arXiv subjects

Hao Wu

Publications and source records attributed to Hao Wu.

At least 73 records · Page 4Linked to original sources

arXiv2026

FutureWorlds: Learning Robotic World Models from Alternative Futures

Robotic world models predict action-conditioned future scenes, providing a foundation for understanding action outcomes. However, turning alternative predictions into useful learning signals remains challenging: similar candidates limit informative quality comparisons, while diverging trajectories require persistent maintenance of their individual histories. We introduce FutureWorlds, a framework that unifies candidate construction, history maintenance, and learning from relative quality. Built on a multimodal discrete autoregressive model, FutureWorlds uses diverse beam search during reinforcement learning to construct candidate futures that balance confidence and diversity. Candidate-specific bounded memory preserves scene states and ensures that generation and policy scoring use matching histories. We further propose MemSPO (Memory-Conditioned Search-Guided Policy Optimization), which converts video trajectory rewards into group-relative advantages to optimize the world model. On RT-1, BridgeV2, and RoboCasa, FutureWorlds reduces LPIPS for 32-frame predictions by 14.78%, 20.84%, and 9.12%, respectively, relative to the strongest baseline on each dataset. Under fixed evaluation configurations, only 200 MemSPO updates further improve generation quality and support continued prediction beyond the training horizon. Memory ablations, decoding sensitivity analysis, and optical-flow evaluation show that these gains extend beyond visual quality to more accurate motion prediction and more consistent object states. Project page and code: https://github.com/Alexander-wu/FutureWorlds.

cs.CV↗

arXiv2026

MWOP: Modality-aware Width-wise Operation Pruning for Efficient MLLMs

Multimodal large language models (MLLMs) incur substantial inference costs when processing long visual-textual sequences. While existing operation compression methods exploit modality-level redundancy, they largely treat computation within attention heads and shared feed-forward network (FFN) channels as unified units, leaving finer-grained redundancy underexplored. We find that redundancy varies both across modality-interaction paths within the same attention head and across visual and textual executions of the same FFN channel. Based on these findings, we propose Modality-aware Width-wise Operation Pruning (MWOP), which independently prunes visual-to-visual (V2V), text-to-visual (T2V), and text-to-text (T2T) attention paths within each layer, and separately selects FFN channels for visual and textual inputs. A first-order Taylor criterion guides the pruning process, with FFN importance re-evaluated after attention pruning and LoRA-based recovery training. To translate the resulting fine-grained sparsity into practical acceleration, we further develop path-sparse Triton attention kernels and compact visual-side FFN execution. MWOP preserves the token sequence while reducing attention and FFN computation, making it complementary to token compression and enabling simultaneous reduction of sequence length and per-token computation. On LLaVA-OneVision-7B, MWOP alone achieves a $1.6\times$ prefill speedup with 99.7\% average performance retention across 12 benchmarks. Combined with two representative token compression methods, it further increases their prefill speedups from $2.0\times$ and $1.9\times$ to $2.9\times$ and $2.7\times$, respectively. Results on Qwen2.5-VL-7B further demonstrate its applicability across architectures. The code is available at https://github.com/EIT-NLP/MWOP.

