EnergyRoute: energy-based uncertainty routing for selective retrieval and large language model assistance in the...

Hierarchical classification of biotechnology R&D projects is critical for strategic planning and national R&D resource allocation, yet remains challenging due to fine-grained hierarchical taxonomies, severe class imbalance, and the cost of large language model (LLM) inference at scale. We propose EnergyRoute, an uncertainty-aware routing framework that distributes computation across three tiers: (i) a fine-tuned transformer encoder for confident predictions, (ii) retrieval-augmented k-nearest-neighbor fusion for moderately uncertain cases, and (iii) evidence-grounded LLM classification for the most difficult cases. Routing decisions are guided by the Helmholtz free energy of the classifier's logit distribution, which serves as a calibration-free uncertainty measure. On a Korean biotechnology project dataset comprising 82,316 training instances and 301 leaf classes with a [Formula: see text] class-imbalance ratio, EnergyRoute achieves a leaf-level micro-F1 of 0.862 and a macro-F1 of 0.778-a statistically significant improvement over the fine-tuned encoder baseline-while routing only 20% of samples to the LLM. Cross-dataset evaluation on the English Web of Science benchmark (142 classes) provides evidence of cross-dataset generalizability. Compared with existing hierarchical text classification methods and confidence-based cascading, EnergyRoute yields the best performance-cost trade-off, reducing LLM usage approximately fivefold relative to full LLM pipelines.