SkillAudit shows how agent skills can be improved without hidden tests by comparing with-skill and without-skill executions, but only when correctness leaves an observable trace.
A practical reading of learner-based concept drift detection: when SPC, windowing, and ensemble methods help, when they disappoint, and why real streams refuse to behave like benchmark streams.
A systematic study of hierarchical VLA agents shows that robot reliability depends less on simply adding hierarchy than on how planning, control, memory, observation, and handoff are orchestrated.
ChemCoTBench-V2 shows why chemical AI evaluation has to inspect intermediate molecular and reaction states, not just final answers.
LongSpace shows why long-video AI needs spatial memory, not just larger context windows.
AgentRedBench shows that enterprise AI-agent risk is less about naughty chat prompts and more about untrusted SaaS content quietly steering authorized tool actions.
A mechanism-first reading of why EEG-to-music reconstruction improves when models preserve electrode-level structure before alignment and generation.
IterCAD shows why reliable CAD automation depends less on one-shot generation and more on closed-loop execution, visual feedback, and survivor-bias-free evaluation.
A mechanism-first reading of an LLM-orchestrated stroke-care conformance pipeline, and what it teaches operators about turning unstructured policy into auditable process checks.
Two recent papers show why serious claims about AI cognition require evidence of internal organization, not just fluent behavior or attractive labels.