Advanced Planning and Scheduling (APS) systems have become indispensable for modern manufacturing operations, enabling optimized resource allocation and production efficiency in increasingly complex and dynamic environments. While algorithms for solving abstracted scheduling problems have been extensively investigated, the critical prerequisite of specifying manufacturing requirements into formal constraints remains manual and labor-intensive. Although recent advances of generative models, particularly Large Language Models (LLMs), show promise in automating constraint specification from heterogeneous raw manufacturing data, their direct application faces challenges due to natural language ambiguity, non-deterministic outputs, and limited domain-specific knowledge. This paper presents a constraint-centric architecture that regulates LLMs to perform reliable automated constraint specification for production scheduling. The architecture defines a hierarchical structural space organized across three levels, implemented through domain-specific representation to ensure precision and reliability while maintaining flexibility. Furthermore, an automated production scenario adaptation algorithm is designed and deployed to efficiently customize the architecture for specific manufacturing configurations. Experimental results demonstrate that the proposed approach successfully balances the generative capabilities of LLMs with the reliability requirements of manufacturing systems, significantly outperforming pure LLM-based approaches in constraint specification tasks.

Machine learning (ML) has become a cornerstone of key applications at ByteDance. As model complexity and data volumes surge, data management for large-scale ML workloads faces substantial challenges, particularly with recent advances in large recommendation models (LRMs) and large multimodal models (LMMs). Traditional approaches exhibit limitations in storage efficiency, metadata scalability, update mechanisms, and integration with ML frameworks. To address these challenges, we propose Magnus, a holistic data management system built upon Apache Iceberg. Magnus integrates innovative optimizations across resource-efficient storage formats optimized for large wide tables and multimodal data, built-in support for vector and inverted indexes to accelerate data retrieval, scalable metadata planning with Git-like branching and tagging capabilities, and high-performance update/upsert based on lightweight merge-on-read (MOR) strategies. Additionally, Magnus provides native support and specialized enhancement for LRM and LMM training workloads. Experimental results demonstrate significant performance gains in real-world ML scenarios. Magnus has been deployed at ByteDance for over five years, enabling robust and efficient data infrastructure for large-scale ML workloads.

Abstract. The List-3-Coloring Problem is to decide, given a graph [Formula: see text] and a list [Formula: see text] of colors assigned to each vertex [Formula: see text] of [Formula: see text], whether [Formula: see text] admits a proper coloring [Formula: see text] with [Formula: see text] for every vertex [Formula: see text] of [Formula: see text], and the 3-Coloring Problem is the List-3-Coloring Problem on instances with [Formula: see text] for every vertex [Formula: see text] of [Formula: see text]. The List-3-Coloring Problem is a classical NP -complete problem, and it is well-known that while restricted to [Formula: see text]- free graphs (meaning graphs with no induced subgraph isomorphic to a fixed graph [Formula: see text]), it remains NP -complete unless [Formula: see text] is isomorphic to an induced subgraph of a path. However, the current state of art is far from proving this to be sufficient for a polynomial time algorithm; in fact, the complexity of the 3-Coloring Problem on [Formula: see text]-free graphs (where [Formula: see text] denotes the eight-vertex path) is unknown. Here we consider a variant of the List-3-Coloring Problem called the Ordered Graph List-3-Coloring Problem, where the input is an ordered graph, that is, a graph along with a linear order on its vertex set. For ordered graphs [Formula: see text] and [Formula: see text], we say [Formula: see text] is [Formula: see text]- free if [Formula: see text] is not isomorphic to an induced subgraph of [Formula: see text] with the isomorphism preserving the linear order. We prove, assuming [Formula: see text] to be an ordered graph, a nearly complete dichotomy for the Ordered Graph List-3-Coloring Problem restricted to [Formula: see text]-free ordered graphs. In particular, we show that the problem can be solved in polynomial time if [Formula: see text] has at most one edge, and remains NP -complete if [Formula: see text] has at least three edges. Moreover, in the case where [Formula: see text] has exactly two edges, we give a complete dichotomy when the two edges of [Formula: see text] share an end, and prove several NP -completeness results when the two edges of [Formula: see text] do not share an end, narrowing the open cases down to three very special types of two-edge ordered graphs.

Autoregressive (AR) automatic speech recognition (ASR) models predict each output token conditioning on the previous ones, which slows down their inference speed. On the other hand, non-autoregressive (NAR) models predict tokens independently and simultaneously within a constant number of decoding iterations, which brings high inference speed. However, NAR models generally have lower accuracy than AR models. In this work, we propose AR pretraining to the NAR encoder to reduce the accuracy gap between AR and NAR models. The experiment results show that our AR-pretrained MaskCTC reaches the same accuracy as AR Conformer on Aishell-1 (both 4.9% CER) and reduce the performance gap with AR Conformer on LibriSpeech by relatively 50%. Moreover, our AR-pretrained MaskCTC only needs single decoding iteration, which reduces inference time by 50%. We also investigate multiple masking strategies in training the masked language model of MaskCTC.

For a positive integer $r$ and graphs $G$ and $H$, we denote by $G+H$ the disjoint union of $G$ and $H$ and by $rH$ the union of $r$ mutually disjoint copies of $H$. Also, we say $G$ is $H$ -free if $H$ is not isomorphic to an induced subgraph of $G$. We use $P_t$ to denote the path on $t$ vertices. For a fixed positive integer $k$, the List-$k$-Coloring Problem is to decide, given a graph $G$ and a list $L(v)\subseteq \{1,\ldots,k\}$ of colors assigned to each vertex $v$ of $G$, whether $G$ admits a proper coloring $\phi$ with $\phi(v)\in L(v)$ for every vertex $v$ of $G$, and the $k$-Coloring Problem is the List-$k$-Coloring Problem restricted to instances with $L(v)=\{1,\ldots, k\}$ for every vertex $v$ of $G$. We prove that, for every positive integer $r$, the List-$5$-Coloring Problem restricted to $rP_3$-free graphs can be solved in polynomial time. Together with known results, this gives a complete dichotomy for the complexity of the List-5-Coloring Problem restricted to $H$-free graphs: For every graph $H$, assuming P$\neq$NP, the List-5-Coloring Problem restricted to $H$-free graphs can be solved in polynomial time if and only if, $H$ is an induced subgraph of either $rP_3$ or $P_5+rP_1$ for some positive integer $r$. As a hardness counterpart, we also show that the $k$-Coloring Problem restricted to $rP_4$-free graphs is NP-complete for all $k\geq 5$ and $r\geq 2$.

The input to a PPNC instance is integers n and p, and a non-negative real weighting of the edges of the clique Kn on the vertex set {1,..., n}. We are asked to find a set of p disjoint cycles spanning {1,..., n} and subject to this such that the sum of the weights of the edges is minimized. We provide an efficient approximation algorithm for the metric version of this problem which has an approximation ratio of 4 if p ≤ n/5 and an approximation ratio of 51 for larger p. For p > n/5, our algorithm uses a subroutine which approximately solves the Mod 3 T-join With Slack problem. The input to an instance of Mod k T-join with Slack consists of integers n and B, a non-negative weighting of the edges of the clique Kn, and a label l(v) from {0,1,..., k - 1} on each vertex of Kn. We are asked to find the minimum weight spanning forest F from amongst those satisfying ∑T∈F((∑v∈V(T)l(v)) mod k) ≤ B. If k = 2 and B = 0 this is the well-studied T-join problem which can be solved exactly in polynomial time.