Searcharxiv⌕ Search

arXiv subjects

Hao Zhang

Publications and source records attributed to Hao Zhang.

At least 37 records · Page 2Linked to original sources
arXiv2026

JPO: Juris Policy Optimization for Structured Legal Reasoning in Criminal Judgment Prediction

Criminal judgment prediction requires models to infer statutory articles, charges, and sentencing outcomes from case facts. Unlike standard classification tasks, it involves a structured reasoning process in which statutes should be matched with facts, charges should be justified by statutes, and sentencing outcomes should remain consistent with charges. Existing approaches optimize final labels, and while some have attempted to evaluate reasoning quality, their evaluations are indirect, often relying on LLM-generated rubrics that reflect model-internal preferences rather than the inherent logical structure of legal adjudication. We propose Juris Policy Optimization (JPO), a post-training framework for structured legal reasoning in Chinese criminal judgment prediction. JPO first uses teacher-generated rationales to supervise a standardized four-step reasoning process, and then applies reinforcement learning with a composite reward over legal prediction quality, reasoning structure completeness, and cross-step consistency. JPO further introduces token-level advantage reweighting and adaptive clipping for legally salient reasoning segments. Experiments on multiple open-source language models and three Chinese legal benchmarks show that JPO consistently improves both judgment prediction and reasoning quality over supervised fine-tuning and reinforcement learning baselines.

cs.CL↗
arXiv2026

RF-Agent: Hierarchical Language-Agent Control for Instruction-Conditioned Active Spectrum Sensing

Active radio-frequency (RF) sensing must acquire evidence under receiver limits while following confirmation, recovery, and stopping instructions. \newhl{We propose RF-Agent, a closed-loop architecture combining language supervision, episode memory, deterministic RF execution, perception feedback, and evidence-based reporting.} An independent auditor checks trajectory compliance. We derive acquisition and supervisory-request bounds, characterize full-history token complexity, and bound joint success by output validity and evidence capacity. \newhl{ActiveRF-AD tests instruction changes at fixed states and evaluates joint success, requiring compliance and RF correctness in the same episode.} \newhl{Training on different instructions at shared sensing states raises counterfactual-pair accuracy from 24.03\% to 96.22\% and Core joint success from 65.13\% to 72.12\%. Of the 6.99-point joint-success gain, 6.73 points reflect a decrease in the fraction of all episodes that are RF-correct but noncompliant.} \newhl{At Core budget three, paired-training hierarchical control achieves 66.60\% joint success versus 9.17\% for direct-action control.} \newhl{In a separate comparison at the same budget, RF-Agent improves joint success by 2.31 points and reduces RF acquisitions by 12.3\% relative to an initial-state planner, at increased inference cost.} Cross-backbone diagnostics cover three model families. These results show how instruction-responsive control improves task completion beyond what RF accuracy alone reveals.

eess.SP↗
arXiv2026

Deep Learning Reforms Image Matching: A Survey and Outlook

Image matching, which establishes correspondences between two images to recover 3D structure and camera geometry, is a cornerstone of computer vision and underpins a wide range of applications, including visual localization, 3D reconstruction, and simultaneous localization and mapping (SLAM). Traditional pipelines, composed of a detector-descriptor, a feature matcher, an outlier filter, and a geometric estimator, falter in challenging scenarios. Recent advances in deep learning have substantially improved both their robustness and accuracy. This survey reviews how deep learning has progressively transformed the classical image matching pipeline. Our taxonomy is aligned with the traditional pipeline and covers two aspects: i) replacing individual steps with learnable alternatives, including learnable detector-descriptors, outlier filters, and geometric estimators; and ii) merging multiple steps into end-to-end learnable modules, including middle-end sparse matchers, end-to-end semi-dense/dense matchers, and pose regressors. We first examine the design principles, advantages, and limitations of both aspects, and then benchmark representative methods on relative pose estimation, homography estimation, matching accuracy assessment, visual localization, and 3D reconstruction. Finally, we discuss open challenges and directions for future research. By systematically categorizing and evaluating learning-based methods, this survey offers a clear overview of how image matching is evolving and where further progress is needed. The project repository is available at https://github.com/ZizhuoLi/awesome-image-matching-survey.

