arXiv subjects

Tao Zhang

Publications and source records attributed to Tao Zhang.

At least 37 records · Page 2Linked to original sources

arXiv2026

Magic-W0: A Structured World-Action Foundation Model for Physical Intelligence

World-action models (WAMs) augment robot policies with action-conditioned environment dynamics, yet existing approaches largely rely on future observation reconstruction or generic latent prediction and lack structured, control-oriented world representations tightly coupled with action generation. We introduce Magic-W0, a world-action foundation model that jointly models structured physical state evolution and continuous actions. Magic-W0 represents interaction as a Structured World Transition consisting of Current State, Transition, and Future State. Current State combines vision-language context with Current 3D Geometry; Transition is represented by 3D Motion capturing action-induced three-dimensional changes; and Future State is represented by Future Semantics describing task-relevant outcomes. To couple prediction and control, we propose a layer-aligned world-action interaction architecture in which evolving action hypotheses condition world-transition prediction, while predicted world representations continuously inform action generation. Magic-W0 is pre-trained on large-scale egocentric human manipulation, UMI, real-robot, and simulation data, with latent supervision for geometry, 3D motion, and future semantics from pre-trained visual models. Inference-time interventions show that structured world representations respond systematically to changes in candidate actions and that action-related information propagates through shared 3D representations into future semantic predictions. On RoboDojo-Sim, Magic-W0 achieves an average Score of 27.10, the highest among the compared WAMs. Across multiple real-robot tasks, it also demonstrates strong downstream performance after fine-tuning with limited downstream data, supporting generalization and rapid adaptation.

cs.RO↗

arXiv2026

Exact Clifford Optimality for Two-Sided Locally Randomized Classical Shadows

Entangling measurements can reduce the statistical cost of learning many-body correlations. We study Pauli strings with full support on a specified $k$-qubit region using classical shadows with two independent, uniformly random product single-qubit Clifford layers around a controllable unitary acting on that region, followed by computational-basis readout. In this architecture, we prove that a circuit $U_*$ of two collective Pauli rotations attains the exact minimum squared shadow norm over all Clifford choices of the controllable unitary for every $k\geq1$, namely $C_k^{\rm opt}=4\cdot 9^k/(3\cdot 5^k+3^k-6)$ for odd $k$ and $C_k^{\rm opt}=4\cdot 9^k/(3\cdot 5^k-3^k+2)$ for even $k$. The asymptotic optimum $(4/3)\cdot(9/5)^k$ establishes Wu et al.'s exponential base as optimal and improves their prefactor from $2$ to $4/3$. For $1\leq k\leq6$, exact rational certificates prove that the same optimum holds over the entire unitary group $\mathrm U(2^k)$, including all non-Clifford unitaries, within the same measurement architecture. These finite-size results motivate the conjecture that $U_*$ is optimal over all unitaries for arbitrary $k$ within this architecture.

quant-ph↗

arXiv2026

Neural Correlation Learning for Quantum-Enhanced Sensing with Time-Independently Driven Rydberg Atom Arrays

Quantum-enhanced sensing typically relies on preparing specific entangled states. However, engineering these states in scalable many-body systems remains challenging due to the difficulty of designing robust preparation protocols and accounting for realistic noise. We show that evolving a simple product state under native time-independent Hamiltonian dynamics in Rydberg atom arrays generates complex many-body correlations that can potentially enable quantum-enhanced parameter estimation reaching the Heisenberg limit (HL). The key challenge is then shifted to extracting the metrological information encoded in spatial correlations of measurement patterns. We introduce a neural correlation learning framework that combines Bayesian inference with neural networks trained on calibration data. Operating in a two-stage calibration-and-sensing protocol, we numerically demonstrate that this framework effectively extracts many-body correlations to saturate the Cramér--Rao bound dictated by the classical Fisher information. Consequently, it achieves quantum-enhanced sensitivity surpassing the standard quantum limit (SQL). Furthermore, the learned estimator is robust to realistic noise, and remains accurate in the absence of rare measurement patterns. Our results establish a hardware-efficient and scalable paradigm for quantum sensing in interacting many-body systems without requiring engineered entangled states.

