arXiv subjects

Zhipeng Zhang

Publications and source records attributed to Zhipeng Zhang.

At least 19 recordsLinked to original sources

arXiv2026

MOON: Learning When to Forget and How to Move On in Vision-and-Language Navigation

Recent vision-and-language navigation (VLN) policies act on their visual history, and an incorrect action can take them off the instructed route. Paired interventions on four VLN policies show that clearing the history doubles pooled success at the onset of a deviation but lowers it on route, while later in a deviation, removing only the off-route frames helps much less. After a deviation, the retained history keeps the policy from turning where a turn is needed, which we call behavioral inertia. We propose MOON, a framework that equips a frozen VLN policy with two lightweight modules, adding less than 1% of its parameters. A deviation judge decides when to forget, raising an alarm that clears the history when it predicts that the agent is off route. A recovery policy, trained with GRPO from alarm states, learns how to move on, turning more of the successes found only by sampling into a single greedy rollout. Experiments on R2R-CE and RxR-CE show that MOON reaches the highest R2R-CE success rate among the compared methods and raises the success rate of its frozen base by 5.2 and 6.0 points, and an ablation with a random trigger indicates that this gain depends on when the history is cleared. Without retraining, MOON also improves zero-shot object-goal navigation on HM3D-OVON. Together, these results suggest that deviation recovery depends on knowing when to forget, not only how to act. Code is available at https://github.com/AutoLab-SAI-SJTU/MOON.

cs.RO↗

arXiv2026

Stable but Wrong: When Learning Stabilizes Away from the Truth

Stable training is often treated as evidence that learning is succeeding, but stability characterizes optimization behavior rather than correctness relative to an external objective. We study what happens when the signal being optimized remains persistently biased. We define Stable but Wrong (SBW) as a learning state in which optimization remains active or stabilizes while the learned outcome stays systematically displaced from an independently defined objective. A minimal theoretical model shows that persistent systematic bias can separate a stable convergence point from the true objective even under standard conditions for stable convergence. We then introduce static and dynamic bias in reinforcement learning, supervised learning, and continual fine-tuning of large language models. Across these settings, optimization of the current learning signal continues or stabilizes while correctness relative to the external objective declines or remains displaced. Finally, two controlled challenge-correction experiments with external ground truth test whether a prespecified error can be exposed and changed. In controlled scientific discovery and CIFAR-10, targeted evidence corrects or reduces established errors more effectively than equal-budget general evidence, with no evident harm to paired correct states under the tested protocols. These experiments do not provide a general method for automatically discovering unknown errors. Overall, SBW identifies a basic limit of convergence and other internal optimization metrics as evidence of correctness in systems that learn through persistent bias and feedback.

cs.LG↗

arXiv2026

Ion Heating in Pair-loaded Relativistic Turbulence

Pair-loaded relativistic turbulence is expected in compact high-energy astrophysical environments and may sustain plasma heating. Using self-consistent three-dimensional radiative particle-in-cell simulations, we identify the mechanism of ion heating mediated by pair-enhanced radiative feedback in relativistic turbulence. We derive the theoretical criterion for the transition to a regime dominated by radiative feedback and show that pair loading promotes the growth and persistence of large, connected optically thick structures. These long-lived structures sustain intense localized ion heating, repeatedly generating prominent hot-ion clumps and substantially enhancing the mean ion temperature. Our results establish a physical link between pair creation and sustained ion heating, providing a theoretical basis for investigating intermittent hot-ion production in black-hole coronae and gamma-ray burst outflows.

physics.plasm-ph↗

arXiv2026

MCSplat: Multi-View Photometric and Geometric Consistent Feed-Forward Gaussian Splatting for Driving Scenes

