Search arXiv⌕ Search

arXiv subjects

Ji Li

Publications and source records attributed to Ji Li.

At least 19 recordsLinked to original sources
arXiv2026

MemT-1.0: Towards Memory-Native Evolution

Memory-native evolution requires making memory an intrinsic capability of foundation models, enabling them to continuously accumulate and utilize information. To this end, we introduce MemT-1.0, a new architecture of memory foundation models that internalize memory as part of model computation. Building on Metis, MemT-1.0 employs Gated DeltaNet-2 with a stable token-level memory update mechanism within each turn. We further introduce Mixture-of-Storage to scale native memory capacity. Following a deliberate backbone preparation, we develop a two-stage training pipeline consisting of continued pre-training and post-training, supported by 34B tokens of training data and efficient training and inference infrastructure. We release a model family comprising MemT-1.0-Nano and MemT-1.0-Flash, and conduct a comprehensive evaluation of memory and general capabilities. The results show that they outperform the parametric memory baselines on average and remain competitive with retrieval-augmented methods, while largely preserving general capabilities. At a history length of 128K tokens, MemT-1.0-Nano achieves approximately $1.86\times$ end-to-end inference speedup and a $66\times$ reduction in persistent storage relative to full context KV cache. By internalizing retrieved procedural experience, MemT-1.0 improves task success on agentic tasks, providing a practical step toward memory-native evolution.

cs.CL↗
arXiv2026

Inflak: A Unified Orchestration Architecture for Human-Agent Interaction

LLM-powered agents increasingly support open-ended activities, yet orchestrating when and how agents interact with humans remains challenging. Existing approaches often provide interaction mechanisms tied to particular tasks or execution patterns, lacking a unified architecture for diverse and evolving interaction needs. We introduce Inflak, a unified four-layer architecture for human-agent interaction orchestration that separates state reconstruction and routing, interaction and execution planning, interaction design, and interface rendering. Its layered design with progressive resolution enables general applicability, runtime flexibility, module composability, and orchestration maintainability. We implement Inflak as agent-native, skill-based plugins within a general-purpose agent framework and evaluate it through a gallery of various application scenarios and interaction paradigms, technical performance testing, and a comparative implementation case study. The evaluations show that Inflak can serve as a general and extensible foundation for orchestrating diverse human-agent interaction flows.

cs.HC↗
arXiv2026

A geometry-independent endpoint estimate for the Littlewood--Paley $g$-function generated by the Poisson semigroup

We prove a weak $(1,1)$ estimate for the full Poisson Littlewood--Paley $g$-function on every complete Riemannian manifold without boundary, with an explicit universal constant approximately equal to $3.33606$. The estimate requires no curvature restriction, volume doubling, heat kernel bound, or stochastic completeness. The proof combines a precise Poisson maximal inequality with a convex entropy identity; local elliptic regularity identifies the extension to arbitrary real $L^1$ data. As an application, we obtain weak $(1,1)$ estimates for certain spectral multipliers represented by Poisson time derivatives with coefficients in $L^2((0,\infty),dt/t)$.

math.CA↗
arXiv2026

LensErase: A Foundation Model for Erase Anything

We introduce LensErase, a general-purpose mask-based Erase Anything foundation model designed to cover a broad spectrum of real-world erase scenarios. LensErase supports diverse objects, visual text, watermarks, associated visual effects, multiple-object removal, and a wide range of object scales, aspect ratios, and image resolutions. First, to enable such broad coverage, we construct LensErase-100M, a large-scale dataset containing 101.07 million paired source-mask-target examples spanning diverse erase scenarios. Second, we pretrain LensErase from the text-to-image model Lens, adapting its strong generative prior to mask-based erasing. To efficiently incorporate the source image and erase mask, we propose Mask-Aware RoPE, a 3D rotary positional encoding that directly injects mask information into the source-image tokens without introducing a separate mask token sequence. Finally, we refine the model through supervised fine-tuning and reinforcement learning with contract-guided, region-aware rewards, and apply TraceFusion at inference to preserve high-fidelity content outside the effective edit region. Across six public benchmarks and the proposed LensErase-Bench, LensErase achieves substantial improvements under both reference-based objective metrics and reference-free VLM evaluation. We hope LensErase can serve as a foundational model for erase-related applications and systems.

