Yuyang Wei

Zehao Chen, Tianxiang Ai, Yifei Li et al.

Ensemble learning of LLMs has emerged as a promising alternative to enhance performance, but existing approaches typically treat models as black boxes, combining the inputs or final outputs while overlooking the rich internal representations and interactions across models.In this work, we introduce LLMBoost, a novel ensemble fine-tuning framework that breaks this barrier by explicitly leveraging intermediate states of LLMs. Inspired by the boosting paradigm, LLMBoost incorporates three key innovations. First, a cross-model attention mechanism enables successor models to access and fuse hidden states from predecessors, facilitating hierarchical error correction and knowledge transfer. Second, a chain training paradigm progressively fine-tunes connected models with an error-suppression objective, ensuring that each model rectifies the mispredictions of its predecessor with minimal additional computation. Third, a near-parallel inference paradigm design pipelines hidden states across models layer by layer, achieving inference efficiency approaching single-model decoding. We further establish the theoretical foundations of LLMBoost, proving that sequential integration guarantees monotonic improvements under bounded correction assumptions. Extensive experiments on commonsense reasoning and arithmetic reasoning tasks demonstrate that LLMBoost consistently boosts accuracy while reducing inference latency.

Yuyang Wei, Yijun Yu, Minxue Pan et al.

Microservice architecture refers to the use of numerous small-scale and independently deployed services, instead of encapsulating all functions into one monolith. It has been a challenge in software engineering to decompose a monolithic system into smaller parts. In this paper, we propose the Feature Table approach, a structured approach to service decomposition based on the correlation between functional features and microservices: (1) we defined the concept of {\em Feature Cards} and 12 instances of such cards; (2) we formulated {\em Decomposition Rules} to decompose monolithic applications; (3) we designed the {\em Feature Table Analysis Tool} to provide semi-automatic analysis for identification of microservices; and (4) we formulated {\em Mapping Rules} to help developers implement microservice candidates. We performed a case study on Cargo Tracking System to validate our microservice-oriented decomposition approach. Cargo Tracking System is a typical case that has been decomposed by other related methods (dataflow-driven approach, Service Cutter, and API Analysis). Through comparison with the related methods in terms of specific coupling and cohesion metrics, the results show that the proposed Feature Table approach can deliver more reasonable microservice candidates, which are feasible in implementation with semi-automatic support.