Bo Bai

Selim F. Yilmaz, Xueyan Niu, Bo Bai et al.

We consider the image transmission problem over a noisy wireless channel via deep learning-based joint source-channel coding (DeepJSCC) along with a denoising diffusion probabilistic model (DDPM) at the receiver. Specifically, we are interested in the perception-distortion trade-off in the practical finite block length regime, in which separate source and channel coding can be highly suboptimal. We introduce a novel scheme, where the conventional DeepJSCC encoder targets transmitting a lower resolution version of the image, which later can be refined thanks to the generative model available at the receiver. In particular, we utilize the range-null space decomposition of the target image; DeepJSCC transmits the range-space of the image, while DDPM progressively refines its null space contents. Through extensive experiments, we demonstrate significant improvements in distortion and perceptual quality of reconstructed images compared to standard DeepJSCC and the state-of-the-art generative learning-based method.

Meng Qin, Chaorui Zhang, Yu Gao et al.

Graph partitioning (GP) is a classic problem that divides the node set of a graph into densely-connected blocks. Following the IEEE HPEC Graph Challenge and recent advances in pre-training techniques (e.g., large-language models), we propose PR-GPT (Pre-trained & Refined Graph ParTitioning) based on a novel pre-training & refinement paradigm. We first conduct the offline pre-training of a deep graph learning (DGL) model on small synthetic graphs with various topology properties. By using the inductive inference of DGL, one can directly generalize the pre-trained model (with frozen model parameters) to large graphs and derive feasible GP results. We also use the derived partition as a good initialization of an efficient GP method (e.g., InfoMap) to further refine the quality of partitioning. In this setting, the online generalization and refinement of PR-GPT can not only benefit from the transfer ability regarding quality but also ensure high inference efficiency without re-training. Based on a mechanism of reducing the scale of a graph to be processed by the refinement method, PR-GPT also has the potential to support streaming GP. Experiments on the Graph Challenge benchmark demonstrate that PR-GPT can ensure faster GP on large-scale graphs without significant quality degradation, compared with running a refinement method from scratch. We will make our code public at https://github.com/KuroginQin/PRGPT.

Benyuan Sun, Jin Dai, Zihao Liang et al.

Pre-training over mixtured multi-task, multi-domain, and multi-modal data remains an open challenge in vision perception pre-training. In this paper, we propose GPPF, a General Perception Pre-training Framework, that pre-trains a task-level dynamic network, which is composed by knowledge "legos" in each layers, on labeled multi-task and multi-domain datasets. By inspecting humans' innate ability to learn in complex environment, we recognize and transfer three critical elements to deep networks: (1) simultaneous exposure to diverse cross-task and cross-domain information in each batch. (2) partitioned knowledge storage in separate lego units driven by knowledge sharing. (3) sparse activation of a subset of lego units for both pre-training and downstream tasks. Noteworthy, the joint training of disparate vision tasks is non-trivial due to their differences in input shapes, loss functions, output formats, data distributions, etc. Therefore, we innovatively develop a plug-and-play multi-task training algorithm, which supports Single Iteration Multiple Tasks (SIMT) concurrently training. SIMT lays the foundation of pre-training with large-scale multi-task multi-domain datasets and is proved essential for stable training in our GPPF experiments. Excitingly, the exhaustive experiments show that, our GPPF-R50 model achieves significant improvements of 2.5-5.8 over a strong baseline of the 8 pre-training tasks in GPPF-15M and harvests a range of SOTAs over the 22 downstream tasks with similar computation budgets. We also validate the generalization ability of GPPF to SOTA vision transformers with consistent improvements. These solid experimental results fully prove the effective knowledge learning, storing, sharing, and transfer provided by our novel GPPF framework.

Meng Qin, Chaorui Zhang, Bo Bai et al.