Driving-scene reconstruction via feed-forward Gaussian splatting from sparse multi-view images remains challenging due to two factors. First, cross-camera photometric inconsistency, caused by variations in exposure, ISP, and illumination, can be incorrectly absorbed into Gaussian geometry. Second, unreliable local geometric supervision arises from normal-based constraints near depth discontinuities. To address these issues, we present \textbf{MCSplat}, a multi-view consistent feed-forward Gaussian splatting framework for driving scenes. MCSplat constructs a canonical appearance space from depth-guided multi-camera correspondences to decouple geometry learning from camera-specific appearance. A multi-scale Bilateral Grid Transformer restores raw camera appearance from canonical renderings, enabling photometric supervision without contaminating Gaussian geometry. To improve local geometric reliability, we estimate Fisher-information-based uncertainty from photometric sensitivity and use it to adaptively weight surface-oriented self-supervision. Extensive experiments demonstrate improved reconstruction quality and generalization across driving datasets. Geometry-oriented evaluations further reveal enhanced cross-view depth consistency and local surface-normal agreement, while qualitative results show fewer geometric artifacts and more coherent structures across overlapping views.

cs.CV↗

arXiv2026

Acting Through Registers

World Action Models (WAMs) use generative visual representations to condition robot actions, yet how these representations organize action-relevant information remains poorly understood. We identify persistent high-norm spatial outliers in WAM representations and find that their action relevance persists even after normalization and K/V projection attenuate their extreme magnitude. Motivated by this observation, we introduce RegisterWAM, which adds learnable non-spatial registers to the world model and uses them directly for action conditioning. Registers alone do not fully eliminate spatial outliers; pairing them with observation flow-matching (FM) loss supervision suppresses spatial outliers while concentrating high-norm responses in registers. Beyond this reorganization, observation FM loss supervision also improves action learning without registers, motivating an additional noised-observation FM objective that learns current-scene recovery from corrupted inputs. The resulting model achieves 81.64\% success on LIBERO-Plus without additional embodied pretraining data or foundation models, reaches 93.0\% average success on RoboTwin 2.0, and improves real-world manipulation performance.

cs.RO↗

arXiv2026

WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction

Vision-language-action (VLA) policies excel at generalist robotic manipulation but generate actions without explicitly modeling their environmental consequences. World action models (WAMs) incorporate future prediction, yet most rely on costly video generation, while latent alternatives either use predicted futures only as auxiliary targets or infer latent actions from suboptimal visual transitions through costly large-scale video pretraining. We introduce WAVE-0 (World-model-Augmented VLA Execution), a framework that bridges VLAs and WAMs through a lightweight latent world model. By connecting the semantic understanding of pretrained VLAs with explicit future prediction, WAVE-0 allows anticipated action consequences to directly guide action generation. Built on a pretrained VLA, its 257M world model translates VLM-generated latent action plans into predicted future representations that condition the action expert, unifying latent planning, future prediction, and action generation. The world model is trained directly on ground-truth robot actions with a counterfactual ranking objective, requiring neither pixel reconstruction nor large-scale video pretraining. WAVE-0 achieves success rates of 91.47% on RoboTwin 2.0 and 98.7% on LIBERO, along with 88.0%, 61.4%, and 75.0% on LIBERO-Plus, LIBERO-Pro, and LIBERO-Para, respectively. On real-world tasks, it improves the success rate of $π_{0.5}$ from 38.3% to 73.3%. These results demonstrate strong manipulation performance and generalization across diverse distribution shifts.

cs.RO↗

arXiv2026

Towards Long-horizon Embodied Agents with Tool-Aligned Vision-Language-Action Models

Vision-language-action (VLA) models are effective robot action executors, but they remain limited on long-horizon tasks due to the dual burden of extended closed-loop planning and diverse physical operations. We therefore propose VLAs-as-Tools, a strategy that distributes this burden across a high-level vision language model (VLM) agent for temporal reasoning and a family of specialized VLA tools for diverse local physical operations. The VLM handles scene analysis, global planning, and recovery, while each VLA tool executes a bounded subtask. To tightly couple agent planning with VLA tool execution in long-horizon tasks, we introduce a VLA tool-family interface that exposes explicit tool selection and in-execution progress feedback, enabling efficient event-triggered agent replanning without continuous agent polling. To obtain diverse specialized VLA tools that faithfully follow agent invocations, we further propose Tool-Aligned Post-Training (TAPT), which constructs invocation-aligned training units for instruction following and adopts tool-family residual adapters for efficient tool specialization. Across five simulated benchmarks, VLAs-as-Tools improves the success rate of pi05 by 8.9 percentage points on average and enhances invocation fidelity by 15.0 points in Non-biased Rate. On three real-world Franka tasks, it improves success rate from 16.7% to 58.3% under matched data and compute. Code will be released.

