Hongbin Zhang

Nankai Lin, Hongbin Zhang, Menglan Shen et al.

Grammatical error correction (GEC) is a challenging task of natural language processing techniques. While more attempts are being made in this approach for universal languages like English or Chinese, relatively little work has been done for low-resource languages for the lack of large annotated corpora. In low-resource languages, the current unsupervised GEC based on language model scoring performs well. However, the pre-trained language model is still to be explored in this context. This study proposes a BERT-based unsupervised GEC framework, where GEC is viewed as multi-class classification task. The framework contains three modules: data flow construction module, sentence perplexity scoring module, and error detecting and correcting module. We propose a novel scoring method for pseudo-perplexity to evaluate a sentence's probable correctness and construct a Tagalog corpus for Tagalog GEC research. It obtains competitive performance on the Tagalog corpus we construct and open-source Indonesian corpus and it demonstrates that our framework is complementary to baseline method for low-resource GEC task.

Liutong Han, Zhiyuan Tan, Hongbin Zhang et al.

The use of intrinsic functions to leverage hardware-specific capabilities is a crucial approach for optimizing library performance. Many mainstream libraries implement a large number of vectorized algorithms on Arm or x86 SIMD (Single-Instruction, Multiple-Data) intrinsic functions. Translating existing vectorized intrinsic code into the intrinsics of an emerging architecture is a practical and effective approach. However, current cross-architecture translation largely relies on manual rewriting or rule-based mapping methods, which are both time-consuming and prone to errors. We present \texttt{IntrinTrans}, a LLM-based agent that utilizes compile-and-test feedback to translate intrinsic code across architectures automatically, and further optimizes the generated intrinsics using register-usage information derived from liveness analysis. To evaluate the effectiveness of our method, we used \texttt{IntrinTrans} to translate the open-source benchmark from Arm Neon Intrinsic to the emerging RISC-V Vector (RVV) Intrinsic implementation and compared its performance with that of the native RVV implementation. Our experiments show that advanced LLMs can generate semantically correct RVV Intrinsic functions with only a finite number of iterations. Depending on the base LLMs, the pass rate ranges from 47% to 100%, achieving performance similar to the native implementation (0.85x to 1.28x).

Dianhao Zheng, Hongbin Zhang, J. Andrew Zhang et al.

This note investigates the stability of both linear and nonlinear switched systems with average dwell time. Two new analysis methods are proposed. Different from existing approaches, the proposed methods take into account the sequence in which the subsystems are switched. Depending on the predecessor or successor subsystems to be considered, sequence-based average preceding dwell time (SBAPDT) and sequence-based average subsequence dwell time (SBASDT) approaches are proposed and discussed for both continuous and discrete time systems. These proposed methods, when considering the switch sequence, have the potential to further reduce the conservativeness of the existing approaches. A comparative numerical example is also given to demonstrate the advantages of the proposed approaches.

Fengyao Bai, Hongbin Zhang, Zhitao Chen et al.

High-throughput inference serving is essential for applications built on large language models (LLMs). Existing serving frameworks reduce request-level and batch-level bubbles through batching and scheduling, but often overlook bubbles within each decode iteration. Tokens generated in the same iteration may incur different costs because they depend on KV caches of different lengths; tokens with long KV caches can become bottlenecks and delay the next iteration. We propose AlignedServe, an LLM serving framework built around prefix-aware batching. It groups requests with similar KV-cache lengths into the same batch to reduce iteration-level bubbles. To support this policy efficiently, AlignedServe uses large CPU memory to maintain sufficient in-flight requests for batching and applies a batch-level scheduling policy to reduce batch-level bubbles. It also introduces a GPU-Prefetch-For-GPU architecture, where one GPU prefetches KV cache for another to reduce CPU-to-GPU transfer latency. Experiments on synthetic and application workloads show that AlignedServe improves decoding throughput by up to 1.98 times and reduces latency by up to 7.4 times over state-of-the-art systems.

Hongbin Zhang, Chaozheng Wang, Kehai Chen et al.

