Subliminal Signals in Preference Labels

As LLM-as-a-judge becomes increasingly widespread, self-preference -- the tendency of a judge to favor its own outputs -- raises growing concerns about evaluation reliability. However, this bias has been studied predominantly on generated text, where stylistic features and response quality are inevitably conflated. As a result, existing measurements cannot separate genuine self-preference from these confounds. We address this limitation by changing the object of evaluation: instead of judging generated text, ten LLMs assess sets of narrative constraints selected from a shared pool, which carry no stylistic fingerprint yet retain a recoverable model-specific signature. Two experiments on this task yield complementary findings. Under blind evaluation, self-preference disappears, with a small effect remaining in the opposite direction once selection quality and judge severity are controlled. Under matched quality, however, self- and other-labels alone -- without naming any model -- shift scores bidirectionally. LLM judges inflate scores for self-labeled selections and deflate those for other-labeled ones regardless of the selection's actual source. We make two contributions: 1) authorship attribution is a distinct driver of evaluation bias, and 2) ground-truth-free tasks can serve as controlled instruments for studying LLM judge behavior.

LLM judges are widely used to evaluate model outputs, but their verdicts can be unreliable: a judge may favor the worse answer for its position, length, or other surface features. When a judge is wrong, is the information needed to judge correctly absent from the model, or present in its internal representations but not reflected in the output? We study this across 64 open-weight evaluators and 14 datasets, including causal interventions on 41 judges (editing activations mid-run to see whether the verdict changes). On LLMBar, built so the superficially better answer is the worse one, the verdicts of 50 judges agree with human labels only 0.456 of the time, even after averaging both answer orders. Yet a small probe on the same judges' activations, with no weight updates, reaches 0.846, and 0.686 once surface features such as length and position are residualized out (0.507 with shuffled labels). The gap holds across eight benchmarks and model families, but is not universal: a score of how well surface features alone predict the human label, computed before any probe is trained, predicts the size of the gain (Spearman rho = 0.90). On rubric tasks that score one answer at a time, leaving no surface cue to exploit, reading the internals gives no advantage. The interventions also show that editing activations mid-network already changes the verdict, before it can be read off directly, and locate the pathways carrying position and length bias. At the same label budget, the recovered signal lets a judge flag cases where it is likely wrong and yields better labels for preference learning. A wrong verdict, then, does not mean the judge lacks the information, and a simple diagnostic shows when it is worth recovering.

Large Language Models (LLMs) are increasingly used as judges for scalable evaluation, yet such LLM--as--a--Judge systems exhibit systematic biases that are decoupled from semantic quality, most notably verbosity bias. Meanwhile, human supervision is costly and typically selective, yielding reliable positive judgments but leaving most outputs unlabelled and potentially mixed in quality. We formulate LLM evaluation under selective human supervision as a positive--unlabelled learning problem and propose a geometric auditing framework based on Partial Optimal Transport. By aligning a small set of human--verified positives with a reliable subset of unlabelled outputs in a fixed embedding space, our method identifies human--consistent preferences and corrects biased judges without retraining. Experiments demonstrate improved alignment with human preferences, increased robustness to presentation biases, and interpretable confidence estimates, offering a scalable and statistically grounded alternative to existing LLM--as--a--judge pipelines.