cs.RO↗

arXiv2026

Exploring the Limits of Combinatorial Generalization in Embodied Manipulation

In embodied manipulation, Vision Language Action (VLA) models can execute atomic skills, but long-horizon manipulation requires recombining these skills without demonstrations for every skill combination. The full instruction route treats skill chaining as long-horizon imitation. In a controlled study, full instruction training on limited skill combination sequences reaches 40.0% ordered success on covered combinations but only 7.7% on recombination, suggesting memorization of sequence templates rather than robust skill recombination. We introduce Hierarchical Subtask Execution (HSE), a diagnostic framework that exposes skill recombination through a shared modular interface over the planner, trigger, and policy, decomposing long-horizon failures into separable modes. We introduce RoboCombine, a benchmark for skill recombination without combined skill training trajectories, with metrics for ordered success and transition conversion across subtask boundaries. RoboCombine reveals that atomic competence alone does not ensure ordered execution. Failures concentrate at transitions where downstream subtasks launch from handoff states from preceding subtasks rather than canonical atomic starts. We identify this mismatch as the start state gap. To address it, we introduce Pose Robust Task Initialization (PRTI), a lightweight atomic task post-training strategy that broadens start state coverage beyond canonical starts. Under oracle timed HSE on RoboCombine, PRTI reaches 18.16% ordered success, compared with 4.21% under random perturbation and 9.47% under a robot-reset oracle, without combined skill training trajectories. These results show that combinatorial generalization requires handoff robust atomic skills that convert first stage progress into reliable downstream execution, not merely decomposition into known subtasks. Code: https://github.com/AutoLab-SAI-SJTU/RoboCombine.

cs.RO↗

arXiv2026

Drive3R: Block-wise Streaming Geometric Foundation Models for Autonomous Driving

High-fidelity closed-loop simulation for autonomous driving increasingly relies on dense multi-view 3D reconstruction. While recent feed-forward geometric foundation models have revolutionized per-scene optimization, applying them to continuous, long-horizon driving sequences is computationally intractable due to GPU memory limits and the quadratic complexity of global attention. Existing scaling strategies are fundamentally limited, as causal streaming suffers from compounding scale drift and trajectory deviation, while rigid "chunk-then-align" heuristics fragment global context and produce trajectory discontinuities. To address these limitations, we introduce Drive3R, a block-wise streaming geometric foundation model. Its block-wise causal attention resolves the scalability bottleneck of fully bidirectional attention while avoiding the loss of local future context in purely causal streaming. Its confidence-aware re-anchoring further mitigates long-range drift by detecting unreliable streaming states, re-initializing the KV cache, and stitching refreshed segments through SE(3) alignment. Extensive experiments demonstrate that Drive3R establishes a new state-of-the-art on nuScenes, KITTI, and Waymo. Specifically, on nuScenes, it improves over the recent DVGT-2 baseline from 0.1806 to 0.1081 in AbsRel, from 0.6750 to 0.5344 in point-cloud Accuracy, and from 1.5812 to 0.9092 in ATE.

cs.CV↗

arXiv2026

SparseWAM: Driving World-Action Models are Worth 16 Register Tokens

World Action Models (WAMs) transfer scene-dynamics priors to planning by jointly forecasting future scenes and ego trajectories. However, WAMs built on video generators must generate dense future videos at inference, incurring prohibitive latency, while efficient variants discard the world branch at deployment and thus forgo test-time imagination. Even when dense video latents are replaced with patch-level DINO or JEPA features, these representations still involve long token sequences and encode decision-irrelevant information. We therefore propose SparseWAM, which forecasts in a sparse register space of only 16 tokens per frame, compressed from frozen DINOv2 features and shaped by joint planning-and-scoring pretraining to concentrate decision-relevant information. Within this space, SparseWAM employs a modality-decoupled Mixture-of-Transformers with a causal attention mask that lets action tokens condition on imagined future registers, retaining explicit imagine-then-act reasoning at deployment. Its compactness further enables low-cost multimodal sampling. As a result, our SparseWAM achieves 92.6 PDMS on NAVSIM v1, 90.1 EPDMS on NAVSIM v2, and 39.6 on NavHard at only 164 ms latency, roughly 1/5 that of DriveVA, demonstrating the efficiency of the sparse register space. Code will be released upon acceptance.