Many network applications can be formulated as NP-hard combinatorial optimization problems of community detection (CD). Due to the NP-hardness, to balance the CD quality and efficiency remains a challenge. Most existing CD methods are transductive, which are independently optimized only for the CD on a single graph. Some of these methods use advanced machine learning techniques to obtain high-quality CD results but usually have high complexity. Other approaches use fast heuristic approximation to ensure low runtime but may suffer from quality degradation. In contrast to these transductive methods, we propose an alternative inductive community detection (ICD) method across graphs of a system or scenario to alleviate the NP-hard challenge. ICD first conducts the offline training of an adversarial dual GNN on historical graphs to capture key properties of the system. The trained model is then directly generalized to new unseen graphs for online CD without additional optimization, where a better trade-off between quality and efficiency can be achieved. ICD can also capture the permutation invariant community labels in the offline training and tackle the online CD on new graphs with non-fixed number of nodes and communities. Experiments on a set of benchmarks demonstrate that ICD can achieve a significant trade-off between quality and efficiency over various baselines.

Xingyuan Chen, Wenwei Kuang, Lei Deng et al.

The pre-training cost of large language models (LLMs) is prohibitive. One cutting-edge approach to reduce the cost is zero-shot weight transfer, also known as model growth for some cases, which magically transfers the weights trained in a small model to a large model. However, there are still some theoretical mysteries behind the weight transfer. In this paper, inspired by prior applications of mean field theory to neural network dynamics, we introduce a mean field ansatz to provide a theoretical explanation for weight transfer. Specifically, we propose the row-column (RC) ansatz under the mean field point of view, which describes the measure structure of the weights in the neural network (NN) and admits a close measure dynamic. Thus, the weights of different sizes NN admit a common distribution under proper assumptions, and weight transfer methods can be viewed as sampling methods. We empirically validate the RC ansatz by exploring simple MLP examples and LLMs such as GPT-3 and Llama-3.1. We show the mean-field point of view is adequate under suitable assumptions which can provide theoretical support for zero-shot weight transfer.

Xueyan Niu, Bo Bai, Wei Han et al.

Post-training of large language models routinely interleaves supervised fine-tuning (SFT) with reinforcement learning (RL). These two methods have different objectives: SFT minimizes the cross-entropy loss between model outputs and expert responses, while RL maximizes reward signals derived from human preferences or rule-based verifiers. Modern reasoning models have widely adopted the practice of alternating SFT and RL training. However, there is no theoretical account of whether they can be decoupled. We prove that decoupling is impossible in either order: (1) SFT-then-RL coupling: RL increases SFT loss under SFT optimality and (2) RL-then-SFT coupling: SFT lowers the reward achieved by RL. Experiments on Qwen3-0.6B confirm the predicted degradation, verifying that SFT and RL cannot be separated without loss of prior performance in the post-training

Zerui Yang, Weichuan Wang, Yanwei Xu et al.

Existing NL2SQL systems face two critical limitations: (1) they rely on in-context learning with only correct examples, overlooking the rich signal in historical error-fix pairs that could guide more robust self-correction; and (2) test-time scaling approaches often decompose questions arbitrarily, producing near-identical SQL candidates across runs and diminishing ensemble gains. Moreover, these methods suffer from a stark accuracy-efficiency trade-off: high performance demands excessive computation, while fast variants compromise quality. We present Memo-SQL, a training-free framework that addresses these issues through two simple ideas: structured decomposition and experience-aware self-correction. Instead of leaving decomposition to chance, we apply three clear strategies, entity-wise, hierarchical, and atomic sequential, to encourage diverse reasoning. For correction, we build a dynamic memory of both successful queries and historical error-fix pairs, and use retrieval-augmented prompting to bring relevant examples into context at inference time, no fine-tuning or external APIs required. On BIRD, Memo-SQL achieves 68.5% execution accuracy, setting a new state of the art among open, zero-fine-tuning methods, while using over 10 times fewer resources than prior TTS approaches.

Bohan Li, Yiming Liu, Xueyan Niu et al.

