Searcharxiv⌕ Search

arXiv subjects

Feng Zhang

Publications and source records attributed to Feng Zhang.

At least 19 recordsLinked to original sources
arXiv2026

Which Long Documents Matter? A Controlled Study of Data Selection for Context Extension

Natural long documents are widely used for context extension, but which selection pipelines are effective remains unclear. We conduct a controlled comparison on Qwen2.5-3B at 128K under matched 6B-token continued-training budgets. Random sampling of natural long documents (random-long) outperforms repacking at longer evaluation lengths. Within the same candidate pool, selecting documents with high LongPPL key-token density yields further gains, unlike the evaluated generic-quality and LLM-judge selectors. We introduce truncLongPPL to reduce scoring cost by computing key-token density over a truncated prefix. At 128K, scoring the first half of the reference span yields downstream performance close to full-span scoring with a measured $2.28\times$ end-to-end scoring speedup. Across three training seeds, truncLongPPL achieves the highest mean scores among evaluated pipelines, exceeding random-long by 4.54 points on validation HELMET and 3.28 on held-out RULER, with gains most consistent in retrieval. At a 256K target context, it gains 2.82 average RULER points over random-long on Qwen2.5-3B. In short/long mixed training across Qwen2.5-3B, Qwen3-8B, and OLMo~3-7B, it gains 1.47--7.31 average RULER points over random-long mixing, with seven-task short-context averages within 0.23 points of the corresponding baselines. After supervised fine-tuning (SFT), Qwen3-8B and OLMo~3-7B models trained with selected long documents outperform their random-long counterparts by 1.13--1.67 points on LongBench v2 and 0.62--1.44 points on MRCR.

cs.CL↗
arXiv2026

AFCS: Atomic-Fact-Conditioned Synthesis for Factual Knowledge Pretraining

Synthetic data is increasingly used in language-model pretraining. Existing knowledge-corpus synthesis pipelines, including WRAP and Nemotron-CC, usually take a source passage or document segment as input and generate rewritten or QA-style samples from that text. This can improve surface form and prompt diversity, but it does not explicitly require fine-grained source facts -- especially secondary entities, list/table entries, and precise temporal attributes -- to appear in the synthesized corpus. We propose AFCS (Atomic-Fact-Conditioned Synthesis), a pipeline that first extracts source-supported atomic facts and then uses each retained fact to drive generation. AFCS uses structure-aware extraction over the document tree, source-grounded verification, local context expansion, and multi-format generation. In continued-pretraining experiments with OLMo-3-7B and a 32B base on the same FineWiki source corpus, AFCS outperforms WRAP and Nemotron-CC on SimpleQA, SimpleQA Verified, PopQA, and HotpotQA. Granularity ablations show that fact-level synthesis improves over chunk-level, paragraph-level, and entity-level synthesis; and a 100B-token mid-training integration improves factual benchmarks and raises the 12-task general benchmark average by +1.09.

cs.CL↗
arXiv2026

Autoregressive Retriever: Improving Query Understanding from Item Feedback for Universal Multimodal Retrieval

Universal multimodal retrieval typically encodes a query once and ranks independently indexed items by embedding similarity. This design supports efficient search, but leaves the query representation unchanged even when retrieved items could help clarify the information need. We introduce the AutoRegressive Retriever (ARR), a multimodal retrieval model that learns both to select informative items and to use their content to refine subsequent retrieval. ARR alternates between retrieving an item and updating the query embedding, then uses the final embedding to rank the collection. Supervised fine-tuning teaches the encoder to use feedback through stepwise contrastive supervision. Reinforcement learning treats feedback items as actions and optimizes their selection using the final reciprocal rank of a relevant item. A query-side adapter enables this optimization against a fixed item index. ARR demonstrates strong retrieval performance on both in-domain and zero-shot benchmarks, outperforming the compared baselines on average. Further analyses show that feedback improves retrieval at inference time and that training with feedback also improves the initial query embedding, before any item is observed.

cs.IR↗
arXiv2026

Quantifying geographic domain shift to decouple the geospatial transferability of human mobility flow generation models