quant-ph↗

arXiv2026

Protecting Sensitive Data in Image Synthesis via PAC-Private Adaptation for Diffusion Models

Synthetic data are increasingly used as an alternative to sharing sensitive records. However, synthetic data generation does not guarantee privacy, as diffusion models trained or adapted on sensitive data remain susceptible to reconstruction attacks. Moreover, while approaches that use differential privacy (DP), such as DP-SGD, achieve provably private diffusion model training, the repeated gradient clipping and noise injection they require result in significant utility loss. An important limitation of DP-based privacy is that, although it has a provable relationship to reconstruction privacy (RP), that relationship is indirect. RP is defined in terms of limiting how much an adversary's posterior distribution over sensitive data differs from the prior, whereas DP provides guarantees by bounding the sensitivity of outputs to changes in individual records. This indirection is an important source of the utility loss. To address this, we propose a PAC-private diffusion model adaptation to achieve reconstruction privacy. Since PAC-privacy is defined directly with respect to posterior advantage over the prior, it directly implicates RP. To obtain scalable PAC privatization in high dimensions, we first learn a compact data-dependent diffusion model component using LoRA or Textual Inversion, and then calibrate anisotropic Gaussian noise from the covariance of repeated mechanism outputs. Unlike DP-SGD, our method perturbs the learned component only once after optimization, thereby avoiding privacy composition across gradient updates. We evaluate the framework on few-shot concept personalization and full-dataset image synthesis, and show that the proposed approach better preserves subject identity, generation quality, and downstream classification accuracy than DP while achieving the same reconstruction privacy.

cs.LG↗

arXiv2026

Wells exact sequences for $n$-Lie algebras

We study derivations and automorphisms of arbitrary non-abelian extensions of $n$-Lie algebras through their full crossed-product data. Their relative center is a canonical coefficient module, and a third cohomology class obstructs completion of data satisfying the mixed Filippov identities. When this class vanishes, a relative second cohomology space classifies the remaining extension data. We prove a non-abelian good criterion, construct the Lie algebra and group of compatible pairs, and obtain derivation and automorphism Wells sequences with explicit lifting equations. The derivation obstruction is a Lie algebra $1$-cocycle; the additive automorphism obstruction is a crossed homomorphism. We distinguish the ordinary center from the relative center, and outer derivations of the kernel from the mixed action required. Sparse crossed products are treated, including the representation class of their second mixed operation and the additional equations governing changes of sparse sections.

math.RA↗

arXiv2026

DAMP: Decay-Aware Mixed-Precision Recurrent-State Quantization

Complex reasoning and agentic applications increasingly rely on long-context inference, where growing KV caches increase both memory usage and decoding overhead. Hybrid models reduce these costs by combining Softmax Attention with Gated DeltaNet (GDN) or Kimi Delta Attention (KDA), which maintain fixed-size recurrent states. These states are commonly stored in FP32 and consume substantial GPU memory, while their updates are limited by memory bandwidth. Quantization can reduce both storage footprint and memory traffic, but we find that uniform INT8 and FP8 degrade complex reasoning accuracy, while INT4 and NVFP4 collapse it to near zero. To our knowledge, this is the first study of post-training recurrent-state quantization for GDN and KDA. Our analysis reveals that outliers in GDN and KDA states are concentrated in particular key channels and value dimensions. Learned decay influences how much quantization error is retained. We find that largely the same GDN heads and KDA key channels exhibit slow decay across tasks. Based on these insights, we propose DAMP, which jointly considers quantization error and decay-based error retention to select high-risk key channels offline. Under a fixed storage budget, it retains these channels in FP16 and stores the remainder in INT8. We evaluate DAMP on Qwen3.6-35B, Kimi-Linear-48B and Kimi-K3 across six reasoning and code generation benchmarks. At 9.9 bits per state value, DAMP maintains average accuracy close to FP32. In SGLang, DAMP reduces recurrent-state storage by 69.1%, accelerates the recurrent-state update kernel by up to 2.59x , and lowers full-model time per output token by up to 19.0%.

