arXiv subjects

Zirong Zeng

Publications and source records attributed to Zirong Zeng.

13 recordsLinked to original sources

arXiv2026

SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD

Full-parameter post-training of trillion-parameter-scale MoE models introduces substantial system-level challenges for large-scale distributed training, including severe memory pressure, non-overlapped communication overhead, and inefficient kernel execution. While most large-scale LLM training systems are built around GPU-based clusters, this report presents an end-to-end optimization practice on the Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, we develop a hierarchical optimization framework spanning model-level parallelism, computation-communication orchestration, and low-level kernel execution. The resulting system achieves 34.22% Model FLOPs Utilization (MFU) with a 2.93x improvement over the open-source baseline recipe while maintaining training stability. Building on this optimized infrastructure, we further establish a CPT and SFT workflow for complex Operations Research (OR) tasks. We refer to the integrated framework as SLAI T-Rex. Using DeepSeek-V4-Flash, we develop OR-oriented CPT and SFT data pipelines that combine collected domain resources with solver-verified synthetic optimization documents. The resulting dataset contains 10K high-quality SFT samples spanning four task categories and three problem representations. The specialized model achieves the highest average zero-shot Pass@1 score among the evaluated models, reaching 71.81% and outperforming GPT-5.4-Mini and the base DeepSeek-V4-Flash model by 3.98 and 11.27 percentage points, respectively. Overall, this work demonstrates a full-stack pathway from efficient trillion-parameter model post-training on Ascend infra to domain-specialized Flash models for solver-grounded mathematical modeling, advancing frontier-model systems for complex reasoning.

cs.CL↗

arXiv2026

An Extensible and Verifiable Language for Query Rewrite Rules

Logical query plan rewriting transforms a relational database query into an equivalent but more efficient form and is crucial to the performance of database-backed applications. In existing systems, rewrite rules are typically implemented manually, tightly coupled to specific execution engines, and often lack formal correctness guarantees. Consequently, developing a new engine requires reimplementing both legacy and new rules, incurring significant engineering cost, limiting portability, and every new implementation is an opportunity for introducing new bugs. We introduce Rulescript, an engine-agnostic domain-specific language (DSL) for developing query rewrite rules. Rulescript separates rule definition from execution infrastructure via a relational algebra-inspired core language and an explicit decomposition of rules into matching and transformation phases. Developers express rewrites by pattern-matching query plans using Rulescript's core operators and constructing semantically equivalent transformed plans, with all rewrites automatically verified formally to ensure correctness. Rulescript is extensible: users can define custom operators in terms of the core language to capture engine-specific semantics. To integrate with an existing system, developers need only implement a lightweight adapter that maps Rulescript's core and custom operators to the operators implemented in the target engine. We evaluate Rulescript by reimplementing 33 rewrite rules from Apache Calcite and extending the language with several custom operators. To demonstrate portability, we automatically deploy these rules to CockroachDB and Apache Data Fusion, two engines with substantially different backends. Our results show that Rulescript enables "write once, deploy everywhere" paradigm for query plan rewriting, with minimal effort required to deploy previously written rules on a new data engine.

cs.DB↗

arXiv2026

GatedCLIP: Gated Multimodal Fusion for Hateful Memes Detection

Detecting hateful content in multimodal memes presents unique challenges, as harmful messages often emerge from the complex interplay between benign images and text. We propose GatedCLIP, a Vision-Language model that enhances CLIP's multimodal capabilities with specialized architectural improvements for hateful memes detection. Our approach introduces learned projection heads that map CLIP embeddings to a task-optimized semantic space, a dynamic gated fusion mechanism that adaptively weights visual and textual features, and a contrastive learning objective that maintains cross-modal semantic alignment. Experiments on the Hateful Memes dataset demonstrate that GatedCLIP achieves an AUROC of 0.66, substantially outperforming the CLIP baseline (AUROC 0.49) while maintaining computational efficiency with only 350K trainable parameters.

cs.CV↗

arXiv2025

FrontierCS: Evolving Challenges for Evolving Intelligence

We introduce FrontierCS, a benchmark of 156 open-ended problems across diverse areas of computer science, designed and reviewed by experts, including CS PhDs and top-tier competitive programming participants and problem setters. Unlike existing benchmarks that focus on tasks with known optimal solutions, FrontierCS targets problems where the optimal solution is unknown, but the quality of a solution can be objectively evaluated. Models solve these tasks by implementing executable programs rather than outputting a direct answer. FrontierCS includes algorithmic problems, which are often NP-hard variants of competitive programming problems with objective partial scoring, and research problems with the same property. For each problem we provide an expert reference solution and an automatic evaluator. Combining open-ended design, measurable progress, and expert curation, FrontierCS provides a benchmark at the frontier of computer-science difficulty. Empirically, we find that frontier reasoning models still lag far behind human experts on both the algorithmic and research tracks, that increasing reasoning budgets alone does not close this gap, and that models often over-optimize for generating merely workable code instead of discovering high-quality algorithms and system designs.