Human mobility serves as an essential proxy for understanding social, economic, and environmental dynamics in urban systems. Geospatial transferability, which measures a model's capability in a new location or unseen region, is a critical dimension for comparing different human mobility generation models. However, few studies have studied the intrinsic characteristics of geospatial transferability. To this end, this study systematically investigates the geospatial transferability of four representative human mobility generation models using a large-scale benchmark dataset of census tract level commuting flows across 2265 counties in the United States. Inspired by the domain adaptation theory in machine learning, we introduce geographic domain shift to describe the intrinsic differences in geographic feature distributions and spatial structures between source and target regions, which may jointly affect model transferability. Moreover, we propose two metrics, mutual information and spatial shift, to quantify the geographic domain shift. To examine their associations with model transferability, we employ linear mixed-effects regression to analyze the associations between geographic domain shifts and transferability. Our results reveal substantial spatial heterogeneity and asymmetry in transfer performance across regions. Both information shift and spatial shift exhibit statistically significant and complementary explanatory power. This indicates that geospatial transferability depends not only on model design but also on intrinsic geographic differences. These findings provide a novel methodological framework for evaluating and improving the geospatial transferability of human mobility generation models and support more robust and fair human mobility data synthesis across diverse regions. It also offers insights on spatial transferability for GeoAI model development.

cs.AI↗
arXiv2026

Robust Tensor Completion via Reflective Convolution Nuclear Norm Minimization

Robust tensor completion recovers multidimensional data from partial observations corrupted by sparse gross errors. Existing convolutional low-rank models typically construct translated copies using circular continuation, which introduces artificial wrap-around neighborhoods for finite nonperiodic data. We propose reflective convolution nuclear norm minimization (RCNNM), which replaces circular shifts with endpoint-nonrepeating reflection. The resulting lifting has nonuniform entry multiplicities and satisfies a weighted Gram identity that supports both the recovery analysis and the optimization method. Under random sampling and sparse corruption, we establish high-probability exact recovery of the underlying tensor and sparse errors, together with stability under bounded dense perturbations. We further develop a two-block ADMM with a closed-form entrywise tensor update, while singular-value thresholding is implemented through the smaller right Gram matrix. Experiments on synthetic tensors, BSDS color images, and CAVE multispectral images show that RCNNM consistently improves over its circular-lifting counterpart, with the clearest gains near image boundaries. In particular, average boundary-PSNR improvements reach 3.26 dB while global reconstruction quality remains competitive.

cs.CV↗
arXiv2026

A 20-kV Optical-Injection-Enhanced 4H-SiC Reverse-Conducting IGBT With Snapback-Free Operation and Low On-State Voltage

This letter proposes a 20-kV 4H-SiC optical-injection-enhanced reverse-conducting IGBT (OIE-RC-IGBT) with snapback-free operation and low on-state voltage. Light-emitting diodes (LEDs) are integrated into the gate structure to provide optical injection during forward conduction. The emitted light generates a high concentration of photogenerated carriers in the carrier-storage layer (CSL) and drift region, thereby reducing the resistance of the JFET and drift regions and increasing the device current. Consequently, the collector-side P+/N junction is turned on earlier, effectively suppressing the snapback effect and facilitating the transition to bipolar conduction. After bipolar conduction is established, the photogenerated carriers further enhance conductivity modulation, resulting in a reduced on-state voltage. TCAD simulation results show that, compared with the conventional RC-IGBT (C-RC-IGBT), the proposed device achieves snapback-free operation while reducing the required number of MOS cells from 25 to 5. Moreover, the on-state voltage, turn-off fall time, and turn-off energy loss are reduced by 22.37%, 25.06%, and 19.69%, respectively. These results demonstrate the potential of optical injection for improving the conduction and turn-off performance of ultrahigh-voltage SiC RC-IGBTs.

physics.app-ph↗
arXiv2026

Hunyuan-A13B Technical Report

We present Hunyuan-A13B, an open-source large language model based on a Mixture-of-Experts architecture. It contains 80 billion total parameters but activates only 13 billion during inference, balancing model capability, computational efficiency, and deployment cost. The model is pretrained on a rigorously filtered 20T-token corpus with enhanced STEM data curation, improving factual reliability and reasoning ability. High-quality supervised fine-tuning and large-scale reinforcement learning further enhance its overall performance. Hunyuan-A13B also introduces a dual-mode Chain-of-Thought framework that adapts reasoning depth to task complexity: fast thinking for routine queries and slow thinking for complex, multi-step problems. Evaluations show competitive performance across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models. Its high inference throughput makes it suitable for latency-sensitive applications. We release Hunyuan-A13B to support open research and practical LLM deployment.

cs.AI↗
arXiv2026

Guaranteed Low-Rank Tensor Recovery from Modewise Measurements via Normalized Block-Weighted Riemannian Gradient Descent

We consider the recovery of low-multilinear-rank tensors from linear measurements and propose an adaptive block-weighted modewise Riemannian gradient descent method. The method combines memory-efficient modewise measurements with a normalized adaptive weighting strategy for the core and factor components of the Riemannian gradient. The weighting improves convergence without increasing the multilinear-rank bound of the search direction or the size of the reduced core used for retraction. Under the tensor restricted isometry property and a suitable initialization, we establish local linear convergence and derive sampling guarantees for sub-Gaussian and subsampled orthogonal with random sign (SORS) measurements. Numerical experiments on synthetic low-Tucker-rank tensors show that the proposed method reduces iteration counts and computational time while maintaining reliable recovery performance, especially near the recovery threshold and for structured SORS measurements.

cs.LG↗
arXiv2026

VEQDB: A Compact and Reconstructible Multi-Device Tokamak Equilibrium Database

Tokamak equilibria are commonly exchanged as gridded G-EQDSK files whose conventions, resolutions, and machine-specific formats impede cross-device comparisons and data-driven modeling. Here, we present VEQDB, an open, compact, and reconstructible fixed-boundary equilibrium database built on continuous MXH--Chebyshev geometry and independent physical-profile roots. By decoupling authoritative equilibrium physics from rectangular meshes, VEQDB enables continuous evaluation and metric differentiation at arbitrary application-demanded resolutions. Backed by an automated numerical validation pipeline, VEQDB is structured as an extensible repository for ongoing community expansion. Its inaugural release provides 13,291 accepted equilibria across 267 conventional and spherical tokamaks, encompassing parameter-sampled Grad--Shafranov solutions, G-EQDSK projections spanning EAST, MAST-U, and ITER scales, and controlled variation families with explicit provenance. Benchmark projections reproduce normalized flux maps with RMS errors between $1.09 \times 10^{-3}$ and $1.45 \times 10^{-3}$, while compact JSON representations achieve an 89--96-fold size reduction relative to standard $129 \times 129$ G-EQDSK files. The complete initial release occupies 41~MB in raw JSON and 18~MB in compressed archives, and all records successfully passed independent reload and evaluation tests. VEQDB establishes an extensible, provenance-preserving foundation for equilibrium studies, reduced-order surrogate modeling, and cross-machine workflows.

physics.plasm-ph↗
arXiv2026

Revisiting Reinforcement Learning with Verifiable Rewards from a Contrastive Perspective

Group Relative Policy Optimization (GRPO) is one of the most widely adopted RLVR algorithms for post-training large language models on reasoning tasks. We first show that GRPO admits an equivalent discriminative reformulation, in which policy optimization maximizes the expected score gap between verified positive and negative rollouts. This reformulation reveals two objective-level limitations: likelihood-misaligned surrogate scores, in which clipped ratio-based scores are optimized rather than the sequence likelihoods that govern generation, and score-insensitive credit assignment, in which rollout-level credit does not reflect the current score gaps between positive and negative rollouts. To address these limitations, we propose ConSPO, a Contrastive Sequence-level Policy Optimization method that uses length-normalized sequence log-probabilities as rollout scores and contrasts verified positive rollouts against negative distractors within the same group. ConSPO optimizes a group-wise InfoNCE-style objective to adaptively strengthen updates for poorly separated positives and high-scoring negatives, together with a curriculum-scheduled margin that preserves separation pressure as training progresses. Experiments across diverse settings show that ConSPO outperforms strong baselines on challenging reasoning benchmarks.

cs.LG↗
arXiv2026

RAVE: Re-Allocating Visual Attention in Large Multimodal Models

Large multimodal models (LMMs) inherit the self-attention mechanism of pretrained language backbones, yet standard attention can exhibit suboptimal allocation, including cross-modal misallocation between textual and visual evidence and intra-visual imbalance among visual tokens. We propose RAVE (Re-Allocating Visual Attention), a lightweight pair-gating mechanism that adds a learned query-key bias to pre-softmax attention scores over visual keys, derived from pre-RoPE query and key features. RAVE requires no architectural modification to the backbone and can be trained end-to-end with the rest of the model. Across a suite of multimodal benchmarks, RAVE improves over standard attention by an average of 3 points, with the largest gains on perception-intensive tasks -- including multilingual OCR, chart understanding, document VQA, and scene text VQA -- where accurate visual grounding is critical.

cs.CV↗
arXiv2026

Robust low-rank tensor completion via factorized weighted tensor schatten-p norm minimization

Low-rank tensor factorization provides a flexible framework for completing multidimensional data from incomplete and corrupted observations. However, unweighted spectral regularizers impose a common shrinkage profile across singular components, which may excessively attenuate dominant low-rank components, and factorized variants either lack component-specific weighting or require costly singular value decompositions (SVDs). This paper proposes two weighted Schatten-$p$ tensor factorization models, termed \WSpTFI{} and \WSpTFII{}, under the tensor-tensor product (t-product) framework to address these limitations. \WSpTFI{} is motivated by a factorized weighted tensor Schatten-$p$ norm identity and permits flexible, possibly asymmetric factor exponents. \WSpTFII{} constructs a regularizer from transform-domain column-pair energies, yielding SVD-free main factor updates and a column-pruning mechanism for reducing redundant rank components. This paper further develops an iteratively reweighted alternating direction method of multipliers (ADMM)-type scheme for \WSpTFI{} and an iteratively reweighted least squares (IRLS)--block successive upper-bound minimization (BSUM) scheme for \WSpTFII{}. Theoretical analysis establishes the weighted factorization relation and provides a conditional limiting Karush--Kuhn--Tucker (KKT) characterization for \WSpTFI{}. For \WSpTFII{}, the actual damped quadratic block updates yield a quantitative sufficient-decrease mechanism for the fixed-$δ$ smoothed factor objective. This implies asymptotic regularity, and every accumulation point of the fixed-dimensional tail is stationary. Experiments on synthetic tensor completion, color-image restoration, hyperspectral inpainting, and printed-circuit-board defect detection demonstrate competitive reconstruction quality and robustness under various degradation conditions.

cs.CV↗
arXiv2026

Semi-Tensor Product-Based Multi-Term Randomized T-SVD and Its Visual Applications

Tensor singular value decomposition (T-SVD), which is built upon the tensor-tensor product (t-product), has emerged as a powerful tool for processing high-dimensional visual data such as color images and videos. However, the standard t-product imposes strict dimensional compatibility constraints. Although extensions based on the semi-tensor product (STP) relax this restriction, their single-term formulations still suffer from limited approximation accuracy. Moreover, these deterministic methods incur high computational costs when processing large-scale tensor data. To address these issues, this paper introduces a novel semi-tensor product for third-order tensors under the t-product framework induced by arbitrary invertible linear transforms. The resulting tensor semi-tensor product breaks the rigid dimension matching requirement of the standard t-product, while retaining the closed-form property of T-SVD. Based on this construction, we develop a multi-term semi-tensor product singular value decomposition (MSTP-SVD), which integrates multiple orthogonal decomposition terms to significantly improve low-rank approximation accuracy compared with single-term schemes. To reduce the computational cost of multi-term modeling, we incorporate randomized projection and power iteration techniques into the MSTP-SVD framework, yielding an accelerated multi-term randomized semi-tensor product SVD (MRSTP-SVD) algorithm that achieves a balance between reconstruction accuracy and computational efficiency. Experiments on image and video compression and completion tasks demonstrate the effectiveness of the proposed method.

cs.LG↗
arXiv2026

Two-parameter variational estimates for averages over tori

One-parameter variational inequalities are well developed, whereas their multi-parameter counterparts remain much less understood. We explore two-parameter variational inequalities for averages over tori in $\mathbb{R}^3$. To capture the underlying two-parameter structure, we introduce a local two-parameter $r$-variation norm that combines rectangular increments with variations along the boundary. The resulting variation operator pointwise dominates the corresponding two-parameter local maximal function and, unlike the maximal function, also captures oscillation across the two parameters. We establish sharp $L^p$--$L^q$ bounds for this variation operator up to endpoints. For comparison, we also obtain sharp $L^p$ bounds up to endpoints for the corresponding local one-parameter variation operator, revealing a genuine difference between the one- and two-parameter boundedness regions. The proof combines square function estimates for two-parameter propagators with local smoothing estimates through mixed-norm interpolation.

math.CA↗
arXiv2026

Minute-Scale High-Fidelity Gyrokinetic Simulations with Portability from Laptop to Supercomputer

