Researchers have published three papers addressing critical challenges in large language model deployment, focusing on privacy, infrastructure, and hardware verification.
According to arxiv.org, a paper accepted at ICML 2026 reveals new privacy vulnerabilities in split learning architectures for LLMs. The research presents “Patched Model Inversion with Dual-Sided Initialization (PIDI),” a two-stage attack that targets both private input prompts and output responses in split-LLM settings. To counter these threats, the researchers propose “Adapter-based DualGuard with Mutual Information Defense (ADMI),” which the paper states provides “strong empirical privacy protection with minimal impact on task performance.”
A second paper, presented at PEARC ‘26, introduces STREAM (Smart Tiered Routing Engine for AI Models), a middleware system addressing the fragmented LLM inference landscape. According to arxiv.org, STREAM combines local, HPC, and cloud inference with a three-tier routing architecture. The system achieves sub-second time-to-first-token (0.54s median on HPC), representing a “21.1x” improvement over batch mode’s 11.40s, according to the paper. The research reports that Llama 3.2 3B achieves “85.1% free-tier retention on a 1,200-query benchmark spanning ten domains.”
Finally, arxiv.org reports on HierSVA, a suite for LLM-driven hierarchical hardware formal verification. The paper presents a dataset of 342 modules and benchmark testing twelve recent LLMs, finding a “67.1%” module-level compile rate, with generated assertions detecting “70.2% of eligible injected faults.”