Zhenbo Yu

Yunhao Li, Zhenbo Yu, Yucheng Zhu et al.

Human motion synthesis is a long-standing problem with various applications in digital twins and the Metaverse. However, modern deep learning based motion synthesis approaches barely consider the physical plausibility of synthesized motions and consequently they usually produce unrealistic human motions. In order to solve this problem, we propose a system ``Skeleton2Humanoid'' which performs physics-oriented motion correction at test time by regularizing synthesized skeleton motions in a physics simulator. Concretely, our system consists of three sequential stages: (I) test time motion synthesis network adaptation, (II) skeleton to humanoid matching and (III) motion imitation based on reinforcement learning (RL). Stage I introduces a test time adaptation strategy, which improves the physical plausibility of synthesized human skeleton motions by optimizing skeleton joint locations. Stage II performs an analytical inverse kinematics strategy, which converts the optimized human skeleton motions to humanoid robot motions in a physics simulator, then the converted humanoid robot motions can be served as reference motions for the RL policy to imitate. Stage III introduces a curriculum residual force control policy, which drives the humanoid robot to mimic complex converted reference motions in accordance with the physical law. We verify our system on a typical human motion synthesis task, motion-in-betweening. Experiments on the challenging LaFAN1 dataset show our system can outperform prior methods significantly in terms of both physical plausibility and accuracy. Code will be released for research purposes at: https://github.com/michaelliyunhao/Skeleton2Humanoid

Jinxian Liu, Ye Chen, Bingbing Ni et al.

Humans have a strong intuitive understanding of physical processes such as fluid falling by just a glimpse of such a scene picture, i.e., quickly derived from our immersive visual experiences in memory. This work achieves such a photo-to-fluid-dynamics reconstruction functionality learned from unannotated videos, without any supervision of ground-truth fluid dynamics. In a nutshell, a differentiable Euler simulator modeled with a ConvNet-based pressure projection solver, is integrated with a volumetric renderer, supporting end-to-end/coherent differentiable dynamic simulation and rendering. By endowing each sampled point with a fluid volume value, we derive a NeRF-like differentiable renderer dedicated from fluid data; and thanks to this volume-augmented representation, fluid dynamics could be inversely inferred from the error signal between the rendered result and ground-truth video frame (i.e., inverse rendering). Experiments on our generated Fluid Fall datasets and DPI Dam Break dataset are conducted to demonstrate both effectiveness and generalization ability of our method.

Yuelin Hu, Zhenbo Yu, Zhengxue Cheng et al.

Hybrid post-training usually combines supervised fine-tuning and reinforcement learning, but fixed mixing schedules cannot adapt when the relative noise of the two signals changes over time. We propose GAC, a noise-aware controller that derives an adaptive mixing weight from online estimates of gradient variance and disagreement between the two training signals. The method adds smoothing, prior guidance, and bounded updates while reusing existing training tensors. Experiments on math, code, science, and logic benchmarks show that GAC consistently improves hybrid post-training over strong fixed and rule-based baselines, with larger gains at larger model scales and less than 1% training overhead.

Jian Jin, Zhenbo Yu, Yang Shen et al.

Customized text-to-image generation renders user-specified concepts into novel contexts based on textual prompts. Scaling the number of concepts in customized generation meets a broader demand for user creation, whereas existing methods face challenges with generation quality and computational efficiency. In this paper, we propose LaTexBlend, a novel framework for effectively and efficiently scaling multi-concept customized generation. The core idea of LaTexBlend is to represent single concepts and blend multiple concepts within a Latent Textual space, which is positioned after the text encoder and a linear projection. LaTexBlend customizes each concept individually, storing them in a concept bank with a compact representation of latent textual features that captures sufficient concept information to ensure high fidelity. At inference, concepts from the bank can be freely and seamlessly combined in the latent textual space, offering two key merits for multi-concept generation: 1) excellent scalability, and 2) significant reduction of denoising deviation, preserving coherent layouts. Extensive experiments demonstrate that LaTexBlend can flexibly integrate multiple customized concepts with harmonious structures and high subject fidelity, substantially outperforming baselines in both generation quality and computational efficiency. Our code will be publicly available.

Yuelin Hu, Zhenbo Yu, Zhengxue Cheng et al.

