Yuanzhi Liang — Research updateshttps://akira-l.github.io/feed.xml2026-07-31T00:00:00+08:00Yuanzhi LiangFrom World Action Models to Embodied Brains: A Roadmap for Open-World Physical Intelligencehttps://akira-l.github.io/publications/embodied-brains/2026-07-31T00:00:00+08:002026-07-13T00:00:00ZYuanzhi LiangThis 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: A Systematic Alignment Framework for High-Fidelity, Controllable, and Robust Video Generationhttps://akira-l.github.io/publications/teleboost/2026-07-31T00:00:00+08:002026-02-07T00:00:00ZYuanzhi LiangTeleBoost 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.Integrating reinforcement learning with visual generative models: foundations and advanceshttps://akira-l.github.io/publications/rl-vgm/2026-07-31T00:00:00+08:002026-01-29T00:00:00ZYuanzhi LiangThis 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.Seeing What Matters: Visual Preference Policy Optimization for Visual Generationhttps://akira-l.github.io/publications/vipo/2026-07-31T00:00:00+08:002026-01-01T00:00:00ZYuanzhi LiangViPO 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.Reward-Aware Trajectory Shaping for Few-step Visual Generationhttps://akira-l.github.io/publications/rats/2026-07-31T00:00:00+08:002026-01-01T00:00:00ZYuanzhi LiangRATS 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.Rethinking Reward Signals in Video GRPO: When Scores Become Targetshttps://akira-l.github.io/publications/taros/2026-07-31T00:00:00+08:002026-01-01T00:00:00ZYuanzhi LiangTaRoS 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.Learning to Credit the Right Steps: Objective-aware Process Optimization for Visual Generationhttps://akira-l.github.io/publications/otca/2026-07-31T00:00:00+08:002026-01-01T00:00:00ZYuanzhi LiangOTCA 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.Learning What to Trust: Bayesian Prior-Guided Optimization for Visual Generationhttps://akira-l.github.io/publications/bpgo/2026-07-31T00:00:00+08:002026-01-01T00:00:00ZYuanzhi LiangBPGO 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: Rethinking Supervision Granularity for 3D Human Motion Generationhttps://akira-l.github.io/publications/laxmotion/2026-07-31T00:00:00+08:002026-01-01T00:00:00ZYuanzhi LiangLaxMotion 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: Towards Dynamic Multimodal Synthesis with a 4D World Modelhttps://akira-l.github.io/publications/teleworld/2026-07-31T00:00:00+08:002025-12-31T00:00:00ZYuanzhi LiangTeleWorld 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: Unifying 3D Human Motion Synthesis Across Diverse Interaction Contextshttps://akira-l.github.io/publications/uni-inter/2026-07-31T00:00:00+08:002025-12-14T00:00:00ZYuanzhi LiangUni-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: Interleaved Learning for Dynamic Motion Synthesis in the Wildhttps://akira-l.github.io/publications/intersyn/2026-07-31T00:00:00+08:002025-01-01T00:00:00ZYuanzhi LiangInterSyn 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 1.0: A Unified Framework for Controllable and Consistent Video Generationhttps://akira-l.github.io/publications/vast/2026-07-31T00:00:00+08:002024-12-21T00:00:00ZYuanzhi LiangVAST 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.Penalizing the Hard Example But Not Too Much: A Strong Baseline for Fine-Grained Visual Classificationhttps://akira-l.github.io/publications/mhem/2026-07-31T00:00:00+08:002024-05-01T00:00:00ZYuanzhi LiangMHEM 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: Evaluation of Social Interaction Competencies in LLM-Driven Agentshttps://akira-l.github.io/publications/anteval/2026-07-31T00:00:00+08:002024-01-12T00:00:00ZYuanzhi LiangAntEval 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: Improving Video Captioning by Compounding Imageshttps://akira-l.github.io/publications/icocap/2026-07-31T00:00:00+08:002024-01-01T00:00:00ZYuanzhi LiangIcoCap 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: Training-Free Long Video Generation with SpectralBlend Temporal Attentionhttps://akira-l.github.io/publications/freelong/2026-07-31T00:00:00+08:002024-01-01T00:00:00ZYuanzhi LiangFreeLong 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: Multimodality-Aware Autoencoder-based Affordance Learning for 3D Articulated Objectshttps://akira-l.github.io/publications/maal/2026-07-31T00:00:00+08:002023-01-01T00:00:00ZYuanzhi LiangMAAL 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: Semantic Energized Co-speech Gesture Generationhttps://akira-l.github.io/publications/seeg/2026-07-31T00:00:00+08:002022-01-01T00:00:00ZYuanzhi LiangSEEG 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.A Simple Episodic Linear Probe Improves Visual Recognition in the Wildhttps://akira-l.github.io/publications/elp/2026-07-31T00:00:00+08:002022-01-01T00:00:00ZYuanzhi LiangELP 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.