On-policy self-distillation (OPSD) is an emerging LLM post-training paradigm in which the model serves as its own teacher: conditioned on privileged information such as a reference trace or hint, the same policy provides dense token-level supervision on its own rollouts. However, recent studies show that OPSD degrades complex reasoning by suppressing predictive uncertainty, which supports exploration and hypothesis revision. Our token-level analysis shows that this failure arises from applying a uniform direction of teacher supervision across tokens with different uncertainty levels: conformity to the privileged self-teacher suppresses exploration at high entropy, while deviation from the teacher degrades step accuracy at low entropy. Accordingly, we propose \textbf{Direction-Adaptive Self-Distillation} (\textbf{DASD}), which reframes privileged self-distillation from uniform teacher imitation into entropy-routed directional supervision: high-entropy tokens are pushed away from the privileged teacher to preserve exploration, while low-entropy tokens are pulled toward the teacher to stabilize step-level execution. Across six mathematical reasoning benchmarks, DASD achieves the best macro Avg@16 over strong RLVR and self-distillation baselines. Pass@$k$, reasoning-health, and generalization analyses show that these average gains come from preserving exploration without sacrificing step-level execution.

Hongbin Zhang, Ning Gao, Yuqin Dai et al.

Proactive task-oriented dialogue (TOD), such as outbound sales, demands a persuasive agent that actively probes the user's concerns and steers the conversation toward acceptance within a bounded number of turns. Yet post-trained LLMs are inherently conservative, and reward-shaping RL (e.g., GRPO) struggles since it only re-weights what an already passive policy samples. We show that conditioning on the user's latent concerns unlocks proactive capability that no amount of sampling can undermine, establishing these concerns as a pivotal training-time signal. To operationalize this finding, we build the \textbf{Cognitive User Simulator}, which models each user as a stratified persona comprising observable external traits and hidden internal concerns. The simulator produces faithful and diverse interactions, while emitting per-turn state dynamics that track persuasion progress. We then introduce \textbf{Simulator-Induced Asymmetric-View Policy Optimization}, which converts the modeled concerns and the simulation state transition into complementary training objectives: (1) \emph{Asymmetric On-Policy Self-Distillation} that transfers concern-aware behavior from a privileged view of the same policy into its deployable, conversation-only view; and (2) \emph{State-Transition Policy Refinement} ...

Hongbin Zhang, Yu Yang, Feng Wu et al.

Optimizing the assortment of products to display to customers is a key to increasing revenue for both offline and online retailers. To trade-off between exploring customers' preference and exploiting customers' choices learned from data, in this paper, by adopting the Multi-Nomial Logit (MNL) choice model to capture customers' choices over products, we study the problem of optimizing assortments over a planning horizon $T$ for maximizing the profit of the retailer. To make the problem setting more practical, we consider both the inventory constraint and the limited switches constraint, where the retailer cannot use up the resource inventory before time $T$ and is forbidden to switch the assortment shown to customers too many times. Such a setting suits the case when an online retailer wants to dynamically optimize the assortment selection for a population of customers. We develop an efficient UCB-like algorithm to optimize the assortments while learning customers' choices from data. We prove that our algorithm can achieve a sub-linear regret bound $\tilde{O}\left(T^{1-α/2}\right)$ if $O(T^α)$ switches are allowed. %, and our regret bound is optimal with respect to $T$. Extensive numerical experiments show that our algorithm outperforms baselines and the gap between our algorithm's performance and the theoretical upper bound is small.

Daobin Zhu, Xiangdong Su, Hongbin Zhang

Connectionist temporal classification (CTC) and attention-based encoder decoder (AED) joint training has been widely applied in automatic speech recognition (ASR). Unlike most hybrid models that separately calculate the CTC and AED losses, our proposed integrated-CTC utilizes the attention mechanism of AED to guide the output of CTC. In this paper, we employ two fusion methods, namely direct addition of logits (DAL) and preserving the maximum probability (PMP). We achieve dimensional consistency by adaptively affine transforming the attention results to match the dimensions of CTC. To accelerate model convergence and improve accuracy, we introduce auxiliary loss regularization for accelerated convergence. Experimental results demonstrate that the DAL method performs better in attention rescoring, while the PMP method excels in CTC prefix beam search and greedy search.

Yixuan Zhang, Zhiyuan Li, Weijia He et al.

Target-oriented discovery under limited evaluation budgets requires making reliable progress in high-dimensional, heterogeneous design spaces where each new measurement is costly, whether experimental or high-fidelity simulation. We present an information-theoretic framework for target-oriented adaptive sampling that reframes optimization as trajectory discovery: instead of approximating the full response surface, the method maintains and refines a low-entropy information state that concentrates search on target-relevant directions. The approach couples data, model beliefs, and physics/structure priors through dimension-aware information budgeting, adaptive bootstrapped distillation over a heterogeneous surrogate reservoir, and structure-aware candidate manifold analysis with Kalman-inspired multi-model fusion to balance consensus-driven exploitation and disagreement-driven exploration. Evaluated under a single unified protocol without dataset-specific tuning, the framework improves sample efficiency and reliability across 14 single- and multi-objective materials design tasks spanning candidate pools from $600$ to $4 \times 10^6$ and feature dimensions from $10$ to $10^3$, typically reaching top-performing regions within 100 evaluations. Complementary 20-dimensional synthetic benchmarks (Ackley, Rastrigin, Schwefel) further demonstrate robustness to rugged and multimodal landscapes.

