Organizations: National Yang Ming Chiao Tung University, Taiwan · Stockmark, Japan · National Institute of Advanced Industrial Science and Technology (AIST), Japan · National Institute of Informatics (NII), Japan
Aligning large language models with expert judgment is especially difficult in subjective evaluation tasks, where experts may disagree, rely on tacit criteria, and change their judgments over time. In this paper, we study expert alignment as a way to understand this difficulty. Using expert evaluations and follow-up questionnaires, we examine how different forms of expert information affect alignment and what this reveals about subjective judgment. Our findings show four consistent patterns. First, alignment difficulty varies substantially across experts, suggesting that expert evaluation styles differ widely in their distance from a model's prior behavior. Second, explicit criteria and reasoning do not always improve alignment, indicating that expert judgment is not fully captured by verbalized rules. Third, editing is sensitive to both the number and the identity of examples, with small numbers of edits providing useful but unstable gains. Fourth, alignment difficulty differs across evaluation dimensions: dimensions grounded more directly in proposal content are easier to align, while dimensions requiring external knowledge or value-based judgment remain harder. Taken together, these results suggest that expert alignment is difficult not only because of model limitations, but also because subjective evaluation is inherently heterogeneous, partly tacit, dimension-dependent, and temporally unstable.
Evaluation and alignment pipelines for large language models increasingly rely on LLM-based judges, whose behavior is guided by natural-language rubrics and validated on benchmarks. We identify a previously under-recognized vulnerability in this workflow, which we term Rubric-Induced Preference Drift (RIPD). Even when rubric edits pass benchmark validation, they can still produce systematic and directional shifts in a judge's preferences on target domains. Because rubrics serve as a high-level decision interface, such drift can emerge from seemingly natural, criterion-preserving edits and remain difficult to detect through aggregate benchmark metrics or limited spot-checking. We further show this vulnerability can be exploited through rubric-based preference attacks, in which benchmark-compliant rubric edits steer judgments away from a fixed human or trusted reference on target domains, systematically inducing RIPD and reducing target-domain accuracy up to 9.5% (helpfulness) and 27.9% (harmlessness). When these judgments are used to generate preference labels for downstream post-training, the induced bias propagates through alignment pipelines and becomes internalized in trained policies. This leads to persistent and systematic drift in model behavior. Overall, our findings highlight evaluation rubrics as a sensitive and manipulable control interface, revealing a system-level alignment risk that extends beyond evaluator reliability alone. The code is available at: https://github.com/ZDCSlab/Rubrics-as-an-Attack-Surface. Warning: Certain sections may contain potentially harmful content that may not be appropriate for all readers.
LLM judges now score most open-ended NLP output, and their mutual agreement is routinely read as evidence that the scores can be trusted. That reading is unsafe: judges may agree because they capture quality, or because they share the same blind spots, and agreement statistics alone cannot tell these apart. We develop a geometric test that can. Treating each judge's scores as a vector, we measure spread, effective rank, the angle to human scores, and the judge-judge, judge-human, and human-human agreement triple for 42 judges on two community-built Indic benchmarks covering four domains and eight languages. Every comparison is reference-matched: a judge and a held-out rater are scored against the same two-rater mean, since an averaged reference otherwise flatters judges by several degrees. On subjective rubrics, judges agree with one another as much as humans do yet reach only 58-66% of human agreement and often concentrate on an axis humans do not weight. On the one rubric with a verifiable answer, most of that gap closes. Ensembles converge on the judges' shared axis rather than the human one, and training widens scores without rotating them. Inter-judge agreement is evidence of human alignment only after this check passes.
An LLM judge evaluates outputs at scale. Experts should label only where it is least sure. Its natural escalation signal conflates two uncertainties: aleatoric, real disagreement in the expert pool, which labels cannot reduce, and epistemic, the judge's ignorance, which labels do reduce. A small Bayesian model separates them: a regression on labels already collected learns how far to trust a black-box judge's prediction. Both components follow as simple formulas, with no sampling or further judge calls. The components isolate on a real LLM judge against exactly known truth, and stated confidence is no guide to its actual error. On real human disagreement (ChaosNLI) the epistemic ranking removes 83% more error than total uncertainty for the same expert labels, though simply escalating the least-labelled items does as well there. We demonstrate we can estimate where a judge is ignorant rather than where experts genuinely disagree, and propose using this to direct expert labelling. Code and data are available at https://github.com/composo-ai/judge-uncertainty-decomposition.