Large language models acquire substantial factual and procedural knowledge when trained on large datasets. A growing body of work is concerned with understanding, modifying, and auditing this knowledge. Many of these works explore the localization of knowledge: whether a fact, skill, or competence associated with a task is localized in a small set of parameters within the model or distributed across them. There is a mixed perspective regarding the state of the art. Some studies identify compact representational or causal localization of facts, including mid-layer feed-forward modules and knowledge neurons, while others show that restricting the parameter updates allows one to change factual associations. Causal tracing scores can fail to accurately locate the best edit site, and superposition can make individual neurons polysemantic, so localization does not have to align with any isolated units. We believe that part of this disagreement is definitional. The question ``Is knowledge localized?'' conflates three distinct properties. Representational localization is whether knowledge can be retrieved from a restricted number of representational units; causal localization is whether interventions on a restricted set of units will affect the behavior; and editable localization is whether a restricted region of the model's native parameters can be changed to affect the knowledge. In this survey, we aim to disambiguate the existing works based on the kind of evidence they actually present and show that the evidence available is uneven in its distribution across the three pairwise relations. We also provide a measurement checklist for describing a localization claim in terms of the notion, units, representation system, and measurement procedure, along with a research program toward a cohesive, representation-inclusive account of the localization of knowledge.

Multimodal large language model (MLLM) agents connect perceptual inputs to tools and external actions, making safety a property of trajectories and retained state. This curated survey uses a constrained partially observable Markov decision process to organize six risk families: grounding failure, observation manipulation, miscalibrated authority, inference-time hijacking, unsafe commitment, and reusable-state corruption. We compare safeguards by mechanism, access requirements, and evidence scope, separating direct multimodal execution from transferred model-level and text-agent findings. The synthesis emphasizes image–text, GUI/web, and selected embodied settings; it is neither exhaustive nor independently coded. We distinguish response scores from executed violations and later-task reactivation, and propose containment and restoration outcomes for recovery evaluation. These distinctions motivate execution safeguards, calibrated autonomy, and integrity checks for reusable agent state.

Persistent self-improvement enables language-model agents to adapt from experience, but its update process competes with inference for limited device memory, energy, and thermal headroom. Existing surveys largely examine self-improvement mechanisms and on-device learning separately, providing limited guidance on how update reliability and device costs should be assessed together. We introduce a unified taxonomy that organizes persistent updates by component, execution locus, feedback, loop closure, timescale, and budget, and distinguishes recursive changes to evaluators or update machinery. Across the included works, we compare updates to memory and skills, prompts and code, weights, and evaluators or acceptance rules, and propose accounting and reporting rules that relate retained gains to device costs, including rejected candidates, verification, and rollback. The examined device studies primarily support memory maintenance or capability recovery, with limited joint validation of sustained capability gains and complete update costs. These gaps motivate controlled studies on edge devices, focusing on verification and acceptance under energy budgets, retention under bounded storage, and persistent safety to establish when self-improvement justifies its resource cost.