Towards Next Era of Multi-objective Optimization: Large Language Models as Architects of Evolutionary Operators

Multi-objective optimization problems (MOPs) are prevalent in various real-world applications, necessitating sophisticated solutions that balance conflicting objectives. Traditional evolutionary algorithms (EAs), while effective, often rely on domain-specific expert knowledge and iterative tuning, which can impede innovation when encountering novel MOPs. Very recently, the emergence of Large Language Models (LLMs) has revolutionized software engineering by enabling the autonomous development and refinement of programs. Capitalizing on this advancement, we propose a new LLM-based framework for evolving EA operators, designed to address a wide array of MOPs. This framework facilitates the production of EA operators without the extensive demands for expert intervention, thereby streamlining the design process. To validate the efficacy of our approach, we have conducted extensive empirical studies across various categories of MOPs. The results demonstrate the robustness and superior performance of our LLM-evolved operators.

Optimizing Large Model Training through Overlapped Activation Recomputation

Large model training has been using recomputation to alleviate the memory pressure and pipelining to exploit the parallelism of data, tensor, and devices. The existing recomputation approaches may incur up to 40% overhead when training real-world models, e.g., the GPT model with 22B parameters. This is because they are executed on demand in the critical training path. In this paper, we design a new recomputation framework, Lynx, to reduce the overhead by overlapping the recomputation with communication occurring in training pipelines. It consists of an optimal scheduling algorithm (OPT) and a heuristic-based scheduling algorithm (HEU). OPT achieves a global optimum but suffers from a long search time. HEU was designed based on our observation that there are identical structures in large DNN models so that we can apply the same scheduling policy to all identical structures. HEU achieves a local optimum but reduces the search time by 99% compared to OPT. Our comprehensive evaluation using GPT models with 1.3B-20B parameters shows that both OPT and HEU outperform the state-of-the-art recomputation approaches (e.g., Megatron-LM and Checkmake) by 1.02-1.53x. HEU achieves a similar performance as OPT with a search time of 0.16s on average.

Seed-TTS: A Family of High-Quality Versatile Speech Generation Models

We introduce Seed-TTS, a family of large-scale autoregressive text-to-speech (TTS) models capable of generating speech that is virtually indistinguishable from human speech. Seed-TTS serves as a foundation model for speech generation and excels in speech in-context learning, achieving performance in speaker similarity and naturalness that matches ground truth human speech in both objective and subjective evaluations. With fine-tuning, we achieve even higher subjective scores across these metrics. Seed-TTS offers superior controllability over various speech attributes such as emotion and is capable of generating highly expressive and diverse speech for speakers in the wild. Furthermore, we propose a self-distillation method for speech factorization, as well as a reinforcement learning approach to enhance model robustness, speaker similarity, and controllability. We additionally present a non-autoregressive (NAR) variant of the Seed-TTS model, named $\text{Seed-TTS}_\text{DiT}$, which utilizes a fully diffusion-based architecture. Unlike previous NAR-based TTS systems, $\text{Seed-TTS}_\text{DiT}$ does not depend on pre-estimated phoneme durations and performs speech generation through end-to-end processing. We demonstrate that this variant achieves comparable performance to the language model-based variant and showcase its effectiveness in speech editing. We encourage readers to listen to demos at \url{https://bytedancespeech.github.io/seedtts_tech_report}.

SciQAG: A Framework for Auto-Generated Scientific Question Answering Dataset with Fine-grained Evaluation

The use of question-answer (QA) pairs for training and evaluating large language models (LLMs) has attracted considerable attention. Yet few available QA datasets are based on knowledge from the scientific literature. Here we bridge this gap by presenting Automatic Generation of Scientific Question Answers (SciQAG), a framework for automatic generation and evaluation of scientific QA pairs sourced from published scientific literature. We fine-tune an open-source LLM to generate \num{960000} scientific QA pairs from full-text scientific papers and propose a five-dimensional metric to evaluate the quality of the generated QA pairs. We show via LLM-based evaluation that the generated QA pairs consistently achieve an average score of 2.5 out of 3 across five dimensions, indicating that our framework can distill key knowledge from papers into high-quality QA pairs at scale. We make the dataset, models, and evaluation codes publicly available.

Hypergraph Self-supervised Learning with Sampling-efficient Signals

Self-supervised learning (SSL) provides a promising alternative for representation learning on hypergraphs without costly labels. However, existing hypergraph SSL models are mostly based on contrastive methods with the instance-level discrimination strategy, suffering from two significant limitations: (1) They select negative samples arbitrarily, which is unreliable in deciding similar and dissimilar pairs, causing training bias. (2) They often require a large number of negative samples, resulting in expensive computational costs. To address the above issues, we propose SE-HSSL, a hypergraph SSL framework with three sampling-efficient self-supervised signals. Specifically, we introduce two sampling-free objectives leveraging the canonical correlation analysis as the node-level and group-level self-supervised signals. Additionally, we develop a novel hierarchical membership-level contrast objective motivated by the cascading overlap relationship in hypergraphs, which can further reduce membership sampling bias and improve the efficiency of sample utilization. Through comprehensive experiments on 7 real-world hypergraphs, we demonstrate the superiority of our approach over the state-of-the-art method in terms of both effectiveness and efficiency.

