Large language models (LLMs) demonstrate superior reasoning capabilities compared to small language models (SLMs), but incur substantially higher costs. We propose COllaborative REAsoner (COREA), a system that cascades an SLM with an LLM to achieve a balance between accuracy and cost in complex reasoning tasks. COREA first attempts to answer questions using the SLM, which outputs both an answer and a verbalized confidence score. Questions with confidence below a predefined threshold are deferred to the LLM for more accurate resolution. We introduce a reinforcement learning-based training algorithm that aligns the SLM’s confidence through an additional confidence calibration reward. Extensive experiments demonstrate that our method jointly improves the SLM’s reasoning ability and confidence calibration across diverse datasets and model backbones. Compared to using the LLM alone, COREA reduces cost by 21.5% and 16.8% on out-of-domain math and non-math datasets, respectively, with only an absolute pass@1 drop within 2%.

Aspect-based sentiment analysis (ABSA) has attracted broad attention due to its commercial value. Natural Language Generation-based (NLG) approaches dominate the recent advance in ABSA tasks. However, current NLG practices are inefficient because most of them directly employ an autoregressive generation framework that cannot efficiently generate location information and semantic representations of ABSA targets. In this paper, we propose a novel framework, namely Efficient Hybrid Generation (EHG) to revolutionize traditions. Specifically, we leverage an Efficient Hybrid Transformer to generate the location and semantic information of ABSA targets in parallel. Besides, we design a novel global hybrid loss function in combination with bipartite matching to achieve end-to-end model training. Extensive experiments demonstrate that our proposed EHG framework outperforms current state-of-the-art methods in almost all cases and outperforms existing NLG-based methods in terms of inference efficiency.