Papers

Novel memory architectures for long-horizon agent tasks, enabling smooth retrieval and natural temporal dynamics.

CE2P (Counterexample-to-Prompt) translates formal verification failures into structured LLM feedback. Key finding: CE2P benefit is inversely correlated with model capability—weaker models gain 34-39pp while GPT-4o needs no structured feedback.

Evolutionary approaches to adversarial testing where attack and defense strategies improve through competitive pressure.

Formal framework for ROI-based decision routing in multi-tier verification systems.

Expected Q2 2026

Comprehensive threat model for AI agents covering input attacks, state attacks, tool attacks, planning attacks, and coordination attacks. Empirical evaluation across agent architectures.

Expected Q2 2026

Using interpretability methods to detect when agentic systems are under adversarial attack. Extends metacognitive probing to security applications.

Expected Q3 2026

Methods for uncertainty quantification that track actual accuracy. Activation-based estimation, propagation through multi-step reasoning, and calibration without ground truth.