Seeing Text in the Dark: Algorithm and Benchmark

Localizing text in low-light environments is challenging due to visual degradations. Although a straightforward solution involves a two-stage pipeline with low-light image enhancement (LLE) as the initial step followed by detector, LLE is primarily designed for human vision instead of machine and can accumulate errors. In this work, we propose an efficient and effective single-stage approach for localizing text in dark that circumvents the need for LLE. We introduce a constrained learning module as an auxiliary mechanism during the training stage of the text detector. This module is designed to guide the text detector in preserving textual spatial features amidst feature map resizing, thus minimizing the loss of spatial information in texts under low-light visual degradations. Specifically, we incorporate spatial reconstruction and spatial semantic constraints within this module to ensure the text detector acquires essential positional and contextual range knowledge. Our approach enhances the original text detector's ability to identify text's local topological features using a dynamic snake feature pyramid network and adopts a bottom-up contour shaping strategy with a novel rectangular accumulation technique for accurate delineation of streamlined text features. In addition, we present a comprehensive low-light dataset for arbitrary-shaped text, encompassing diverse scenes and languages. Notably, our method achieves state-of-the-art results on this low-light dataset and exhibits comparable performance on standard normal light datasets. The code and dataset will be released.

Construction of Functional Materials Knowledge Graph in Multidisciplinary Materials Science via Large Language Model

The convergence of materials science and artificial intelligence has unlocked new opportunities for gathering, analyzing, and generating novel materials sourced from extensive scientific literature. Despite the potential benefits, persistent challenges such as manual annotation, precise extraction, and traceability issues remain. Large language models have emerged as promising solutions to address these obstacles. This paper introduces Functional Materials Knowledge Graph (FMKG), a multidisciplinary materials science knowledge graph. Through the utilization of advanced natural language processing techniques, extracting millions of entities to form triples from a corpus comprising all high-quality research papers published in the last decade. It organizes unstructured information into nine distinct labels, covering Name, Formula, Acronym, Structure/Phase, Properties, Descriptor, Synthesis, Characterization Method, Application, and Domain, seamlessly integrating papers' Digital Object Identifiers. As the latest structured database for functional materials, FMKG acts as a powerful catalyst for expediting the development of functional materials and a fundation for building a more comprehensive material knowledge graph using full paper text. Furthermore, our research lays the groundwork for practical text-mining-based knowledge management systems, not only in intricate materials systems but also applicable to other specialized domains.

STG-Mamba: Spatial-Temporal Graph Learning via Selective State Space Model

Spatial-Temporal Graph (STG) data is characterized as dynamic, heterogenous, and non-stationary, leading to the continuous challenge of spatial-temporal graph learning. In the past few years, various GNN-based methods have been proposed to solely focus on mimicking the relationships among node individuals of the STG network, ignoring the significance of modeling the intrinsic features that exist in STG system over time. In contrast, modern Selective State Space Models (SSSMs) present a new approach which treat STG Network as a system, and meticulously explore the STG system's dynamic state evolution across temporal dimension. In this work, we introduce Spatial-Temporal Graph Mamba (STG-Mamba) as the first exploration of leveraging the powerful selective state space models for STG learning by treating STG Network as a system, and employing the Graph Selective State Space Block (GS3B) to precisely characterize the dynamic evolution of STG networks. STG-Mamba is formulated as an Encoder-Decoder architecture, which takes GS3B as the basic module, for efficient sequential data modeling. Furthermore, to strengthen GNN's ability of modeling STG data under the setting of SSSMs, we propose Kalman Filtering Graph Neural Networks (KFGN) for adaptive graph structure upgrading. KFGN smoothly fits in the context of selective state space evolution, and at the same time keeps linear complexity. Extensive empirical studies are conducted on three benchmark STG forecasting datasets, demonstrating the performance superiority and computational efficiency of STG-Mamba. It not only surpasses existing state-of-the-art methods in terms of STG forecasting performance, but also effectively alleviate the computational bottleneck of large-scale graph networks in reducing the computational cost of FLOPs and test inference time.

What is Wrong with End-to-End Learning for Phase Retrieval?

For nonlinear inverse problems that are prevalent in imaging science, symmetries in the forward model are common. When data-driven deep learning approaches are used to solve such problems, these intrinsic symmetries can cause substantial learning difficulties. In this paper, we explain how such difficulties arise and, more importantly, how to overcome them by preprocessing the training set before any learning, i.e., symmetry breaking. We take far-field phase retrieval (FFPR), which is central to many areas of scientific imaging, as an example and show that symmetric breaking can substantially improve data-driven learning. We also formulate the mathematical principle of symmetry breaking.

How Multimodal Integration Boost the Performance of LLM for Optimization: Case Study on Capacitated Vehicle Routing Problems

Recently, large language models (LLMs) have notably positioned them as capable tools for addressing complex optimization challenges. Despite this recognition, a predominant limitation of existing LLM-based optimization methods is their struggle to capture the relationships among decision variables when relying exclusively on numerical text prompts, especially in high-dimensional problems. Keeping this in mind, we first propose to enhance the optimization performance using multimodal LLM capable of processing both textual and visual prompts for deeper insights of the processed optimization problem. This integration allows for a more comprehensive understanding of optimization problems, akin to human cognitive processes. We have developed a multimodal LLM-based optimization framework that simulates human problem-solving workflows, thereby offering a more nuanced and effective analysis. The efficacy of this method is evaluated through extensive empirical studies focused on a well-known combinatorial optimization problem, i.e., capacitated vehicle routing problem. The results are compared against those obtained from the LLM-based optimization algorithms that rely solely on textual prompts, demonstrating the significant advantages of our multimodal approach.