Large Language Model Soundness: A Survey of Knowledge Grounding, Trust Assessment, and Behavioral Control

Deliverable D4.1 (WP4 midterm state-of-the-art report) of the ARMADA MSCA Doctoral Network, funded by Horizon Europe under Grant Agreement No. 101168951 and by the Swiss State Secretariat for Education, Research and Innovation (SERI) under contract No 25.00154. Large language models now mediate search, programming, decision support, and scientific work, but fluent output is not necessarily trustworthy. We call a system sound when its behavior is reliable under immaterial variation, its claims are supported by appropriate evidence, and its confidence reflects the likelihood that those claims are correct. We survey work from 2021 to 2026, with earlier foundational results, and organize it around three questions. Reliable Knowledge Grounding asks whether an answer rests on the right evidence and uses that evidence faithfully, covering retrieval decisions, knowledge conflicts, and contextual hallucination. Verification, Robustness, and Confidence asks when an answer deserves trust, covering methods for checking answers, testing their stability, quantifying uncertainty, and identifying alignment-induced overconfidence. Post-Training and Inference-Time Control asks how sound behavior can be induced and enforced through post-training and controlled during generation. We judge each method by what it contributes to soundness. We then ask how those methods transfer to high-stakes domains, other languages, and multimodal evidence, and how the same checks apply when models plan and call tools over long horizons.