Dinghao Pan

Kmin⁡(ρi​Ai,clip​(ρi,1−ε,1+ε)​Ai)\displaystyle\;\frac{1}{K}\sum_{i=1}^{K}\min\!\bigl(\rho_{i}A_{i},\,\mathrm{clip}(\rho_{i},1-\varepsilon,1+\varepsilon)A_{i}\bigr) −β​DKL​(πθ∥πref)\displaystyle\;-\beta\,D_{\mathrm{KL}}\!\bigl(\pi_{\theta}\|\pi_{\mathrm{ref}}\bigr) Here, ρi=πθ​(oi|x)πθo​l​d​(oi|x)\rho_{i}=\frac{\pi_{\theta}(o_{i}|x)}{\pi_{\theta_{old}}(o_{i}|x)} represents the quantification of policy change. ε\varepsilon controls the clipping threshold to ensure stable parameter updates of the policy model. 5 Results 5.1 Experimental Datasets and Settings We perform experiments on the DrugComb dataset (Tiktinsky et al., 2022), a biomedical corpus designed for n-ary drug combination extraction. It contains 1634 manually annotated abstracts, each mentioning 2 to 15 drugs, and categorizes relations into three types: POS, OTHER, and NO_COMB. To demonstrate the applicability of the RexDrug, we further apply the same training pipeline to the DDI13 dataset (Herrero-Zazo et al., 2013), a widely used benchmark for drug–drug interaction extraction, defining four binary relation types: Mechanism, Effect, Advice, and Int. Sentences without interactions are labeled NO_COMB. Additional dataset statistics and preprocessing details are provided in Supplementary Section S2. Table 2: Performance comparison on the DrugComb dataset. Bold indicates the best result and underline the second best. Results are reported as mean ±\pm standard deviation over 5 runs. w/o reasoning" denotes direct extraction without generating a reasoning trace, while "w/ reasoning" requires a reasoning trace. DAPT denotes continued domain-adaptive pretraining. "–" indicates that the value was not reported. PURE* and RCFIND* are classification baselines that use human-annotated entity mentions, which simplifies the task. Category Model F1(pos,exact) F1(pos,partial) F1(any,exact) F1(any,partial) PURE* PubmedBERT+DAPT 61.8 67.7 69.4 77.5 RCFIND* PubmedBERT 72.0 74.9 80.3 83.3 Seq