Long-form explanations for readers who want the intuition quickly and the research boundary accurately. Each note links back to the primary paper, evidence locations, and citation exports.
TeleBoost explains why high-quality video alignment depends on staged supervised shaping, reward-driven reinforcement learning, preference refinement, diagnostics, and systems engineering working together.
A practical explanation of why sequence-level rewards are too coarse for visual generation, and how Visual Preference Policy Optimization turns perceptual structure into localized learning signals.
TaRoS examines how fixed reward components can become optimization targets rather than useful feedback, then reshapes video GRPO signals by component, group sparsity, and saturation.
OTCA reframes diffusion-model alignment as process credit assignment, decomposing an outcome reward over a denoising trajectory and allocating multiple objectives where they are most useful.
An intuitive and technical guide to Bayesian Prior-Guided Optimization, which treats reward confidence as part of visual policy optimization instead of trusting every group comparison equally.
Uni-Inter maps three previously separate 3D interaction settings into a shared volumetric language, then predicts probabilistic human motion while preserving each task’s entities and constraints.
Moderate Hard Example Modulation emphasizes difficult fine-grained samples while bounding their relative influence so that noise and memorized exceptions do not dominate training.
AntEval evaluates LLM-driven agents at the process level: whether information actually moves between participants and whether intent is expressed in a human-like way.
IcoCap combines image semantics with video features and selects captions that match the resulting visual content instead of preserving noisy original labels.
MAAL studies low-sample affordance learning by combining 3D object and robot modalities, retaining successful actions, and scoring candidates through reconstruction rather than appearance alone.
SEEG separates rhythmic and semantic cues, constrains generated motion with semantic information, and shows why gesture evaluation needs content-aware data and metrics.
Episodic Linear Probe repeatedly resets a simple classifier during training to distinguish what the classifier remembers from what the current representation makes easy to learn.
VrR-VG uses a counterfactual dataset test: remove image pixels, measure which relationships remain predictable from object labels and geometry, and refocus learning on visual evidence.