Diffusion models have achieved remarkable success in generating high quality image and video data. More recently, they have also been used for image compression with high perceptual quality. In this paper, we present a novel approach to extreme video compression leveraging the predictive power of diffusion-based generative models at the decoder. The conditional diffusion model takes several neural compressed frames and generates subsequent frames. When the reconstruction quality drops below the desired level, new frames are encoded to restart prediction. The entire video is sequentially encoded to achieve a visually pleasing reconstruction, considering perceptual quality metrics such as the learned perceptual image patch similarity (LPIPS) and the Frechet video distance (FVD), at bit rates as low as 0.02 bits per pixel (bpp). Experimental results demonstrate the effectiveness of the proposed scheme compared to standard codecs such as H.264 and H.265 in the low bpp regime. The results showcase the potential of exploiting the temporal relations in video data using generative models. Code is available at: https://github.com/ElesionKyrie/Extreme-Video-Compression-With-Prediction-Using-Pre-trainded-Diffusion-Models-

Zhengyi Jiang, Wenhao Liu, Zhongyi Huang et al.

Analog error-correcting codes (Analog ECCs) for approximate vector-matrix multiplication have been extensively studied as means to achieve fault-tolerant in-memory computation. The theoretical foundations for such coding schemes, particularly the characterization of their correction capabilities via the height profile, have been well established in recent literature. In this paper, we focus on the case of single-error detection Analog ECCs. Among several open problems related to this case proposed by Ron M. Roth in [1], Problem 1 asks: "Identify the values of $k$ and $n$ for which every linear $[n, k]$ code $\mathcal{C}$ over $\mathbb{R}$ satisfies: $$\mathsf{h}_1(\mathcal{C}):=\max_{\boldsymbol{c}\in \mathcal{C}\setminus{\{\boldsymbol{0}\}}}\mathsf{h}_1(\boldsymbol{c})\geq \Big\lceil \frac{k}{n-k} \Big\rceil.\text{"}$$ Here, for any $\boldsymbol{x}\in\mathbb{R}^n$, $\mathsf{h}_1(\boldsymbol{x})$ represents the ratio between the largest and second largest absolute values of $\boldsymbol{x}$'s entries. As the simplest special case of Problem 1 (with $n-k=2$), the following problem was posed as Problem 2 in [1]: "Must every $(n-2)$-dimensional subspace of $\mathbb{R}^n$, $n$ even, contain a nonzero vector in which the ratio between the largest and second largest absolute values of its entries is at least $(n/2)-1$?" These problems directly pertain to the lower bounds on the single-error detection threshold for Analog ECCs: Problem 1 corresponds to arbitrary $n-k$ and Problem 2 corresponds to $n-k=2$. In this paper, we provide an affirmative answer to Problem 2 and a rigorous proof using theories related to convex optimization. Furthermore, we extend our analytical method to show that the lower bound in Problem 1 is tight for the case where $n-k$ divides $k$. Our results fill the gap in the lower bound theory of thresholds for single-error detection in Analog ECCs.

Weizhi Fei, Xueyan Niu, Pingyi Zhou et al. · tsinghua

Transformer-based Large Language Models (LLMs) often impose limitations on the length of the text input to ensure the generation of fluent and relevant responses. This constraint restricts their applicability in scenarios involving long texts. We propose a novel semantic compression method that enables generalization to texts that are 6-8 times longer, without incurring significant computational costs or requiring fine-tuning. Our proposed framework draws inspiration from source coding in information theory and employs a pre-trained model to reduce the semantic redundancy of long inputs before passing them to the LLMs for downstream tasks. Experimental results demonstrate that our method effectively extends the context window of LLMs across a range of tasks including question answering, summarization, few-shot learning, and information retrieval. Furthermore, the proposed semantic compression method exhibits consistent fluency in text generation while reducing the associated computational overhead.

Xueyan Niu, Bo Bai, Lei Deng et al.

