Search arXiv⌕ Search

arXiv subjects

Qi Zhang

Publications and source records attributed to Qi Zhang.

At least 91 records · Page 5Linked to original sources
arXiv2026

ReDraft, Don't Just Distill: Reference-Driven Revision for Continual VLLM Post-Training

Continual post-training of large multimodal models should add new capabilities while preserving those from pre-training, and the two goals pull in opposite directions. SFT gives explicit target supervision that learns a task from near-zero accuracy, but its off-policy targets move the model far enough to cause forgetting; on-policy methods such as RLVR and self-distillation preserve policy proximity yet supply little signal when the policy cannot yet solve the task. We introduce ReDraft (Reference-Driven Revision and Fine-Tuning), which obtains both from the model's own failures: using an expert response only as a reference, it has the model revise its own incorrect rollout, keeps the revision only if a verifier accepts it, and fine-tunes on what survives. Each retained target is therefore explicit, yet still close to the current policy. Across Counting, Clock Reading, and Jigsaw on Qwen2.5-VL-3B/7B, two of them with near-zero accuracy, ReDraft gains 56.9 points on the target task against SFT's 52.9 while cutting prior-task loss from 16.6 to 1.5 points (11.3x less forgetting), and improves on OPSD along both axes (19.3 gain, 6.2 loss). Data- and parameter-space analyses match the design: revised targets are more probable under the base model, and the updates they induce stay compact and follow SFT's direction more closely than OPSD's. Together, these results show that revising the model's own rollout rather than directly imitating an expert trajectory can reconcile cold-start acquisition with prior-capability retention.

cs.AI↗
arXiv2026

Decoupling Disease, Covariates, and Individual Variability: A Unified Disentanglement Framework for Medical Image Classification

Accurately isolating disease-related features from confounding covariates (e.g., age, gender, site) and individual variations remains a fundamental challenge in medical image classification. Traditional regression-based approaches may ignore non-linear relations between image features and true covariates. To overcome this issue, we present a generalized Medical Imaging Disentanglement Learning (MedIDL) framework. MedIDL maps image features into three mutually orthogonal latent spaces through specialized disentanglement heads: a disease classification head guided by a supervised loss, a covariate-alignment head constrained by cross-subject similarity matching, and a Gaussian head absorbing individual variations. We evaluated our framework across 7 datasets encompassing diverse imaging modalities. MedIDL outperforms state-of-the-art supervised and self-supervised classification methods in accuracy across all datasets. Association analyses demonstrate that MedIDL successfully isolates target-specific latent representations. Gradient-based interpretability mappings localize pathognomonic patterns aligning with established clinical literature.

cs.CV↗
arXiv2026

CCTU: A Benchmark for Tool Use under Complex Constraints

Solving problems through tool use under explicit constraints constitutes a highly challenging yet unavoidable scenario for large language models (LLMs), requiring capabilities such as function calling, instruction following, and self-refinement. However, progress has been hindered by the absence of dedicated evaluations. To address this, we introduce CCTU, a benchmark for evaluating LLM tool use under complex constraints. CCTU is grounded in a taxonomy of 12 constraint categories spanning four dimensions (i.e., resource, behavior, toolset, and response). The benchmark comprises 200 carefully curated and challenging test cases across diverse tool-use scenarios, each involving an average of seven constraint types and an average prompt length exceeding 4,700 tokens. To enable reliable evaluation, we develop an executable constraint validation module that performs step-level validation and enforces compliance during multi-turn interactions between models and their environments. We evaluate nine state-of-the-art LLMs in both thinking and non-thinking modes. Results indicate that when strict adherence to all constraints is required, no model achieves a task completion rate above 20\%. Further analysis reveals that models violate constraints in over 50\% of cases, particularly in the resource and response dimensions. Moreover, LLMs demonstrate limited capacity for self-refinement even after receiving detailed feedback on constraint violations, highlighting a critical bottleneck in the development of robust tool-use agents. To facilitate future research, we release the data and code.