cs.CV↗
arXiv2026

JetSpec: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting

Speculative decoding (SD) accelerates autoregressive Large Language Models (LLMs) by drafting multiple tokens and verifying them in parallel, but it faces a scaling limitation: increasing the draft budget improves speed only when acceptance remains high and drafting overhead stays low. This ceiling has been difficult to break because prior head-based SD methods face a causality-efficiency dilemma. Autoregressive drafters produce path-conditioned candidates that are effective for tree speculative decoding with higher acceptance length, but their drafting cost grows with tree depth. Bidirectional block-diffusion drafters generate all positions in one pass, but their branch-agnostic marginals can form individually plausible yet mutually inconsistent trees, wasting budget and reducing acceptance. We propose JetSpec, a head-based SD framework that combines one-forward drafting efficiency with branch-wise causal conditioning. JetSpec trains a causal parallel draft head over fused hidden states from the frozen target model, producing candidate trees whose scores align with the target model's autoregressive factorization. This enables JetSpec to convert larger draft budgets into longer accepted prefixes and higher end-to-end speedup. Across math, coding, and chat benchmarks on dense and MoE Qwen3 models, JetSpec consistently outperforms bidirectional-head and tree-based SD baselines. On H100 GPUs, JetSpec achieves up to 9.64x speedup on MATH-500 and 4.58x on open-ended conversational workloads, with further latency gains demonstrated through vLLM integration under realistic serving loads. Our code and models are available at https://github.com/hao-ai-lab/JetSpec.

cs.CL↗
arXiv2026

CSIR: Contextually and Socially Informed Robots for Efficient Person Goal Navigation

We address the Person Goal Navigation (PersonNav) problem, enabling robots to search for people under realistic constraints on information accessibility and language uncertainty. This tackles the current solutions revolving around rigid person finding systems requiring exact knowledge of the individual for item-delivery in indoor settings. While the focus in literature is on learned methods using unavailable public or sparse data due to human privacy. We propose a planning framework that combines distance and semantic information about the recipient (habits and intent) weighted by the trust of the user-provided information. A synthetic benchmark of scenarios is also developed including actors, items, and requests to evaluate performance before real-world deployment aiming to simulate natural language human-robot interactions. Results show our informed search outperforms classical distance-based graph baselines, while semantics alone lead to ungrounded, sporadic search. Our method achieves strong gains over baselines, with an LLM-based variant performing comparably, and its explicit belief representation naturally supports future Bayesian filtering. Hardware tests demonstrate our method supports real-world embodiment able to leverage between semantic and distance information in a real setting. This work moves toward more intelligent mobile service agents capable of human-like, informed search in realistic environments to be leveraged in day-to-day use. https://anonymous.4open.science/r/personnavsite-4ED2/index.html.

cs.RO↗
arXiv2026

World Embedding Benchmark

Physical fidelity has received increasing attention in world models and video generation, yet how video representations encode physical information remains less understood. We introduce the World Embedding Benchmark, comprising 8,000 controlled simulation cases from 80 families spanning fluid mechanics, solid mechanics, dynamics, and optics & electromagnetism. Each case pairs a rendered video with simulation-derived physical annotations, supporting three complementary tasks: text-video retrieval, physical-property regression, and multiple-choice video-description pair classification. We use these tasks to distinguish cross-modal physical alignment from the recoverability of quantitative physical information. Evaluated pre-trained omnimodal embedding models show weak retrieval and near-chance within-family pair classification, while lightweight probes recover useful physical information from frozen video embeddings. Continual contrastive training with physics-specific video-text pairs improves retrieval and pair classification but degrades physical-property regression, revealing a trade-off between alignment and quantitative information recoverability. Finally, we use the embeddings to retrieve reference videos for retrieval-augmented generation with MiniMax-H3. Retrieved references improve the physical fidelity of generated videos, with stronger retrieval models yielding larger gains in our experiments. Together, these findings highlight the need to evaluate physical alignment and property recoverability jointly, and demonstrate the utility of physical representations for improving video generation.

