arXiv subjects

Junming Zhang

Publications and source records attributed to Junming Zhang.

At least 19 recordsLinked to original sources

arXiv2026

Lyapunov spectrum, Harder--Narasimhan filtration and thin monodromy

We study the equality case of the inequalities between the Lyapunov spectrum and the Harder--Narasimhan slopes of a semi-simple local system with non-expanding cusp monodromy over a finite-area hyperbolic surface. This gives a nonlattice-transversality alternative of the monodromy group. On the other hand, we generalize Filip's Assumptions A and B to give classes of polarized variation of Hodge structures that achieve the equality. As an application, we recover the thinness of C-32 in $\mathrm{Sp}_6\mathbb{R}$. Besides, we provide an upper bound, called average Cartan spectrum, that dominates the Lyapunov spectrum and study its rigidity.

math.DG↗

arXiv2026

FGD2HF: Frequency-guided Dual-domain Hypergraph Fusion for Semantic Segmentation

Frequency-domain information is beneficial for semantic segmentation. However, existing frequency-based methods often suffer from insufficient frequency utilization, inefficient fusion, and high computational complexity. To address these issues, we propose FGD2HF, a dual-domain fusion network that leverages frequency features to guide spatial-domain fusion and enhancement. Specifically, we introduce a novel Hybrid-Attention Encoder (NAHA) that jointly models local and global relationships through neighborhood and hypergraph attention, enhancing the network's ability to extract effective features. We further introduce a Wavelet Transform and Enhancement (WTE) module to extract and refine low- and high-frequency cues from preceding-stage features, followed by Frequency-guided Dual-domain Fusion (FGDF) modules that use frequency cues to condition current-stage spatial grouping and feature propagation. This design incorporates both global and local frequency guidance into spatial feature modulation, enhancement and fusion with low computational cost. Extensive comparisons with numerous state-of-the-art semantic segmentation methods demonstrate that our method is highly competitive. Our full FGD2HF-B model achieves 48.7% mIoU on ADE20K (+0.3% over SegMAN-B with 18% fewer GFLOPs), 83.2% mIoU on Cityscapes (comparable to SegMAN-B with 21% fewer GFLOPs), and 1.1% higher mIoU than VWFormer-T on SBD with 21% fewer GFLOPs, while also achieving significant improvements in boundary segmentation accuracy.

cs.CV↗

arXiv2026

Non-maximal quasi-isometric $\mathrm{PU}_{2,n+1}$-representations via alternating surfaces in complex pseudo-hyperbolic spaces

For every admissible non-maximal Toledo invariant $t$, we construct a locus of irreducible representations of a closed genus-$g$ surface group into $\mathrm{PU}_{2,n+1}$ with Toledo invariant $t$ whose orbit maps are quasi-isometric embeddings. These loci have real codimension $10(g-1)$ or $10(g-1)+2(n-1)$ in the character variety depending on the Toledo value. Our construction uses $4$-cyclic Higgs bundles and their correspondence with $\partial$-alternating surfaces in complex pseudo-hyperbolic spaces. The associated equivariant minimal maps into the symmetric space are bi-Lipschitz embeddings. We further construct cocompact domains of discontinuity in the Shilov boundary. Their quotients are smooth fiber bundles over the surface with fiber homeomorphic to $S^{2n-1}\times S^{2n-1}$.

math.DG↗

arXiv2026

CollabFlow: Recursive Self-Improvement of Agent Collaboration

Recursive self-improvement (RSI) lets a system improve from its own outcomes; in LLM-based multi-agent systems, Agents refine one another within a task, and outcomes improve how they collaborate across tasks. However, existing multi-agent collaboration leaves this loop open: collaboration is pre-defined at the operator level, topology-only learning keeps verbatim exchange that propagates errors, and reward maximization on a system's own outcomes concentrates on a few teams. To address these challenges, we propose CollabFlow, an RSI system of Learned Agent Collaboration: a trainable Collab-Director constructs teams of complete Agents, a frozen executor runs them, and each round's outcomes retrain the director. Within each round, the edges of a collaboration graph carry protocols of Evidence-Conditioned Communication: a receiver adopts a differing answer only when the sender's evidence is stronger by a margin, so the director learns who communicates and how. Across rounds, we further propose Collaborative Trajectory Balance (CTB), a flow-based objective that credits each team once across its construction orders and targets a reward-proportional distribution over teams, so several good teams stay in play. We also bound how far this self-generated target moves between rounds, which shrinks as records accumulate. On twelve datasets, CollabFlow outperforms all baselines and keeps improving across rounds. Code is available at https://anonymous.4open.science/r/CollabFlow-631E.

cs.MA↗

arXiv2026

HABIT-Bench: Evaluating Habit Induction from Longitudinal Weak Evidence in Agent Memory

In long-term interactions with large language model (LLM) agents, users reveal habits through their choices and actions. These habits capture implicit user information that agents must infer across conversations. However, individual observations may provide only partial evidence about a habit or when it applies. Retaining past interactions alone does not ensure that agents can infer habits and apply them appropriately. We introduce HABIT-Bench to evaluate how well agent memory systems infer habits from weak evidence and apply them to later requests. We construct synthetic user histories from task-oriented dialogue data and controlled habit graphs. The histories span hundreds of sessions across food, finance, software, and travel. Four-choice tasks assess agents' ability to infer user habits and apply them appropriately. They include temporary exceptions, changes in habits, and suggestions that users have not accepted. Sampled tasks undergo independent human review. We evaluate eight memory methods with Qwen3-235B-A22B and GLM-5.3-Flash. The best-performing memory method achieves 33.61\% macro accuracy across the four domains. Performance remains limited even when annotated evidence is supplied directly. These results expose a persistent gap between access to past interactions and the ability to infer and appropriately use user habits.

cs.AI↗

arXiv2026

Deep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models

Large language models need effective ways to acquire and use external knowledge beyond what they retain from pretraining. Continued pretraining risks catastrophic forgetting, while retrieval at inference time can add substantial overhead. External parametric memory avoids online retrieval, but constructing its training data requires extensive offline search, and output-level interpolation provides only shallow fusion. We introduce Deep Fusion Memory (DFM), a parametric memory architecture that reduces this construction workload and incorporates learned memory into intermediate hidden states. A causal memory predictor learns to generate memory from the preceding context, using retrieved chunk embeddings from the corpus as supervision. The predicted memory is shared across tokens within each chunk and fused into their hidden states through gated cross-attention. Memory prediction and fusion are trained end to end with the backbone frozen, and inference requires no explicit retrieval. Compared with existing parametric memory approaches, which often construct token-level memory targets, DFM requires $32\times$ fewer retrieval queries and $2048\times$ fewer retrieved items to construct its training data. Across six QA tasks, DFM performs better over the strongest baseline by 12.71% on Mistral-7B-v0.3 and 11.19% on Qwen3-8B-Base. Beyond QA, domain adaptation yields average gains of 6.51 and 4.38 points over Qwen3-4B-Base and Qwen3-8B-Base, respectively.

cs.CL↗

arXiv2026

Search-G1: Co-Evolving Representation-Based Rewards for Selective Grounded Search

Search-augmented language agents should retrieve external information only when necessary and ground their answers in retrieved evidence. Yet terminal correctness assigns equal credit to direct answers from parametric knowledge, evidence-grounded answers, and answers recalled after redundant search. Process supervision and LLM judges provide richer feedback but can require high-cost annotation or inference; low-cost, non-contrastive policy confidence signals do not directly identify whether retrieved evidence changes an answer. We propose Search-G1, a reward framework that couples retrieval necessity with evidence reliance through two intervention-calibrated hidden-state readouts for selective grounded search. A prompt readout predicts closed-book sufficiency, whose complement proxies policy-relative retrieval necessity; an answer-commit readout predicts evidence reliance as sensitivity to evidence deletion conditional on the realized reasoning prefix. A correctness-first reward gates evidence-reliance credit by retrieval necessity, rewards sufficient direct answers, and penalizes repeated search. After calibration, scoring needs neither process annotations nor LLM-judge inference. As policy-relative targets, trajectory distributions, and representations may drift, Search-G1 periodically regenerates targets and refits its readouts. Across four QA benchmarks and two scales, Search-G1 achieves competitive task accuracy, with stronger counterfactual evidence dependence and fewer searches in the main comparisons, complemented by independent human audits of answer support.

cs.CL↗

arXiv2026

AF-Muon: An AdamW-Free Muon Optimizer for Tied-Embedding Models

Muon improves large-scale training by applying a spectral-norm steepest-descent update to matrix parameters, but practical models also contain parameter blocks that do not fit dense-matrix geometry. One important case is the tied vocabulary table, which appears in language models and other token generators and can receive multiple structurally different gradient sources, from sparse input lookups to dense output-classifier updates. In the reference recipe these blocks are handed to an auxiliary AdamW optimizer, which restores second-moment state and updates the aliased table as a generic tensor. We propose AF-Muon, an AdamW-free extension of Muon that keeps the Muon matrix update for hidden weight matrices while using a support-aware finite-cap linear minimization oracle for tied vocabulary tables and an RMS-normalized update for one-dimensional auxiliary parameters. AF-Muon therefore trains every parameter class with a single first-moment buffer and no second-moment state, saving around 20% optimizer-state memory relative to Hybrid Muon in our benchmark. Across nine tied-token settings - decoder-only language models from 124M to 1B parameters, a fully shared T5-style encoder-decoder, and ImageGPT-style image-token, protein, and sparse-MoE variants, spanning text, image, and protein-sequence data - AF-Muon improves mean validation loss and perplexity over both Hybrid Muon and a SCION-style Sign endpoint. Long-horizon runs and hyperparameter sensitivity studies confirm the gain is robust, and identical-momentum diagnostics attribute it to the finite cap, which preserves more within-row magnitude than Sign while bounding the coordinate concentration of row-RMS. These results identify tied vocabulary tables as a distinct optimizer geometry and yield a robust AdamW-free Muon variant across models, modalities, and architectures, with about 1% step-time overhead in matched training.

cs.LG↗

arXiv2026

ADSEL: Adaptive Dual Self-Expression Learning for EEG Feature Selection via Incomplete Multi-Dimensional Emotion Labels

EEG based multi-dimension emotion recognition has attracted substantial research interest in affective computing. However, the high dimensionality of EEG features, coupled with limited sample sizes, frequently leads to classifier overfitting and high computational complexity. Feature selection constitutes a critical strategy for mitigating these challenges. However, most existing EEG feature selection methods assume complete multi-dimensional emotion labels. In practice, open acquisition environment and the inherent subjectivity of emotion perception often result in incomplete label data, which can compromise model generalization. Additionally, existing feature selection methods for handling incomplete multi-dimensional labels primarily focus on correlations among various dimensions during label recovery, neglecting the correlation between samples in the label space and their interaction with various dimensions. To address these issues, we propose a novel incomplete multi-dimensional emotion feature selection framework integrating Adaptive Dual Self-Expression Learning (ADSEL) with least squares regression. ADSEL could establish a bidirectional pathway between sample-level and dimension-level self-expression learning processes within the label space. It could facilitate the cross-sharing of learned information between these processes, enabling the simultaneous exploitation of effective information across both samples and dimensions for label reconstruction. Consequently, ADSEL could enhance label recovery accuracy and effectively identifies the optimal EEG feature subset for multi-dimensional emotion recognition. ADSEL was evaluated against fourteen state-of-the-art feature selection methods on three public EEG datasets with multi-dimensional emotion labels. Experimental results demonstrate that ADSEL could achieve superior performance under conditions of partial label absence.

