Original research at the intersection of adversarial AI, benchmark design, legal reasoning, neurosymbolic learning, and the limits of what current models can actually do.
Null State Suppression, Epistemic Freezing, and AI Trust Score Sabotage in Moderation Systems. Theorizes and empirically evidences a novel adversarial abuse tactic in modern AI-driven content moderation ecosystems: the use of mass reporting as an ontological weapon. Coordinated botnet-enabled adversaries weaponize engagement to poison classifier confidence at the point of content deployment, resulting in recursive trust score decay, velocity suppression, and long-term entity declassification.
AI coding agents are benchmarked as fully-autonomous systems — but real-world use is interactive. Users correct and reject agent outputs 44% of the time. Agents seek clarification 1–2% of the time. Dialogue-SWEBench closes that gap: 500 real SWE-Bench problems resolved entirely through multi-turn dialogue with a persona-grounded user simulator. Better coding models are not always better dialogue models.
Lawyer-client consultation is a critical starting point for legal services. DLawBench evaluates whether LLMs can conduct real legal consultation: eliciting facts, correcting client misframes, and writing defensible memos. Built from 461 real court opinions across Chinese and U.S. law, with four client personality types. The best-performing model achieves only 0.562 in consultation-grounded legal reasoning.
Neurosymbolic models integrate neural networks with symbolic reasoning for robust and interpretable AI. EM-NeSy recasts probabilistic NeSy learning as an instance of the Expectation-Maximization algorithm — unlocking the full potential of EM for NeSy learning with no differentiability requirements on the symbolic side.
Mechanistic analysis of how alignment training produces moral indifference rather than moral reasoning in large language models. Examines the gap between surface-level safety compliance and genuine ethical reasoning capability — and what that gap means for deployment in high-stakes domains.
AI models, once deployed, learn essentially nothing. Their mode of operation is fixed. Confronting the data wall: quality text data on the internet is finite. A model trained on all of it cannot exceed the frontier of what humans have already written — it can only recombine it. Scaling model size past this ceiling yields diminishing returns by definition, not by current technical limitation.
Industrial computer vision needs data before it can build trust, and trust before users will tolerate the imperfections that come with early data. GenAI promises to break that deadlock — but the domain gap between human-centric generative models and featureless industrial parts runs deeper than expected. The models know "rusty" as a descriptor of dogs, not generators.
Operational infrastructure — scaling frameworks for AI-native companies. Five standards for scaling without breaking: the frameworks that separate companies that survive hypergrowth from the ones that collapse under it. Built from first principles, not consultant decks.