cs.CV↗
arXiv2026

AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards

Generating information-rich visual artifacts such as slides, posters, and dashboards as code is appealing because the outputs are structured, editable, and verifiable. Yet code models cannot see how their choices of layout and color come together on the canvas, often limiting spatial planning and aesthetic quality. Image generators excel at visual design but can misrender requested content, making their outputs unreliable targets for direct reproduction. We therefore introduce AesCode, which generates code grounded in the content prompt while using an image generated from the same prompt as an aesthetic guidance. To do so, AesCode represents the target as a design graph that describes the whole canvas and makes each property attributable. From this graph, we derive complementary cross-modal rewards: deterministic verifiers score properties parsable from code and rendering, while a VLM judge assesses the non-parsable ones with a sample-specific Visual Graph Rubric (VGR) tied to graph elements and relations. We optimize these decoupled rewards through reinforcement learning with per-channel normalization to jointly improve spatial planning, aesthetic quality, and verifiable correctness. To evaluate this setting, we build a benchmark that scores verifiable correctness and visual quality separately, with visual scores highly consistent with human judgments. We find that reference images substantially improve the visual quality of code generated by existing models, and ablations show that AesCode benefits from references both as model input and as a source of visual rewards. Relative to same-scale Qwen3-VL baselines given the same references, AesCode-8B and AesCode-32B improve visual quality by 31.1 and 22.4 percentage points, reaching scores comparable to those of closed-source frontier models such as GPT-5.5 and Claude Opus 4.8.

cs.AI↗
arXiv2026

Compactness and essential norms for two-weight fractional Riesz transforms on curves

We prove a compact two-weight $T1$ theorem for the $α$-fractional Riesz vector, $0\leqα<n$, when one measure is supported on a compact regular $C^{1,\mathrm{Dini}}$ curve $Γ$. The measures may have common point masses. Under the global boundedness assumptions, we give a two-sided estimate for the essential norm in terms of Muckenhoupt and testing constants computed after restricting both measures to small cubes. Compactness is equivalent to the uniform vanishing of these local bounds as the cube side length tends to zero. The smooth truncations have a unique weak operator limit and converge in operator norm in the compact case. For arclength measure and distance power weights, we determine the compactness ranges for Green gradients on bounded tubes and for Newtonian gradients and Biot--Savart fields on the whole space. The whole-space operators in these ranges are compact but are not Hilbert--Schmidt. When $α=n-1$ and both measures have positive continuous arclength densities $a,b$ on the same closed curve, we obtain the exact essential norm $πc_{n-1,n}\max_Γ\sqrt{ab}$.

math.CA↗
arXiv2026

On the endpoint estimate for discrete spherical average over sparse sequences

Let $d\ge5$, and let $A_λ$ denote the normalized discrete spherical average on $\mathbb{Z}^d$ with squared radius $λ$. For a strictly increasing sequence $(μ_k)$ of positive integers, set $λ_k=μ_k!$ and consider the lacunary maximal operator $A_\star f:=\sup_k|A_{λ_k}f|$. For each fixed $γ>0$, we prove that the estimate \[ \#\{x:A_\star f(x)>α\} \le K\sum_x\frac{|f(x)|}α \left(1+\log^+\frac{|f(x)|}α\right)^γ\] holds with some finite $K$ for all $α>0$ and finitely supported $f$ if and only if \[ \sup_{N\ge2}\frac{\#\{k:μ_k\le N\}}{(1+\log N)^γ}<\infty. \] The summand is interpreted as zero when $f(x)=0$. The case $γ=1$ answers the $\ell\log\ell$ endpoint question of Kesler, Lacey and Mena in the formulation stated above. For $μ_k=2^{2^k}$, these estimates hold for every fixed $γ>0$, but $A_\star$ is not of weak type $(1,1)$. For factorial blocks formed from arbitrary subsets of dyadic grids at widely separated scales, we characterize restricted weak type $(p,p)$, $1 p_*$, and unbounded on $\ell^p$ for $1\le p<p_*$.