Increasing the size of a Transformer does not always lead to enhanced performance. This phenomenon cannot be explained by the empirical scaling laws. Furthermore, the model's enhanced performance is closely associated with its memorization of the training samples. We present a theoretical framework that sheds light on the memorization during pre-training of transformer-based language models. We model the behavior of Transformers with associative memories using Hopfield networks, such that each transformer block effectively conducts an approximate nearest-neighbor search. In particular, the energy function in modern continuous Hopfield networks serves as an explanation for the attention mechanism, which we approximate with a distance-based energy function. By observing that the softmax function corresponds to the gradient of the LogSumExp function in the energy, and employing the majorization-minimization technique, we construct a global energy function designed to capture the layered architecture. We demonstrate a dependency between the model size and the dataset size for the model to achieve optimal performance, and we show that the achievable cross-entropy loss is bounded from below.

Bo Bai

Large language models (LLMs) have demonstrated remarkable capabilities in numerous real-world applications. While the vast majority of research conducted from an experimental perspective is progressing rapidly, it demands substantial computational power, data, and other resources. Therefore, how to open the black-box of LLMs from a theoretical standpoint has become a critical challenge. This paper takes the theory of rate-distortion function, directed information, and Granger causality as its starting point to investigate the information-theoretic principles behind LLMs, leading to the development of semantic information theory for LLMs, where the fundamental unit is token, rather than bits that lacks any semantic meaning. By defining the probabilistic model of LLMs, we discuss structure-agnostic information-theoretic measures, such as the directed rate-distortion function in pre-training, the directed rate-reward function in post-training, and the semantic information flow in inference phase. This paper also delves deeply into the theory of token-level semantic embedding and the information-theoretically optimal vectorization method. Thereafter, we propose a general definition of autoregression LLM, where the Transformer architecture and its performance such as ELBO, generalization error bound, memory capacity, and semantic information measures can be derived theoretically. Other architectures, such as Mamba/Mamba2 and LLaDA, are also discussed in our framework. Consequently, this paper provides a theoretical framework for understanding LLMs from the perspective of semantic information theory, which also offers the necessary theoretical tools for further in-depth research.

Weizhi Fei, Xueyan Niu, Guoqing Xie et al. · tsinghua

Although applications involving long-context inputs are crucial for the effective utilization of large language models (LLMs), they also result in increased computational costs and reduced performance. To address this challenge, we propose an efficient, training-free prompt compression method that retains key information within compressed prompts. We identify specific attention heads in transformer-based LLMs, which we designate as evaluator heads, that are capable of selecting tokens in long inputs that are most significant for inference. Building on this discovery, we develop EHPC, an Evaluator Head-based Prompt Compression method, which enables LLMs to rapidly "skim through" input prompts by leveraging only the first few layers with evaluator heads during the pre-filling stage, subsequently passing only the important tokens to the model for inference. EHPC achieves state-of-the-art results across two mainstream benchmarks: prompt compression and long-context inference acceleration. Consequently, it effectively reduces the complexity and costs associated with commercial API calls. We further demonstrate that EHPC attains competitive results compared to key-value cache-based acceleration methods, thereby highlighting its potential to enhance the efficiency of LLMs for long-context tasks.

Yikuan Xia, Jiazun Chen, Yirui Zhan et al.

This paper presents the db3 team's winning solution for the Meta CRAG-MM Challenge 2025 at KDD Cup'25. Addressing the challenge's unique multi-modal, multi-turn question answering benchmark (CRAG-MM), we developed a comprehensive framework that integrates tailored retrieval pipelines for different tasks with a unified LLM-tuning approach for hallucination control. Our solution features (1) domain-specific retrieval pipelines handling image-indexed knowledge graphs, web sources, and multi-turn conversations; and (2) advanced refusal training using SFT, DPO, and RL. The system achieved 2nd place in Task 1, 2nd place in Task 2, and 1st place in Task 3, securing the grand prize for excellence in ego-centric queries through superior handling of first-person perspective challenges.

