A mechanism-first reading of Visual Funnel, a training-free method showing that multimodal models need structured intermediate context—not just tighter crops—to read visual details correctly.
A practical reading of how recursive LLM agents converge, drift, or wander depending less on the model than on the loop we force it to run.
A mechanism-first reading of GPG, a Transformer-aware policy-gradient framework that turns output segments into trainable macro-actions for LLM agents.
A mechanism-first reading of ExaCraft, an AI education system that treats learner behavior—not just learner profiles—as the missing layer of personalized examples.
A mechanism-first reading of ImplicitRDP, showing why force-aware robot policies need causal structure, not just extra sensor channels.
A mechanism-first reading of why reinforcement learning for text-to-3D generation needs specialized rewards, token-level optimization, reasoning-heavy benchmarks, and coarse-to-fine training.
SceneMaker shows why open-set 3D scene generation needs separate priors for de-occlusion, geometry, and pose instead of forcing one pipeline to guess everything at once.
A mechanism-first reading of Suzume-chan shows why embodied RAG may matter less as a robot novelty and more as a practical interface for capturing, preserving, and replaying expert knowledge.
A mechanism-first reading of DaSH, a hierarchy-aware dataset selection method that treats data procurement as source diagnosis rather than sample hoarding.
A mechanism-first reading of how evidence scoring and selective multi-agent debate can make health misinformation detection more disciplined, cheaper, and less theatrically wrong.