Benchmark-Defined Adaptive Candidate Reranking for Indonesian Multi-Hop Question Answering

Multi-hop question answering (MHQA) requires recovering a complete set of supporting documents, yet most retrieval-augmented pipelines apply one evidence budget and one selection strategy to every question. This paper studies benchmark-defined adaptive candidate reranking for Indonesian MHQA in a closed setting in which each question is paired with ten candidate documents. A routing classifier assigns each question to a two-gold-document branch (HotpotQA), processed by two-stage sequential reranking, or to a four-gold-document branch (a bridge-comparison subset of 2WikiMultihopQA), processed by parallel ranking-based reranking, using lightweight XLM-RoBERTa cross-encoders. All components are trained without stemming and selected on a held-out validation split. On 8,000 Indonesian-translated evaluation questions, the framework recovers the complete supporting-document set for 95.40% of questions with 3.00 documents per question and 97.69% document precision. In evidence recovery, it significantly outperforms a single-model fixed-K=4 reranker (+2.60 points; McNemar's exact test p < 10−17); relative to top-4 selection with the same branch-specific scorers, it trades 0.89 points of containment for 23.55 points higher document precision and one document fewer per question. Ablations attribute the largest effect to sequential conditioning (+14.00 points at an equal budget of two documents) and show that routing errors cost 0.03 points within the benchmark, that transfer initialization mainly accelerates convergence, and that Sastrawi stemming degrades the scorers. Class-wise probes and a cross-source test show that the routing classifier relies on a film-topic shortcut and misroutes 10.78% of unseen two-document questions; it is therefore a benchmark-defined controller, not a detector of reasoning complexity, and the results do not establish generalization beyond the evaluated benchmark. With GPT-5-mini as the generator, adaptive evidence uses 68.5% fewer context tokens than all ten candidates and 26.8% fewer than fixed-K=4 selection at 9 to 20 ms of additional selection latency per question, but no improvement in answer accuracy was found: answer accuracy remains statistically indistinguishable across evidence contexts and within the run-to-run variance of the generator.