Haolin Pan, Jinyuan Dong, Hongbin Zhang et al.

Learning effective numerical representations, or embeddings, of programs is a fundamental prerequisite for applying machine learning to automate and enhance compiler optimization. Prevailing paradigms, however, present a dilemma. Static representations, derived from source code or intermediate representation (IR), are efficient and deterministic but offer limited insight into how a program will behave or evolve under complex code transformations. Conversely, dynamic representations, which rely on runtime profiling, provide profound insights into performance bottlenecks but are often impractical for large-scale tasks due to prohibitive overhead and inherent non-determinism. This paper transcends this trade-off by proposing a novel quasi-dynamic framework for program representation. The core insight is to model a program's optimization sensitivity. We introduce the Program Behavior Spectrum, a new representation generated by probing a program's IR with a diverse set of optimization sequences and quantifying the resulting changes in its static features. To effectively encode this high-dimensional, continuous spectrum, we pioneer a compositional learning approach. Product Quantization is employed to discretize the continuous reaction vectors into structured, compositional sub-words. Subsequently, a multi-task Transformer model, termed PQ-BERT, is pre-trained to learn the deep contextual grammar of these behavioral codes. Comprehensive experiments on two representative compiler optimization tasks -- Best Pass Prediction and -Oz Benefit Prediction -- demonstrate that our method outperforms state-of-the-art static baselines. Our code is publicly available at https://github.com/Panhaolin2001/PREP/.

Hongbin Zhang, Kehai Chen, Xuefen Bai et al.

Large language models (LLMs) have become a standard for multilingual evaluation, yet they exhibit a severe systematic translationese bias. In this paper, translationese bias is characterized as LLMs systematically favoring machine-translated text over human-authored references, particularly in low-resource languages. We attribute this bias to spurious correlations with (i) latent manifold alignment with English and (ii) cross-lingual predictability. To mitigate this bias, we propose DIBJudge, a robust fine-tuning framework that learns a minimally sufficient, judgment-critical representation via variational information compression, while explicitly isolating spurious factors into the dedicated bias branch. Furthermore, we incorporate a cross-covariance penalty that explicitly suppresses statistical dependence between robust and bias representations, thereby encouraging effective disentanglement. Extensive evaluations on multilingual reward modeling benchmarks and a dedicated translationese bias evaluation suite demonstrate that the proposed DIBJudge consistently outperforms strong baselines and substantially mitigates translationese bias.

Hongbin Zhang, Shihao Gao, Yang Liu et al.

In recent years, end-to-end Large Language Model (LLM) technology has shown substantial advantages across various domains. As critical system software and infrastructure, compilers are responsible for transforming source code into target code. While LLMs have been leveraged to assist in compiler development and maintenance, their potential as an end-to-end compiler remains largely unexplored. This paper explores the feasibility of LLM as a Compiler (LaaC) and its future directions. We designed the CompilerEval dataset and framework specifically to evaluate the capabilities of mainstream LLMs in source code comprehension and assembly code generation. In the evaluation, we analyzed various errors, explored multiple methods to improve LLM-generated code, and evaluated cross-platform compilation capabilities. Experimental results demonstrate that LLMs exhibit basic capabilities as compilers but currently achieve low compilation success rates. By optimizing prompts, scaling up the model, and incorporating reasoning methods, the quality of assembly code generated by LLMs can be significantly enhanced. Based on these findings, we maintain an optimistic outlook for LaaC and propose practical architectural designs and future research directions. We believe that with targeted training, knowledge-rich prompts, and specialized infrastructure, LaaC has the potential to generate high-quality assembly code and drive a paradigm shift in the field of compilation.

Hongbin Zhang, Taosheng Wei, Jiazhi Jiang et al.