cs.CV↗
arXiv2026

LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization

Recent world action models (WAMs) reuse pretrained video VAEs whose encoder latents directly condition downstream action policies. Quantization must therefore preserve not only reconstruction fidelity but also the policy-facing latent contract expected by the frozen policy. Direct NVFP4 leaves W4A4 quantization error uncompensated, whereas joint quantization-aware training (QAT) can recover reconstruction by moving this representation. On Wan2.1, joint QAT nearly matches FP32 VBench-7 (0.7403 versus 0.7409), yet LIBERO success collapses from 95.5% to 10.5%. Controlled decoder-only experiments show that activation quantize-dequantize operations alter the reconstruction signal and decoder Jacobian, redirecting the gradient returned to the encoder and inducing persistent latent drift. Based on this mechanism, we introduce LatentQuant, a two-stage NVFP4 QAT framework that first aligns the quantized encoder with its high-precision counterpart, then freezes it while adapting the decoder. Across Wan2.1 and Wan2.2, LatentQuant preserves near-baseline control and high reconstruction quality, achieving 95.75% success on LIBERO and 68.8% on RoboTwin. On NVIDIA B300 GPUs, NVFP4 execution achieves 1.17x-1.26x end-to-end VAE speedups over BF16 cuDNN.

cs.AI↗
arXiv2026

Nanoscale Electroviscous Lift Force

About forty years ago, it was predicted that a charged particle, moving parallel to a charged wall in an electrolyte, should experience a lift force that, contrarily to electrostatic forces, is not screened at large distances. Up to now, such electroviscous lift force has not been directly measured. Here, we use Atomic Force Microscopy to directly measure the electroviscous lift force and quantify its dependency with the distance to the wall, the translation velocity or the particle's size. Observing that existing theories exhibit large discrepancies with our experimental observations, we develop an analytical approach combining lubrication theory to a previously introduced formalism for small screening length. The experimentally observed lift forces are in good agreement with our theoretical predictions and reveal, for the first time, a saturation of the lift force for increasing velocities. Altogether, our results characterize, through direct measurements and analytical approach, the properties of electroviscous forces between charged particles in viscous electrolytes in non-equilibrium conditions.

cond-mat.soft↗
arXiv2026

FastKernels: Benchmarking GPU Kernel Generation in Production

LLM-based agents for GPU kernel generation are advancing rapidly, but the benchmarks they optimize against evaluate kernels in isolation, with synthetic inputs and weak baselines, rewarding sandbox speedups that break or vanish in real inference systems. We introduce FastKernels, a benchmark of 384 tasks drawn from 47 representative architectures across 8 categories, whose kernels suffice to reimplement 94.6% (472/499) of HuggingFace Transformers architectures with outputs matching the native implementations. Each task mirrors the interface of the corresponding production module and is scored against the kernels production frameworks ship, and tasks form a compositional hierarchy, from primitives to full models, in which higher-level modules import lower-level ones. Candidates are scored at the kernel level and end to end inside the models they come from, on the production execution path, and MacroEval aggregates calibrated correctness, coverage, and speedup into a leaderboard. Seeding it with five representative agents (6,900 agent-hours), we find that kernel-level speedups of $1.6$-$6.6\times$ shrink to at most $1.25\times$ end to end, only 20% of winning kernel sets run correctly as-is, and kernel-level scores mis-rank agents: Claude Code matches or beats KDA at every level in isolation, yet KDA scores $3\times$ higher end to end. Code is available at https://github.com/Snowflake-AI-Research/fastkernels.