cs.CV↗

arXiv2026

Scalable Algorithm for Dynamic Quasi-clique Detection

Identifying dense subgraphs known as quasi-cliques is pivotal in numerous graph mining tasks across domains such as social networks, biology, and e-commerce. While prior work has developed efficient algorithms for quasi-clique detection in static graphs, real-world networks are inherently dynamic, where edges appear and disappear continuously. This renders static methods inefficient and ill-suited for real-time analysis. In this paper, we initiate the study of the Dynamic Maximum Quasi-Clique Problem (DMQCP), which aims to maintain and update the largest quasi-clique in a graph under streaming graph updates. We propose DMI, a novel MinHash-based dynamic framework that supports fast, high-quality approximate maintenance of quasi-cliques. DMI leverages two update-efficient hashing schemes, i.e., $l$-buffered $k$-MinHash and Bottom-$k$ MinHash, to maintain candidate quasi-cliques incrementally. To ensure robustness and reduce bias, we further design a batch reconstruction strategy to periodically rebuild the candidate set, guaranteeing both stability and adaptability under frequent updates. Extensive experiments on real-world and synthetic datasets show that DMI achieves up to four orders of magnitude speedup over static baselines, while preserving solution quality. As a side product, we also propose a framework NSF that primarily uses the neighbor-search technique to maintain quasi-clique candidates while edge updating. This work establishes the first efficient algorithmic framework for dynamic quasi-clique extraction, enabling scalable and real-time dense subgraph mining in evolving networks.

cs.SI↗

arXiv2026

UMI-WM: Interactive World Modeling for Bimanual Manipulation from Wrist Cameras

Universal Manipulation Interface (UMI) pairs wrist-camera videos with camera trajectories and gripper openings, providing scalable, robot-free data for learning how bimanual actions change a shared scene. Existing interactive world models, however, typically rely on text or single-camera control and struggle to follow each arm's actions while maintaining consistency between the two wrist views. We introduce UMI-WM, an action-conditioned bimanual world model that jointly generates both wrist views from observation history, per-arm relative camera poses, and gripper widths, with cross-view consistency, action following, and visual fidelity. UMI-WM injects each arm's actions into its own wrist stream, while sparse hub tokens exchange scene information between the two streams. Camera-aware geometry conditioning and camera-guided noise further support per-view motion control and cross-view geometric exchange. We adapt the model for block-causal generation and distill the 30-step Teacher into a Student with four denoising steps. We train UMI-WM on self-collected open-world and in-studio UMI recordings combined with public manipulation data. To evaluate action following, we use an inverse dynamics model (IDM) to recover relative camera poses and gripper widths from generated videos, and introduce Bimanual Action Fidelity (BAF) to measure how closely these recovered actions match the conditioning actions. UMI-WM achieves the best results on all five reference-video fidelity metrics and the best or competitive performance on VBench and WorldArena. The Teacher and Student outperform all compared baselines in action following, while the four-step Student generates paired wrist videos 6.9 times faster than the Teacher.

cs.RO↗

arXiv2026

RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models

Existing robot video world models are typically trained with low-level objectives such as reconstruction and perceptual similarity, which are poorly aligned with the capabilities that matter most for robot decision making, including instruction following, manipulation success, and physical plausibility. They also suffer from error accumulation in long-horizon autoregressive prediction. We present RoboAlign-R1, a framework that combines reward-aligned post-training with stabilized long-horizon inference for robot video world models. We construct RobotWorldBench, a benchmark of 10,000 annotated video-instruction pairs collected from four robot data sources, and train a multimodal teacher judge, RoboAlign-Judge, to provide fine-grained six-dimensional evaluation of generated videos. We then distill the teacher into a lightweight student reward model for efficient reinforcement-learning-based post-training. To reduce long-horizon rollout drift, we further introduce Sliding Window Re-encoding (SWR), a training-free inference strategy that periodically refreshes the generation context. Under our in-domain evaluation protocol, RoboAlign-R1 improves the aggregate six-dimension score by 10.1% over the strongest baseline, including gains of 7.5% on Manipulation Accuracy and 4.6% on Instruction Following; these ranking improvements are further supported by an external VLM-based cross-check and a blinded human study. Meanwhile, SWR improves long-horizon prediction quality with only about 1% additional latency, yielding a 2.8% gain in SSIM and a 9.8% reduction in LPIPS. Together, these results show that reward-aligned post-training and stabilized long-horizon decoding improve task consistency, physical realism, and long-horizon prediction quality in robot video world models.