cs.LG↗

arXiv2025

Global dissipative solutions of the 3D Naiver-Stokes and MHD equations

For any divergence free initial data in $H^\frac12$, we prove the existence of infinitely many dissipative solutions to both the 3D Navier-Stokes and MHD equations, whose energy profiles are continuous and decreasing on $[0,\infty)$. If the initial data is only $L^2$, our construction yields infinitely many solutions with continuous energy, but not necessarily decreasing. Our theorem does not hold in the case of zero viscosity as this would violate the weak-strong uniqueness principle due to Lions. This was achieved by designing a convex integration scheme that takes advantage of the dissipative term.

math.AP↗

arXiv2024

Non-Leray-Hopf solutions to 3D stochastic hyper-viscous Navier-stokes equations: beyond the Lions exponents

We consider the 3D stochastic Navier-Stokes equations (NSE) on torus where the viscosity exponent can be larger than the Lions exponent 5/4. For arbitrarily prescribed divergence-free initial data in $L^{2}_x$, we construct infinitely many probabilistically strong and analytically weak solutions in the class $L^{r}_ΩL_{t}^γW_{x}^{s,p}$, where $r\geq1$ and $(s, γ, p)$ lie in two supercritical regimes with respect to the Ladyžhenskaya-Prodi-Serrin (LPS) criteria.It shows that even in the high viscosity regime beyond the Lions exponent, though solutions are unique in the Leray-Hopf class, the uniqueness fails in the mixed Lebesgue spaces and, actually, there exist infinitely manly non-Leray-Hopf solutions which can be very close to the Leray-Hopf solutions. Furthermore, we prove the vanishing noise limit result, which relates together the stochastic solutions and the deterministic solutions constructed by Buckmaster-Vicol [4] and the recent work [23].

math.AP↗

arXiv2024

Existence and non-uniqueness of weak solutions with continuous energy to the 3D deterministic and stochastic Navier-Stokes equations

The continuity of the kinetic energy is an important property of incompressible viscous fluid flows. We show that for any prescribed finite energy divergence-free initial data there exist infinitely many global in time weak solutions with smooth energy profiles to both the 3D deterministic and stochastic incompressible Navier-Stokes equations. In the stochastic case the constructed solutions are probabilistically strong. Our proof introduces a new backward convex integration scheme with delicate selections of initial relaxed solutions, backward time intervals, and energy profiles. Our initial relaxed solutions satisfy a new time-dependent frequency truncated NSE, different from the usual approximations as it decreases the large Reynolds error near the initial time, which plays a key role in the construction.

math.AP↗

arXiv2022

Non-uniqueness for the hypo-viscous compressible Navier-Stokes equations

We study the Cauchy problem for the isentropic hypo-viscous compressible Navier-Stokes equations (CNS) under general pressure laws in all dimensions $d\geq 2$. For all hypo-viscosities $(-Δ)^α$ with $α\in (0,1)$, we prove that there exist infinitely many weak solutions with the same initial data. This provides the first non-uniqueness result of weak solutions to viscous compressible fluid. Our proof features new constructions of building blocks for both the density and momentum, which respect the compressible structure. It also applies to the compressible Euler equations and the hypo-viscous incompressible Navier-Stokes equations (INS). In particular, in view of the Ladyženskaja-Prodi-Serrin criteria, the obtained non-uniqueness of $L^2_tC_x$ weak solutions to the hypo-viscous INS is sharp, and reveals that $α=1$ is the sharp viscosity threshold for the well-posedness in $L^2_tC_x$. Furthermore, we prove that the Hölder continuous weak solutions to the compressible Euler equations may be obtained as a strong vanishing viscosity limit of a sequence of weak solutions to the hypo-viscous CNS.

math.AP↗

arXiv2022

Sharp non-uniqueness of weak solutions to 3D magnetohydrodynamic equations

We prove the non-uniqueness of weak solutions to 3D hyper viscous and resistive MHD in the class $L^γ_tW^{s,p}_x$, where the exponents $(s,γ,p)$ lie in two supercritical regimes. The result reveals that the scaling-invariant Ladyženskaja-Prodi-Serrin (LPS) condition is the right criterion to detect non-uniqueness, even in the highly viscous and resistive regime beyond the Lions exponent. In particular, for the classical viscous and resistive MHD, the non-uniqueness is sharp near the endpoint $(0,2,\infty)$ of the LPS condition. Moreover, the constructed weak solutions admit the partial regularity outside a small fractal singular set in time with zero $\mathcal{H}^{η_*}$-Hausdorff dimension, where $η_*$ can be any given small positive constant. Furthermore, we prove the strong vanishing viscosity and resistivity result, which yields the failure of Taylor's conjecture along some subsequence of weak solutions to the hyper viscous and resistive MHD beyond the Lions exponent.