cs.LG↗

arXiv2026

QuantMLA: Function-Aligned Dual-Path Quantization for Low-Bit MLA KV Caching

Multi-Head Latent Attention (MLA) enables expressive multi-head attention with compact caches for its content and decoupled RoPE paths, yet cache memory still scales linearly with context length and batch size. In this work, we establish a systematic model of MLA's dual-path quantization errors, characterizing their distinct effects on attention-output distortion and explaining the pronounced amplification of RoPE-path errors. Guided by this analysis, we introduce QuantMLA, a function-aligned framework for low-bit dual-path quantization. We derive path-specific transformation spaces that preserve full-precision computation while remaining fully fusible into model parameters offline, eliminating online transformation overhead. Within these spaces, QuantMLA learns path-specific transformations with function-aligned objectives: attention-output reconstruction captures the content path's coupled matching and aggregation errors, while positional QK reconstruction preserves the RoPE-induced component of the attention logits and admits a theoretical bound on output distortion. Across four MLA model families, QuantMLA enables, to our knowledge, the first reported joint INT4 caching of the content and RoPE caches with minimal accuracy degradation. Further compressing the content cache to INT2 while retaining the RoPE key cache at INT4 maintains competitive performance on challenging reasoning and code benchmarks. We develop a native low-bit MLA attention kernel that integrates unpacking and dequantization directly into attention computation. The physical cache layout provides 3.59x compression at 128K context, while a cache-pressure serving workload achieves 5.168x higher whole-job output throughput than BF16. The code will be released upon acceptance.

cs.LG↗

arXiv2026

Soliton resolution conjecture for the Calogero--Moser derivative nonlinear Schrödinger equation in the scaling-critical space

We prove the soliton resolution conjecture for the Calogero--Moser derivative nonlinear Schrödinger equation in the scaling-critical space $L^2_+(\mathbb R)$. For a natural class of initial data including a broad spectral Dini---moment class, global flow--Lax admissible solutions decompose into finitely many explicit modulated solitons plus a dispersive radiation, while finite-time blow-up solutions resolve into quantized zero-carrier $R$-bubbles and a strongly convergent endpoint remainder. This work extends soliton resolution to the optimal critical regularity. Previous results had required weighted $H^{1,1}$ initial data, and in the global case the solution was additionally required to remain in $H^{1,1}$ for all time, which amounts to an extra spatial decay assumption. One key ingredient in our argument is a new modular operator-theoretic framework that separates the discrete and continuous spectral channels and identifies the radiation profile through the distorted Fourier transform. A second ingredient is the development of several independent rigidity criteria for the discrete soliton profiles, which avoids inverse scattering and handles embedded eigenvalues. Our framework also yields a unified description of both global and finite-time asymptotics, with exact mass partition and mass-defect quantization.

math.AP↗

arXiv2026

Mean Field Games and Control on Large Expander Graphs

This paper investigates mean field games on sparse networks. In the case of large expander graphs, the limit topologies are analyzed using the graphexon framework, which characterizes sparse connections. We prove that the associated sequence of discrete averaging operators converges strongly to a continuous operator and this is illustrated in the development of infinite limits of Gabber-Galil-Margulis expander graphs \cite{gabber1981explicit}. These properties enable the formulation and existence proof of equilibria for linear-quadratic mean field games in which each agent is identified by a spatial network label $α\in X$ and only interacts with the neighborhood average characterized by the operator $\mathcal{G}$, i.e., the average state of a limited number of connected neighbors in a large expander graph. Furthermore, algebraic conditions induced by the spectral gap of $\mathcal{G}$ for the global asymptotic stability of the closed-loop system are established, and parameter thresholds that give rise to a Turing-type topological instability are established.