Offline LLM inference seeks to maximize request processing under fixed budgets, making commodity GPU servers a promising choice. However, prior work typically considers offloading and parallelism in isolation, resulting in suboptimal performance. In this paper, we propose PipeMax, a high-throughput LLM inference system that integrates pipeline parallelism with offloading to overcome interconnect and memory constraints on GPU servers. Particularly, pipeline parallelism naturally incurs low communication overhead and keeps only one batch active on each GPU at a time, which enables offloading the KV cache of inactive batches. By coordinating computation with offloading data movement, PipeMax effectively expands GPU memory capacity and sustains large-batch execution. Experiments show that PipeMax achieves up to 2.51x higher throughput than vLLM, and up to 1.42x and 1.38x higher throughput than state-of-the-art high-throughput LLM systems, respectively, on an 8-GPU node.

Xin Li, Yeying Jin, Xin Jin et al.

This paper reviews the NTIRE 2025 Challenge on Day and Night Raindrop Removal for Dual-Focused Images. This challenge received a wide range of impressive solutions, which are developed and evaluated using our collected real-world Raindrop Clarity dataset. Unlike existing deraining datasets, our Raindrop Clarity dataset is more diverse and challenging in degradation types and contents, which includes day raindrop-focused, day background-focused, night raindrop-focused, and night background-focused degradations. This dataset is divided into three subsets for competition: 14,139 images for training, 240 images for validation, and 731 images for testing. The primary objective of this challenge is to establish a new and powerful benchmark for the task of removing raindrops under varying lighting and focus conditions. There are a total of 361 participants in the competition, and 32 teams submitting valid solutions and fact sheets for the final testing phase. These submissions achieved state-of-the-art (SOTA) performance on the Raindrop Clarity dataset. The project can be found at https://lixinustc.github.io/CVPR-NTIRE2025-RainDrop-Competition.github.io/.

Hongbin Zhang, Kehai Chen, Xuefeng Bai et al.

Large language models (LLMs) have succeeded remarkably in multilingual translation tasks. However, the inherent translation mechanisms of LLMs remain poorly understood, largely due to sophisticated architectures and vast parameter scales. In response to this issue, this study explores the translation mechanism of LLM from the perspective of computational components (e.g., attention heads and MLPs). Path patching is utilized to explore causal relationships between components, detecting those crucial for translation tasks and subsequently analyzing their behavioral patterns in human-interpretable terms. Comprehensive analysis reveals that translation is predominantly facilitated by a sparse subset of specialized attention heads (less than 5\%), which extract source language, indicator, and positional features. MLPs subsequently integrate and process these features by transiting towards English-centric latent representations. Notably, building on the above findings, targeted fine-tuning of only 64 heads achieves translation improvement comparable to full-parameter tuning while preserving general capabilities.

Hongbin Zhang, Kehai Chen, Xuefeng Bai et al.

Large language models (LLMs) have exhibited impressive multilingual reasoning capabilities, driven by extensive multilingual pre-training corpora and instruction fine-tuning data. However, a performance gap exists between high- and low-resource language reasoning tasks due to the language imbalance in the pre-training corpus, which is exacerbated by evaluation bias in existing reasoning benchmarks lacking low-resource language coverage. To alleviate this issue, we propose LinguaLIFT, a two-stage instruction tuning framework for advancing low-resource language reasoning. LinguaLIFT employs a language alignment layer to capture multilingual alignment in a code-switched tuning way without requiring multilingual instruction or parallel data, thereby transferring the cross-lingual reasoning capabilities to low-resource languages through English-only instruction tuning data. To comprehensively evaluate the multilingual reasoning capabilities, we introduce the Multilingual Math World Problem (MMWP) benchmark, which spans 21 low-resource, 17 medium-resource, and 10 high-resource languages. Experimental results show that LinguaLIFT outperforms several competitive baselines across MMWP and four widely used benchmarks.

Hongbin Zhang, Kehai Chen, Xuefeng Bai et al.

Reward models (RMs) are crucial for aligning large language models (LLMs) with diverse cultures. Consequently, evaluating their cultural awareness is essential for further advancing global alignment of LLMs. However, existing RM evaluations fall short in assessing cultural awareness due to the scarcity of culturally relevant evaluation datasets. To fill this gap, we propose Cultural Awareness Reward modeling Benchmark (CARB), covering 10 distinct cultures across 4 cultural domains. Our extensive evaluation of state-of-the-art RMs reveals their deficiencies in modeling cultural awareness and demonstrates a positive correlation between performance on CARB and downstream multilingual cultural alignment tasks. Further analysis identifies the spurious correlations within culture-aware reward modeling, wherein RM's scoring relies predominantly on surface-level features rather than authentic cultural nuance understanding. To address these, we propose Think-as-Locals to elicit deeper culturally grounded reasoning from generative RMs via reinforcement learning from verifiable rewards (RLVR) and employ well-designed rewards to ensure accurate preference judgments and high-quality structured evaluation criteria generation. Experimental results validate its efficacy in mitigating spurious features interference and advancing culture-aware reward modeling.

