The proposed presentation will address a critical challenge in modern AI: adapting foundation models to individual user needs efficiently. As generative AI systems with large numbers of parameters become ubiquitous, personalization is essential for practical deployment across diverse applications. This talk will be highly relevant to ICASSP attendees, bridging signal processing, machine learning, and efficient algorithm design. The technical content will cover parameter-efficient fine-tuning techniques, with particular focus on Low-Rank Adaptation (LoRA). I will explain how freezing foundation models and training only low-rank adapters enables cost-effective personalization. The presentation will detail LoRA implementation strategies, including for text, speech and image generation applications, demonstrating how these methods achieve state-of-the-art results with minimal computational overhead. Beyond single-task personalization, a key frontier is enabling models to handle multiple specialized tasks efficiently. The talk will introduce advanced adapter merging techniques that address this challenge by combining multiple task-specific LoRA adapters into unified models. I will present three complementary research contributions from our lab. 1) Compositional Multi-tasking (EMNLP'25), which merges adapters for complex operations like translated summarization in a single inference pass. 2) LoRA.rar (ICCV'25), which uses hypernetwork-based merging to combine subject and style adapters for image generation with 4000x speedup. 3) D2C (ICASSP'26), a data-driven clustering method that identifies suitable groupings of task-specific adapters using minimal number of examples. It then merges adapters within each cluster to create compact multi-task adapters deployable on resource-constrained devices. Such research direction is motivated by the practical limitations of current foundation models: full fine-tuning is computationally prohibitive, while "one-size-fits-all" models fail to capture individual user preferences and task-specific requirements. The need for efficient, personalized AI is particularly acute in resource-constrained environments where memory and compute are scarce, yet users demand sophisticated multi-task capabilities. This talk will inspire the signal processing community by showcasing how efficient parameter adaptation techniques can democratize access to powerful AI systems. Attendees will gain practical insights into LoRA-based personalization, understanding adapter merging strategies, and applying these methods to real-world applications. The presentation will balance theoretical foundations with empirical results, providing both researchers and practitioners with actionable knowledge for advancing personalized generative AI.

Large language models (LLMs) in the telecommunications domain are accelerating the digital transformation of emerging industries like low-altitude economies and IoT. However, AIGC (AI-Generated Content) faces critical safety risks—such as value misalignment, privacy leakage, and prompt injection. The unique business environment of the communication industry requires security solutions to operate with extremely low resource overhead, creating an urgent need for lightweight, low-cost security approaches. Our solution: We introduce AIGuardrail, a zero-training, plug-and-play LLM safety guardrail derived from industrial deployment practices. By pioneering a "Security-as-a-Skill" paradigm, leveraging agent skills and prompt engineering, it establishes a skill-driven automated lifecycle for security compliance, drastically reducing reliance on expert personnel and intricate coding. Key Innovations & Methodology AIGuardrail integrates non-intrusively into AIGC processing pipelines, augmenting security without disrupting core business logic. By embedding structured safety guidelines and few-shot examples into system prompts, it enables low-latency first-token judgments for inference-time safety checks. Harnessing the natural language orchestration of Agent Skills, complex compliance requirements are encapsulated as modular "Skills" spanning prompt authoring, optimization, testing, and deployment. Business users simply describe emerging risks in natural language, triggering the skill engine to auto-update safeguards for "hot-swapping" policies. The safety adjudication of user inputs hinges on five pivotal elements: (1) Role assignment: explicitly designate the model as a safety auditor and constrain it to perform only safety-review tasks. (2) Safety guidelines: the guidelines cover known unsafe content in current AIGC applications (e.g., illegal content, IP infringement, privacy leakage, injection attacks) and instruct the LLM to identify potentially non-compliant queries or malicious attack intent in user inputs. For emerging threats or product-specific policies, AIGuardrail supports dynamic, in-runtime insertion and updates of safety guidelines. (3) Global principle: To address challenges such as moderation in low-resource languages, code-injection risks, and false positives triggered by English abbreviations, AIGuardrail adopts a globally scoped prompting scheme that enables the model to construct a consistent safety decision criterion prior to inference. For example, by composing structured prompts with few-shot examples, we elicit the model’s inherent multilingual capability and enable safety screening for inputs written in low-resource languages. (4) Moderation principle: AIGuardrail employs a chain-of-thought (CoT)–based hierarchical moderation mechanism，the moderation flow follows the priorities below. Block, Allow, Review. (5) Output format: AIGuardrail adopts a first-token safety-moderation policy. The system completes compliance determination early in decoding, significantly reducing the latency and computational overhead induced by deep reasoning, making it well-suited for high-throughput, low-latency industrial deployments. Experimental Results and Industrial Impact: AIGuardrail pioneers an agent-skill-driven paradigm for LLM security, automating a closed-loop from compliance specification to defensive execution by externalizing safeguards onto the model's inference. Deployed across 20+ production systems (including Qwen and DeepSeek variants) for six months, our solution comprehensively outperforms the current SOTA—Qwen3Guard—in core metrics: the overall detection rate reaches 90.2% (8-12% higher), with a False Positive Rate (FPR) of only 0.08% (∼10% lower)