APTER Improves LLM Reasoning with Expert-Grounded Rubrics

As large language models (LLMs) are increasingly deployed in professional settings, they must meet specific domain constraints, incorporate critical evidence, and provide comprehensive reasoning, moving beyond merely fluent responses. Current post-training methods often rely on broad preferences or outcome-level verification. While some rubric-based approaches exist, they typically generate rubrics independently for each query, which can lead to the omission of crucial requirements and inconsistency across samples, hindering the diagnosis and targeted correction of persistent capability gaps. To overcome these limitations, researchers propose APTER (Adaptive Post-Training with Expert-Grounded Rubrics). This framework systematically integrates structured domain knowledge into a fine-grained process for evaluation, optimization, and diagnosis, particularly for complex reasoning tasks in specialized domains. APTER's expert-grounded rubric construction begins with a framework of criteria established by domain experts, where each criterion represents a stable professional capability. For each query, APTER selects relevant criteria and instantiates them into query-level rubrics, linking them back to their source criteria. This transforms reusable expert criteria into executable, query-level supervision without needing reference answers. The adaptive post-training component uses these rubric verdicts as both optimization signals and criterion-level diagnostic indicators. By aggregating low-scoring verdicts based on criterion ID, persistent deficiencies are revealed, triggering targeted supervised fine-tuning updates during reinforcement learning. Experiments in mathematical reasoning and medical question answering demonstrated consistent performance gains across both domains, with APTER improving scores by up to 15.86 and 8.04 points, respectively, over base models across three model generations.