cs.CV↗

arXiv2026

Low Mach number limit of a two-phase flow model in $\mathbb{R}^3$

We study the simultaneous low Mach number limit of a compressible Navier--Stokes--Euler two-phase system in $\mathbb R^3$, in which the pressure terms in both phases are scaled by $\varepsilon^{-2}$. For sufficiently small $H^3$-perturbations around the constant equilibrium, we establish the global well-posedness of the scaled compressible system and derive global energy-dissipation estimates that are uniform with respect to the Mach number $\varepsilon$. A key feature of the analysis is the degenerate dissipation structure: viscosity acts only on the Navier--Stokes phase, while the drag coupling transfers dissipation to the inviscid Euler phase through the relaxation mode. We then prove the global well-posedness and large-time decay of the limiting incompressible two-phase system, and show that the relative velocity $u-ω$ decays faster than the full velocity pair. Finally, for the well-prepared initial data, we introduce pressure-corrected acoustic variables to remove the singular pressure mismatch and establish a global-in-time $H^2$-error. As a result, the scaled compressible solutions converge uniformly in time to the corresponding solution of the limiting incompressible system.

math.AP↗

arXiv2026

Enhancing the Angular Resolution of Large Array of imaging atmospheric Cherenkov Telescope (LACT) at Ultra-High Energies

The Large Array of Imaging Atmospheric Cherenkov Telescopes (LACT) is dedicated to high-resolution morphological studies of PeVatrons. In this work, we present a fundamental investigation into stereoscopic direction reconstruction for the LACT array, specifically addressing the challenges of ultra-high-energy observations. We demonstrate that the standard Hillas parameterization introduces a significant reconstruction bias under severe image leakage. To mitigate this, we introduce an approach utilizing a 2D Gaussian fit, achieving an exceptional angular resolution of better than $0.06^\circ$ at $100\text{ TeV}$ within the central $0^\circ\text{--}1^\circ$ offset bin, and maintaining better than $0.12^\circ$ across offsets up to $4^{\circ}$. Building on this robust baseline, we evaluate advanced weighting schemes by utilizing a LightGBM-based quantile regression model to independently estimate single-image quality. Applying these quality-based weights yields a consistent improvement of $0.02^\circ$ to $0.03^\circ$ for high-energy, large-offset events using both the \textit{HillasWeightedSum} and \textit{HillasWeightedDisp} methods. Finally, to establish a theoretical performance ceiling, we explore a pixel-wise likelihood reconstruction technique utilizing Neural Ratio Estimation. While its practical realization depends heavily on minimizing the gap between Monte Carlo simulations and observational data, this exploratory approach demonstrates the potential to yield an overall improvement of approximately 15\% to 40\% at $100~\rm TeV$ across the entire field of view. Such high angular resolution is critical for disentangling complex emission regions and mapping the internal structures of PeVatrons.

astro-ph.HE↗

arXiv2026

HelloWorld: Towards Practical Applications of Generative Driving World Models

Driving world models provide a promising route toward scalable counterfactual data generation and interactive simulation beyond recorded driving logs. Realizing this potential requires a system that can generalize across diverse scenes, respond faithfully to prescribed controls, generate coherent multi-sensor observations, and operate efficiently under repeated inference. We present \textbf{HelloWorld}, a 2B driving world model system designed around these requirements. HelloWorld progressively specializes broad visual and motion priors from heterogeneous video data into controllable driving generation using ego pose, HD maps, and 3D boxes. A block-causal generation interface, together with adaptation to self-generated context, aligns the model with sequential simulation. The system further supports synchronized seven-camera RGB generation and conditional LiDAR synthesis, and is distilled toward few-step inference for efficient deployment. Experiments evaluate visual quality, control fidelity, cross-view consistency, robustness under repeated generation, inference efficiency, and LiDAR synthesis. Together, HelloWorld provides a unified framework for scalable driving data generation and interactive simulation.