cs.HC↗

arXiv2026

Non-Hitchin Borel Anosov representations from surface groups to $\mathrm{SL}_{3k}\mathbb{R}$

We use the Labourie--Wentworth's formula and the thermodynamic formalism to show that, along the slice constructed by Bronstein--Davalo, the logarithmic top-eigenvalue length spectrum has a uniformly positive second variation near the Barbot representation. As a major application, we show that every closed surface group admits a non-Hitchin Borel Anosov representation into $\mathrm{SL}_{3k}\mathbb{R}$ for every $k\geqslant 1$. In particular, we obtain the first such examples in the even dimensions $6k$. We also study the local behavior of related objects of this slice around the Barbot representation, including the Lyapunov exponent of the flat bundle, the Hausdorff dimension of the limit set, and the Hilbert entropy of the representation.

math.GT↗

arXiv2026

EStream: Fast and Memory-Efficient MoE Prefill through Expert Virtualization on Mobile NPUs

Mobile vendors and application developers increasingly deploy LLMs on smartphones for diverse prefill-only services. Yet current systems rely mainly on dense models whose regular computation maps efficiently to mobile NPUs, leaving more capable MoEs underused. MoE prefill does not fit mobile NPUs: NPU graphs are fixed at compile time, yet MoE picks experts at runtime; and one request touches most experts, more than a phone can hold in memory. We present EStream, which resolves both by separating what the NPU must fix from what MoE decides at runtime. A single compiled expert graph serves every expert, with each expert's routed tokens and weight address bound at call time, so dynamic MoE execution runs entirely on the NPU without padding or CPU/GPU fallback. Expert virtualization keeps the expert pool in UFS flash storage and pages it through a fixed-size NPU-addressable arena, group by group, with loading hidden behind computation, so memory is bounded by the arena rather than by the model. It further introduces a hardware-aware configuration algorithm that automatically configures the UFS--NPU pipeline and maximizes loading--computation overlap. Across 18 comparative settings covering three 7B--16B MoEs and 256--4,096-token prompts, we evaluate EStream on a commercial Snapdragon smartphone. Compared to the fastest baseline at each setting, EStream achieves a 2.25--27.57X pure-prefill TTFT speedup and reduces peak physical memory by 1.19--12.29X. EStream further scales to MoE models with up to 46.7B parameters.

cs.DC↗

arXiv2026

An Ultrathin Laterally Conductive Mesh Interphase Enables Spatially Extended Zinc Deposition for Aqueous Zinc Batteries

Zn metal anodes often suffer from nonuniform interfacial reactions during cycling, resulting in uneven deposition and dendrite growth. Existing artificial interphases can mitigate side reactions or regulate nucleation, but rarely achieve regulation of the interfacial electron/field distribution to sustain uniform deposition at the evolving Zn/electrolyte interface. Here, we develop an Au mesh interphase (AuMI), an ultrathin two-dimensional Au aerogel network that couples lateral electron redistribution with open pathways for ions. During Zn plating/stripping, the conductive AuMI distributes electron transport across the Zn surface, while its porous mesh preserves Zn$^{2+}$ access, enabling more uniform interfacial reactions. Experiments and simulations show that AuMI homogenizes the interfacial electric field and current distribution, promotes more uniform Zn plating/stripping, and limits dendrite growth. As a result, this regulated interfacial reaction mode enables AuMI Zn symmetric cells to operate stably for 3000 h at 1 mA cm$^{-2}$/1 mAh cm$^{-2}$ and for 1100 h at 10 mA cm$^{-2}$/10 mAh cm$^{-2}$, while AuMI Zn||NVO (NaV$_3$O$_8$\cdot$1.5H$_2$O) full cells retain 80.8 % capacity after over 5000 cycles at 1 A g$^{-1}$. These findings highlight the importance of combining ultrathin architecture, lateral electron transport, and open Zn2+ access in artificial interphases for stable aqueous Zn metal anodes.

