Ensembles of Small Language Models Outperform Single LLMs in Malware Analysis

Researchers have demonstrated that orchestrating multiple small language models (SLMs) can effectively match or surpass the performance of single, larger language models (LLMs) in malware analysis. By employing various architectures like multi-agent pipelines, adversarial debates, and hierarchical consultations, ensembles of SLMs showed significant improvements. A hybrid system combining Qwen3-4B with Foundation-Sec-8B achieved 35.30% accuracy on Meta's CyberSecEval Malware Analysis benchmark, outperforming specialized baselines and ungrounded frontier models, though a grounded Gemini configuration still led with 38.22% accuracy. AI

IMPACT Demonstrates that smaller, more accessible models can be effectively combined to achieve high performance in specialized domains like cybersecurity.

RANK_REASON The cluster contains an academic paper detailing novel research findings on LLM orchestration. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• CyberSecEval Malware Analysis benchmark
• Foundation-Sec-8B
• Gemini
• Meta
• Qwen3-4B

• paper
• model release
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →