Zhiwei Zhang

Minhua Lin, Juncheng Wu, Zijun Wang et al.

LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it remains unclear whether a model's base capability in task-solving predicts its capabilities in harness self-evolution: which models produce useful harness updates, and which actually benefit from them? We analyze two harness self-evolution capabilities: (i) harness-updating, the capability to produce useful persistent harness updates from execution evidence; (ii) harness-benefit, the capability to benefit from updated harnesses during task solving. Our analysis reveals two findings. First, harness-updating is flat in base capability: models from different capability tiers produce harness updates that lead to surprisingly similar gains; even Qwen3.5-9B's updates yield gains comparable to those of Claude Opus~4.6. Second, harness-benefit is non-monotonic in base capability: weak-tier models benefit little from updated harnesses, mid-tier models benefit most, and strong-tier models benefit less than mid-tier. We trace low gains at the weak tier to two failure modes: weak-tier models may fail to activate relevant harness artifacts, or activate them but fail to follow them faithfully. These findings suggest investing capability budget in the task-solving agent rather than the evolver, and targeting harness invocation and long-horizon instruction following in agent training. Our source code is publicly available at https://github.com/A-EVO-Lab/a-evolve/tree/release/harness-evolution.

Zhiwei Zhang, Zhizhong Zhang, Qian Yu et al.

3D panoptic segmentation is a challenging perception task that requires both semantic segmentation and instance segmentation. In this task, we notice that images could provide rich texture, color, and discriminative information, which can complement LiDAR data for evident performance improvement, but their fusion remains a challenging problem. To this end, we propose LCPS, the first LiDAR-Camera Panoptic Segmentation network. In our approach, we conduct LiDAR-Camera fusion in three stages: 1) an Asynchronous Compensation Pixel Alignment (ACPA) module that calibrates the coordinate misalignment caused by asynchronous problems between sensors; 2) a Semantic-Aware Region Alignment (SARA) module that extends the one-to-one point-pixel mapping to one-to-many semantic relations; 3) a Point-to-Voxel feature Propagation (PVP) module that integrates both geometric and semantic fusion information for the entire point cloud. Our fusion strategy improves about 6.9% PQ performance over the LiDAR-only baseline on NuScenes dataset. Extensive quantitative and qualitative experiments further demonstrate the effectiveness of our novel framework. The code will be released at https://github.com/zhangzw12319/lcps.git.

Moshe Y. Vardi, Zhiwei Zhang

Determining the satisfiability of Boolean constraint-satisfaction problems with different types of constraints, that is hybrid constraints, is a well-studied problem with important applications. We study here a new application of hybrid Boolean constraints, which arises in quantum computing. The problem relates to constrained perfect matching in edge-colored graphs. While general-purpose hybrid constraint solvers can be powerful, we show that direct encodings of the constrained-matching problem as hybrid constraints scale poorly and special techniques are still needed. We propose a novel encoding based on Tutte's Theorem in graph theory as well as optimization techniques. Empirical results demonstrate that our encoding, in suitable languages with advanced SAT solvers, scales significantly better than a number of competing approaches on constrained-matching benchmarks. Our study identifies the necessity of designing problem-specific encodings when applying powerful general-purpose constraint solvers.

Anastasios Kyrillidis, Moshe Y. Vardi, Zhiwei Zhang

Boolean MaxSAT, as well as generalized formulations such as Min-MaxSAT and Max-hybrid-SAT, are fundamental optimization problems in Boolean reasoning. Existing methods for MaxSAT have been successful in solving benchmarks in CNF format. They lack, however, the ability to handle 1) (non-CNF) hybrid constraints, such as XORs and 2) generalized MaxSAT problems natively. To address this issue, we propose a novel dynamic-programming approach for solving generalized MaxSAT problems with hybrid constraints -- called \emph{Dynamic-Programming-MaxSAT} or DPMS for short -- based on Algebraic Decision Diagrams (ADDs). With the power of ADDs and the (graded) project-join-tree builder, our versatile framework admits many generalizations of CNF-MaxSAT, such as MaxSAT, Min-MaxSAT, and MinSAT with hybrid constraints. Moreover, DPMS scales provably well on instances with low width. Empirical results indicate that DPMS is able to solve certain problems quickly, where other algorithms based on various techniques all fail. Hence, DPMS is a promising framework and opens a new line of research that invites more investigation in the future.

Minhua Lin, Zhiwei Zhang, Hanqing Lu et al.

Memory-augmented LLM agents maintain external memory banks to support long-horizon interaction, yet most existing systems treat construction, retrieval, and utilization as isolated subroutines. This creates two coupled challenges: strategic blindness on the forward path of the memory cycle, where construction and retrieval are driven by local heuristics rather than explicit strategic reasoning, and sparse, delayed supervision on the backward path, where downstream failures rarely translate into direct repairs of the memory bank. To address these challenges, we propose MemMA, a plug-and-play multi-agent framework that coordinates the memory cycle along both the forward and backward paths. On the forward path, a Meta-Thinker produces structured guidance that steers a Memory Manager during construction and directs a Query Reasoner during iterative retrieval. On the backward path, MemMA introduces in-situ self-evolving memory construction, which synthesizes probe QA pairs, verifies the current memory, and converts failures into repair actions before the memory is finalized. Extensive experiments on LoCoMo show that MemMA consistently outperforms existing baselines across multiple LLM backbones and improves three different storage backends in a plug-and-play manner. Our code is publicly available at https://github.com/ventr1c/memma.

Yi Xin, Siqi Luo, Qi Qin et al.

Diffusion Multi-modal Large Language Models (dMLLMs) have recently emerged as a novel architecture unifying image generation and understanding. However, developing effective and efficient Test-Time Scaling (TTS) methods to unlock their full generative potential remains an underexplored challenge. To address this, we propose dMLLM-TTS, a novel framework operating on two complementary scaling axes: (1) trajectory exploration scaling to enhance the diversity of generated hypotheses, and (2) iterative refinement scaling for stable generation. Conventional TTS approaches typically perform linear search across these two dimensions, incurring substantial computational costs of O(NT) and requiring an external verifier for best-of-N selection. To overcome these limitations, we propose two innovations. First, we design an efficient hierarchical search algorithm with O(N+T) complexity that adaptively expands and prunes sampling trajectories. Second, we introduce a self-verified feedback mechanism that leverages the dMLLMs' intrinsic image understanding capabilities to assess text-image alignment, eliminating the need for external verifier. Extensive experiments on the GenEval benchmark across three representative dMLLMs (e.g., Lumina-DiMOO, MMaDA, Muddit) show that our framework substantially improves generation quality while achieving up to 6x greater efficiency than linear search. Project page: https://github.com/Alpha-VLLM/Lumina-DiMOO.

Yunuo Cen, Zhiwei Zhang, Xuanyao Fong

Although state-of-the-art (SOTA) SAT solvers based on conflict-driven clause learning (CDCL) have achieved remarkable engineering success, their sequential nature limits the parallelism that may be extracted for acceleration on platforms such as the graphics processing unit (GPU). In this work, we propose FastFourierSAT, a highly parallel hybrid SAT solver based on gradient-driven continuous local search (CLS). This is realized by a novel parallel algorithm inspired by the Fast Fourier Transform (FFT)-based convolution for computing the elementary symmetric polynomials (ESPs), which is the major computational task in previous CLS methods. The complexity of our algorithm matches the best previous result. Furthermore, the substantial parallelism inherent in our algorithm can leverage the GPU for acceleration, demonstrating significant improvement over the previous CLS approaches. We also propose to incorporate the restart heuristics in CLS to improve search efficiency. We compare our approach with the SOTA parallel SAT solvers on several benchmarks. Our results show that FastFourierSAT computes the gradient 100+ times faster than previous prototypes implemented on CPU. Moreover, FastFourierSAT solves most instances and demonstrates promising performance on larger-size instances.

Huaisheng Zhu, Guoji Fu, Zhimeng Guo et al.

Graph Neural Networks (GNNs) have shown great power in various domains. However, their predictions may inherit societal biases on sensitive attributes, limiting their adoption in real-world applications. Although many efforts have been taken for fair GNNs, most existing works just adopt widely used fairness techniques in machine learning to graph domains and ignore or don't have a thorough understanding of the message passing mechanism with fairness constraints, which is a distinctive feature of GNNs. To fill the gap, we propose a novel fairness-aware message passing framework GMMD, which is derived from an optimization problem that considers both graph smoothness and representation fairness. GMMD can be intuitively interpreted as encouraging a node to aggregate representations of other nodes from different sensitive groups while subtracting representations of other nodes from the same sensitive group, resulting in fair representations. We also provide a theoretical analysis to justify that GMMD can guarantee fairness, which leads to a simpler and theory-guided variant GMMD-S. Extensive experiments on graph benchmarks show that our proposed framework can significantly improve the fairness of various backbone GNN models while maintaining high accuracy.

Zhiwei Zhang, Xinyi Du, Weihao Wang et al.

Traffic forecasting is a cornerstone of intelligent transportation systems. While existing research has made significant progress in short-term prediction, long-term forecasting remains a largely uncharted and challenging frontier. Extending the prediction horizon intensifies two critical issues: escalating computational resource consumption and increasingly complex spatial-temporal dependencies. Current approaches, which rely on spatial-temporal graphs and process temporal and spatial dimensions separately, suffer from snapshot-stacking inflation and cross-step fragmentation. To overcome these limitations, we propose \textit{VisiFold}. Our framework introduces a novel temporal folding graph that consolidates a sequence of temporal snapshots into a single graph. Furthermore, we present a node visibility mechanism that incorporates node-level masking and subgraph sampling to overcome the computational bottleneck imposed by large node counts. Extensive experiments show that VisiFold not only drastically reduces resource consumption but also outperforms existing baselines in long-term forecasting tasks. Remarkably, even with a high mask ratio of 80\%, VisiFold maintains its performance advantage. By effectively breaking the resource constraints in both temporal and spatial dimensions, our work paves the way for more realistic long-term traffic forecasting. The code is available at~ https://github.com/PlanckChang/VisiFold.

Zhiwei Zhang, Xingyuan Zeng, Xinkai Kong et al.

Agricultural parcel extraction plays an important role in remote sensing-based agricultural monitoring, supporting parcel surveying, precision management, and ecological assessment. However, existing public benchmarks mainly focus on regular and relatively flat farmland scenes. In contrast, terraced parcels in mountainous regions exhibit stepped terrain, pronounced elevation variation, irregular boundaries, and strong cross-regional heterogeneity, making parcel extraction a more challenging problem that jointly requires visual recognition, semantic discrimination, and terrain-aware geometric understanding. Although recent studies have advanced visual parcel benchmarks and image-text farmland understanding, a unified benchmark for complex terraced parcel extraction under aligned image-text-DEM settings remains absent. To fill this gap, we present GTPBD-MM, the first multimodal benchmark for global terraced parcel extraction. Built upon GTPBD, GTPBD-MM integrates high-resolution optical imagery, structured text descriptions, and DEM data, and supports systematic evaluation under Image-only, Image+Text, and Image+Text+DEM settings. We further propose Elevation and Text guided Terraced parcel network (ETTerra), a multimodal baseline for terraced parcel delineation. Extensive experiments demonstrate that textual semantics and terrain geometry provide complementary cues beyond visual appearance alone, yielding more accurate, coherent, and structurally consistent delineation results in complex terraced scenes.

Zhiwei Zhang, Yuliang Liu

The recent success of ChatGPT and GPT-4 has drawn widespread attention to multimodal dialogue systems. However, there is a lack of datasets in the academic community that can effectively evaluate the multimodal generation capabilities of Visual Language Models (VLMs) in textual-visual chat tasks. In this paper, we address this gap by introducing two novel multimodal datasets: the synthetic CLEVR-ATVC dataset (620K) and the manually pictured Fruit-ATVC dataset (50K). These datasets incorporate both visual and text-based inputs and outputs. Furthermore, to facilitate the accountability of multimodal systems in rejecting human requests, similar to language-based ChatGPT conversations, we introduce specific rules as supervisory signals within the datasets. This allows the trained VLM to provide a yes or no answer after engaging in visual and textual reasoning, accompanied by a language explanation to clarify the reasons behind the inability to execute the given human instruction. Our proposed method involves a two-stage training procedure, which includes training the image auto-encoder and the auto-regressive transformer from scratch. The first stage employs a discrete variational autoencoder (dVAE) to compress each image into concise tokens, which are then combined with text tokens into a single data stream. This stream is subsequently fed into the decoder-based transformer to generate visual re-creations and textual feedback in the second stage. We conduct comprehensive analyses of experimental results, focusing on re-created image quality, answer accuracy, and the model's behavior when faced with uncertainty and imperfect user queries. Through our explorations and findings, we aim to contribute valuable insights into the accountability of textual-visual generative models.

Xiaomin Li, Jianheng Hou, Zheyuan Deng et al.

Large reasoning models (LRMs) increasingly expose chain-of-thought-like reasoning for transparency, verification, and deliberate problem solving. This creates a safety blind spot: harmful or policy-violating content may appear in reasoning traces even when final answers appear safe. We test whether final-answer safety is a sufficient proxy for the full reasoning-answer trajectory by scoring both stages under a unified twenty-principle safety rubric. Using prompts from seven public harmfulness and jailbreak sources, plus four out-of-distribution (OOD) sources, we evaluate 15 open-weight and API-based LRMs across 41K prompts per model. Reasoning traces consistently reveal additional safety risks beyond final answers, especially in high-severity stage-wise failures: leak cases, where unsafe reasoning precedes a safe-looking answer, and escape cases, where benign-looking reasoning precedes an unsafe final response. Principle-level analysis shows that risk concentrates in misinformation, legal compliance, discrimination, physical harm, and psychological harm. We further propose adaptive multi-principle steering, a white-box test-time mitigation that learns one unsafe-to-safe activation direction per safety principle and activates only directions whose current hidden state is closer to the unsafe than safe centroid. On three steerable open reasoning models, adaptive steering reduces unsafe counts in both reasoning traces and final answers on held-out and OOD benchmarks. DeepSeek-R1-Qwen-7B achieves a 40.8% average unsafe-count reduction while retaining 97.7% macro-averaged accuracy on BBH, GSM8K, and MMLU. These results suggest that LRM safety should be evaluated and mitigated over the full exposed reasoning-answer trajectory, not only at the final-answer stage.

Zhiwei Zhang, Fali Wang, Xiaomin Li et al.

Large language models (LLMs) have shown remarkable proficiency in generating text, benefiting from extensive training on vast textual corpora. However, LLMs may also acquire unwanted behaviors from the diverse and sensitive nature of their training data, which can include copyrighted and private content. Machine unlearning has been introduced as a viable solution to remove the influence of such problematic content without the need for costly and time-consuming retraining. This process aims to erase specific knowledge from LLMs while preserving as much model utility as possible. Despite the effectiveness of current unlearning methods, little attention has been given to whether existing unlearning methods for LLMs truly achieve forgetting or merely hide the knowledge, which current unlearning benchmarks fail to detect. This paper reveals that applying quantization to models that have undergone unlearning can restore the "forgotten" information. To thoroughly evaluate this phenomenon, we conduct comprehensive experiments using various quantization techniques across multiple precision levels. We find that for unlearning methods with utility constraints, the unlearned model retains an average of 21\% of the intended forgotten knowledge in full precision, which significantly increases to 83\% after 4-bit quantization. ... Our code is available at: \href{https://github.com/zzwjames/FailureLLMUnlearning}{https://github.com/zzwjames/FailureLLMUnlearning}.

Songtao Liu, Hongwu Peng, Zhiwei Zhang et al.

Long-context inference in large language models is bottlenecked by Key--Value (KV) cache loading during the decoding stage, where the sequential nature of generation requires repeatedly transferring the KV cache from off-chip High-Bandwidth Memory (HBM) to on-chip Static Random-Access Memory (SRAM) at each step. While Multi-Head Latent Attention (MLA) significantly reduces the total KV cache size, it suffers from a sharding bottleneck during distributed decoding via Tensor Parallelism (TP). Since its single latent head cannot be partitioned, each device is forced to redundantly load the complete KV cache for every token, consuming excessive memory traffic and diminishing TP benefits like weight sharding. In this work, we propose Multi-Head Low-Rank Attention (MLRA), which enables partitionable latent states for efficient 4-way TP decoding. Extensive experiments show that MLRA achieves state-of-the-art perplexity and downstream task performance, while also delivering a 2.8$\times$ decoding speedup over MLA. Code is available at https://github.com/SongtaoLiu0823/MLRA. Pretrained weights, along with the training and evaluation data, are available at https://huggingface.co/Soughing/MLRA.

Biao Wu, Yanda Li, Zhiwei Zhang et al.

Mobile agents are essential for automating tasks in complex and dynamic mobile environments. As foundation models evolve, the demands for agents that can adapt in real-time and process multimodal data have grown. This survey provides a comprehensive review of mobile agent technologies, focusing on recent advancements that enhance real-time adaptability and multimodal interaction. Recent evaluation benchmarks have been developed better to capture the static and interactive environments of mobile tasks, offering more accurate assessments of agents' performance. We then categorize these advancements into two main approaches: prompt-based methods, which utilize large language models (LLMs) for instruction-based task execution, and training-based methods, which fine-tune multimodal models for mobile-specific applications. Additionally, we explore complementary technologies that augment agent performance. By discussing key challenges and outlining future research directions, this survey offers valuable insights for advancing mobile agent technologies. A comprehensive resource list is available at https://github.com/aialt/awesome-mobile-agents

Zhiwei Zhang, Xinyi Du, Xuanchi Guo et al.