cond-mat.mtrl-sci↗

arXiv2026

Curvature-Aware Zeroth-Order Optimization for Memory-Efficient Test-Time Adaptation

Test-time adaptation (TTA) aims to enhance the cross-domain performance of pre-trained models by adapting to unlabeled test data. While most existing TTA methods rely on backpropagation (BP) for finetuning, BP-free methods such as zeroth-order (ZO) methods are more desired in practical on-device scenarios. ZO methods rely only on forward computation, which can largely reduce the complexity and memory overhead of on-device deployment. However, ZO methods suffer from much higher variance compared with first-order methods in estimating the gradient. To address this, we propose an improved ZO method to substantially boost the performance of ZO optimization based TTA. First, we provide an observation to reveal the persistent low-rank Hessian structure of the loss during the adaptation process. Based on this insight, we then propose a loss-landscape curvature-aware zeroth-order (CAZO) method, which leverages a sliding-average estimation of the diagonal Hessian to construct a covariance matrix for anisotropic perturbation sampling. CAZO operates by freezing pretrained weights and optimizing minimal adapter parameters via forward-only passes based gradient estimation, which can substantially reduce the memory overhead compared to BP-based methods. Extensive experiments demonstrate that CAZO significantly outperforms existing TTA methods, achieving state-of-the-art performance while maintaining an excellent balance between accuracy and memory efficiency. Code is available at https://github.com/Hollyming/CAZO.

cs.CV↗

arXiv2026

Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

Decoder-only language models entangle long-term memory and reasoning in a single parameter set, making it difficult to scale memory capacity independently. Memory Decoder introduces a parametric long-term memory module but only studies it at a relatively small scale. In this work, we present Memory Decoder at Scale, scaling memory models up to 6.9B parameters and pretraining them on 300B tokens. At this data scale, the combined cost of indexing and search makes a standard Faiss pipeline infeasible. We address this bottleneck with a distributed pipeline for Faiss indexing and retrieval, together with sparse, batch-wise loading of kNN distributions. Across model scales, we find that allocating more parameters to memory yields a better parameter-performance tradeoff than scaling the base model alone. On 17 benchmarks, pairing a 6.9B general memory with Pythia-410M raises its average score from 29.86 to 37.34, surpassing Pythia-12B (37.24) with 39% fewer total parameters. For Qwen3 Base models ranging from 0.6B to 14B, 1.7B domain memories improve the average score across the three domains by more than 9 points at every scale. Overall, our results demonstrate that independently scaling pretrained memory offers a more parameter efficient path to improving language model performance.

cs.CL↗

arXiv2026

M2I2HA: Multi-modal Object Detection Based on Intra- and Inter-Modal Hypergraph Attention

Recent advances in multi-modal detection have significantly improved detection accuracy in challenging environments (e.g., low light, overexposure). By integrating RGB with modalities such as thermal and depth, multi-modal fusion increases data redundancy and system robustness. However, significant challenges remain in effectively extracting task-relevant information both within and across modalities, as well as in achieving precise cross-modal alignment. While CNNs excel at feature extraction, they are limited by constrained receptive fields, strong inductive biases, and difficulty in capturing long-range dependencies. Transformer-based models offer global context but suffer from quadratic computational complexity and are confined to pairwise correlation modeling. Mamba and other State Space Models (SSMs), on the other hand, are hindered by their sequential scanning mechanism, which flattens 2D spatial structures into 1D sequences, disrupting topological relationships and limiting the modeling of complex higher-order dependencies. To address these issues, we propose a multi-modal perception network based on hypergraph theory called M2I2HA. Our architecture includes an Intra-Hypergraph Enhancement module to capture global many-to-many high-order relationships within each modality, and an Inter-Hypergraph Fusion module to align, enhance, and fuse cross-modal features by bridging configuration and spatial gaps between data sources. We further introduce a M2-FullPAD module to enable adaptive multi-level fusion of multi-modal enhanced features within the network, meanwhile enhancing data distribution and flow across the architecture. Extensive object detection experiments on multiple public datasets against baselines demonstrate that M2I2HA achieves state-of-the-art performance in multi-modal object detection tasks.