cs.RO↗

arXiv2026

Retraction-Based Gradient Projection Algorithms on Manifolds

We introduce a framework for retraction-based convex optimization on Riemannian manifolds, which includes a notion of retraction-specific convex sets and retraction-based gradient projection algorithms. The standard theory of gradient projection algorithms generalizes easily to this framework. Within this framework, we establish convergence results for retraction-based gradient projection algorithms with various stepsize rules. As an application, we use our framework to study the weighted low-rank approximation. We also provide numerical validation of our convergence results on the image completion task.

math.OC↗

arXiv2026

TileBench: A Controlled Benchmark for Performance Evaluation and Bottleneck Diagnosis of Tile-Based Programming Models

Tile-based programming models, such as Triton and cuTile, aim to simplify high-performance kernel development, but their practical performance, tuning behavior, and usability remain difficult to compare systematically. We present TileBench, a controlled benchmark for evaluating Triton and cuTile on NVIDIA B200 GPUs under matched operator semantics and comparable implementation structures. TileBench contains 45 operators covering diverse AI-kernel patterns and memory/computation behaviors. Each task provides a PyTorch reference, verified Triton and cuTile implementations, standardized data-types (dtype) and input-size sweeps, default and autotuned configurations, roofline-based metrics, and profiling-guided diagnosis. Our evaluation shows that performance gaps are workload-dependent: cuTile excels on a small cluster of Tensor-Core/TMA-friendly kernels, while Triton is stronger on many irregular, streaming, and bandwidth-bound operators. We further evaluate LLM-generated cuTile and Triton kernels and find that Triton is consistently more token-efficient than cuTile under the same iterative refinement protocol. TileBench is publicly available at https://github.com/Deep-Learning-Profiling-Tools/Tilebench.

cs.PF↗

arXiv2026

EBRL: Asynchronous Embodied RL by Multi-Grained Resource Management

Embodied reinforcement learning (RL) improves model capabilities with a pipeline of environment simulation, action generation, and model updates. These stages show heterogeneous CPU and GPU demands, making efficient resource utilization difficult. Recent systems overlap rollout (simulation and generation) with training for efficiency, but exclusive GPU allocation and synchronized barrier in rollout still leave substantial hardware resource waste. In this paper, we present EBRL, an asynchronous embodied RL training system with two core techniques. The asynchronous pipelined scheduler overlaps rollout and training, pipelines simulation and generation across environment groups, and carries out each environment independently, eliminating synchronization stalls. The fine-grained resource manager pools CPU cores and GPU streaming multiprocessors, and uses stage profiles and runtime feedback to adjust resource quotas and batch sizes to meet the shifting demands among stages. We implement EBRL on RLinf and evaluate it with four embodied policies and four simulation benchmarks across heterogeneous GPU testbeds. Experiments show that EBRL achieves 1.30-3.47 times the end-to-end rollout throughput and 2.5 times of training convergency compared to the SOTA embodied RL systems.

cs.LG↗

arXiv2026

The Naito--Okuda--Ouyang Biconcave Red Blood Cell Profile as a Finite-Energy Unbranched Weak Immersion with Point-Force Residues

The axisymmetric profile $\sinψ(ρ)=c_0ρ\log(ρ/ρ_B)$ introduced by Naito, Okuda, and Ouyang is an exact solution of the unconstrained Helfrich shape equation for every $ρ>0$, but its mean curvature diverges logarithmically at the dimple center. A covariant local analysis shows that in Cartesian graph coordinates the immersion is nondegenerate and belongs to $W^{2,p}$ for every finite $p$; it is $C^{1,α}$ for every $α<1$, but neither $C^{1,1}$ nor $C^2$. The conformal factor has a finite nonzero limit, so the point has branch multiplicity one and is not a branch point in the sense used in weak-immersion compactness theory. The Helfrich energy is finite, whereas the complete Euler--Lagrange operator carries the distributional residue $\E_{\CH}=4πc_0δ_p$ under the convention specified below. Hence the profile is a classical solution on the punctured surface and a finite-energy weak immersion, but it is not a free weak critical point of the unforced Canham--Helfrich functional. Stationarity is restored after adding the corresponding point-force potential or under a pinned-point constraint. Area and volume multipliers cannot cancel the Dirac residue. The apparent sign discrepancy between the shape equation used in the original papers and the modern convention is explained by the opposite definition of mean curvature. These results delimit precisely what can and cannot be inferred from modern existence and regularity theorems for Canham--Helfrich minimizers.