math.AP↗

arXiv2022

Sharp non-uniqueness for the 3D hyperdissipative Navier-Stokes equations: above the Lions exponent

We study the 3D hyperdissipative Navier-Stokes equations on the torus, where the viscosity exponent $α$ can be larger than the Lions exponent $5/4$. It is well-known that, due to Lions [55], for any $L^2$ divergence-free initial data, there exist unique smooth Leray-Hopf solutions when $α\geq 5/4$. We prove that even in this high dissipative regime, the uniqueness would fail in the supercritical spaces $L^γ_tW^{s,p}_x$, in view of the generalized Ladyženskaja-Prodi-Serrin condition. The non-uniqueness is proved in the strong sense and, in particular, yields the sharpness at two endpoints $(3/p+1-2α, \infty, p)$ and $(2α/γ+1-2α, γ, \infty)$. Moreover, the constructed solutions are allowed to coincide with the unique Leray-Hopf solutions near the initial time and, more delicately, admit the partial regularity outside a fractal set of singular times with zero Hausdorff $\mathcal{H}^{η_*}$ measure, where $η_*>0$ is any given small positive constant. These results also provide the sharp non-uniqueness in the supercritical Lebesgue and Besov spaces. Furthermore, the strong vanishing viscosity result is obtained for the hyperdissipative Navier-Stokes equations.

math.AP↗

arXiv2022

Non-uniqueness of weak solutions to 3D magnetohydrodynamic equations

We prove the non-uniqueness of weak solutions to 3D magnetohydrodynamic (MHD for short) equations. The constructed weak solutions do not conserve the magnetic helicity and can be close to any given smooth, divergence-free and mean-free velocity and magnetic fields. Furthermore, we prove that the weak solutions constructed by Beekie-Buckmaster-Vicol [2] for the ideal MHD can be obtained as a strong vanishing viscosity and resistivity limit of a sequence of weak solutions to MHD equations. This shows that, in contrast to the weak ideal limits, Taylor's conjecture does not hold along the vanishing viscosity and resistivity limits. Unlike in the context of the NSE [13] and the ideal MHD [2], new types of velocity and magnetic flows, featuring both the refined spatial and temporal intermittency, are constructed to respect the geometry of MHD and to control the strong viscosity and resistivity. Compatible algebraic structure is derived in the convex integration scheme. More interestingly, the new intermittent flows indeed enable us to prove the aforementioned results for the hyper-viscous and hyper-resistive MHD equations up to the sharp exponent $5/4$, which coincides exactly with the Lions exponent for 3D hyper-viscous NSE.

math.AP↗

arXiv2022

A Boundary Value Problem for a Class of Anisotropic Stochastic Degenerate Parabolic-Hyperbolic Equations

We establish the well-posedness of an initial-boundary value problem of mixed type for a stochastic nonlinear parabolic-hyperbolic equation on a space domain $\cO=\cO'\X\cO''$ where a Neumann boundary condition is imposed on $\po\cO'\X\cO"$, the hyperbolic boundary, and a Dirichlet condition is imposed on $\cO'\X\po\cO"$, the parabolic boundary. Among other points to be highlighted in our analysis of this problem we mention the new strong trace theorem for the special class of stochastic nonlinear parabolic-hyperbolic equations studied here, which is decisive for the uniqueness of the kinetic solution, and the new averaging lemma for the referred class of equations which is a vital part of the proof of the strong trace property. We also provide a detailed analysis of the approximate nondegenerate problems, which is also made here for the first time, as far as the authors know, whose solutions we prove to converge to the solution of our initial-boundary value problem.

math.AP↗

arXiv2020

The Strong Trace Property and the Neumann Problem for Stochastic Conservation Laws

We establish the well-posedness of the Neumann problem for stochastic conservation laws with multiplicative noise. As a major step for establishing the uniqueness of the kinetic solution to the referred problem we establish the new strong trace property for stochastic conservation laws. Existence of kinetic solutions is proved through the vanishing viscosity method and the detailed analysis of the corresponding stochastic parabolic problem is also made here for the first time, as far as the authors know.

math.AP↗