Meta-analyses are central to evidence-based medicine, but their production remains slow because reviewers must search PubMed and screen large candidate sets under PICOS eligibility constraints. We introduce MetaBench, a large-scale benchmark for early-stage PubMed title/abstract citation screening in meta-analysis study selection, formulated as a two-stage reranking and Top-N screening problem with effort-aware evaluation. MetaBench contains 7,801 PubMed meta-analyses, 120,930 PMID-linked included studies, and reproducible candidate pools with 4.2M unique title/abstract records constructed from replayed PubMed strategies or expert-validated training-only reconstructions. To improve evaluation fidelity, we introduce a coverage-based tiering protocol and reserve Tier 1 (1,297 topics) as a high-integrity test set. MetaBench evaluates candidate reranking with Recall/nDCG and Effort@R/NormEffort@R, and budgeted Top-N screening with sensitivity, specificity, precision, and F-scores. The main benchmark intentionally focuses on scalable title/abstract screening rather than full-text eligibility adjudication; PMCID metadata supports an auxiliary full-text track on an OA-constrained subset. Experiments show that MetaBench supervision improves dense reranking, reduces screening effort, and strengthens LLM-based inclusion decisions. An anonymized codebase and dataset sample are included with this ARR submission; the full dataset and codebase will be publicly released upon acceptance.

Large-scale generative models like DeepSeek-R1 and OpenAI-O1 benefit substantially from chain-of-thought (CoT) reasoning, yet pushing their performance typically requires vast data, large model sizes, and full-parameter fine-tuning. While parameter-efficient fine-tuning (PEFT) helps reduce cost, most existing approaches primarily address domain adaptation or layer-wise allocation rather than explicitly tailoring data and parameters to different response demands. Inspired by “Thinking, Fast and Slow,” which characterizes two distinct modes of thought—System 1 (fast, intuitive, often automatic) and System 2 (slower, more deliberative and analytic)—we draw an analogy that different “subregions” of an LLM’s parameters might similarly specialize for tasks that demand quick, intuitive responses versus those requiring multi-step logical reasoning. Therefore, we propose LoRA-PAR, a dual-system LoRA framework that partitions both data and parameters by System 1 or System 2 demands, using fewer yet more focused parameters for each task. Specifically, we classify task data via multi-model role-playing and voting, and partition parameters based on importance scoring, then adopt a two-stage fine-tuning strategy of training System 1 tasks with supervised fine-tuning (SFT) to enhance knowledge and intuition and refine System 2 tasks with reinforcement learning (RL) to reinforce deeper logical deliberation next. Extensive experiments show that the two-stage fine-tuning strategy, SFT and RL, lowers active parameter usage while matching or surpassing SOTA PEFT baselines.