SearcharxivSearch

arXiv subjects

Zizhen Zhu

Publications and source records attributed to Zizhen Zhu.

2 recordsLinked to original sources

Confidence-Calibrated Small-Large Language Model Collaboration for Cost-Efficient Reasoning

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%.

cs.CL

LESS: Efficient Log Storage System Based on Learned Model and Minimum Attribute Tree

In recent years, cyber attacks have become increasingly sophisticated and persistent. Detection and investigation based on the provenance graph can effectively mitigate cyber intrusion. However, in the long time span of defenses, the sheer size of the provenance graph will pose significant challenges to the storage systems. Faced with long-term storage tasks, existing methods are unable to simultaneously achieve lossless information, efficient compression, and fast query support. In this paper, we propose a novel provenance graph storage system, LESS, which consumes smaller storage space and supports faster storage and queries compared to current approaches. We innovatively partition the provenance graph into two distinct components, the graph structure and attribute, and store them separately. Based on their respective characteristics, we devise two appropriate storage schemes: the provenance graph structure storage method based on machine learning and the use of the minimal spanning tree to store the graph attributes. Compared with the state-of-the-art approach, LEONARD, LESS reduces 6.29 times in storage time, while also achieving a 5.24 times reduction in disk usage and an 18.3 times faster query speed while using only 11.5% of the memory on DARPA TC dataset.

cs.CR