Confidence-Calibrated Small-Large Language Model Collaboration for Cost-Efficient Reasoning
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914368089751552 |
|---|---|
| author | Zhang, Chuang Zhu, Zizhen Wei, Yihao Tian, Bing Liu, Junyi Wang, Henan Wang, Xavier Liu, Yaxiao |
| author_facet | Zhang, Chuang Zhu, Zizhen Wei, Yihao Tian, Bing Liu, Junyi Wang, Henan Wang, Xavier Liu, Yaxiao |
| contents | 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%. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_03752 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Confidence-Calibrated Small-Large Language Model Collaboration for Cost-Efficient Reasoning Zhang, Chuang Zhu, Zizhen Wei, Yihao Tian, Bing Liu, Junyi Wang, Henan Wang, Xavier Liu, Yaxiao Computation and Language Artificial Intelligence 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%. |
| title | Confidence-Calibrated Small-Large Language Model Collaboration for Cost-Efficient Reasoning |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2603.03752 |