cs.CV↗

arXiv2026

RoboFollow: Unveiling the Instruction Following Mirage in Embodied Agents

Modern embodied agents achieve impressive success rates, yet their actual instruction-following ability is far weaker than these numbers suggest. We trace this illusion to a structural property we term low scene entropy: when a visual scene admits only one valid task, language becomes redundant and a policy can score highly while barely using it. We introduce RoboFollow, a diagnostic benchmark with three principles: (1) High Scene Entropy: each training scene supports multiple kinematically distinct task branches, making vision alone insufficient and forcing reliance on language. (2) Hierarchical Diagnostic Protocol: a four-level protocol (L0--L3) progressively perturbs visual layout and semantics, probing whether equivalent instructions yield consistent behavior and distinct ones yield discriminable behavior across spatial relations, attributes, trajectory constraints, and logic. (3) Confound-Controlled Diagnosis: we simplify interaction objects, restrict actions to the trained repertoire and report stage-wise Intent and Execution scores, isolating comprehension from motor execution. Evaluation of nine VLA and WAM policies shows that strong L0 performance, where attained, does not reliably transfer to L1--L3 under our fine-tuning setup. Representative mitigations, including stronger VLM backbones, QA co-training, LangForce, and Classifier-Free Guidance, all fail to close this gap. RoboFollow exposes genuine instruction following as a critical, overlooked bottleneck. Code and dataset are available at https://github.com/AutoLab-SAI-SJTU/RoboFollow and https://huggingface.co/datasets/AutoLab-SJTU/robofollow-data.

cs.RO↗

arXiv2026

Libra: Taming Attention Workload Skew in Long-Context LLM Training with Bounded Sequence Pool

Long-context LLM training suffers from a load-balancing problem that sequence packing does not solve. Packing samples into fixed-token sequences balances memory and linear-cost operators, but the dominant attention cost scales with the sum of squared sequence lengths. Thus, equally sized packed sequences drawn from a long-tailed corpus can carry substantially different attention workloads, creating data-parallel stragglers and pipeline bubbles. Existing approaches either balance at the granularity of sequences or microbatches, where an outlier can dominate an assignment, or disaggregate attention over a global worker pool whose communication domain grows with the data-parallel (DP) degree. We present Libra, which operationalizes the law of large numbers (LLN) as a scaling principle for load balancing: the attention-balancing pool need not grow with the DP degree. Libra groups packed sequences and their CP groups into fixed-size sequence pools. As DP scales out, Libra adds pools rather than enlarging each one, bounding every attention exchange. Variance-Reduced Sequence Placement makes this effective for finite, long-tailed workloads by co-locating sequences with complementary attention workloads to reduce residual inter-pool skew. Within each pool, Tiled Attention Pooling dispatches sequence-head SH-Tiles across GPUs, while a pipelined runtime overlaps tile exchange with attention. Libra exposes a drop-in context-parallel attention operator and a pluggable data sampler, requiring no changes to model layers, optimizers, or pipeline schedules. On three production Qwen3 models (8B, 30B, 235B) and 256K- and 1M-token production workloads, Libra improves end-to-end training throughput over the strongest evaluated baseline (WLB-LLM) by 44% on average and up to 68% at 256 GPUs. Libra has run for hundreds of thousands of GPU-hours in production on jobs spanning 32K to 1M tokens.

cs.DC↗

arXiv2026

Low Mach number limit of the compressible Euler--Vlasov--Fokker--Planck system in the whole space