Multivariate time series forecasting is crucial across a wide range of domains. While presenting notable progress for the Transformer architecture, iTransformer still lags behind the latest MLP-based models. We attribute this performance gap to unstable inter-channel relationships. To bridge this gap, we propose EMAformer, a simple yet effective model that enhances the Transformer with an auxiliary embedding suite, akin to armor that reinforces its ability. By introducing three key inductive biases, i.e., \textit{global stability}, \textit{phase sensitivity}, and \textit{cross-axis specificity}, EMAformer unlocks the further potential of the Transformer architecture, achieving state-of-the-art performance on 12 real-world benchmarks and reducing forecasting errors by an average of 2.73\% in MSE and 5.15\% in MAE. This significantly advances the practical applicability of Transformer-based approaches for multivariate time series forecasting. The code is available on https://github.com/PlanckChang/EMAformer.

Minhua Lin, Hanqing Lu, Zhan Shi et al.

As Large Language Models (LLMs) move from curated training sets into open-ended real-world environments, a fundamental limitation emerges: static training cannot keep pace with continual deployment environment change. Scaling training-time and inference-time compute improves static capability but does not close this train-deploy gap. We argue that addressing this limitation requires a new scaling axis-evolution. Existing deployment-time adaptation methods, whether parametric fine-tuning or heuristic memory accumulation, lack the strategic agency needed to diagnose failures and produce durable improvements. Our position is that agentic evolution represents the inevitable future of LLM adaptation, elevating evolution itself from a fixed pipeline to an autonomous evolver agent. We instantiate this vision in a general framework, A-Evolve, which treats deployment-time improvement as a deliberate, goal-directed optimization process over persistent system state. We further propose the evolution-scaling hypothesis: the capacity for adaptation scales with the compute allocated to evolution, positioning agentic evolution as a scalable path toward sustained, open-ended adaptation in the real world.

Weiying Wang, Zhiwei Zhang

Modern RISC vector processors rely on the synergy of multi-lane parallelism and chaining to achieve high sustained throughput, yet their achieved performance often falls substantially short of the theoretical performance bound due to microarchitectural inefficiencies. In this work, we take the open-source RVV processor Ara as the target platform and analyze the sources of its sustained-throughput loss and optimize the design accordingly. We first establish an ideal multi-lane chaining execution model as a microarchitectural reference for the ideal steady-state progression of the vector backend. Based on this model, we attribute Ara's key bottlenecks to inefficiencies along three critical execution paths: memory-side inefficiencies in data supply and transaction issuance, control-side inefficiencies caused by conservative dependence management and issue control, and operand-delivery inefficiencies caused by access conflicts and result-propagation overhead. To address these bottlenecks, we propose a coordinated set of microarchitectural optimizations. Experimental results show that, without increasing raw memory bandwidth or changing the main processor configuration, Ara-Opt achieves a geometric-mean speedup of 1.33x over baseline Ara. Under roofline-based normalization, the geometric-mean gap-closed ratio reaches 12.2%. In particular, scal, axpy, ger, and gemm achieve speedups of approximately 2.41x, 1.60x, 1.52x, and 1.42x, with corresponding gap-closed ratios of 93.7%, 88.9%, 78.3%, and 59.3%, respectively. These results show that the proposed method can effectively recover sustained-throughput capability lost to microarchitectural inefficiencies in Ara under essentially unchanged hardware resource constraints, and move the implementation points of regular streaming and high-throughput workloads significantly closer to the theoretical performance bound.

Fali Wang, Jihai Chen, Shuhua Yang et al.

Test-Time Scaling (TTS) improves large language models (LLMs) by allocating additional computation during inference, typically through parallel, sequential, or hybrid scaling. However, prior studies often assume fixed collaboration architectures (e.g., topologies) and single-model usage, overlooking that optimal architectures and model combinations can vary across tasks. Therefore, we study the novel problem of searching for compute-optimal model combinations and architectures in TTS under a fixed budget. We formalize it as a multi-LLM collaboration graph, where nodes encode roles and LLM model assignments, and edges capture information flow. This problem is challenging because (i) the combinatorial search space is prohibitively large, and (ii) task-specific requirements demand tailored designs. To address these, we reformulate the problem as probabilistic graph optimization and, through pilot experiments, derive three empirical insights into TTS collaboration graphs. Guided by these insights, we propose Agent-REINFORCE, an LLM-agent-augmented framework that mirrors the REINFORCE pipeline by mapping sampling-gradient-update to sampling-feedback-update, where feedback serves as a textual gradient to update the probabilistic graph and efficiently search for optimal multi-LLM collaboration graphs. Experiments show that Agent-REINFORCE outperforms both traditional and LLM-based baselines in sample efficiency and search performance, and effectively identifies optimal graphs under joint objectives of accuracy and inference latency.

Xuyin Qi, Zeyu Zhang, Aaron Berliano Handoko et al.

Prostate cancer, a growing global health concern, necessitates precise diagnostic tools, with Magnetic Resonance Imaging (MRI) offering high-resolution soft tissue imaging that significantly enhances diagnostic accuracy. Recent advancements in explainable AI and representation learning have significantly improved prostate cancer diagnosis by enabling automated and precise lesion classification. However, existing explainable AI methods, particularly those based on frameworks like generative adversarial networks (GANs), are predominantly developed for natural image generation, and their application to medical imaging often leads to suboptimal performance due to the unique characteristics and complexity of medical image. To address these challenges, our paper introduces three key contributions. First, we propose ProjectedEx, a generative framework that provides interpretable, multi-attribute explanations, effectively linking medical image features to classifier decisions. Second, we enhance the encoder module by incorporating feature pyramids, which enables multiscale feedback to refine the latent space and improves the quality of generated explanations. Additionally, we conduct comprehensive experiments on both the generator and classifier, demonstrating the clinical relevance and effectiveness of ProjectedEx in enhancing interpretability and supporting the adoption of AI in medical settings. Code will be released at https://github.com/Richardqiyi/ProjectedEx

Zhiwei Zhang, Fei Zhao, Rui Wang et al.