math.AP↗

arXiv2026

Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation

Generalizable embodied manipulation remains difficult to achieve through pretraining alone, due to unseen physical conditions in the real world. We argue that robots need to learn from their own physical interactions on the fly during real-world deployment and use this knowledge to inform subsequent actions. We present Zeva, the first framework that enables in-context learning from a robot's own physical interaction experience while keeping the policy model frozen. Zeva employs a Causal Interaction Extractor to encode an executed action and its induced state change into a causal interaction signal, which is stored in a dual-timescale causal memory. For subsequent actions, relevant causal interaction signals are retrieved from memory and injected into the frozen policy model as context. Experiments in simulation and real-world manipulation demonstrate that Zeva achieves the best performance among the compared frontier VLAs and WAMs and, more importantly, enables self-evolution during deployment without gradient updates. Its success rate continues to improve as the robot accumulates interaction experience. Furthermore, the acquired interaction experience can generalize across tasks.

cs.RO↗

arXiv2026

Zeva-Ego: Egocentric Mid-Training with In-Context Causal Learning for Robot Manipulation

Egocentric video offers a scalable source of physical interaction experience, yet translating it into robot-executable knowledge and enabling continual adaptation remain challenging. We introduce Zeva-Ego, a unified framework that learns physical priors from human experience and evolves through robot interaction. An Action-Centric Encoder (ACE) converts egocentric visual transitions into action-centered supervision for VLA mid-training, while In-Context Causal Learning (ICCL) enables parameter-free adaptation from action-effect feedback at deployment. Scaling Ego data to 10K hours improves RoboTwin success from 63.8% to 75.3%, matching 2K hours of robot demonstrations (74.7%), corresponding to an empirical data ratio of roughly 4-5:1. With accumulated interaction experience, ICCL further improves success from 58% to 89% within four attempts without parameter updates. These results demonstrate a scalable path toward embodied intelligence that learns from human experience and continuously improves through its own interaction.

cs.RO↗

arXiv2026

Mass Distribution of M31 from its Rotation Curve up to 167 kpc

The mass distribution of the Andromeda galaxy (M31) is fundamental for understanding its dark matter content, assembly history, and dynamical role within the Local Group. However, estimates of M31's total mass remain quite uncertain, partly because of limited dynamical constraints at large galactocentric radii. Moreover, M31's recent merger history has produced abundant tidal substructures and may have left its outer halo out of dynamical equilibrium, complicating equilibrium-based mass estimates. We use over 8,500 tracers with line-of-sight velocities to constrain the rotation curve (RC) of M31 over R=2--167 kpc. In the disk, we derive the RC from a clean tracer sample using the Jeans equation and find good agreement with previous H I measurements. In the halo, we use Gaussian mixture modeling to account for tidal substructures and Galactic foreground contamination and constrain the kinematics of the residual dynamically hot halo. Assuming this halo population can be treated as an equilibrium tracer, we infer the outer RC and constrain M31's mass distribution for different velocity anisotropies beta and dark matter halo profiles. The NFW and Dekel--Zhao (DZ) models yield broadly consistent total masses ranging from 1.12(+0.11/-0.10) to 1.63(+0.21/-0.17)x10^12 Msun for beta=0.3--0.7. In contrast, the Einasto profile favors substantially lower masses of 0.41(+0.09/-0.06) to 0.86(+0.21/-0.16)x10^12 Msun, highlighting significant dependence on the adopted halo profile, although Einasto solutions are more in tension with independent mass estimates based on M31 satellite galaxies than the NFW and DZ models at fixed beta. Overall, our analysis provides a comprehensive measurement of M31's mass distribution under the assumption of dynamical equilibrium, while also quantifying the systematic uncertainties associated with halo velocity anisotropy and the choice of dark matter density profile.