math.OC↗

arXiv2026

LatCom: Cross-Agent Latent Compression for Efficient Multi-Agent Collaboration

LLM-based multi-agent systems (MAS) increasingly use latent collaboration to avoid the information loss and repeated encoding-decoding overhead of natural-language communication. However, directly forwarding all sender latents makes the receiver-side context scale with both the number of agents and the reasoning length, increasing computation, memory usage, and collaboration latency. A natural solution is latent compression. But we find that cross-agent redundancy remains unresolved in existing latent compression approaches, which typically compress each sender independently and then concatenate the results. We propose LatCom, a cross-agent latent compression framework for efficient multi-agent latent collaboration. LatCom maps multiple sender latents into a fixed number of receiver-readable and task-relevant slots. Rather than reconstructing all sender hidden states, it optimizes the compressed latents for receiver-side task utility. LatCom trains the compressor in two stages: single-sender readability learning first establishes a latent interface interpretable by the frozen receiver, and multi-sender fusion learning then trains the compressor to fuse complementary evidence and remove redundancy across agents. Experiments on multiple benchmarks with Qwen3-4B show that LatCom achieves an average 2.46x inference speed-up over LatentMAS and reduces output token usage by 70.3% while maintaining comparable average accuracy.

cs.CL↗

arXiv2026

Bridging User Feedback and System Diagnosis: Reproducing Mobile Performance Issues from Reviews

Mobile application performance is a vital factor for user experience. Yet, performance issues are notoriously difficult to detect in development environments, where they often manifest less conspicuously, making their diagnosis more challenging. In this setting, app reviews from end users across diverse devices and usage contexts can provide timely and context-rich information about emerging performance issues. However, unlike structured bug reports, app reviews are written by end-users and tend to be more ambiguous, with individual reviews often providing only partial descriptions of the underlying issue. To bridge this gap, we present RevPerf, the first approach to automatically reproduce mobile application performance issues by leveraging and synthesizing information from app reviews. RevPerf retrieves complementary reviews via semantic retrieval and uses prompt engineering to integrate them, enriching the original review with performance issue details. An execution agent is then employed to generate and execute commands to reproduce the issue. After executing all necessary steps, the system incorporates multifaceted detection methods to identify performance issues by monitoring Android logs, GUI changes, and system resource utilization during the reproduction process. Experimental results demonstrate that our proposed framework achieves a 72.73% success rate in reproducing performance issues on the constructed dataset, outperforming the best baseline by 27.28 percentage points.

cs.SE↗

arXiv2026

From Experience to Expertise: Adoption-Aware Memory Learning for Data-Scarce NPU Kernel Synthesis

High-performance kernels underpin efficient accelerator execution but require expert tuning and lengthy manual optimization cycles. LLM coding agents promise automation, yet their CUDA knowledge transfers poorly to data-scarce domain-specific architectures (DSAs) such as NPUs, whose execution models and memory hierarchies differ substantially from those of GPUs. To address this transfer gap, post-training methods adapt LLMs to NPU programming but depend on scarce expert data and substantial training compute. Memory-learning agents instead adapt through external memory, but their uniform credit assignment gives adopted and unused experiences the same reward target, potentially biasing subsequent retrieval rankings. Moreover, when learned values guide only retrieval, high-value experiences that generalize across operators must be retrieved repeatedly rather than retained in context, thereby increasing retrieval overhead and weakening cross-task guidance. We therefore present SAGE, a persistent self-improving agent for NPU kernel synthesis. Adoption-Traced Utility estimation (ATU) combines explicit adoption records with kernel evaluation outcomes for adoption-aware credit assignment. Utility-Gated Consolidation (UGC) uses positive utility and repeated adoption across operators to select and abstract reusable rules into a bounded resident context. On NPUKernelBench, SAGE achieves a 95.5% execution rate versus 84.1% for the strongest controlled baseline, with 86.9% of solved operators outperforming torch_npu. With GLM-5.3, SAGE achieves a 43.99x speedup over the torch_npu reference on sparse flash attention. These results show that adoption-aware credit assignment and selective consolidation enable agents to accumulate and reuse hardware-specific knowledge across tasks.

