Bridge-Aware Reinforced Compositional Exploration and Decomposition for Biomedical Multi-Hop Answering

Source abstract

Biomedical question answering requires multi-hop reasoning over implicit relationships—e.g., identifying gene-associated diseases to retrieve target drugs. However, closed-book frontier language models achieve only ~14.57% precision on two-hop queries under single-answer prompting and answers drawn from parametric knowledge lack verifiable evidence, limiting their clinical utility. In this paper we propose Bridge-Aware Reinforced Compositional Exploration and Decomposition (BRACED). A planner trained with Group Relative Policy Optimization emits typed sub-queries, and a deterministic type-aware two-hop traversal over PrimeKG converts them into an answer set. Since the traversal is deterministic, the candidate set is a function of the emitted sub-queries alone, so the reinforcement signal optimizes the decomposition independently of other learned components, while the final answers are drawn directly from the retrieved entities. On the BioHopR benchmark, BRACED reaches a micro-F1 of 0.3783 and a macro-F1 of 0.6197, improving over the same planner trained with supervised fine-tuning alone by 10.2% and over an agentic baseline by a factor of 2.4. Two typed hops already reach 95.1% of gold answers, and reordering the same candidates raises micro-F1 by 12.0% without retraining, showing that performance is bounded by candidate ranking under a fixed cut-off rather than by decomposition quality.