Local & private deployment
Design model adaptation and operating patterns that keep sensitive workloads closer to their owners.
Research programme · 03
Model Security: Local LLMs & Trustworthy Agents
How can locally deployed models and agents be evaluated, aligned, and controlled before they act?
Safety for LLMs and agentic systems, spanning local and private deployment, jailbreak robustness, alignment, and repeatable evaluation harnesses.
6 selected research outputs
Research focus
Three questions shape this programme.
This is the working agenda: a precise statement of the problems the lab is building toward, not a claim that every component is already solved.
Robust alignment
Stress-test preferences, prompts, and safety policies under adversarial or noisy feedback.
Agent evaluation harnesses
Build repeatable tests around tools, memory, plans, and failure recovery before agentic systems are deployed.
Selected publications & outcomes
Evidence behind the direction.
Peer-reviewed papers and clearly labelled arXiv preprints are shown with their source records. See the complete Google Scholar profile ↗ for the full publication history.
AutoRAS: Learning Robust Agentic Systems with Primitive Representations
Develops robust agentic systems around primitive representations.
Primary source CodeSafety at Scale: A Comprehensive Survey of Large Model and Agent Safety
Organises the threat models, methods, and open questions across large-model and agent safety.
Primary source arXivShort-length Adversarial Training Helps LLMs Defend Long-length Jailbreak Attacks: Theoretical and Empirical Evidence
Provides theoretical and empirical evidence on adversarial training for long-length jailbreak defence.
Primary source CodeAdversarial Preference Learning for Robust LLM Alignment
Uses adversarial preference learning to strengthen LLM alignment.
Primary sourceAn LLM Can Fool Itself: A Prompt-Based Adversarial Attack
Demonstrates a prompt-based adversarial attack in which an LLM can be induced to fool itself.
Primary source CodePerplexity-aware Correction for Robust Alignment with Noisy Preferences
Uses perplexity-aware correction to improve alignment when preference data are noisy.
Primary source CodeThe full agenda
Explore all four research layers.
Students · collaborators · partners
Help move this programme from a hard question to reliable practice.
RISE-AI welcomes people who value rigorous foundations, reproducible systems, and generous teamwork.
See openings and how to apply →