cs.CL↗
arXiv2026

xHAP: Cross-Modal Attention for Haptic Feedback Estimation in the Tactile Internet

The Tactile Internet requires ultra-low latency and high-fidelity haptic feedback to enable immersive teleoperation. A key challenge is to ensure ultra-reliable and low-latency transmission of haptic packets under channel variations and potential network outages. To address these issues, one approach relies on local estimation of haptic feedback at the operator side. However, designing an accurate estimator that can faithfully reproduce the true haptic forces remains a significant challenge. In this paper, we propose a novel deep learning architecture, xHAP, based on cross-modal attention to estimate haptic feedback. xHAP fuses information from two distinct data streams: the teleoperator's historical force feedback and the operator's control action sequence. We employ modality-specific encoders to learn temporal representations, followed by a cross-attention layer where the teleoperator haptic data attend to the operator input. This fusion allows the model to selectively focus on the most relevant operator sensory data when predicting the teleoperator's haptic feedback. The proposed architecture reduces the mean-squared error by more than two orders of magnitude compared to existing methods and lowers the SNR requirement for reliable transmission by $10~\mathrm{dB}$ at an error threshold of $0.1$ in a 3GPP UMa scenario. Additionally, it increases coverage by $138\%$ and supports $59.6\%$ more haptic users even under 10 dB lower SNR compared to the baseline.

eess.SP↗
arXiv2026

MotionJEPA: Preventing Temporal Feature Collapse by Capturing Visual Changes in Latent Space

Joint Embedding Predictive Architectures (JEPAs) are a promising paradigm for learning task-agnostic latent world models without visual reconstruction. However, standard JEPA training exhibits a strong inductive bias towards slow features, causing feature suppression and the collapse of latent representation. While inverse dynamics provides temporal anti-collapse, it relies on action labels and offers little incentive to embed general, unlabeled dynamics. We introduce Difference Image and Single image embedding Regularization (DISReg), a novel regularizer that builds on an inverse-dynamics-style module that predicts temporal difference image embeddings without any pixel reconstruction loss, encouraging balanced static and dynamic feature learning. DISReg consists of a static term that shapes the distribution of the image embedding and encourages slow features, and a dynamic term, which, unlike direct regularization on the embedding, imposes no constraint on the image embedding's shape or distribution and instead only incentivizes that dynamic features be present. By integrating this regularizer into a standard JEPA, we establish our new architecture, MotionJEPA. Latent probing demonstrates that MotionJEPA produces more complete representations than other methods, and our trajectory analysis shows it maintains geometrically simple latent embeddings with low curvature. We further show that MotionJEPA improves downstream planning success under static-background distractors across four environments.

cs.CV↗
arXiv2026

SafeStage: Evaluating Safety Before, During, and After Vision-Language-Conditioned Robot Manipulation

Vision-language-conditioned robot policies integrate perception, language understanding, and control for general-purpose manipulation. However, existing evaluations often focus on task success, isolated physical constraints, semantic refusal, or realized physical damage, providing limited insight into where safety fails during closed-loop manipulation. We introduce SafeStage, a lifecycle-structured benchmark for evaluating manipulation safety before, during, and after task execution. SafeStage contains 97 purpose-built risk scenarios organized into three stages. Initial-State Hazards captures safety-relevant relations that must be resolved before manipulating the target. Execution-Time Safety evaluates unsafe contacts, trajectories, region entries, and object interactions during execution. Final-State Hazards capture unstable or otherwise unsafe conditions remaining after nominal task completion. The benchmark evaluates realized interactions using event-based and state-based checks and reports native task success independently from stage-specific safety outcomes. We evaluate representative direct-action Vision-Language-Action (VLA) policies and policies with world-model-based policies under a common closed-loop protocol. Our results demonstrate that nominal task completion frequently coexists with safety violations and that different policies exhibit distinct failure profiles across the three stages. By separating task success from safety and localizing when violations occur, SafeStage provides a unified diagnostic testbed for evaluating and improving vision-language-conditioned robot manipulation policies.