cs.LG↗
arXiv2026

4Director: Controlling Video World Models with Rigid 3D Geometry

Precise control over camera and object motion is essential for professional video production. Existing methods control objects only coarsely, through image-plane cues that are ambiguous in depth and rotation or through 3D tracks and blobs that lack complete geometry and lose consistency across viewpoint changes. We introduce 4Director, a video world model conditioned on an explicit 4D scene representation: each object is reconstructed once from the input image as a canonical mesh and moved by one prescribed rigid transformation per frame. This representation provides an intuitive 3D control interface and prevents unobserved geometry from being regenerated independently in every frame. We render the controlled scene as a depth video and introduce a Motion Adapter that transforms this geometric scaffold into video while synthesizing view-consistent appearance, illumination, and non-rigid dynamics. For training, we construct RealCOD-Rigid, a new dataset of 20,774 clips annotated with rigid 3D scenes by our automatic pipeline. We further introduce Identity-Gated IoU (IG-IoU), which jointly evaluates adherence to prescribed object motion and preservation of object identity. Experiments demonstrate that 4Director consistently outperforms prior methods in visual quality and in camera and object control.

cs.CV↗
arXiv2026

Agora: Git as Shared Memory for Collective AutoResearch

Research agents working in separate sessions need to know what others have tried and which results they can build on. Agora stores their contributions as an append-only directed acyclic graph (DAG) in Git. Each commit records a result, insight, hypothesis, verification, or report and links it to prior work. Searchable views show leading results, neglected branches, and verification status; diversity-aware recommendations suggest experiments beyond the current leaders. We report a run of nearly 12 days in which 13 language-model workers, with no assigned tasks or central planner, used Agora to solve a weight-transfer problem. Given 141 pretrained donor models and a frozen 119.6M-parameter attention--SSM hybrid whose dimensions match no donor, the workers had to initialize the target without training data or gradient updates. They published 1,703 contributions and reduced the development evaluator score from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. The best method compresses donor next-token statistics into the target's embedding and output head, then adds short-range context through sparse edits to attention, feed-forward, and state-space blocks. Its 145-commit ancestry spans 15 accounts. Participants also posted 165 verifications of 95 targets, each by an account other than the target's author, with no reported failures. The run documents how agents reused and verified shared work. Measuring the effect on discovery per unit of compute requires a matched comparison.

cs.LG↗
arXiv2026

How Can Recommendation Feedback Evolve Agent Memory?

Content-generation agents continuously receive impressions, clicks, conversions, and negative feedback from recommendation systems, providing real-world outcome signals for memory evolution. However, these signals are delayed and noisy, confounded by audience composition, placement, and recommendation policies, and may result from the combined influence of multiple memories, making accurate attribution difficult. Existing methods rely primarily on immediate feedback or semantic retrieval and therefore struggle to reliably translate recommendation outcomes into memory fitness. To address this challenge, we propose TIDE (Trajectory-Informed Directed Memory Evolution), an external memory evolution framework driven by delayed recommendation feedback. We further introduce Memory Evolution Gain (MEG), which measures the utility improvement of evolved memory over a no memory baseline on strictly future tasks. TIDE treats memory as a capacity-constrained population of experiences: temporal and semantic credit assignment estimates contextual fitness, while responsibility credit distributes outcome signals according to the memories referenced during generation. These signals are then used to reinforce, crossover, mutate, or evict memories. On an e-commerce membership marketing content-generation agent, TIDE achieves a +7.75-percentage-point MEG in offline temporal replay and significantly improves both unique click-through rate (UCTR) and activation rate in an online A/B test. On a delayed-label benchmark, TIDE achieves the lowest mean absolute error (MAE) and root mean squared error (RMSE) and the highest MEG among the compared methods, demonstrating its effectiveness.

cs.AI↗
arXiv2026

A macroscopic-shadow-corrected lattice Boltzmann method for fast, time-accurate simulation of low-Reynolds-number transient flows

