Three research papers published on arXiv on June 18, 2026, examine different aspects of large language model (LLM) behavior and interfaces.
According to arXiv.org (arXiv:2606.17443), researchers studied brand dynamics in LLM recommendations using skincare products across GPT-4o-mini, Claude Sonnet, and Gemini 3 Flash. The study found what researchers termed a “Conditional Monopoly” where well-known brands received recommendations 100% of the time (IAI = 10.0) when products had identical specifications, though this advantage disappeared with less than a +0.1-star rating difference. The research also found that authority-style marketing language, including fabricated clinical-evidence claims, could break this monopoly at a “Bias Surplus Value” equal to +0.17 rating points.
A second paper (arXiv:2606.14517) revealed a vulnerability in LLM-based guardrails, according to arXiv.org. Researchers demonstrated that attackers could inject crafted data to trap guardrails in extended reasoning loops, creating denial-of-service attacks. The study reported that payloads optimized on open-source models transferred to eight leading model backbones including Claude, GPT, Gemini, DeepSeek, and Qwen, achieving 13–63× token amplification and up to 148× latency amplification in real-world agent deployments.
Finally, arXiv.org reported on PromptMN (arXiv:2606.17164), a pseudo-prompting domain-specific language using %-prefixed typed directives. The paper stated that feasibility was evaluated across Claude Fable 5, Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5 without fine-tuning.