cs.CV↗

arXiv2026

Real-space identification of distinct magnetic configurations in a candidate d-wave altermagnet

Altermagnetism is an emerging class of magnetic order characterized by momentum-dependent spin-split electronic structures despite vanishing net magnetization. Although momentum-space signatures consistent with altermagnetism have been reported in a growing number of materials, their relationship to the underlying real-space magnetic configurations remains incompletely understood, because similar spin-split electronic structures can arise from distinct magnetic orders. In the candidate d-wave altermagnet KV2Se2O, the magnetic origin of the observed momentum-dependent spin splitting has remained controversial. Here, we employ spin-polarized scanning tunnelling microscopy combined with magnetic-field-dependent quasiparticle interference imaging to determine the magnetic configuration of KV2Se2O at the atomic scale. Spin-resolved quasiparticle interference reveals a checkerboard-like antiparallel spin texture within the V2O layer and determines its interlayer spin arrangement across unit-cell step edges. Remarkably, we identify both C-type and G-type magnetic configurations, both of which generate similar spin-split electronic structures at the single-layer level but correspond to d-wave altermagnetic and conventional antiferromagnetic orders, respectively. These observations reveal a complex magnetic landscape arising from nearly degenerate magnetic states. Our results establish a direct connection between momentum-space spin splitting and real-space magnetic order, providing a framework for identifying the microscopic origin of spin-split electronic structures in altermagnetic materials.

cond-mat.mtrl-sci↗

arXiv2026

PepALD: Macrocyclic Peptide Generation via Autoregressive Latent Diffusion

Macrocyclic peptides are promising therapeutic candidates for intracellular targets, but their design requires simultaneous control over non-natural monomer chemistry, ring topology, membrane permeability, and target binding. Existing SMILES- or HELM-string generative models either operate in long atom-level sequence spaces or treat monomers as symbolic tokens with limited chemical grounding. We introduce PepALD, an Autoregressive Latent Diffusion (ALD) foundation model for \textit{de novo} macrocyclic peptide generation. The model represents HELM monomers with structured chemical embeddings, generates each residue through context-conditioned diffusion in chemically informed latent space, predicts R-group-aware ring closures during autoregressive generation, and aligns the denoiser to affinity rewards using winner-protected diffusion-adapted preference optimization. In silico experiments demonstrate PepALD's generation quality and reward-optimization performance against representative peptide generation baselines.

cs.LG↗

arXiv2026

Infinitesimal Rigidity of Cyclic Surfaces and Alternating Surfaces

We study the infinitesimal rigidity of equivariant minimal maps from the universal cover of a smooth oriented surface (possibly non-compact) into a Riemannian symmetric space, focusing on representations arising from cyclic harmonic bundles. By developing a unified Lie-theoretic framework that connects cyclic surfaces and cyclic harmonic bundles over Riemann surfaces, we prove the infinitesimal rigidity for irreducible cyclic surfaces under admissible smooth variations, including both compactly supported deformations and $L^p$-integrable variations on non-compact surfaces. As a geometric application, we introduce $n$-alternating surfaces in $\mathbb H^{p,q}$ and establish their correspondence with a special class of cyclic surfaces. This yields an infinitesimal rigidity theorem that conceptually unifies and extends known rigidity results for maximal space-like surfaces, alternating holomorphic curves, and $A$-surfaces in certain $\mathbb H^{p,q}$.

math.DG↗