Although there are many important contributions on compressible and incompressible fluid-particle interaction models respectively, how to connect the two-type fluid-particle models via the low Mach number limit remains a challenging open problem. In this paper, we resolve it for the compressible isentropic fluid-particle model (Euler--Vlasov--Fokker--Planck (Euler--VFP) system) in the whole space $\mathbb{R}^3$. First, we establish the global-in-time {\it a priori estimates} of strong solutions that are uniform with respect to the Mach number $\varepsilon$ near the global Maxwellian. The proof relies on a refined energy method that combines the relaxation structure $b^\varepsilon-u^\varepsilon$ induced by the fluid-particle interaction and the symmetrized acoustic structure of the compressible Euler part in the model. Under the assumption of well-prepared initial data, we derive a {\it global-in-time} uniform error estimate in the $H^2$ framework between the solution of the compressible Euler--VFP system and that of the limiting incompressible Euler--VFP system. A key point is to introduce the corrected acoustic variable $ q^\varepsilon-\varepsilon [P'(1)]^{-1}π$, which captures the pressure corrector in the low Mach number limit. This also allows us to exploit the exact cancellation of the singular acoustic terms and to close the {\it global-in-time} error estimate. The damping term $b^\varepsilon-u^\varepsilon$, which is absent in the pure Euler equations, plays an essential role in recovering the relative velocity dissipation and in controlling the coupled fluid-particle dynamics. As a consequence, we prove the low Mach number limit of the compressible Euler--VFP system with the convergence rate $\mathcal O(\varepsilon)$ in the time-continuous $H^2$ topology.

math.AP↗

arXiv2026

Low Mach number limit for the Navier--Stokes--Korteweg equations with a stationary force

In this paper, we investigate the low Mach number limit for the three-dimensional compressible Navier--Stokes--Korteweg equations in the whole space under a small stationary external force. We first construct a family of small stationary solutions uniformly with respect to the Mach number $ε$ and prove that both the stationary density fluctuation and the compressible component of the stationary velocity are of order $ε^2$. For ill-prepared non-stationary perturbations around these stationary solutions, we establish the existence and uniqueness of global strong solution by combining uniform high-order energy estimates with a low-frequency Besov estimate and a Kawashima-type compensating functional. The main difficulty is that Korteweg tensor not only changes the elliptic structure of the stationary problem, but also modifies the dispersive mechanism of the acoustic modes. In Korteweg-symmetric variables, the associated spectral projections are uniformly bounded zero-order Fourier multipliers, while the acoustic-capillary phase is wave-like at low frequencies and Schrödinger-like at high frequencies. Since the source terms generated by the stationary coefficients are generally not integrable in time, we decompose the Duhamel source according to its time-integrability and frequency behavior. Dyadic dispersive estimates, high-frequency damping estimates, and maximal regularity for the heat equation yield the global-in-time convergence rate $ε^{\min\{1/r,\,1/2-1/p\}}$ in the mixed Besov norms $L^r(0,\infty;\dot B^s_{p,1})$. As a consequence, Besov embeddings also yield quantitative convergence in the mixed Lebesgue norms $L^r(0,\infty;L^p)$.

math.AP↗

arXiv2026

Compressible Navier--Stokes equations with a potential force: global well-posedness and optimal time-decay rates for arbitrarily large $L^2$ initial data

We study the Cauchy problem for the three-dimensional barotropic compressible Navier--Stokes equations with a time-independent potential force near a spatially nonconstant stationary state. The potential is controlled in unweighted homogeneous Besov spaces; in particular, no polynomial spatial-weight condition involving $(1+|x|)^j\nabla^jϕ$ is imposed. For initial data relative to the stationary state that are sufficiently small in $\dot H^{\frac12-δ}\cap\dot H^3$, we establish the existence and uniqueness of a global strong solution in $H^3$, while allowing the initial $L^2$ norm to be arbitrarily large. If the initial data are bounded in $\dot B^s_{2,\infty}$ for $s\in[-\frac32,-1)$, then the solution and its first spatial derivative decay at the optimal rates $(1+t)^{-\frac{k-s}{2}}$ with $k=0$ and $1$, respectively. The analysis relies on refined homogeneous energy estimates and a frequency-localized description for the dissipative and asymptotic structures of the system.

math.AP↗