math.CA↗
arXiv2026

KITE: KV-Invariant Transformer Expansion for Efficient Agentic LLM Scaling

Scaling a language model is not only a question of final quality: the architectural choice determines how much computation is spent during training, prompt processing, and autoregressive decoding to achieve certain model quality. An ideal model architecture should lower all above computation costs to facilitate scaling to a larger model, while ensure the larger model indeed outperforms smaller baselines. We introduce KV-Invariant Transformer Expansion (KITE), a scaling paradigm that achieves this goal. It trains the model from a smaller size to a larger size (i.e., saving training costs via upcycling), while places newly added parameters in regions that do not affect attention KV. Consequently, during inference, prefilling KV only relies on the smaller part of the model, so the inference costs are saved. As a concrete instantiation, we present Step Scale Transformer (SST), a two-tower decoder in which one tower produces KV and the other reads them. At comparable cumulative training compute, SST, a 67B MoE model with 2.15B active body parameters per decode token, achieves lower training loss than 47B and 63B MoE Transformers with 1.48B and 2.02B active body parameters, respectively, while reducing estimated inference cost by 6.7% and 31.6%.

cs.LG↗
arXiv2026

WAVE-Go: World-Model Navigation with Adaptive Execution for Wheel-Legged Robots

World models can anticipate the consequences of navigation actions, but predicted action sequences may become invalid during execution, especially when wheel-legged robots encounter dynamic obstacles or change locomotion modes. We propose WAVE-Go, an image-goal navigation framework that separates world-action prediction from interruptible command execution. Its executor adaptively selects an action prefix and cancels pending commands when updated observations invalidate execution. A conditional-risk formulation specifies prefix selection under an estimated cumulative failure budget, while posture and locomotion-mode transitions require clearance, stability, and task-evidence checks. In the reported navigation evaluation, WAVE-Go achieves 74.1% in-distribution success and 63.3% dynamic out-of-distribution success, exceeding the strongest baseline by 4.7 and 7.7 percentage points, respectively, while reducing collisions from 4.4 to 2.9 per 100 m. Compared with interruptible fixed four-command execution, WAVE-Go raises success by 4.0 percentage points while reducing replanning frequency by 51.2% and collision rate by 6.5%. Execution ablations also show that runtime interruption improves success, collision rate, and reaction latency at the cost of additional replanning. These results support adaptive, interruptible execution as a means of balancing navigation performance and planning overhead. Code is available at https://github.com/vigorlee/wave-go.

cs.RO↗
arXiv2026

Commutators with two matrix weights

Let $U,V$ be matrix $\mathcal A_p$ weights, $1 d$ and $\mathcal S^{d,\infty}$ membership by $\ell^r$ and weak $\ell^d$ conditions on the two Hilbert--Schmidt cube oscillations. For $0<r\leq d$, $\mathcal S^r$ membership is equivalent to $B$ being constant almost everywhere. The weak endpoint condition implies a weighted first-order Sobolev estimate for $C^1$ symbols, with a converse for constant or uniformly elliptic weights.

math.CA↗
arXiv2026

Lacunary $δ$-Discretised Spherical Maximal Operators

We study the lacunary analogue of the $δ$-discretised spherical maximal operators introduced by Hickman and Jančar, for $δ\in (0, 1/2)$, and establish the boundedness on $L^p$ for all $1 < p < \infty$, along with the endpoint weak-type estimate $H^1 \to L^{1,\infty}$. We also prove the corresponding $L^p$ boundedness for the multi-parameter variant. The constants in these bounds are uniform in $δ$, and thus, by taking the limit $δ\to 0^+$, our results recover the classical boundedness of the lacunary spherical maximal function.

