Forma shows that synthetic clinical cases become more controllable when the relationships that must survive generation are specified explicitly and audited after generation.
HiMA-MDD shows that evidence permissions, decision ownership, bounded revision, and provenance can matter more than simply adding specialist agents.
CamoDocs shows why RAG poisoning defenses must look beyond query overlap and compact embedding clusters—and why aggressive evidence removal can undermine the retrieval system they protect.
ReCAST shows how explicit de-obfuscation supervision can improve SMS risk detection without adding a restoration model to every production request.
MACE shows that event-linking accuracy can improve by fixing the evidence used to build candidate sets before replacing or retraining the downstream linker.
A benchmark of output tampering shows why internal anomaly signals, visible model failure, and accurate self-diagnosis should be treated as separate production metrics.
A causal foundation model can make repeated partial-identification queries fast without pretending observational data determine one causal answer.
A child–language-model comparison suggests that data efficiency depends not only on how much input a learner receives, but on whether later experience becomes progressively more valuable.
A signer-independent test can expose hidden robustness losses in sign-language translation, but it can still reward memorisation when target sentences repeat across splits.
A theoretical result separates the modest representation cost of adding languages from the training-budget and evaluation choices that can make multilingual quality deteriorate in practice.