Large language models (LLMs) can call tools effectively, yet they remain brittle in multi-turn execution: following a tool call error, smaller models often degenerate into repetitive invalid re-invocations, failing to interpret error feedback and self-correct. This brittleness hinders reliable real-world deployment, where the execution errors are inherently inevitable during tool interaction procedures. We identify a key limitation of current approaches: standard reinforcement learning (RL) treats errors as sparse negative rewards, providing no guidance on how to recover, while pre-collected synthetic error-correction datasets suffer from distribution mismatch with the model's on-policy error modes. To bridge this gap, we propose Fission-GRPO, a framework that converts execution errors into corrective supervision within the RL training loop. Our core mechanism fissions each failed trajectory into a new training instance by augmenting it with diagnostic feedback from a finetuned Error Simulator, then resampling recovery rollouts on-policy. This enables the model to learn from the precise errors it makes during exploration, rather than from static, pre-collected error cases. On the BFCL v4 Multi-Turn, Fission-GRPO improves the error recovery rate of Qwen3-8B by 5.7% absolute, crucially, yielding a 4% overall accuracy gain (42.75% to 46.75%) over GRPO and outperforming specialized tool-use agents.

Yi Xin, Jianjiang Yang, Siqi Luo et al.

Pre-trained vision models (PVMs) have demonstrated remarkable adaptability across a wide range of downstream vision tasks, showcasing exceptional performance. However, as these models scale to billions or even trillions of parameters, conventional full fine-tuning has become increasingly impractical due to its high computational and storage demands. To address these challenges, parameter-efficient fine-tuning (PEFT) has emerged as a promising alternative, aiming to achieve performance comparable to full fine-tuning while making minimal adjustments to the model parameters. This paper presents a comprehensive survey of the latest advancements in the visual PEFT field, systematically reviewing current methodologies and categorizing them into four primary categories: addition-based, partial-based, unified-based, and multi-task tuning. In addition, this paper offers an in-depth analysis of widely used visual datasets and real-world applications where PEFT methods have been successfully applied. Furthermore, this paper introduces the V-PEFT Bench, a unified benchmark designed to standardize the evaluation of PEFT methods across a diverse set of vision tasks, ensuring consistency and fairness in comparison. Finally, the paper outlines potential directions for future research to propel advances in the PEFT field. A comprehensive collection of resources is available at https://github.com/synbol/Awesome-Parameter-Efficient-Transfer-Learning.

Xuyin Qi, Zeyu Zhang, Canxuan Gang et al.

Data augmentation is essential in medical imaging for improving classification accuracy, lesion detection, and organ segmentation under limited data conditions. However, two significant challenges remain. First, a pronounced domain gap between natural photographs and medical images can distort critical disease features. Second, augmentation studies in medical imaging are fragmented and limited to single tasks or architectures, leaving the benefits of advanced mix-based strategies unclear. To address these challenges, we propose a unified evaluation framework with six mix-based augmentation methods integrated with both convolutional and transformer backbones on brain tumour MRI and eye disease fundus datasets. Our contributions are threefold. (1) We introduce MediAug, a comprehensive and reproducible benchmark for advanced data augmentation in medical imaging. (2) We systematically evaluate MixUp, YOCO, CropMix, CutMix, AugMix, and SnapMix with ResNet-50 and ViT-B backbones. (3) We demonstrate through extensive experiments that MixUp yields the greatest improvement on the brain tumor classification task for ResNet-50 with 79.19% accuracy and SnapMix yields the greatest improvement for ViT-B with 99.44% accuracy, and that YOCO yields the greatest improvement on the eye disease classification task for ResNet-50 with 91.60% accuracy and CutMix yields the greatest improvement for ViT-B with 97.94% accuracy. Code will be available at https://github.com/AIGeeksGroup/MediAug.

Tianxiong Zhong, Xingye Tian, Boyuan Jiang et al.

Modern video generation frameworks based on Latent Diffusion Models suffer from inefficiencies in tokenization due to the Frame-Proportional Information Assumption. Existing tokenizers provide fixed temporal compression rates, causing the computational cost of the diffusion model to scale linearly with the frame rate. The paper proposes the Duration-Proportional Information Assumption: the upper bound on the information capacity of a video is proportional to the duration rather than the number of frames. Based on this insight, the paper introduces VFRTok, a Transformer-based video tokenizer, that enables variable frame rate encoding and decoding through asymmetric frame rate training between the encoder and decoder. Furthermore, the paper proposes Partial Rotary Position Embeddings (RoPE) to decouple position and content modeling, which groups correlated patches into unified tokens. The Partial RoPE effectively improves content-awareness, enhancing the video generation capability. Benefiting from the compact and continuous spatio-temporal representation, VFRTok achieves competitive reconstruction quality and state-of-the-art generation fidelity while using only 1/8 tokens compared to existing tokenizers. The code and weights are released at: https://github.com/KwaiVGI/VFRTok.

Zhiwei Zhang, Shaojun E, Fandong Meng et al.

Traffic forecasting plays a key role in Intelligent Transportation Systems, and significant strides have been made in this field. However, most existing methods can only predict up to four hours in the future, which doesn't quite meet real-world demands. we identify that the prediction horizon is limited to a few hours mainly due to the separation of temporal and spatial factors, which results in high complexity. Drawing inspiration from Albert Einstein's relativity theory, which suggests space and time are unified and inseparable, we introduce Extralonger, which unifies temporal and spatial factors. Extralonger notably extends the prediction horizon to a week on real-world benchmarks, demonstrating superior efficiency in the training time, inference time, and memory usage. It sets new standards in long-term and extra-long-term scenarios. The code is available at https://github.com/PlanckChang/Extralonger.

Zhiqiang Sheng, Xumeng Han, Zhiwei Zhang et al.

Multimodal generative models have made significant strides in image editing, demonstrating impressive performance on a variety of static tasks. However, their proficiency typically does not extend to complex scenarios requiring dynamic reasoning, leaving them ill-equipped to model the coherent, intermediate logical pathways that constitute a multi-step evolution from an initial state to a final one. This capacity is crucial for unlocking a deeper level of procedural and causal understanding in visual manipulation. To systematically measure this critical limitation, we introduce InEdit-Bench, the first evaluation benchmark dedicated to reasoning over intermediate pathways in image editing. InEdit-Bench comprises meticulously annotated test cases covering four fundamental task categories: state transition, dynamic process, temporal sequence, and scientific simulation. Additionally, to enable fine-grained evaluation, we propose a set of assessment criteria to evaluate the logical coherence and visual naturalness of the generated pathways, as well as the model's fidelity to specified path constraints. Our comprehensive evaluation of 14 representative image editing models on InEdit-Bench reveals significant and widespread shortcomings in this domain. By providing a standardized and challenging benchmark, we aim for InEdit-Bench to catalyze research and steer development towards more dynamic, reason-aware, and intelligent multimodal generative models.

