Our research focuses on the foundational science of trust decisions-verification, adversarial robustness, interpretability, and formal guarantees for AI systems.
Memory systems for AI agents modeled as continuous fields rather than discrete key-value stores. Enables smooth interpolation, natural decay, and principled attention over long contexts.
A framework for verified code generation using the CE2P hypothesis: Confidence, Error rate, Effort, and Progress are sufficient signals to verify correctness without ground truth.
Evolutionary algorithms for adversarial testing of trust systems. Systems must constantly evolve to survive against co-evolving adversaries.
Formal mathematical framework for ROI-based decision routing in multi-tier trust systems. Proves convergence of self-learning rules and optimality of cascade architectures.
Using sparse autoencoders to discover interpretable "trust circuits" in large language models. Mechanistic interpretability applied to AI safety and sandbagging detection.
Graph neural networks for modeling trust propagation through entity networks. Applications to KYC/AML in financial services.
Security implications of persistent memory in AI agents. Attack vectors include memory poisoning, extraction, and manipulation.
Formal verification of trust properties in multi-agent systems.
Using world models to simulate adversarial scenarios for trust testing.
Inference optimization for trust decision systems. Cascade-aware batching, confidence-based early exit.
Type to search across all pages and posts
Press ↑ ↓ to navigate, Enter to select