cs.LG↗

arXiv2026

A combinatorial approach to point-flat incidences over finite fields

We establish new bounds for incidences between a point set P and a family L of n-flats in PG(n+d,q). For fixed dimensions, our bound on the incidence discrepancy has an explicit piecewise-linear exponent in $\log_q|L|$, improving the classical estimate of Haemers and the Kong-Tamo bound in specified ranges of the number of flats. Matching constructions establish sharpness up to constant factors in several parameter ranges. The proof is combinatorial and avoids spectral and Fourier analytic methods. As applications, we obtain improved estimates for rich flats and exceptional orthogonal projections, together with stronger lower bounds for Furstenberg sets in certain ranges where the fraction of prescribed directions is small.

math.CO↗

arXiv2026

Packet-Level In-Network Semantic Adaptation for Unstable Mobile Emergency Networks

Mobile emergency networks can experience independently changing intermediate wireless links on timescales shorter than endpoint feedback can track. When an egress changes after packet emission, feedback affects only later source data, while the on-path node observes the current condition with the affected packet still mutable. This paper presents DINA, a packet-level in-network semantic adaptation method. An image is divided into self-describing spatial packets carrying coordinates, a current representation identifier, and payload. At each eligible node, an offline-trained frozen selector scores compatible operators, immediately transforms the packet, and forwards it without image reconstruction or cross-packet adaptation state. Later nodes can retain or further compact the packet through the same typed compatibility contract. The receiver places available packets by coordinate, fills missing regions with black, and runs a fixed machine task. We realize DINA in a 24-node UAV environment using XDP and AF_XDP. In the primary forest-fire trace, DINA raises deadline tile coverage from 40.4% to 72.6% and classification accuracy from 77.5% to 95.0% relative to forwarding. In an independently trained RescueNet segmentation case, it raises coverage from 65.6% to 91.7% and foreground mIoU from 0.486 to 0.541. Sufficient- and extreme-capacity profiles expose a no-gain boundary and a common task-failure boundary, respectively.

cs.NI↗

arXiv2026

High Dynamic Range Video Reconstruction from Single-Exposure Raw Sequences

Due to the limited dynamic range of conventional image sensors, captured low dynamic range (LDR) video often suffers from highlight clipping and shadow detail loss, making high-quality high dynamic range (HDR) reconstruction from single-exposure sequences highly challenging without alternating exposures or extra hardware. Alternating-exposure HDR methods sacrifice frame rate and struggle with motion alignment, making them impractical for real-world capture. To address this, we propose RawHDRV, an end-to-end framework for single-exposure Raw video HDR reconstruction, that fundamentally exploits the linear response and channel-specific characteristics of Bayer data. Specifically, it features a channel-decomposition temporal alignment and fusion strategy that processes Bayer channels separately to exploit their distinct exposure characteristics, together with exposure-aware weighted fusion. It further incorporates an exposure complementarity mask-guided restoration module that leverages inter-frame exposure redundancy to adaptively fuse reliable information and suppress saturation artifacts, and introduces a mask-guided color loss that combines normalized error constraints with gradient smoothing to enhance highlight recovery. Furthermore, we construct a large-scale mobile Raw-HDR video dataset with per-frame HDR annotations. Experiments show that our method achieves the state-of-the-art results in all metrics, demonstrating superior spatial quality and temporal stability under extreme exposure conditions. The code is available at https://github.com/supeixian/RawHDRV.