Weizhi Fei, Hao Shi, Jing Xu et al. · tsinghua

Efficiently editing knowledge stored in large language models (LLMs) enables model updates without large-scale training. One possible solution is Locate-and-Edit (L\&E), allowing simultaneous modifications of a massive number of facts. However, such editing may compromise the general abilities of LLMs and even result in forgetting edited facts when scaling up to thousands of edits. In this paper, we model existing linear L\&E methods as querying a Key-Value (KV) database. From this perspective, we then propose NeuralDB, an editing framework that explicitly represents the edited facts as a neural KV database equipped with a non-linear gated retrieval module, % In particular, our gated module only operates when inference involves the edited facts, effectively preserving the general abilities of LLMs. Comprehensive experiments involving the editing of 10,000 facts were conducted on the ZsRE and CounterFacts datasets, using GPT2-XL, GPT-J (6B) and Llama-3 (8B). The results demonstrate that NeuralDB not only excels in editing efficacy, generalization, specificity, fluency, and consistency, but also preserves overall performance across six representative text understanding and generation tasks. Further experiments indicate that NeuralDB maintains its effectiveness even when scaled to 100,000 facts (\textbf{50x} more than in prior work).

Yikuan Xia, Jiazun Chen, Yirui Zhan et al.

Large language models (LLMs) excel in question-answering (QA) tasks, and retrieval-augmented generation (RAG) enhances their precision by incorporating external evidence from diverse sources like web pages, databases, and knowledge graphs. However, current RAG methods rely on agent-specific strategies for individual data sources, posing challenges low-resource or black-box environments and complicates operations when evidence is fragmented across sources. To address these limitations, we propose ER-RAG, a framework that unifies evidence integration across heterogeneous data sources using the Entity-Relationship (ER) model. ER-RAG standardizes entity retrieval and relationship querying through ER-based APIs with GET and JOIN operations. It employs a two-stage generation process: first, a preference optimization module selects optimal sources; second, another module constructs API chains based on source schemas. This unified approach allows efficient fine-tuning and seamless integration across diverse data sources. ER-RAG demonstrated its effectiveness by winning all three tracks of the 2024 KDDCup CRAG Challenge, achieving performance on par with commercial RAG pipelines using an 8B LLM backbone. It outperformed hybrid competitors by 3.1% in LLM score and accelerated retrieval by 5.5X.

Weizhi Fei, Xueyan Niu, Guoqing Xie et al.

Current Large Language Models (LLMs) face inherent limitations due to their pre-defined context lengths, which impede their capacity for multi-hop reasoning within extensive textual contexts. While existing techniques like Retrieval-Augmented Generation (RAG) have attempted to bridge this gap by sourcing external information, they fall short when direct answers are not readily available. We introduce a novel approach that re-imagines information retrieval through dynamic in-context editing, inspired by recent breakthroughs in knowledge editing. By treating lengthy contexts as malleable external knowledge, our method interactively gathers and integrates relevant information, thereby enabling LLMs to perform sophisticated reasoning steps. Experimental results demonstrate that our method effectively empowers context-limited LLMs, such as Llama2, to engage in multi-hop reasoning with improved performance, which outperforms state-of-the-art context window extrapolation methods and even compares favorably to more advanced commercial long-context models. Our interactive method not only enhances reasoning capabilities but also mitigates the associated training and computational costs, making it a pragmatic solution for enhancing LLMs' reasoning within expansive contexts.

Lingyi Chen, Shitong Wu, Wenhao Ye et al.