Zhiwei Zhang, Xiaomin Li, Yudi Lin et al.

Large Language Models (LLMs) trained with reinforcement learning and verifiable rewards have achieved strong results on complex reasoning tasks. Recent work extends this paradigm to a multi-agent setting, where a meta-thinking agent proposes plans and monitors progress while a reasoning agent executes subtasks through sequential conversational turns. Despite promising performance, we identify a critical limitation: lazy agent behavior, in which one agent dominates while the other contributes little, undermining collaboration and collapsing the setup to an ineffective single agent. In this paper, we first provide a theoretical analysis showing why lazy behavior naturally arises in multi-agent reasoning. We then introduce a stable and efficient method for measuring causal influence, helping mitigate this issue. Finally, as collaboration intensifies, the reasoning agent risks getting lost in multi-turn interactions and trapped by previous noisy responses. To counter this, we propose a verifiable reward mechanism that encourages deliberation by allowing the reasoning agent to discard noisy outputs, consolidate instructions, and restart its reasoning process when necessary. Extensive experiments demonstrate that our framework alleviates lazy agent behavior and unlocks the full potential of multi-agent framework for complex reasoning tasks.

Zhuoyang Qian, Wei Shi, Xu Lin et al.

Generating scientific manuscripts requires maintaining alignment between narrative reasoning, experimental evidence, and visual artifacts across the document lifecycle. Existing language-model generation pipelines rely on unconstrained text synthesis with validation applied only after generation, often producing structural drift, missing figures or tables, and cross-section inconsistencies. We introduce Story2Proposal, a contract-governed multi-agent framework that converts a research story into a structured manuscript through coordinated agents operating under a persistent shared visual contract. The system organizes architect, writer, refiner, and renderer agents around a contract state that tracks section structure and registered visual elements, while evaluation agents supply feedback in a generate evaluate adapt loop that updates the contract during generation. Experiments on tasks derived from the Jericho research corpus show that Story2Proposal achieved an expert evaluation score of 6.145 versus 3.963 for DirectChat (+2.182) across GPT, Claude, Gemini, and Qwen backbones. Compared with the structured generation baseline Fars, Story2Proposal obtained an average score of 5.705 versus 5.197, indicating improved structural consistency and visual alignment.

Fang Wu, Xu Huang, Weihao Xuan et al.

Reinforcement learning from human feedback (RLHF) has emerged as the standard paradigm for aligning large language models (LLMs) with human preferences. However, reward-based methods built on the Bradley-Terry assumption struggle to capture the non-transitive and heterogeneous nature of real-world preferences. To address this, recent studies have reframed alignment as a two-player Nash game, giving rise to Nash learning from human feedback (NLHF). While this perspective has inspired algorithms such as INPO, ONPO, and EGPO with strong theoretical and empirical guarantees, they remain fundamentally restricted to two-player interactions, creating a single-opponent bias that fails to capture the full complexity of realistic preference structures. In this work, we introduce Multiplayer Nash Preference Optimization (MNPO), a novel framework that generalizes NLHF to the multiplayer regime. It formulates alignment as an $n$-player game, where each policy competes against a population of opponents while being regularized toward a reference model. Our framework establishes well-defined Nash equilibria in multiplayer settings and extends the concept of duality gap to quantify approximation quality. We demonstrate that MNPO inherits the equilibrium guarantees of two-player methods while enabling richer competitive dynamics and improved coverage of diverse preference structures. Through comprehensive empirical evaluation, we show that MNPO consistently outperforms existing NLHF baselines on instruction-following benchmarks, achieving superior alignment quality under heterogeneous annotator conditions and mixed-policy evaluation scenarios. Together, these results establish MNPO as a principled and scalable framework for aligning LLMs with complex, non-transitive human preferences. Code is available at https://github.com/smiles724/MNPO.

Lingling Zhang, Pengpeng Qiao, Zhiwei Zhang et al.

Temporal Graph Neural Networks (TGNs) achieve state-of-the-art performance on dynamic graph tasks, yet existing systems focus exclusively on accelerating training -- at inference time, every new edge triggers $O(|V|)$ embedding updates even though only a small fraction of nodes are affected. We present \textbf{StreamTGN}, the first streaming TGN inference system exploiting the inherent locality of temporal graph updates: in an $L$-layer TGN, a new edge affects only nodes within $L$ hops of the endpoints, typically less than 0.2\% on million-node graphs. StreamTGN maintains persistent GPU-resident node memory and uses dirty-flag propagation to identify the affected set $\mathcal{A}$, reducing per-batch complexity from $O(|V|)$ to $O(|\mathcal{A}|)$ with zero accuracy loss. Drift-aware adaptive rebuild scheduling and batched streaming with relaxed ordering further maximize throughput. Experiments on eight temporal graphs (2K--2.6M nodes) show 4.5$\times$--739$\times$ speedup for TGN and up to 4,207$\times$ for TGAT, with identical accuracy. StreamTGN is orthogonal to training optimizations: combining SWIFT with StreamTGN yields 24$\times$ end-to-end speedup across three architectures (TGN, TGAT, DySAT).

Qingmei Li, Yang Zhang, Zurong Mai et al.

