Mohamed Mahmoud

Automating vulnerability analysis with LLMs is now more practical: RAVEN shows how to generate structured, high-quality reports using RAG, even achieving a 54% quality score on real-world vulnerabilities.

Multimodal retrieval can be boosted by over 10 points on a challenging benchmark by unifying LLM-based query expansion, a reasoning-enhanced retriever, and chain-of-thought reranking.

Multimodal retrieval struggles aren't about the retriever, but the query: distilling multimodal queries with RL unlocks state-of-the-art performance even with text-only retrievers.

LLMs can supercharge multimodal retrieval by iteratively "querying, hypothesizing, and verifying" to bridge visual-text reasoning gaps, yielding a 14-point nDCG@10 boost over the best multimodal model.