math.CA↗
arXiv2026

Non-radial Dunkl multipliers at the $L^2$-Sobolev threshold

We prove a Hörmander multiplier theorem for the Dunkl transform associated with an arbitrary finite reflection group. Uniform $H^σ(\mathbb{R}^N)$ bounds for the normalised dyadic pieces of a measurable symbol, with $σ>\mathbf N/2$ and $\mathbf N$ the homogeneous dimension, imply $L^p(dω)$ boundedness for $1<p<\infty$ and weak type $(1,1)$. No radiality or reflection-group invariance is assumed, and the dyadic pieces may meet the reflecting hyperplanes. The main new estimate controls the Dunkl kernel $E_κ(iξ,x)$ and its first Euclidean derivatives in $x$, uniformly near arbitrary intersections of reflecting hyperplanes. Here $ξ$ lies in a fixed compact annulus. Writing $w_κ$ for the Dunkl weight and $Ω_κ$ for its inhomogeneous counterpart, it gives the bound $CΩ_κ(x)^{-1/2}w_κ(ξ)^{-1/2}$. The proof decomposes each root contribution into Hermitian two-dimensional blocks, applies oscillatory changes of variables near the rootwise transition points, and uses a commutator identity to cancel the non-integrable phase derivative. Chamber lifting also yields endpoint bounds for the pullbacks of the componentwise Coifman--Weiss atomic $H^1$ space and the corresponding BMO space on a fixed chamber.

math.CA↗
arXiv2026

Polynomial corners in finite fields beyond the distinct-degree case

We prove a quantitative polynomial Roth theorem for corners in \(\mathbb F_p^2\) for arbitrary pairs of linearly independent polynomials. More precisely, given a positive integer $d$, there are constants $p_0$ and $C$ (depending only on $p$) so that for every $ p>p_0$, if polynomials \(ϕ_1,ϕ_2\in \mathbb \mathbb{F}_p [y]\) are of degree $\leq d$ vanishing at $0$ and are not linearly dependent, then every \(A\subset\mathbb F_p^2\) with $ |A|\ge C p^{2-1/14} $ contains a nontrivial corner $$ (x_1,x_2),\qquad (x_1+ϕ_1(y),x_2),\qquad (x_1,x_2+ϕ_2(y)) $$ for some \(y\in\mathbb F_p^\times\). This improves the estimate $p^{2-1/16}$ of Han--Lacey--Yang and removes the distinct-degree restriction from their quantitative theorem. The main obstruction is the equal-degree resonant case, where the Jacobian argument of Han--Lacey--Yang degenerates. We adjoin the frequency-independent part of the phase to form an augmented map \(\widetilde F:W\to\mathbb A^3\) from the correlation threefold. We prove that this map is generically finite on every top-dimensional geometric component and has no two-dimensional fibre. Using the associated Artin--Schreier sheaf and Katz--Laumon estimates for Fourier transform of perverse sheaves, we obtain square-root cancellation outside an algebraic exceptional set of dimension at most one and uniformly bounded degree. A separate curve-sum argument gives uniform control on the exceptional set. An \(\ell^2\) matrix estimate adapted to such sets completes the resonant case.

math.CA↗
arXiv2026

Fefferman--Stein type inequalities via area and maximal functions for Schrödinger operators with applications

In this paper, we establish a Fefferman--Stein inequality in terms of area function and non-tangential maximal function associated with the Schrödinger operator $\mathcal{L} = -Δ+ V$ on stratified Lie groups $\mathcal G$, where $Δ$ denotes the sub-Laplacian on $\mathcal G$ and $V$ is a nonnegative locally integrable function. As an application, we extend this inequality to the tensor product $\mathcal G_1 \times \mathcal G_2$ of two stratified Lie groups and develop atomic decompositions associated with the Schrödinger operator for functions in the Orlicz space $L\log^{+}L(\mathcal G_1 \times \mathcal G_2).$ Using these atomic decompositions, we further prove weak-type endpoint estimates for the area integral operator and the Riesz transforms associated with the Schrödinger operator on $L\log^{+}L(\mathcal G_1 \times \mathcal G_2),$ thereby extending the celebrated result of R.\,Fefferman and E.M.\,Stein \cite{FSt1982} to the setting of singular integrals with non-smooth kernels.