Large Multimodal Models (LMMs) has demonstrated capabilities across various domains, but comprehensive benchmarks for agricultural remote sensing (RS) remain scarce. Existing benchmarks designed for agricultural RS scenarios exhibit notable limitations, primarily in terms of insufficient scene diversity in the dataset and oversimplified task design. To bridge this gap, we introduce AgroMind, a comprehensive agricultural remote sensing benchmark covering four task dimensions: spatial perception, object understanding, scene understanding, and scene reasoning, with a total of 13 task types, ranging from crop identification and health monitoring to environmental analysis. We curate a high-quality evaluation set by integrating eight public datasets and one private farmland plot dataset, containing 27,247 QA pairs and 19,615 images. The pipeline begins with multi-source data pre-processing, including collection, format standardization, and annotation refinement. We then generate a diverse set of agriculturally relevant questions through the systematic definition of tasks. Finally, we employ LMMs for inference, generating responses, and performing detailed examinations. We evaluated 20 open-source LMMs and 4 closed-source models on AgroMind. Experiments reveal significant performance gaps, particularly in spatial reasoning and fine-grained recognition, it is notable that human performance lags behind several leading LMMs. By establishing a standardized evaluation framework for agricultural RS, AgroMind reveals the limitations of LMMs in domain knowledge and highlights critical challenges for future work. Data and code can be accessed at https://rssysu.github.io/AgroMind/.

Jie Li, Yulong Xia, Tongxin Yang et al.

In this paper, we introduce a new dataset in the medical field of hypertensive intracerebral hemorrhage (HICH), called HICH-IT, which includes both electronic medical records (EMRs) and head CT images. This dataset is designed to enhance the accuracy of artificial intelligence in the diagnosis and treatment of HICH. This dataset, built upon the foundation of standard text and image data, incorporates specific annotations within the EMRs, extracting key content from the text information, and categorizes the annotation content of imaging data into four types: brain midline, hematoma, left and right cerebral ventricle. HICH-IT aims to be a foundational dataset for feature learning in image segmentation tasks and named entity recognition. To further understand the dataset, we have trained deep learning algorithms to observe the performance. The pretrained models have been released at both www.daip.club and github.com/Deep-AI-Application-DAIP. The dataset has been uploaded to https://github.com/CYBUS123456/HICH-IT-Datasets. Index Terms-HICH, Deep learning, Intraparenchymal hemorrhage, named entity recognition, novel dataset

Lun Quan, Zhiwei Zhang, Xingguang Zhong et al.

The quadrotor is popularly used in challenging environments due to its superior agility and flexibility. In these scenarios, trajectory planning plays a vital role in generating safe motions to avoid obstacles while ensuring flight smoothness. Although many works on quadrotor planning have been proposed, a research gap exists in incorporating self-adaptation into a planning framework to enable a drone to automatically fly slower in denser environments and increase its speed in a safer area. In this paper, we propose an environmental adaptive planner to adjust the flight aggressiveness effectively based on the obstacle distribution and quadrotor state. Firstly, we design an environmental adaptive safety aware method to assign the priority of the surrounding obstacles according to the environmental risk level and instantaneous motion tendency. Then, we apply it into a multi-layered model predictive contouring control (Multi-MPCC) framework to generate adaptive, safe, and dynamical feasible local trajectories. Extensive simulations and real-world experiments verify the efficiency and robustness of our planning framework. Benchmark comparison also shows superior performances of our method with another advanced environmental adaptive planning algorithm. Moreover, we release our planning framework as open-source ros-packages.

Fali Wang, Zhiwei Zhang, Xianren Zhang et al.

Large language models (LLMs) have demonstrated emergent abilities in text generation, question answering, and reasoning, facilitating various tasks and domains. Despite their proficiency in various tasks, LLMs like PaLM 540B and Llama-3.1 405B face limitations due to large parameter sizes and computational demands, often requiring cloud API use which raises privacy concerns, limits real-time applications on edge devices, and increases fine-tuning costs. Additionally, LLMs often underperform in specialized domains such as healthcare and law due to insufficient domain-specific knowledge, necessitating specialized models. Therefore, Small Language Models (SLMs) are increasingly favored for their low inference latency, cost-effectiveness, efficient development, and easy customization and adaptability. These models are particularly well-suited for resource-limited environments and domain knowledge acquisition, addressing LLMs' challenges and proving ideal for applications that require localized data handling for privacy, minimal inference latency for efficiency, and domain knowledge acquisition through lightweight fine-tuning. The rising demand for SLMs has spurred extensive research and development. However, a comprehensive survey investigating issues related to the definition, acquisition, application, enhancement, and reliability of SLM remains lacking, prompting us to conduct a detailed survey on these topics. The definition of SLMs varies widely, thus to standardize, we propose defining SLMs by their capability to perform specialized tasks and suitability for resource-constrained settings, setting boundaries based on the minimal size for emergent abilities and the maximum size sustainable under resource constraints. For other aspects, we provide a taxonomy of relevant models/methods and develop general frameworks for each category to enhance and utilize SLMs effectively.

Zhiwei Zhang, Weizhong Zhang, Yaowei Huang et al.

Traffic forecasting is a challenging task due to the complex spatio-temporal correlations among traffic series. In this paper, we identify an underexplored problem in multivariate traffic series prediction: extreme events. Road congestion and rush hours can result in low correlation in vehicle speeds at various intersections during adjacent time periods. Existing methods generally predict future series based on recent observations and entirely discard training data during the testing phase, rendering them unreliable for forecasting highly nonlinear multivariate time series. To tackle this issue, we propose a test-time compensated representation learning framework comprising a spatio-temporal decomposed data bank and a multi-head spatial transformer model (CompFormer). The former component explicitly separates all training data along the temporal dimension according to periodicity characteristics, while the latter component establishes a connection between recent observations and historical series in the data bank through a spatial attention matrix. This enables the CompFormer to transfer robust features to overcome anomalous events while using fewer computational resources. Our modules can be flexibly integrated with existing forecasting methods through end-to-end training, and we demonstrate their effectiveness on the METR-LA and PEMS-BAY benchmarks. Extensive experimental results show that our method is particularly important in extreme events, and can achieve significant improvements over six strong baselines, with an overall improvement of up to 28.2%.

Xin Tan, Wenbin Wu, Zhiwei Zhang et al.

