Language Is a Security Surface for Agent Guards

New research challenges the superposition theory of polysemanticity, suggesting neural networks are more interpretable than previously believed through recursive generalization patterns.

Researchers propose MirrorMoE to address the exploration dilemma in modern Mixture-of-Experts architectures between fine-grained segmentation and shared experts.

Researchers introduce OmniVerifier-C1, a decomposed graded verifier that provides continuous scoring instead of binary judgments for multimodal verification tasks.

Researchers introduce T2-z, a method for test-time optimization of environment latents in robotic world models that improves prediction accuracy and action success rates across simulations and real-world tasks.

Researchers introduce FDE-Gym, a benchmark simulating real-world AI delivery tasks where agents must clarify ambiguous requirements and discover unstated constraints in customer environments.