Explicit lattice Boltzmann simulations of slow transient flows are constrained by the acoustic time step. Dual-time stepping can remove this restriction, but slow inner convergence has limited reported wall-clock gains to roughly four- to tenfold, without a mechanism that strengthens under refinement. We present the macroscopic-shadow-corrected lattice Boltzmann method (MSC-LBM), which applies defect correction to the unsplit kinetic residual. At each Fourier wavenumber, a Stokes-like system is solved exactly for the conserved residual moments while preserving the second-order backward-differentiation formula (BDF2) fixed point and temporal accuracy. Along the two-relaxation-time axis $Λ=1/4$, $ω^+=ω^-=1$ makes one collision eliminate all non-hydrodynamic perturbations; $η(ν)=|6ν-1|/(6ν+1)$ vanishes at $ν=1/6$, leaving hydrodynamic slow modes whose long-wave shadow is inverted by the corrector. In completed timing campaigns, MSC-LBM achieves 27.95- and 52.71-fold wall-clock speedups at $N=256$, $\mathrm{Re}=1$ under matched-accuracy and $1\%$ gates. The matched speedup rises monotonically to 122.02 at $N=1024$, with gains persisting across $\mathrm{Re}=10^{-4}$--$100$. The three-dimensional D3Q19 extension reaches 27.85 and 8.84 under the same gates at $N=128$, $\mathrm{Re}=1$. A bounce-back-consistent kinetic coarse solver with adaptive relinearisation extends MSC-LBM to fully enclosed cavities, yielding setup-excluding physical-march speedups of 26.86 at $\mathrm{Re}=10$ and 3.84 at $\mathrm{Re}=100$ for $N=128$. An exact per-wavenumber symbol inverse establishes the attainable off-design contraction envelope. The contraction, parameter-sweep, and timing results jointly delimit the demonstrated operating regime: low-to-moderate-Reynolds-number transients for which acoustic stepping need not dictate computational cost.

physics.flu-dyn↗
arXiv2026

Strike a Chord! Modal Kinetic Typography

We introduce modal kinetic typography, which animates a vector glyph to express a semantic concept while keeping it legible. Our key idea is to build motion from the glyph's natural vibration modes. Specifically, a finite-element eigenproblem assembled from the vector outline yields the glyph's softest modes, for the whole letter and for each of its parts, allowing it to bend. The problem's zero-energy solutions, i.e., rigid translations and rotations, are applied in closed form to each part, allowing parts to also move as blocks. To animate the glyph, a frozen video diffusion model supervises only the modes' amplitudes and phases. Our modal approach addresses two weaknesses of prior work. Free-form point optimization under video score distillation (SDS) moves each point and frame independently along noisy gradients, tearing the outline and causing jitter. In contrast, our modes are smooth along the outline and driven by a few whole-cycle harmonics, which restricts these gradients to smooth, seamlessly looping motion. On the other hand, structured alternatives rely on skeletons or keypoints from category-specific priors, whereas our modes come from the glyph itself; the only prior is a list naming each letter's moving parts, generated once for the whole alphabet by a language model. In modal kinetic typography, shape and motion are disentangled by construction: a single base outline is sculpted toward the concept, and the modal drive cannot alter it, so a letter can also be animated without being reshaped. Our method produces more articulated and smoother motion than Dynamic Typography and AniClipart at comparable or better concept alignment, with less glyph tearing than Dynamic Typography, and is preferred by human raters, including in a frozen-shape setting where motion alone must carry the concept. Our results were also preferred over Astra (GPT-6) by human raters.

cs.CV↗
arXiv2026

Where Speech Enhancement Hurts Recognition: An Inference Time Polar Projection Diagnosis

Speech enhancement (SE) can substantially improve perceptual quality, yet enhanced speech does not necessarily improve automatic speech recognition (ASR). Common explanations such as enhancement artifacts and over-suppression remain qualitative and do not localize which enhancement component affects recognition. We study inference-time polar projection, which transforms an STFT mask $M=Ae^{jϕ}$ into $M_{α,γ}=A^αe^{jγϕ}$ and independently varies magnitude strength and estimated phase correction with frozen SE and ASR models. The resulting response curves separate effects that are coupled in ordinary enhanced-vs-noisy comparisons. On VoiceBank+DEMAND with FRCRN, wav2vec2.0 improves from 17.13 to 10.20 WER near full magnitude correction ($α=.85$), while Whisper reaches its best WER under much weaker correction, improving from 6.44 to 5.57 at $α=.25$ and returning to 6.46 at full magnitude. A second enhancer, a robustness-oriented ASR model, and a reverberant WSJ0 mixture set generalize the pattern and show that the preferred magnitude region depends on the recognizer and acoustic condition. Estimated phase correction is not a stable ASR lever in these evaluations: its effect is smaller, setting-dependent, and provides no consistent gain once magnitude strength is fixed.