cs.RO↗
arXiv2026

Instance Segmentation and Fine-grained Classification for Urban Buildings with Adaptive Region Dividing and Spatially-Supervised Contrastive Learning

Accurate instance-level and functional understanding of urban buildings in large-scale point clouds is essential for digital city modeling and urban analysis. However, the extensive spatial coverage of urban scenes leads most existing methods to rely on predefined blocks for training and evaluation, although such partitions are rarely available in real-world applications and introduce additional preprocessing while fragmenting complete building structures. To address this issue, we propose an adaptive region-dividing strategy with unified scene-level evaluation. Specifically, the 3D point cloud is projected onto a bird's-eye-view (BEV) plane, where a pretrained segmentation model is used to detect building regions. The detected bounding boxes are then back-projected to the original point cloud to construct structure-aligned adaptive training blocks, enabling semantically guided dynamic partitioning without manual design. Furthermore, beyond instance-level understanding, few methods have explored fine-grained classification for urban buildings, and thus we also put forward a fine-grained classification model for urban buildings with a spatially-supervised contrastive loss. First, for each segmented building, a point transformer classifier jointly encodes its body and local context using geometric, color, and core-context information. Then, the class-balanced weighted cross-entropy is used to alleviate severe class imbalance. The proposed spatially-supervised contrastive loss further enhances inter-class discriminability by assigning greater weight to spatially proximate, same-category buildings, encouraging compact functional representations while separating easily confused categories. Extensive experiments on UrbanBIS and STPLS3D demonstrate the advantages of the proposed method in building instance segmentation and fine-grained classification compared to existing SOTA methods.

cs.CV↗
arXiv2026

Unleash LLMs Potential for Sequential Recommendation by Coordinating Dual Dynamic Index Mechanism

Owing to the unprecedented capability in semantic understanding and logical reasoning, large language models (LLMs) have shown fantastic potential in developing next-generation sequential recommender systems (RSs). However, existing LLM-based sequential RSs mostly separate index generation from sequential recommendation, leading to insufficient integration between semantic information and collaborative information. On the other hand, the neglect of user-related information hinders LLM-based sequential RSs from exploiting high-order user-item interaction patterns. In this paper, we propose the End-to-End Dual Dynamic (ED$^2$) recommender, the first LLM-based sequential RS which adopts dual dynamic index mechanism, targeting resolving the above limitations simultaneously. The dual dynamic index mechanism can not only assembly index generation and sequential recommendation into a unified LLM-backbone pipeline, but also make it practical for LLM-based sequential recommender to take advantage of user-related information. Specifically, to facilitate the LLM comprehension ability to dual dynamic index, we propose a multigrained token regulator which constructs alignment supervision based on LLMs semantic knowledge across multiple representation granularities. Moreover, the associated user collection data and a series of novel instruction tuning tasks are specially customized to capture the high-order user-item interaction patterns. Extensive experiments on three public datasets demonstrate the superiority of ED$^2$, achieving an average improvement of 19.62% in Hit-Rate and 21.11% in NDCG.

cs.IR↗
arXiv2026

Secure Relay Low-Altitude Networks via Hybrid Fixed-Position and Rotatable Antenna Arrays

In this paper, a relay network with hybrid fixed-position and rotatable antenna arrays is proposed. The deployment of rotatable arrays in conventional relay networks is considered to provide more secure communications for low-altitude economy applications. Specifically, both the base station and the relay station are equipped with fixed-position antenna arrays and rotatable arrays to serve ground users and aerial users, respectively. To address the challenge of multi-user interference, a low-cost reconfigurable intelligent surface is exploited as a candidate path. Accordingly, under constraints on transmit power, user quality of service, rotatable range, and path selection, the objective is to maximize the worst-case secrecy rate (SR) through joint beamforming, power allocation, and rotatable antenna orientation design. First, the SR performance in the \textcolor{black}{single ground-aerial user pair scenario} is investigated, and a step-by-step leakage-based scheme is proposed. Then, the general multi-user scenario is studied, and a Distributional Soft Actor-Critic with Three refinements (DSAC-T)-based learning scheme, which supports hybrid discrete and continuous actions, is proposed to maximize the worst-case SR. Simulation results validate the effectiveness of the proposed schemes. The proposed schemes achieve approximately a twofold improvement in SR performance compared to isotropic antennas. \textcolor{black}{The proposed system can achieve approximately 99.7\% power saving, 55\% antenna saving, and can serve more users.}

