dtControl 2+ε shows how model checking can turn bounded performance loss into substantially simpler controller representations without leaving approximation quality to heuristic judgment.
A fraud pipeline study shows why graph signals, anomaly detection, and LLM investigators should earn narrowly defined decision roles rather than automatic authority.
A repaired information-flow semantics shows how implementation-level checks can provide machine-checked evidence for higher-level confidentiality policies without collapsing policy meaning into enforcement.
ProtoBlend shows that video dataset distillation can avoid repeatedly optimizing stored synthetic videos by selecting, allocating, and blending teacher-scored clips directly.
CogVis shows how multi-query change detection can reuse scene-level temporal evidence while keeping semantic calibration and region verification specific to each query.
MeRLa suggests that RLHF teams can improve post-training not only by changing reward models or optimizers, but by learning a reusable layer that makes existing rewards more informative.
Prox shows that sparse LLM inference can use cheap approximation to decide what deserves exact computation, rather than approximating the final values themselves.
AutoIntervene shows how bidirectional human-robot handoff can turn deployment failures into targeted training data while reducing operator-control requirements.
HyGRL shows how preserving source text inside a knowledge graph and learning a lightweight router can improve multi-hop retrieval without putting an LLM in every search step.
Dependency-guided rollback reframes agent memory incidents as state-recovery problems, with selective repair offering higher benchmark recovery at a measurable cost-versus-recurrence trade-off.