Original research at the intersection of adversarial AI, benchmark design, legal reasoning, neurosymbolic learning, and the limits of what current models can actually do.
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.
Five open-access LLMs hallucinated dangerous drug recommendations 90–99% of the time when asked questions where the historically correct answer is now a banned or withdrawn drug. A five-agent adversarial auditing loop dropped that rate below 5%.
LLMs are undergoing a fundamental transformation from conversational generators into integrated AI systems capable of reasoning, action, memory, and self-improvement. This survey conceptualizes the transition through the Workspace + Skill paradigm.
148 GitHub projects, 15,549 agentic pull requests, one question: does writing an AGENTS.md or CLAUDE.md file actually make Copilot, Claude, and Cursor produce better pull requests? The answer is not a clean yes. 27.7% of projects saw improvement the rest didn't.
Personalized health AI faces a cold-start problem: models need weeks of behavioral data before they can tell constitutional variation from environmentally driven deviation. This framework proposes an exogenous genetic anchor to solve it from day one.
The alignment demands on AI models are increasingly complex. There are familiar trade-offs between helpfulness and safety and within safety itself, independently desirable goals that pull in opposite directions. Virtue-maximizing AI systems may be the most dangerous of all.
Most adversarial-attack research assumes a hand on the input pixels. This work reaches further upstream into the physical act of capture itself. Sound, light, and environmental manipulation as attack vectors against AI perception systems.
Movement sonification turning how you walk into real-time sound has strong experimental backing for stroke rehabilitation, but almost no clinical adoption. This three-phase design framework explains why, and fixes it.
Solves the core architectural bottleneck of LLM knowledge injection. RAG keeps knowledge outside the model shallow, prompt-level, retrieval-bound. Post-training writes into shared parameters deep but destructive. DMoE routes updates to dedicated expert subnetworks.
Text-to-image models can still produce unsafe content, and existing concept-erasing methods over-correct suppressing benign concepts that share a prompt with a harmful one. ForceForget erases precisely, without collateral damage.
Generative AI, LLMs, and agentic systems constitute the most disruptive transformation in the history of software engineering reshaping development processes, required competencies, and professional roles from the ground up.
A novel anomaly detection framework built on Hyperdimensional Computing a brain-inspired paradigm encoding information as high-dimensional binary vectors. Unlike prior HDC approaches relying on reconstruction error, D2H-AD is discriminative and significantly faster at inference.
Fashion sketching lets designers visualise a concept long before any fabric is cut, yet sketch-based fashion image synthesis has stalled for want of large-scale, high-quality paired data. GarmentSketch closes that gap with a domain-bridging generation pipeline.
As generative AI reshapes higher education, institutions require students to declare its use. But binary declarations a single checkbox can't capture how these tools are actually used. This framework builds the taxonomy for declarations that actually mean something.
Continuous physiological monitoring throws off a number every few seconds, but a number means nothing without a reference. Population norms can say whether a reading is typical for a group; they can't say whether it's typical for you. A genomic prior solves that.
Most CAD-generation systems work the way a junior drafter would if forbidden from ever looking back at their own work: one shot, no revision. Professional CAD practice runs the opposite. IterCAD brings iterative, self-correcting refinement to automated 3D design generation.
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.