Browse publications by year. Each paper page includes the abstract, a concise overview, related-work context, sources, and citation downloads.
This roadmap places World Action Models inside a broader physical-intelligence stack: an embodied brain compares possible interventions, a physical harness grounds its requests through tools and controllers, shared contracts connect heterogeneous components, and verified interaction becomes post-training experience.
TeleBoost treats video post-training as a staged, stability-constrained system that connects supervised policy shaping, reward-driven reinforcement learning, and preference-based refinement while diagnosing costly rollouts, compounding temporal failures, and uncertain feedback.
This survey organizes reinforcement learning for image, video, and 3D/4D generation, treating RL not only as a fine-tuning algorithm but as an interface for optimizing non-differentiable, preference-driven, temporal, and high-level objectives.
ViPO turns one scalar reward per generated sample into spatially and temporally structured, pixel-level advantage maps, using pretrained vision features to focus GRPO updates on perceptually important regions while retaining the standard training pipeline.
RATS combines trajectory distillation with preference feedback: horizon matching aligns teacher and student at key denoising stages, while a reward-aware gate strengthens teacher guidance only when the teacher is better under the chosen reward.
TaRoS v4 addresses reward hacking and saturation in video GRPO by organizing multi-aspect feedback with component-level performance assessment and intra-group sparsity, then adaptively downweighting components whose scores have saturated.
OTCA replaces uniform, scalar reward propagation in visual GRPO with structured credit assignment along two axes: which denoising steps matter and which reward objectives should matter at each point in the trajectory.
BPGO treats visual-reward scores as uncertain rather than equally trustworthy: a semantic prior anchors inter-group trust allocation and intra-group renormalization so that GRPO emphasizes confident feedback and suppresses ambiguous signals.
LaxMotion removes direct 3D pose supervision and instead learns 3D motion as a structurally consistent explanation of global trajectories and monocular 2D kinematic cues, supported by view, orientation, and stability regularization.
TeleWorld closes a loop between video generation and 4D reconstruction: generated streams update a persistent spatiotemporal representation that guides later generation, while Macro-from-Micro Planning and distillation target long horizons and real-time latency.
Uni-Inter represents human–human, human–object, and human–scene interactions in one Unified Interactive Volume and predicts motion probabilistically joint by joint, enabling one task-agnostic model to reason over heterogeneous and compound interaction contexts.
InterSyn learns solo and multi-person dynamics together rather than treating them as separate tasks: INS synthesizes both from a first-person interaction perspective, and REC refines relative coordination so characters move synchronously.
VAST is a two-stage Video-As-Storyboard-from-Text framework: StoryForge converts text into storyboards containing human poses and object layouts, and VisionForge turns those structural plans into videos.
MHEM means Moderate Hard Example Modulation: it defines conditions for a loss that emphasizes informative hard samples without continually increasing the influence of extreme cases that a fine-grained classifier may simply memorize.
AntEval evaluates social interaction competency in LLM-driven agents through a multi-agent interaction framework and two quantitative metrics: Information Exchanging Precision (IEP) and Interaction Expressiveness Gap (IEG).
IcoCap changes the content density seen by a video captioner: Image-Video Compounding Strategy (ICS) injects concise image semantics into video samples, while Visual-Semantic Guided Captioning (VGC) adapts caption supervision to the resulting ambiguous compound content.
FreeLong extends pretrained short-video diffusion models without additional training by blending low-frequency global features with high-frequency local-subsequence features during denoising, targeting global consistency and local spatiotemporal detail at once.
MAAL learns affordances for 3D articulated objects with a one-stage autoencoder-based pipeline and a MultiModal Energized Encoder that jointly models object geometry, robot actions, and their interactions while requiring only a small number of positive samples.
SEEG separates beat-related and semantic information with a DEcoupled Mining module, then uses a Semantic Energizing Module and semantic prompter to make generated co-speech gestures express semantics as well as align with speech.
ELP brings linear probing into training: an episodically reinitialized classifier learns on detached features to measure current discriminability, and ELP-SR uses the discrepancy between that probe and the main classifier to adaptively regularize samples.
The paper jointly learns food categories and ingredients with an Attention Fusion Network that emphasizes discriminative regions and a Food-Ingredient Joint Learning module using balance focal loss to address ingredient imbalance.
The method uses a complementary cascaded network with both raindrop→streak and streak→raindrop branches, fuses their outputs with attention, searches deraining blocks with neural architecture search, and introduces the real-world RainDS dataset.
VrR-VG prunes relationships that can be predicted from non-visual statistics, builds a visually relevant scene-graph dataset from Visual Genome, and learns representations that jointly encode instances, attributes, and relationships.