eess.AS↗
arXiv2026

KPI: A Promptable Kernel for Physical Interaction on Humanoids

Humanoids now walk, balance and reach with remarkable generality: one whole-body tracking policy follows references from a human, or from an end-to-end policy. That generality travels in the trajectory, and a trajectory alone carries limited information about the interaction it should produce: at contact, the executing controller determines how the robot behaves. Single-task policies usually reach hard interactions by optimising trajectory and controller together in simulation; general stacks usually assume a preset or hand-chosen controller. We present KPI, a promptable kernel for physical interaction between the trajectory source and an unmodified whole-body tracker. Instead of a controller fixed before the task, the trajectory source sends a contract: per direction, track, comply, or hold a force range. From tracking error and a wrench estimate, the kernel adapts the arms' stiffness, damping, reference and feedforward toward it at contact rate. We demonstrate KPI through an agentic framework: from one instruction, a vision-language agent writes both the reference trajectory and the contract, with no task-specific code. We demonstrate instruction-driven winch operation, door opening, and box transport, alongside scripted surface-interaction experiments. In the winch demonstration, the humanoid is able to turn a crank to hoist a second robot fully off the ground.

cs.RO↗
arXiv2026

SynthCoder: Anti-pattern identification and model training for FIM mode code completion

As a leading application of large language models (LLMs) in software engineering, Fill-in-the-Middle (FIM) mode code completion has drawn wide attention. Training such models requires masking code corpora, yet common strategies tend to be problematic: character-delimited random masking may yield many unrealistic cases (e.g., cutting keywords or identifiers), while purely AST-based masking cannot mask concurrent elements that span across multiple subtrees. These limitations easily create anti-patterns that rarely occur in real-world code completion, diminishing FIM performance. We introduce SynthCoder, which adopts optimized masking strategies better aligned with developers' expectations in FIM. Specifically, we first refine AST-level node masking and add heuristics that better mimic developers' expectations to construct the training corpora. Subsequently, SynthCoder-Seed and SynthCoder-Qwen, built upon Seed-Coder-8B-Base and Qwen2.5-Coder-7B respectively, employ a two-stage training pipeline, i.e., a curriculum-based fine-tuning stage followed by a Direct Preference Optimization (DPO) alignment stage with rejected code sampled preference data. Besides, to suppress erroneous context repetition, we include negative samples that duplicate existing code during DPO, mitigating such failures when the model fails in producing valid completions. Extensive experiments on Santacoder-fim-task, aiXcoder-FIM-Evaluation and CrossCodeEval benchmarks show that the models trained with our mitigation strategies improve over mainstream baselines on Exact Match (EM) and Edit Similarity (ES) for text-based FIM benchmarks, and on Pass@1 for Santacoder-fim-task which has test cases. SynthCoder also yields less code-echo and consumes fewer tokens at inference, yielding higher practical efficiency. Ablation studies further support the contribution of our optimized masking and repetition-suppression mechanisms.

cs.SE↗
arXiv2026

Swapping Quantum Annealing Errors into a Cavity

Quantum annealing is a promising form of quantum computation, but it slows down at a small energy gap $Δ$ during the anneal, and the errors it makes there remain undetected. We introduce cavity quantum annealing, in which a cavity mode of frequency $O(1)\ggΔ$, coupled mid-schedule, swaps these errors into photons. In the $p$-spin model, this reduces the annealing time to nearly the square root of that required without the cavity. Moreover, counting the photons certifies the result: high-photon runs, with an $O(1)$ yield, reach a ground-state fidelity above $99.99\%$ orders of magnitude sooner, a gain that grows exponentially with system size. A pre-loaded photon instead runs the swap in reverse, preparing an excited state. More broadly, our work shows that cavities can enhance computations performed with qubits, paving the way toward high-fidelity quantum state preparation by quantum annealing.

quant-ph↗