astro-ph.GA↗

arXiv2026

Entropy-Driven Initiation and Cytoskeletal Viscoelasticity in Endocytosis: An Onsager Variational Framework

Receptor-mediated endocytosis requires a particle to approach the cell membrane to within a few nanometers before ligand--receptor binding can occur. Existing continuum models often start from an already established contact and do not explicitly describe how crowding particles on the extracellular side influence the distribution of the particle near the membrane. We examine entropic depletion forces as one possible nonspecific contribution to this initial approach. For ideal depletants, the Asakura--Oosawa excluded-volume construction gives an exact depletion potential for the planar geometry before contact. The potential and force vanish continuously at the onset of excluded-volume overlap. This interaction provides a possible contribution to membrane proximity before specific binding, while its extension to curved wrapping geometries requires additional approximation. Within a reduced continuum model, we combine depletion attraction, ligand--receptor binding, membrane deformation, and cytoskeletal viscoelastic dissipation. The viscoelastic contact is formulated through a hereditary integral and a standard linear solid. The kinetic model gives a conditional minimum ligand density for complete engulfment, a finite particle-size window, and a stiffness-dependent upper limit. When the stationary radius lies inside the domain of finite positive wrapping times, the estimated wrapping time has a minimum at a radius that decreases with increasing binding energy density. At fixed viscosity and other independent parameters, the same time approximation predicts slower wrapping as cell stiffness increases. The two positive roots defining the size window merge at a limiting parameter value, which characterizes closure of the admissible size interval. Depletion attraction is interpreted as one possible contribution to particle-membrane association

cond-mat.soft↗

arXiv2026

Persistent Quantum-Enhanced Frequency Sensing with T^{-3/2} Scaling

Quantum sensing uses nonclassical states to improve measurement sensitivity, but the same states that provide metrological gain also decohere more rapidly. This limits the usable interrogation time and, in practice, often precludes improvement in ultimate sensitivity - realized useful quantum advantage has consequently remained rare. Here, we restore persistent quantum advantage by embedding Fock-state enhancement within a quantum heterodyne (Qdyne) protocol, decoupling sensitivity from the decoherence-limited interrogation time τ. Measurements are acquired at short τ where the quantum-enhanced gain is optimal, while precision accumulates with the total measurement time. Demonstrated on the motional mode of a trapped 40Ca+ ion, we observe quantum-enhanced precision that persists to measurement times seven orders of magnitude beyond the dephasing limit, scaling as T^{-3/2} with no indication of saturation. Using the n=3 Fock state, we reach a frequency precision of 0.5uHz relative to an 86MHz carrier, achieving a fractional precision ~6x10^{-15}. This represents a quantum-enhanced gain of 7.1(10) dB over the n=0 state, in agreement with Fisher information predictions. This is the first demonstration of Qdyne beyond solid-state spin-defects. Further, by using the quantum harmonic oscillator to perform frequency mixing, we extend operation beyond 1GHz, two orders of magnitude above the ceiling of pulsed dynamical-decoupling implementations. These results recover quantum advantage at the long timescales required to improve ultimate sensitivity, with direct implications for nanoscale NMR and quantum logic spectroscopy.

quant-ph↗

arXiv2026

Phoneme-guided TTS augmentation for ASR: A unified pipeline and multilingual evaluation