cs.CV↗

arXiv2026

3dgs-sc: a controlled static screen-content benchmark for 3d gaussian splatting

Does high image fidelity imply readable screen content in 3D Gaussian Splatting (3DGS)? We introduce 3DGS-SC, a controlled static screen-content dataset and benchmark for examining this mismatch. Ten procedural scenes provide fixed multi-view splits, exact cameras, screen masks, text boxes, and transcripts. The protocol separates whole-image fidelity, screen-region fidelity, OCR readability, and edge preservation. In the reported five-method comparison, LightGaussian exceeds Mip-Splatting in screen PSNR by only 0.08 dB, yet trails it in OCR accuracy by 12.7 percentage points. Across all ten method pairs, screen-PSNR and OCR orderings disagree in seven cases. These aggregate results expose a method-selection failure of fidelity-only evaluation. Complementing prior text-aware 3DGS research, 3DGS-SC targets controlled monitor interfaces and exact annotations; scene-wise robustness and acquisition effects remain open validation questions.

cs.CV↗

arXiv2026

Co-occurrence Patterns of LoRA Adapters in Production Diffusion Model Inference Services

Low-rank adaptation (LoRA) has become a key technology for serving large-scale personalized large language models and diffusion models in the cloud. However, the co-occurrence patterns, resource contention relationships, and evolutionary regularities of adapters under production inference workloads have not been systematically or quantitatively studied. Based on GenTD26, Alibaba's production diffusion model inference dataset, this paper adopts a graph-theoretic framework to construct an adapter co-occurrence network and conducts a characterization from both static structure and dynamic evolution. Our main findings are as follows. (1) The co-occurrence network is extremely sparse, and adapter usage frequency follows a significant heavy-tailed distribution. (2) Introducing the first adapter incurs a 66.1% execution-latency overhead, with diminishing marginal costs afterwards. (3) Co-occurrence relationships are driven by base models: in 90.6% of multi-adapter requests, all adapters share the same dominant base model; 66.2% of significant co-occurrence edges connect same-model adapter pairs; and in 85.8% of multi-adapter requests, all adapter pairs form significant co-occurrence edges. (4) The adapter ecosystem exhibits a core-periphery bipolar structure, with a weekly Jaccard similarity of 0.696 at the model level and a churn rate of 54.5% for the top-10 hottest models within a 12-hour window. Based on these findings, we propose a preloading strategy built on top-k co-occurrence statistics; offline experiments show that it covers 81.0% of test-set co-occurrence pairs at k=3, and sensitivity analyses across frequency thresholds and time windows verify the robustness of the conclusions. These results provide a data-driven basis for cache preloading, adaptive scheduling, and GPU memory management in LoRA inference services.

cs.DC↗

arXiv2026

Bayesian Filtering in Physical Systems via Test-time Trained Flow Matching

Bayesian filtering provides a principled framework for online state estimation under uncertainty, yet its application to systems with high-dimensional states and complicated posterior distributions remains challenging. Recent generative models, such as flow matching, have shown potential in Bayesian filtering. However, they still rely on particle-based representations of the posterior, which lose the rich information of the full distribution, or tackle a trajectory-level inverse problem that conflicts with the recursive structure of Bayesian filtering. To address this, we propose a new perspective of directly encoding the evolving distribution into flow matching model weights, namely, the Belief Flow Filter (BFF). It is a generative filtering framework that updates model weights via gradient descent at test time to track the posterior evolution. Thereby, BFF bypasses the scalability issue of particle representations or the flexibility limitation of Gaussian assumptions in conventional filters. We theoretically justify that the BFF design is structurally aligned with Bayesian filtering, and its training objective targets the recursive filtering operator. BFF is empirically verified across 5 different physical systems, including ones with chaotic dynamics and highly sparse, non-linear observations. The results show that BFF attains the best score in 8 of 9 metric-benchmark cells across the three standard 1D and 2D PDE benchmarks, and similarly leads on the extreme single-moving-sensor setting and on a real-world-grounded tokamak plasma estimation task, demonstrating its potential to accurately approximate the Bayesian filtering operator in high-dimensional probability space.

stat.ML↗