Research map / Large language models / Agents, code and math
Advancing Intelligent Software Development and Trustworthy Models Through the Synergy of Software Engineering and LLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A doctoral consortium research summary describing the author's work on applying LLMs to software engineering tasks (NER for CVE metadata, CWE vulnerability identification) and applying SE testing and red-teaming methods to improve LLM reliability.
Problem
Integrating LLMs into software engineering raises reliability, safety, and scalability concerns, both for using LLMs to replace labor-intensive SE tools and for using SE techniques to test LLMs.
Approach
Frames a research agenda in two directions: LLMs replacing conventional SE tools (fine-tuned NER, multi-LLM reasoning verification for CWE identification) and SE methods (automated behavioral testing, continuous red-teaming, synthetic data pipelines) for testing and improving LLMs.
Key results
- Sample-efficient LLM-finetuned NER for CVE metadata achieved 90% F1 using 10% of the annotated data needed by prior models.
- Automated NLP behavioral testing with LLM-generated test cases reported a 98% reduction in manual effort.
- Chatbot fine-tuning with filtered synthetic data reported 90.55% verbosity reduction and 142.85% hallucination reduction (as stated).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune an LLM for software metadata extraction | NER for software metadata in CVE reports · CVE reports | F1 score: 90% | Helped |
| Use fine-tuning to reduce the annotated-data requirement | NER for software metadata in CVE reports · CVE reports | annotated data required: 10% | Helped |
| Generate test cases automatically using LLMs | Automated NLP system testing system · NLP model testing | manual effort reduction: 98% | Helped |
| Generate concise training examples from unstructured data | Consumer banking chatbot data pipeline · Consumer banking chatbot | synthetic training examples: 260,000 | Helped |
| Use prompt-based filtering to identify relevant examples | Consumer banking chatbot · Consumer banking examples | filtering accuracy: 99.4% | Helped |
| Fine-tune with filtered synthetic data | Consumer banking chatbot · Consumer banking chatbot | verbosity reduction: 90.55% | Helped |
| Fine-tune with filtered synthetic data | Consumer banking chatbot · Consumer banking chatbot | hallucination reduction: 142.85% | Helped |
Limitations
- Mostly summaries of ongoing or prior work without full methodological detail
- CWE multi-LLM framework is still in progress with no results reported
Experimental setup
Datasets: CVE reports CWE categories unstructured banking data
Metrics: F1 score verbosity reduction hallucination reduction filter accuracy
Closest papers by meaning
- CRAFT: Customizing LLMs by Creating and Retrieving from Specialized Toolsets (ICLR 2024 · similarity 0.86)
- Can LLM Replace Stack Overflow? A Study on Robustness and Reliability of Large Language Model Code Generation (AAAI 2024 · similarity 0.85)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.84)
- RvLLM: LLM Runtime Verification with Domain Knowledge (NeurIPS 2025 · similarity 0.84)
- Position: Trustworthy AI Agents Require the Integration of Large Language Models and Formal Methods (ICML 2025 · similarity 0.83)
- SG-Bench: Evaluating LLM Safety Generalization Across Diverse Tasks and Prompt Types (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.