EnergyRoute: energy-based uncertainty routing for selective retrieval and large language model assistance in the hierarchical classification of biotechnology R&D projects – Scientific Reports

Classifying biotechnology R&D projects into fine-grained categories is make-or-break for strategic planning, funding, and national tech roadmaps. But getting those labels right is tough: taxonomies are deep and nuanced, classes are severely imbalanced, and running large language models (LLMs) on everything is expensive. A new pre-publication manuscript introduces EnergyRoute, a routing framework that blends classic transformers, retrieval, and targeted LLM use—guided by a principled uncertainty signal—to deliver strong accuracy at a fraction of the LLM cost.

Why this problem is hard

• Fine-grained hierarchies: Biotechnology taxonomies include hundreds of leaf classes with subtle distinctions, making small errors cascade up the hierarchy.
• Severe class imbalance: A few categories dominate the data while many long-tail classes are sparsely represented, challenging generalization.
• Cost at scale: End-to-end LLM classification can be accurate but is rarely economical across tens of thousands of project records.

What EnergyRoute proposes

EnergyRoute is an uncertainty-aware, three-tier routing system that assigns just enough computation to each sample:

• Tier 1 — Fast path: A fine-tuned transformer encoder (think BERT-class models) handles confident predictions at low cost.
• Tier 2 — Retrieval assist: For moderately uncertain cases, the system performs retrieval-augmented classification via k-nearest-neighbor fusion, combining model outputs with evidence from similar, labeled projects.
• Tier 3 — LLM with grounding: The most difficult samples are escalated to an instruction-tuned LLM that classifies with explicit references to retrieved evidence, improving reliability and auditability.

The key to routing is an uncertainty score derived from the model’s logit distribution using the Helmholtz free energy. Unlike confidence thresholds on softmax probabilities (which are notoriously miscalibrated), this energy-based measure is calibration-free and more robust, enabling the system to decide—on the fly—whether to trust the encoder, consult retrieval, or invoke the LLM.

Results that stand out

• Dataset: A large Korean biotechnology project corpus with 82,316 training instances and 301 leaf classes (highly imbalanced).
• Accuracy: Leaf-level micro-F1 of 0.862 and macro-F1 of 0.778—both statistically significant gains over a strong fine-tuned encoder baseline.
• Efficiency: Only 20% of samples are routed to the LLM, cutting LLM usage roughly fivefold versus full LLM pipelines while improving accuracy.
• Generalizability: Cross-dataset tests on the English Web of Science benchmark (142 classes) suggest the approach transfers beyond the original corpus.

For context, micro-F1 weights frequent classes more heavily, while macro-F1 gives every class equal voice. Strong scores on both indicate the system performs well on common categories without neglecting rare, long-tail ones—vital for national R&D mapping where niche areas matter.

Why it matters

EnergyRoute shows that smart triage beats brute force. By invoking LLMs only where uncertainty is high and grounding their decisions in retrieved evidence, the framework delivers a superior performance-cost trade-off. For ministries, funding agencies, and enterprise R&D offices, that means better dashboards, faster portfolio analysis, and more trustworthy signals without runaway compute bills.

Under the hood: the energy-based router

EnergyRoute’s secret sauce is how it quantifies uncertainty. Instead of leaning on probability confidence—often overconfident in deep models—it computes the Helmholtz free energy from encoder logits as a sharper, calibration-free uncertainty estimate. In practice:

• Low energy → confident encoder prediction → accept Tier 1 output.
• Moderate energy → consult nearest neighbors and fuse signals (Tier 2).
• High energy → escalate to LLM with retrieved evidence (Tier 3).

This design avoids brittle, hand-tuned thresholds and reduces the risk of overusing LLMs on cases the encoder could already solve.

How it compares to prior art

• Versus hierarchical text classifiers: EnergyRoute pairs competitive leaf-level performance with selective LLM use, improving both accuracy and interpretability through evidence grounding.
• Versus confidence-based cascading: Energy metrics offer sturdier uncertainty estimates than softmax confidence, leading to more precise routing and lower costs.
• Versus full LLM pipelines: Similar or better accuracy with approximately one-fifth the LLM calls on the reported dataset.

Where this could go next

Because EnergyRoute decouples routing from any specific encoder or LLM, it’s poised to benefit from future advances in both. There’s room to:

• Adapt thresholds dynamically per class to further protect long-tail performance.
• Expand retrieval sources for richer evidence trails and stronger LLM grounding.
• Explore multilingual corpora and domain shifts to stress-test generalizability.
• Integrate cost-aware policies (e.g., budget caps) for enterprise deployment.

Caveats

The manuscript is an early, unedited version and may contain errors prior to final publication. Real-world performance will depend on taxonomy design, retrieval quality, and the stability of LLM outputs under prompt or model changes. Nonetheless, the reported gains—higher F1 scores with sharply reduced LLM usage—make a compelling case for energy-based routing as a practical path to scalable, accurate hierarchical classification in biotech R&D and beyond.