The few-shot dilemma: Over-prompting large language models
Published in: 2025 3rd International Conference on Foundation and Large Language Models (FLLM)
Yongjian Tang; Doruk Tuncel; Christian Koerner; Thomas Runkler
Over-prompting occurs when excessive examples in prompts degrade LLM performance, challenging traditional few-shot learning assumptions. To study this dilemma, a prompting framework evaluated random sampling, semantic embedding and TF-IDF vector selection across models like GPT-4o, DeepSeek-V3, and LLaMA-3.1. Results show that too many domain-specific examples paradoxically harm accuracy. Tested on software requirement datasets, gradually increasing stratified and TF-IDF-selected examples revealed optimal prompt sizes. This approach avoids over-prompting while outperforming state-of-the-art requirement classification by 1%.
View at publisher's page
LLM-based agentic systems for software engineering: Challenges and opportunities
Submitted to: SE2026, Gesellschaft für Informatik, Bonn
Yongjian Tang, Thomas Runkler
Despite LLM advancements, complex software engineering tasks demand collaborative, specialized approaches. This paper systematically reviews the emerging paradigm of LLM-based multi-agent systems across the entire software development lifecycle, including requirements engineering, code generation, testing and debugging. It examines model selection, evaluation benchmarks, agentic frameworks and communication protocols. Additionally, the work highlights key challenges like orchestration, human-agent coordination, cost optimization and data collection, offering essential insights into the state of agentic software engineering.
View at publisher's page
Model-driven legacy system modernization at scale
Published in: ReCode '26: Proceedings of the 1st Workshop on Code Translation, Transformation, and Modernization, pages 13 ‑ 18
Tobias Böhm, Jens Guan Su Tien, Mohini Nonnenmann, Tom Schoonbaert, Bart Carpels, Andreas Biesdorf
This experience report presents a four-stage model-driven approach (analysis, enrichment, synthesis, transition) that inserts a technology-agnostic intermediate model between legacy codebases and modern target platforms. Applied to a large industrial .NET application, it semi-automatically migrates user interfaces and page structures to modern web stacks while preserving functional behavior. By consolidating architectural knowledge into explicit models, it improves maintainability and developer experience. Despite layout challenges requiring manual effort, the approach reduces risk, scales effectively and offers reusable patterns.
View at publisher's page
Bridging formal syntax and LLM semantics: Extracting knowledge graphs for legacy code understanding
Published in: 2026 IEEE 23rd International Conference on Software Architecture Companion (ICSA‑C)
Henrik Thillmann; Bernhard Rumpe; Andreas Biesdorf
Established enterprises often rely on legacy systems that are difficult to maintain due to outdated languages and unclear identifiers. However, replacing these systems poses the risk of losing their proven reliability. Large language models can assist with architecture recovery and modernization, but they require precise context management. Since code is a parseable formal language, retrieval can exploit program structure instead of relying on unstructured text. We propose a parsing-based methodology to build a structure-aware knowledge graph for GraphRAG. This approach demonstrates improved retrieval precision and multi-hop architectural reasoning in an industrial case study.
View at publisher's page