eess.SP↗
arXiv2026

CAFE: Self-Improving Search Agents Need Co-Evolving Feedback

Reliable search requires more than acquiring external evidence. An agent must also recognize and recover from errors as its trajectory unfolds. In-trajectory feedback provides a mechanism for such recovery by diagnosing where the search has drifted and redirecting subsequent reasoning steps. This is particularly important in long-horizon search, where an early directional error may receive no immediate corrective signal and can compound across later steps. Making such feedback learnable, however, creates a coupled problem: the agent must learn when to request and use feedback, while the critic must learn corrections from outcome-confounded rollouts as the agent's failure patterns evolve. We introduce CAFE (Coupled Agent--Feedback Evolution), a framework in which a shared-parameter model alternates between search-agent and critic roles. CAFE initializes feedback-conditioned recovery from trajectories built around the base agent's own failures, then couples online and offline optimization. During online RL, a comparative feedback estimate uses a prompt-level call--skip success gap to shape request returns, while feedback-aware advantage shaping reweights token advantages before and after feedback. Offline, rollout-derived preference optimization learns feedback from matched successful and unsuccessful trajectories. On seven agentic search benchmarks, CAFE outperforms the evaluated RL-based search agents on average, retains its gains across all six out-of-domain benchmarks, and reduces answer-level hallucinations. One-sided ablations show that improving only the agent or only the critic eventually plateaus, whereas alternating the two updates continues to improve performance. These findings suggest that a self-improving search agent needs feedback that co-evolves with the policy it guides.

cs.AI↗
arXiv2026

Dynamics-Informed Reinforcement Learning for Agile and Energy-Efficient Locomotion of a Monopedal Hopping Quadcopter

Although aerial-legged robots offer combined agility and efficiency, controlling high-speed hopping under complex hybrid dynamics is challenging. Reinforcement Learning (RL) is promising but prone to energy-inefficient "reward hacking". We propose a Dynamics-Informed RL framework for a monopedal hopping quadcopter. By embedding a target Specific Energy into the reward, we constrain the optimization to a physically viable energy manifold, ensuring stable hopping behaviour. By rewarding the phase-consistent behavior, it can encourage bio-inspired stance-phase impulse. Furthermore, penalizing the electro-mechanical power waste induces the motors generate an efficient impulse. This enables the policy to inject energy strictly during spring restitution without heuristic state machines. MuJoCo simulations validate robust height regulation and forward velocity tracking up to 2.0 m/s despite severe attitude-contact coupling. Ultimately, our approach yields a highly agile hopping gait, reducing energy consumption by 82% and 73% compared to hovering baselines and inefficiency baseline, respectively.

cs.RO↗
arXiv2026

Learning to Exploit Passive Dynamics for Energy-Efficient Target Hopping of a Spring-Legged Quadcopter

Combining aerial thrust with spring-loaded hopping makes monopedal quadcopters promising for locomotion over complex terrain, but heuristic proportional-integral-derivative (PID) tuning limits coordination between active thrust and passive contact dynamics. We present a direct estimated-state-to-motor Proximal Policy Optimization (PPO) policy that commands four motors without an explicit hopping state machine or low-level attitude PID. Its reward combines Energy-Manifold Shaping for mass-normalized vertical-energy tracking and apex-state anchoring with Efficiency Shaping, which uses a history-aware power estimator to penalize general power use, impose an additional airborne-power cost, and penalize airborne near-stationarity. In representative hardware runs, the PPO-based control stack reduced cycle-averaged measured electrical power by 30.7% and mean total normalized thrust by 49.8% relative to the tuned PID-based control stack, while retaining repeatable commanded-height hopping and more concentrated landings. These observations are consistent with improved use of passive dynamics and reduced measured electrical demand.

