Two research papers published on ArXiv on July 13, 2026, present complementary approaches to improving reliability and safety in Large Language Model (LLM) systems through enhanced control mechanisms.
According to the first paper (arXiv:2607.09076), researchers introduced a “neuro-agentic control framework” that addresses the safety risks of using LLMs for autonomous control in critical infrastructure. The framework couples an LLM-based planner, specifically Gemini 2.5 Flash-Lite, with a Time-Series Foundation Model (TimesFM) to enable “physics-grounded autonomous defense” against cyberattacks on operational technology. The paper describes a “Counterfactual Physics Injection” mechanism that simulates proposed LLM interventions before execution, allowing the system to reject unsafe or hallucinatory actions.
Tested on the Secure Water Treatment (SWaT) industrial dataset, according to the paper, the framework prevented five breaches (33.3%) below the threshold compared to LSTM (26.7%) and TCN (13.3%) baselines, with “zero physically invalid (hallucinated) actions executed.”
The second paper (arXiv:2607.08774) introduces CogniConsole, which addresses reliability through “inference-time control” — the computational layer governing task framing and context selection. Through controllability-oriented probes (N=489) in a multi-step interactive environment, the research demonstrated that increasing structural scaffolding “systematically reduces output variance and failure rates under a fixed model architecture,” according to the paper. The researchers argue that many failure modes arise from “under-specified control rather than insufficient capability.”