Synthetic speech can provide additional supervision for automatic speech recognition (ASR), but constructing useful synthetic training data requires choosing both what to synthesize and how to synthesize it. We present a phoneme-guided text-to-speech (TTS) augmentation pipeline for ASR that connects multilingual speech generation with candidate-text selection and reference-speech quality control. Within this pipeline, we propose phoneme-frequency-guided selection (PFGS), which uses phoneme frequencies from real ASR training transcripts to prioritize candidate texts containing common phonetic content. Experiments with separate monolingual ASR systems cover four languages and 13 test sets. With random text selection, the pipeline improves recognition on 11 test sets at one or more synthesis ratios. PFGS further outperforms random selection on nine test sets, with relative word error rate (WER) reductions of up to 19.3%. An ablation with fixed target texts and synthesis counts further shows the benefit of reference-speech filtering. These results support using real-data phoneme statistics to guide the construction of effective synthetic supervision for ASR.

cs.CL↗

arXiv2026

Classification of commutation relation for multi-radial SLE

Locally commuting multiple radial Schramm-Loewner evolutions ($\mathrm{SLE}_κ$) are encoded by partition functions satisfying the radial Belavin-Polyakov-Zamolodchikov (BPZ) equations and a conformal Ward identity with spectral parameters $λ,ν\in\mathbb{R}$. For $κ>0$ and $λ\in\mathbb{R}$, we show that the solution space of the radial BPZ system has dimension $2^n$, where $n$ is the number of variables. We then determine all admissible Ward parameters $ν$ and the exact dimensions of the subspaces selected by the conformal Ward identity, covering both generic and degenerate cases. The classification reveals a parity difference: nonzero rotation-invariant solutions exist for every $λ$ when $n$ is even, but only at finitely many exceptional values when $n$ is odd. When $0<κ\leq4$, $λ>0$ for odd $n$ or $λ>-3/2$ for even $n$, we construct a basis of positive solutions using multiple SLE, providing global realizations of the locally commuting SLEs. At $κ=4$, we identify a family of explicit solutions as partition functions for level lines of a Gaussian free field with suitable boundary data and interior singularities.

math.PR↗

arXiv2026

Extended MF-AFBSDDEswCN and stochastic optimal controls of linear system

The research studies the extended mean-field anticipated forward-backward stochastic delayed differential equations with common noise (extended MF-AFBSDDEswCN) by expanding the domination-monotonicity conditions. We generalize the conditions from a linear setting to a nonlinear one by including nonlinear adjoint functions which are key to guaranteeing the well-posedness of extended MF-AFBSDDEswCN. Utilizing this broader well-posedness framework for extended MF-AFBSDDEswCN, combined with other refined analytical tools, we examine two classes of stochastic optimal control problems. These include a linear-convex problem and a linear-quadratic problem with input constraints that can be both time-dependent and random. Moreover, the control elements regarding the initial values are path-dependent, which will cause some essential difficulties. For each problem, we establish the existence and uniqueness of optimal controls and provide their explicit, closed-form representations.

math.OC↗

arXiv2026

PhysBrain 1.5: From Vision-Language Models to Physical Foundation Models

We present PhysBrain 1.5, a unified model for understanding physical environments, generating actions, and predicting future states. Motivated by the physical loop of observation, interaction, and environmental change, we bring these capabilities into a common learning framework. Starting from a general vision--language model, we encode language responses, end-effector motion, and dense visual targets as discrete sequences and jointly optimize them with autoregressive next-token prediction. Pre-training draws its embodied supervision entirely from human interaction videos, using task-centered episodes to pair semantic and spatial context with recovered motion and subsequent observations. We then adapt the model through supervised fine-tuning on a mixture of human demonstrations, robot trajectories, and simulated experience. Across 28 embodied understanding benchmarks, our 8B model achieves an average score of 72.5, setting a new open-source state of the art and performing on par with leading proprietary models such as GPT-6-Astra and Gemini 3.6 Flash. It achieves the best open-source results on 14 benchmarks while retaining general multimodal capabilities. Beyond these understanding evaluations, qualitative examples show the model's ability to produce end-effector trajectories and predict future scenes through spatially aligned RGB, depth, and robot-mask outputs.

cs.CV↗