Global gyrokinetic particle simulations remain computationally expensive, as they demand both adequate marker statistics and three-dimensional field solvers. In this work, we present a hybrid spectral method within the particle-in-Fourier (PIF) framework and implement it in the electrostatic model of GTC. Charge scatter and field gather are performed between particles and fields on a two-dimensional poloidal mesh, while the corresponding Poisson solver is discretized using radial finite differences and poloidal $m$-harmonics. Truncated spectral transforms are employed to connect multiple representations for fields, avoiding costly particle-grid operations for each individual $m$-harmonic within the particle loop. Benchmarks against conventional particle-in-cell (PIC) simulations successfully reproduce single-$n$ ion temperature gradient (ITG) mode structures and dispersion relations, as well as multi-$n$ nonlinear ITG transport and its regulation by zonal flows. Compared to conventional PIC, the proposed method reduces the effective problem size by more than a factor of 48 and achieves a speedup of over two orders of magnitude for single-$n$ cases. A 2000-step single-$n$ simulation with approximately 2 million markers completes in 78.2 seconds on a laptop GPU, while multi-$n$ turbulence simulation also completes within minutes. Furthermore, the elimination of toroidal particle-shift communication yields promising preliminary scaling performance on multiple NVIDIA A100 GPUs. The numerical scheme is broadly applicable for accelerating particle simulations on platforms ranging from laptops to supercomputers.

physics.plasm-ph↗
arXiv2026

Environment Evolution for Terminal Agents

Scaling interactive and verifiable environments is critical for training terminal agents. As frontier models become more capable, environments synthesized from scratch become less challenging and thus provide limited learning signals. Recent co-evolution methods iteratively synthesize environments near the model's learnable frontier based on weaknesses exposed during rollouts. However, their dependence on on-policy rollouts limits generalization and the continuous provision of learning signals as the model becomes stronger. In this paper, we propose environment evolution, which incrementally increases environment difficulty off-policy and schedules the evolved environments generation by generation during training to provide continuous learning signals. We derive three evolution directions that influence environment difficulty from the multi-turn learning objective and then implement evolution along these directions through a loop-engineered multi-agent harness. Quantitative rollout experiments with Hy4 preview, Claude Opus 5, and GPT-5.6 Sol show that environment evolution consistently produces more difficult environments. We validate its effectiveness on Qwen3.6-27B and Qwen3.6-35B-A3B through simple long-horizon RL training, improving their performance by 14.4 and 18.0 percentage points on Terminal-Bench 2.1, respectively.

cs.AI↗
arXiv2026

PolicyLong: Towards On-Policy Context Extension

Extending LLM context windows is hindered by scarce high-quality long-context data. Recent methods synthesize data with genuine long-range dependencies via information-theoretic verification, selecting contexts that reduce a base model's predictive entropy. However, their single-pass offline construction with a fixed model creates a fundamental off-policy gap: the static screening landscape misaligns with the model's evolving capabilities, causing the training distribution to drift. We propose PolicyLong, shifting data construction towards a dynamic on-policy paradigm. By iteratively re-executing data screening (entropy computation, retrieval, and verification) using the current model, PolicyLong ensures the training distribution tracks evolving capabilities, yielding an emergent self-curriculum. Crucially, both positive and hard negative contexts derive from the current model's entropy landscape, co-evolving what the model learns to exploit and resist. Experiments on RULER, HELMET, and LongBench-v2 (Qwen2.5-3B) show PolicyLong consistently outperforms EntropyLong and NExtLong, with gains growing at longer contexts (e.g., +2.54 at 128K on RULER), confirming the value of on-policy data evolution.

cs.LG↗
arXiv2026

Emergent Noncollinearity and Near-Degenerate Magnetic Superlattices in AT6X6 Kagome Metals

Ferromagnetic AT6X6 Kagome compounds are a popular class of systems in which quantum magnetism with topological features has been observed. These systems allow easy chemical substitution, creating an opportunity to fine-tune their properties. In this paper, we present electronic-structure and magnetic ground-state studies of several AT6X6 compounds with relatively low magnetic-ground-state stability. We find unusual magnetic orderings, including complex spin-spiral states and the formation of magnetic long-range superstructures. While LiFe6Ga6 and TiMn6Ge6 retain collinear AFM ground states with low-energy FM/AFM layer sequences, competing spin-spiral and long-period antiferromagnetic structures in MgFe6Ga6 and a double-spin-spiral ground state in TiFe6Ga6 were determined. Magnetism in all these systems appears local, with adiabatic energy profiles suggesting non-Heisenberg long-range interactions, including a strong biquadratic term. In TiMn6Ge6, we found the conditions for magnetic tunneling. Our results show that, in addition to traditional magnetic topological features in such FM Kagome systems, near-degenerate magnetic superstructures suitable for spintronic switching applications can form naturally. Overall, these systems represent a potentially rich playground for neutron diffraction and spintronics experimental studies.

cond-mat.mtrl-sci↗