Hongbin Zhang, Kehai Chen, Xuefeng Bai et al.

Large language models (LLMs) have showcased impressive multilingual machine translation ability. However, unlike encoder-decoder style models, decoder-only LLMs lack an explicit alignment between source and target contexts. Analyzing contribution scores during generation processes revealed that LLMs can be biased towards previously generated tokens over corresponding source tokens, leading to unfaithful translations. To address this issue, we propose to encourage LLMs to pay more attention to the source context from both source and target perspectives in zeroshot prompting: 1) adjust source context attention weights; 2) suppress irrelevant target prefix influence; Additionally, we propose 3) avoiding over-reliance on the target prefix in instruction tuning. Experimental results from both human-collected unfaithfulness test sets focusing on LLM-generated unfaithful translations and general test sets, verify our methods' effectiveness across multiple language pairs. Further human evaluation shows our method's efficacy in reducing hallucinatory translations and facilitating faithful translation generation.

Han Bao, Jinyong Feng, Nam Dinh et al.

Multiphase flow phenomena have been widely observed in the industrial applications, yet it remains a challenging unsolved problem. Three-dimensional computational fluid dynamics (CFD) approaches resolve of the flow fields on finer spatial and temporal scales, which can complement dedicated experimental study. However, closures must be introduced to reflect the underlying physics in multiphase flow. Among them, the interfacial forces, including drag, lift, turbulent-dispersion and wall-lubrication forces, play an important role in bubble distribution and migration in liquid-vapor two-phase flows. Development of those closures traditionally rely on the experimental data and analytical derivation with simplified assumptions that usually cannot deliver a universal solution across a wide range of flow conditions. In this paper, a data-driven approach, named as feature-similarity measurement (FSM), is developed and applied to improve the simulation capability of two-phase flow with coarse-mesh CFD approach. Interfacial momentum transfer in adiabatic bubbly flow serves as the focus of the present study. Both a mature and a simplified set of interfacial closures are taken as the low-fidelity data. Validation data (including relevant experimental data and validated fine-mesh CFD simulations results) are adopted as high-fidelity data. Qualitative and quantitative analysis are performed in this paper. These reveal that FSM can substantially improve the prediction of the coarse-mesh CFD model, regardless of the choice of interfacial closures, and it provides scalability and consistency across discontinuous flow regimes. It demonstrates that data-driven methods can aid the multiphase flow modeling by exploring the connections between local physical features and simulation errors.

Han Bao, Nam Dinh, Linyu Lin et al.

Current system thermal-hydraulic codes have limited credibility in simulating real plant conditions, especially when the geometry and boundary conditions are extrapolated beyond the range of test facilities. This paper proposes a data-driven approach, Feature Similarity Measurement FFSM), to establish a technical basis to overcome these difficulties by exploring local patterns using machine learning. The underlying local patterns in multiscale data are represented by a set of physical features that embody the information from a physical system of interest, empirical correlations, and the effect of mesh size. After performing a limited number of high-fidelity numerical simulations and a sufficient amount of fast-running coarse-mesh simulations, an error database is built, and deep learning is applied to construct and explore the relationship between the local physical features and simulation errors. Case studies based on mixed convection have been designed for demonstrating the capability of data-driven models in bridging global scale gaps.

Han Bao, Jinyong Feng, Nam Dinh et al.

To realize efficient computational fluid dynamics (CFD) prediction of two-phase flow, a multi-scale framework was proposed in this paper by applying a physics-guided data-driven approach. Instrumental to this framework, Feature Similarity Measurement (FSM) technique was developed for error estimation in two-phase flow simulation using coarse-mesh CFD, to achieve a comparable accuracy as fine-mesh simulations with fast-running feature. By defining physics-guided parameters and variable gradients as physical features, FSM has the capability to capture the underlying local patterns in the coarse-mesh CFD simulation. Massive low-fidelity data and respective high-fidelity data are used to explore the underlying information relevant to the main simulation errors and the effects of phenomenological scaling. By learning from previous simulation data, a surrogate model using deep feedforward neural network (DFNN) can be developed and trained to estimate the simulation error of coarse-mesh CFD. The research documented supports the feasibility of the physics-guided deep learning methods for coarse mesh CFD simulations which has a potential for the efficient industrial design.