The Blahut-Arimoto (BA) algorithm has played a fundamental role in the numerical computation of rate-distortion (RD) functions. This algorithm possesses a desirable monotonic convergence property by alternatively minimizing its Lagrangian with a fixed multiplier. In this paper, we propose a novel modification of the BA algorithm, wherein the multiplier is updated through a one-dimensional root-finding step using a monotonic univariate function, efficiently implemented by Newton's method in each iteration. Consequently, the modified algorithm directly computes the RD function for a given target distortion, without exploring the entire RD curve as in the original BA algorithm. Moreover, this modification presents a versatile framework, applicable to a wide range of problems, including the computation of distortion-rate (DR) functions. Theoretical analysis shows that the outputs of the modified algorithms still converge to the solutions of the RD and DR functions with rate $O(1/n)$, where $n$ is the number of iterations. Additionally, these algorithms provide $\varepsilon$-approximation solutions with $O\left(\frac{MN\log N}{\varepsilon}(1+\log |\log \varepsilon|)\right)$ arithmetic operations, where $M,N$ are the sizes of source and reproduced alphabets respectively. Numerical experiments demonstrate that the modified algorithms exhibit significant acceleration compared with the original BA algorithms and showcase commendable performance across classical source distributions such as discretized Gaussian, Laplacian and uniform sources.

Lingying Huang, Xiaomeng Chen, Wei Huo et al.

In this paper, we surveyed the existing literature studying different approaches and algorithms for the four critical components in the general branch and bound (B&B) algorithm, namely, branching variable selection, node selection, node pruning, and cutting-plane selection. However, the complexity of the B&B algorithm always grows exponentially with respect to the increase of the decision variable dimensions. In order to improve the speed of B&B algorithms, learning techniques have been introduced in this algorithm recently. We further surveyed how machine learning can be used to improve the four critical components in B&B algorithms. In general, a supervised learning method helps to generate a policy that mimics an expert but significantly improves the speed. An unsupervised learning method helps choose different methods based on the features. In addition, models trained with reinforcement learning can beat the expert policy, given enough training and a supervised initialization. Detailed comparisons between different algorithms have been summarized in our survey. Finally, we discussed some future research directions to accelerate and improve the algorithms further in the literature.

Ze Cui, Jing Wang, Shangyin Gao et al.

With the development of deep learning techniques, the combination of deep learning with image compression has drawn lots of attention. Recently, learned image compression methods had exceeded their classical counterparts in terms of rate-distortion performance. However, continuous rate adaptation remains an open question. Some learned image compression methods use multiple networks for multiple rates, while others use one single model at the expense of computational complexity increase and performance degradation. In this paper, we propose a continuously rate adjustable learned image compression framework, Asymmetric Gained Variational Autoencoder (AG-VAE). AG-VAE utilizes a pair of gain units to achieve discrete rate adaptation in one single model with a negligible additional computation. Then, by using exponential interpolation, continuous rate adaptation is achieved without compromising performance. Besides, we propose the asymmetric Gaussian entropy model for more accurate entropy estimation. Exhaustive experiments show that our method achieves comparable quantitative performance with SOTA learned image compression methods and better qualitative performance than classical image codecs. In the ablation study, we confirm the usefulness and superiority of gain units and the asymmetric Gaussian entropy model.

Kai Lei, Meng Qin, Bo Bai et al.

In this paper, we generally formulate the dynamics prediction problem of various network systems (e.g., the prediction of mobility, traffic and topology) as the temporal link prediction task. Different from conventional techniques of temporal link prediction that ignore the potential non-linear characteristics and the informative link weights in the dynamic network, we introduce a novel non-linear model GCN-GAN to tackle the challenging temporal link prediction task of weighted dynamic networks. The proposed model leverages the benefits of the graph convolutional network (GCN), long short-term memory (LSTM) as well as the generative adversarial network (GAN). Thus, the dynamics, topology structure and evolutionary patterns of weighted dynamic networks can be fully exploited to improve the temporal link prediction performance. Concretely, we first utilize GCN to explore the local topological characteristics of each single snapshot and then employ LSTM to characterize the evolving features of the dynamic networks. Moreover, GAN is used to enhance the ability of the model to generate the next weighted network snapshot, which can effectively tackle the sparsity and the wide-value-range problem of edge weights in real-life dynamic networks. To verify the model's effectiveness, we conduct extensive experiments on four datasets of different network systems and application scenarios. The experimental results demonstrate that our model achieves impressive results compared to the state-of-the-art competitors.