Confidence-Calibrated Small-Large Language Model Collaboration for Cost-Efficient Reasoning
This addresses cost efficiency for users of language models in reasoning tasks, but it is incremental as it builds on existing cascading and confidence calibration methods.
The paper tackles the problem of high costs in using large language models (LLMs) for reasoning by proposing a cascaded system that uses a small language model (SLM) with confidence calibration to defer uncertain questions to an LLM, reducing costs by 21.5% and 16.8% on out-of-domain datasets with minimal accuracy loss.
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%.