Answers are grounded in this paper and cite their sources.
Large language models (LLMs) exhibit strong generalization capability, but their inference cost is high and they are not always superior on supervised tasks to locally deployed small language models (SLMs) fine-tuned on labeled data. In contrast, SLMs can achieve strong in-domain performance with efficient deployment, yet often degrade under out-of-distribution (OOD) or cross-domain shifts. Although prior work has studied hybrid inference that invokes LLMs on demand, most approaches rely on trained routers and are primarily designed for generative tasks; under supervised classification with the ID-to-OOD transfer setting, stable and interpretable collaboration without additional router training remains underexplored. To this end, we propose LinkModel, an uncertainty-guided hybrid inference framework. LinkModel follows a Prediction–Estimation–Prediction mechanism: a locally fine-tuned SLM first produces a prediction and estimates uncertainty to obtain a risk signal, and a threshold-gated policy then defers high-risk samples to an LLM for a second-pass prediction. This design improves accuracy and robustness while controlling the LLM invocation ratio and the overall cost. We conduct systematic evaluation and mechanism analysis under the GLUE-X OOD generalization setting; results show that LinkModel achieves consistent end-to-end gains across multiple train–test transfer combinations and substantially mitigates OOD performance degradation.
No takes yet. Share an insight, caveat, or question.
Bai et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: