Iulia Brezeanu

Background

Writes hands-on benchmarks of LLM-pipeline techniques, measuring both quality and latency rather than describing methods abstractly.

Articles in This Vault

• How to Detect Hallucinations in LLMs — four self-consistency methods compared on accuracy and speed
• How to Cut RAG Costs by 80% Using Prompt Compression — AutoCompressors, Selective Context, LongLLMLingua

Key Contributions

• Showed SelfCheckGPT-Prompt best-performing among self-consistency hallucination detectors, at ~0.5 s
• Demonstrated that query-blind compression (Selective Context) discards the answer, while question-aware compression does not — the central point of Prompt Compression

Concepts

• Hallucination Detection · Prompt Compression · RAG · LLM as Judge