cs.RO↗
arXiv2026

Structured flows regulate droplet dynamics in phase-separating fluids

Liquid-liquid phase separation is a fundamental route by which homogeneous fluids self-organize into mesoscale droplets and domains. In passive mixtures, the fate of these structures is largely dictated by curvature-dependent chemical potentials, which drive interfacial relaxation and Ostwald ripening toward progressively coarser morphologies. Here we show that structured flows can drive phase-separating fluids into flow-selected nonequilibrium states, thereby qualitatively rewriting the droplet dynamics. Using an advective Cahn-Hilliard model, we demonstrate that patterned velocity fields deform droplet interfaces along the streamlines of the imposed flow. The resulting curvature variations reshape the chemical-potential landscape that steers diffusive material transport. The resulting fluxes allow structured vortices to capture droplets, select stable droplet sizes, and invert the conventional ripening hierarchy by rendering larger droplets effectively less stable than smaller ones. In many-droplet systems, periodic and disordered vortex fields arrest coarsening and select steady morphologies whose length scales and symmetries are set by the underlying flow structure. Together, these results establish structured flow as a nonequilibrium route for controlling the thermodynamic driving forces that govern phase-separating fluids.

physics.flu-dyn↗
arXiv2026

NovGauge: A Fine-Grained Benchmark for Diagnosing LLMs' Capability in Paper Novelty Assessment

Large language models (LLMs) are increasingly used in peer review at major AI conferences, yet novelty remains a persistent weak point. Existing benchmarks assess novelty as a single holistic score, making it difficult to diagnose which dimension a model misjudges or whether its evidence is faithful. We present NovGauge, a human-anchored benchmark for fine-grained novelty assessment diagnosis. The benchmark contains 619 paper pairs and 50 multi-paper sets, drawn from two expert sources: ICLR reviewer overlap claims and survey co-citations. Instances are independently labeled along three dimensions: task, problem, and method, capturing application goals, technical challenges, and solution approaches. We propose a cascading diagnostic pipeline that verifies per-dimension correctness, evidence grounding, and logical support. Evaluation of 18 LLMs shows hallucination rates ranging from 0% to 39% across dimensions, and among non-hallucinated correct-positive judgments, over 70% cite evidence fails to logically support the stated reason. The best-performing model, GPT-5.5, achieves 43-72% Verified F1 across dimensions, while most models retain less than half of their raw F1 after faithfulness verification. These results suggest that current LLMs remain far from reliable scientific novelty assessment, particularly when correctness is conditioned on faithful evidence grounding.

cs.AI↗
arXiv2026

Fully Coupled Nonlinear Forward-Backward Stochastic Difference Equations: Spectral Contraction and Infinite-Horizon Maximum Principle

This paper develops an explicit spectral-contraction approach to studying fully coupled nonlinear forward--backward stochastic difference equations on infinite horizon and their applications to stochastic control. The estimates for this fully coupled system are assembled into an explicit two-dimensional nonnegative matrix. Its spectral radius yields an explicit sufficient discount threshold, and a corresponding equivalent weighted product norm is constructed to establish the contraction property. Moreover, for an infinite-horizon control problem with convex control constraints and an accumulated discounted running cost, we derive a Pontryagin-type stochastic maximum principle, its equivalent pointwise normal-cone formulation, and a verification theorem. Finally, a recursive risk-adjusted portfolio example is given to demonstrate the applications of our theoretical results, and a projection-type sufficient optimality condition for this example is derived.

