth is that the problem is hard to see, and what you can't see you can't improve. Confident wrong answers cluster in rare cases and vanish inside aggregate accuracy (Why do confident wrong answers hide in standard accuracy metrics?). Temperature zero makes a judge repeat the same verdict on a borderline case every time. That looks stable, but it is one draw from a distribution that may be close to a coin flip (Does setting temperature to zero actually make LLM outputs reliable?).

The corpus suggests two ways forward. One is to measure against human labels and plant known cases as alarms, so boundary errors get caught mechanically (Can deterministic checks protect LLM judges from failure?). The other is to keep improving the judge alongside the thing it grades, since a frozen judge caps what self-improvement can achieve (Why do self-improvement loops plateau without updating the judge?).