math.AP↗
arXiv2026

Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference Optimization

Generating visually appealing images is fundamental to modern text-to-image generation models. A potential solution to better aesthetics is direct preference optimization (DPO), which has been applied to diffusion models to improve general image quality including prompt alignment and aesthetics. Popular DPO methods propagate preference labels from clean image pairs to all the intermediate steps along the two generation trajectories. However, preference labels provided in existing datasets are blended with layout and aesthetic opinions, which would disagree with aesthetic preference. Even if aesthetic labels were provided (at substantial cost), it would be hard for the two-trajectory methods to capture nuanced visual differences at different steps. To improve aesthetics economically, this paper uses existing generic preference data and introduces step-by-step preference optimization (SPO) that discards the propagation strategy and allows fine-grained image details to be assessed. Specifically, at each denoising step, we 1) sample a pool of candidates by denoising from a shared noise latent, 2) use a step-aware preference model to find a suitable win-lose pair to supervise the diffusion model, and 3) randomly select one from the pool to initialize the next denoising step. This strategy ensures that diffusion models focus on the subtle, fine-grained visual differences instead of layout aspect. We find that aesthetics can be significantly enhanced by accumulating these improved minor differences. When fine-tuning Stable Diffusion v1.5 and SDXL, SPO yields significant improvements in aesthetics compared with existing DPO methods while not sacrificing image-text alignment compared with vanilla models. Moreover, SPO converges much faster than DPO methods due to the use of more correct preference labels provided by the step-aware preference model.

cs.CV↗
arXiv2026

ReDeck: Step-Level Render-Grounded Refinement for Document-to-Slide Generation

Document-to-slide generation is challenging because slides are dense editable artifacts that require both faithful content selection and precise spatial layout. Recent slide agents adopt iterative reflection, but typically follow a monolithic "one version, one feedback" loop: a slide or deck is rewritten, rendered afterward, and critiqued only at the turn boundary. This delayed feedback makes local failures such as overflow, overlap, clipping, and off-canvas placement difficult to attribute and repair. We propose ReDeck, a step-level render-grounded refinement framework that decomposes slide revision into atomic edit actions and returns renderer-derived observations after each step, turning refinement into "one edit, one observation." To balance local repair with global quality, ReDeck uses multi-granular feedback: step-level render feedback for spatial errors, a turn-level adaptive critic for semantic and design guidance, and a submission-level gate for hard layout validation. We further introduce DeckQuiz, a benchmark that decouples content fidelity, spatial correctness, and design quality. Across GPT-5.4, Claude-4.6, and Gemini-3.1, ReDeck consistently outperforms existing slide-generation agents, and ablations confirm that feedback timing and granularity are critical for reliable slide refinement.

cs.AI↗
arXiv2026

DesignAsCode: Bridging Structural Editability and Visual Fidelity in Graphic Design Generation

Graphic design generation demands a delicate balance between high visual fidelity and fine-grained structural editability. However, existing approaches typically bifurcate into either non-editable raster image synthesis or abstract layout generation devoid of visual content. Recent combinations of these two approaches attempt to bridge this gap but often suffer from rigid composition schemas and unresolvable visual dissonances (e.g., text-background conflicts) due to their inexpressive representation and open-loop nature. To address these challenges, we propose DesignAsCode, a novel framework that reimagines graphic design as a programmatic synthesis task using HTML/CSS. Specifically, we introduce a Plan-Implement-Reflect pipeline, incorporating a Semantic Planner to construct dynamic, variable-depth element hierarchies and a Visual-Aware Reflection mechanism that optimizes the code to rectify rendering artifacts. Extensive experiments demonstrate that DesignAsCode significantly outperforms baselines in both structural validity and aesthetic quality. Furthermore, our code-native representation unlocks advanced capabilities, including automatic layout retargeting, complex document generation (e.g., resumes), and CSS-based animation. Our project page is available at https://liuziyuan1109.github.io/design-as-code/.

cs.GR↗