math.OC↗
arXiv2026

IAE-VTG: Interaction-Aligned Action-Entity Video Temporal Grounding

Video Temporal Grounding (VTG) localizes the video segment that matches a natural-language query. Many queries describe an action performed by a particular entity. Existing methods often encode the query as a whole or use general video-text interactions, without explicitly checking whether the action and entity occur together. They may therefore select a segment that contains both concepts but not the event described by the query. We propose Interaction Aligned Action-Entity Video Temporal Grounding (IAE-VTG), which models this rela?tionship at both the representation and training assignment levels. First, the Fine-grained Disentangled Interaction Module (FDIM) separates action and entity related query information and aligns it with complementary motion and appearance features. It then combines token-level interactions to build representations that capture the relationship between the action and entity. Second, Interaction-Sensitive Assignment (ISA) adds this interaction evidence to bipartite matching, so training targets are selected using both temporal overlap and semantic compatibility. This reduces supervision from temporally plausible but semantically incorrect proposals. Experiments on QVHighlights, Charades?STA, and TACoS show that IAE-VTG consistently improves strong baselines and achieves competitive or state-of-the-art performance on standard grounding metrics. Additional analyses show that the method is especially effective when similar actions or entities appear at multiple times and produces more reliable assignments for complex events.

cs.CV↗
arXiv2026

StreetDiff: Multi-view Street Scenes Generation via Cross-view Consistent Multi-view Stable Diffusion with Structure Prompts

Multi-view diffusion models have shown strong performance in scenes with strong geometric priors and sparse semantics, such as indoor rooms or simple outdoor environments (e.g., fields, courtyards). However, they often fail to maintain cross-view consistency under camera rotation, especially in structurally complex urban environments. Without explicit modeling of spherical correspondence across views, existing approaches tend to produce object duplication, structural distortion, and layout inconsistency. To address this limitation, we propose StreetDiff, a multi-view diffusion framework that explicitly enforces cross-view alignment during denoising. StreetDiff introduces a Panorama--Perspective Synergy design to decouple global layout reasoning from local detail synthesis, and incorporates a Panorama Alignment Module (PAM) that establishes spherical-projection-based attention constraints across views. By injecting structured alignment constraints without modifying the diffusion backbone, our framework achieves robust cross-view coherence in challenging urban street scene generation tasks. In addition, we construct Street360, a large-scale HDR multi-view urban panorama dataset. Extensive experiments demonstrate that StreetDiff significantly improves structural consistency and visual fidelity compared to prior multi-view diffusion generation methods.

cs.CV↗
arXiv2026

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

We introduce NCP-ArchPreview, a latent-space language model that pushes autoregressive pretraining beyond standard next-token prediction (NTP). Alongside NTP, the model learns through Next Concept Prediction (NCP) to predict discrete concepts that span multiple tokens, introducing an explicit and more challenging concept-level objective while preserving standard token-level autoregressive generation. NCP-ArchPreview builds a latent space by constructing a product-quantized concept vocabulary directly from its hidden states, and subsequently learns to predict future concepts via a dedicated Concept Module. These predicted concepts are then fed back to the token level to guide subsequent generation, with NTP and NCP trained jointly end-to-end. We scale this architecture to 8.9B parameters and train it on 5.73T tokens from the Dolma-3 dataset, marking the largest demonstration of a latent-space language model to date. Remarkably, by consuming only 51.3% of the total training tokens, NCP-ArchPreview achieves the final pretraining loss of OLMo-3-7B. Following full pretraining, it outperforms OLMo-3-7B by 2.45 points on the downstream macro-average, including a notable 5.99-point gain on GSM8K. Controlled experiments isolate a clear progression of performance gains stemming from both the latent architecture and the NCP objective. Furthermore, utilizing only 85% of the standard computation, NCP-ArchPreview approaches the training loss of a strictly parameter-aligned 8.9B baseline. The learned latent space remains highly valuable after the pretraining stage: updating just the 17M-parameter VQ module yields a novel, lightweight interface for domain adaptation, while a simple injection of concept representations into a DFlash2 drafter improves the mean accepted length by 4.17% with negligible overhead.

cs.CL↗