We formalize Rollout Informativeness under a Fixed Budget (RIFB) as the expected non-vanishing policy-gradient mass that a tool-use rollout set injects into Group Relative Policy Optimization (GRPO). We prove that any budget-agnostic independent sampler suffers a collapse rate bounded away from zero for hard prompts regardless of the budget. Motivated by this, we recast intermediate state selection as a monotone submodular maximization problem, where a greedy one-step selector enjoys a 1 minus 1/e approximation guarantee. Our Uncertainty-aware Upper Confidence Bound (UUCB) terms arise as closed-form marginal gains of this objective. This turns the token-level entropy bonus from an empirical trick into an analytic consequence of the formulation. We present InfoTree, a training-time tree-search framework coupling UUCB with a learned Adaptive Budget Allocator (ABA) and an asynchronous Speculative Expansion scheme. ABA rescues prompts whose initial tree is wasted on uniform outcomes, lifting the mixed-outcome ratio from 58.1 percent to 76.3 percent with less than 5 percent budget overhead. Speculative Expansion reduces wall-clock overhead from 14.3 percent to 4.8 percent by tolerating bounded staleness in UUCB scores. Across nine benchmarks spanning math reasoning (AIME 2024 and 2025, MATH-500, OlympiadBench, USAMO), web-search agents (GAIA, HLE-100, BrowseComp-lite), and tool-rich coding and OS agents (APPS-verified, AgentBench-OS), InfoTree outperforms flat GRPO, DeepSearch, Tree-GRPO, AT2PO, CW-GRPO, and RC-GRPO. Head-to-head compositions with Tree-GRPO prefix sharing and CW-GRPO contribution weights deliver further gains, confirming that our selector operates orthogonally to rollout reuse and trajectory re-weighting. A 5 by 5 by 5 robustness grid reveals that over three quarters of the hyperparameter space lies on a performance plateau, confirming UUCB robustness.

Yuelin Hu, Zhenbo Yu, Zhengxue Cheng et al.

Agent-repair leaderboards reorder under evaluator reconfiguration, and a measurable share of the reordering is produced by methods that consult evaluator-derived signal during internal selection of candidate repairs. We document this failure mode on a public leaderboard and release AuditRepairBench, a paired-execution trace corpus of 576,000 registered cells (96,000 executed) that operationalizes evaluator-channel-blocking ranking instability within a declared observability boundary. A modular screening architecture decides pathway-blocking through four interchangeable implementations, a learned influence proxy, a rule-based channel-exposure ratio that uses no trained model, a counterfactual sensitivity proxy, and a sparse human-audit proxy, combined into a screening posterior that feeds a cell-level flip functional, a set-valued label, a stratified system score, and a set-valued leaderboard. The resource is supported by mechanism-anchored validation on an 80-case source-level channel-surgery subset, an independent-discovery protocol under which two annotator groups separated from the pipeline developers discover coupling patterns blinded to the screening design and the frozen ensemble attains pooled AUROC 0.83 on their 79 cases, implementation robustness, uncertainty propagation that raises 95% coverage from 0.81 to 0.95, and forward transfer with pooled community-evaluator Spearman \r{ho} = 0.65. Screening-guided blinding patches reduce rank displacement by 55--74% (mean 62%) at fewer than 50 lines of code, whereas random channel blinding produces at most 7% reduction and generic retraining at most 13%. AuditRepairBench-Lite, a rule-only configuration on a 12,000-cell subset, preserves the leaderboard at Kendall τ = 0.88 under twenty-four GPU-hours and is the primary release artifact at 42 GB.

Yuelin Hu, Zhenbo Yu, Zhengxue Cheng et al.

Many continual-learning methods modify gradients upstream (e.g., projection, penalty rescaling, replay mixing) while treating Adam as a neutral backend. We show this composition has a hidden failure mode. In a high-overlap, non-adaptive 8-domain continual LM, all shared-routing projection baselines collapse close to vanilla forgetting (12.5--12.8 vs. 13.2). A 0.5% replay buffer is the strongest shared alternative but still reaches 11.6, while fixed-strength decoupling falls below vanilla at 14.1. Only adaptive decoupled routing remains stable at 9.4, improving over vanilla by 3.8 units. On a 16-domain stream, its gain over the strongest shared-routing projection baseline grows to 4.5--4.8 units. The failure is largely invisible on clean benchmarks. We explain this effect through Adam's second-moment pathway: in the tested regime, projection induces a 1/(1-alpha) inflation of the old-direction effective learning rate, matching measurements within 8% across eight alpha values. The same conflict appears with penalty methods, replay mixing, and at 7B scale under LoRA. Our fix routes the modified gradient only to the first moment while preserving magnitude-faithful second-moment statistics, with overlap-aware adaptive strength. This simple change is the only tested configuration that consistently avoids collapse across methods, optimizers, and scale.

Ji Yang, Xinxin Zuo, Sen Wang et al.

Given a single image of a general object such as a chair, could we also restore its articulated 3D shape similar to human modeling, so as to animate its plausible articulations and diverse motions? This is an interesting new question that may have numerous downstream augmented reality and virtual reality applications. Comparing with previous efforts on object manipulation, our work goes beyond 2D manipulation and rigid deformation, and involves articulated manipulation. To achieve this goal, we propose an automated approach to build such 3D generic objects from single images and embed articulated skeletons in them. Specifically, our framework starts by reconstructing the 3D object from an input image. Afterwards, to extract skeletons for generic 3D objects, we develop a novel skeleton prediction method with a multi-head structure for skeleton probability field estimation by utilizing the deep implicit functions. A dataset of generic 3D objects with ground-truth annotated skeletons is collected. Empirically our approach is demonstrated with satisfactory performance on public datasets as well as our in-house dataset; our results surpass those of the state-of-the-arts by a noticeable margin on both 3D reconstruction and skeleton prediction.