3D occupancy perception holds a pivotal role in recent vision-centric autonomous driving systems by converting surround-view images into integrated geometric and semantic representations within dense 3D grids. Nevertheless, current models still encounter two main challenges: modeling depth accurately in the 2D-3D view transformation stage, and overcoming the lack of generalizability issues due to sparse LiDAR supervision. To address these issues, this paper presents GEOcc, a Geometric-Enhanced Occupancy network tailored for vision-only surround-view perception. Our approach is three-fold: 1) Integration of explicit lift-based depth prediction and implicit projection-based transformers for depth modeling, enhancing the density and robustness of view transformation. 2) Utilization of mask-based encoder-decoder architecture for fine-grained semantic predictions; 3) Adoption of context-aware self-training loss functions in the pertaining stage to complement LiDAR supervision, involving the re-rendering of 2D depth maps from 3D occupancy features and leveraging image reconstruction loss to obtain denser depth supervision besides sparse LiDAR ground-truths. Our approach achieves State-Of-The-Art performance on the Occ3D-nuScenes dataset with the least image resolution needed and the most weightless image backbone compared with current models, marking an improvement of 3.3% due to our proposed contributions. Comprehensive experimentation also demonstrates the consistent superiority of our method over baselines and alternative approaches.

Zhiwei Zhang, Minhua Lin, Enyan Dai et al.

Graph Neural Networks (GNNs) have shown remarkable performance in various tasks. However, recent works reveal that GNNs are vulnerable to backdoor attacks. Generally, backdoor attack poisons the graph by attaching backdoor triggers and the target class label to a set of nodes in the training graph. A GNN trained on the poisoned graph will then be misled to predict test nodes attached with trigger to the target class. Despite their effectiveness, our empirical analysis shows that triggers generated by existing methods tend to be out-of-distribution (OOD), which significantly differ from the clean data. Hence, these injected triggers can be easily detected and pruned with widely used outlier detection methods in real-world applications. Therefore, in this paper, we study a novel problem of unnoticeable graph backdoor attacks with in-distribution (ID) triggers. To generate ID triggers, we introduce an OOD detector in conjunction with an adversarial learning strategy to generate the attributes of the triggers within distribution. To ensure a high attack success rate with ID triggers, we introduce novel modules designed to enhance trigger memorization by the victim model trained on poisoned graph. Extensive experiments on real-world datasets demonstrate the effectiveness of the proposed method in generating in distribution triggers that can by-pass various defense strategies while maintaining a high attack success rate.

Ke Wu, Kaizhao Zhang, Zhiwei Zhang et al.

Online dense mapping of urban scenes forms a fundamental cornerstone for scene understanding and navigation of autonomous vehicles. Recent advancements in mapping methods are mainly based on NeRF, whose rendering speed is too slow to meet online requirements. 3D Gaussian Splatting (3DGS), with its rendering speed hundreds of times faster than NeRF, holds greater potential in online dense mapping. However, integrating 3DGS into a street-view dense mapping framework still faces two challenges, including incomplete reconstruction due to the absence of geometric information beyond the LiDAR coverage area and extensive computation for reconstruction in large urban scenes. To this end, we propose HGS-Mapping, an online dense mapping framework in unbounded large-scale scenes. To attain complete construction, our framework introduces Hybrid Gaussian Representation, which models different parts of the entire scene using Gaussians with distinct properties. Furthermore, we employ a hybrid Gaussian initialization mechanism and an adaptive update method to achieve high-fidelity and rapid reconstruction. To the best of our knowledge, we are the first to integrate Gaussian representation into online dense mapping of urban scenes. Our approach achieves SOTA reconstruction accuracy while only employing 66% number of Gaussians, leading to 20% faster reconstruction speed.

Zhiwei Zhang

Class imbalance is a critical issue in image classification that significantly affects the performance of deep recognition models. In this work, we first identify a network degeneration dilemma that hinders the model learning by introducing a high linear dependence among the features inputted into the classifier. To overcome this challenge, we propose a novel framework called Whitening-Net to mitigate the degenerate solutions, in which ZCA whitening is integrated before the linear classifier to normalize and decorrelate the batch samples. However, in scenarios with extreme class imbalance, the batch covariance statistic exhibits significant fluctuations, impeding the convergence of the whitening operation. Therefore, we propose two covariance-corrected modules, the Group-based Relatively Balanced Batch Sampler (GRBS) and the Batch Embedded Training (BET), to get more accurate and stable batch covariance, thereby reinforcing the capability of whitening. Our modules can be trained end-to-end without incurring substantial computational costs. Comprehensive empirical evaluations conducted on benchmark datasets, including CIFAR-LT-10/100, ImageNet-LT, and iNaturalist-LT, validate the effectiveness of our proposed approaches.

Zongyu Wu, Yuwei Niu, Hongcheng Gao et al. · tsinghua

Large Vision-Language Models (LVLMs) have shown impressive performance in various tasks. However, LVLMs suffer from hallucination, which hinders their adoption in the real world. Existing studies emphasized that the strong language priors of LVLMs can overpower visual information, causing hallucinations. However, the positive role of language priors is the key to a powerful LVLM. If the language priors are too weak, LVLMs will struggle to leverage rich parameter knowledge and instruction understanding abilities to complete tasks in challenging visual scenarios where visual information alone is insufficient. Therefore, we propose a benchmark called LanP to rethink the impact of Language Priors in LVLMs. It is designed to investigate how strong language priors are in current LVLMs. LanP consists of 170 images and 340 corresponding well-designed questions. Extensive experiments on 25 popular LVLMs reveal that many LVLMs' language priors are not strong enough to effectively aid question answering when objects are partially hidden. Many models, including GPT-4 Turbo, exhibit an accuracy below 0.5 in such a scenario.

Pengpeng Qiao, Zhiwei Zhang, Xinzhou Wang et al.

Various real-world applications rely on in-memory dynamic graphs that must efficiently handle frequent updates while supporting low-latency analytics on evolving structures. Achieving both objectives remains challenging due to the trade-off between update efficiency and traversal locality, particularly under highly skewed degree distributions. This motivates the design of graph indexing schemes optimized for in-memory graph management on modern multi-core CPUs. We present LHGstore, a degree-aware Learned Hierarchical Graph storage that, for the first time, integrates learned indexing into graph management. LHGstore designs a two-level hierarchy that decouples vertex and edge access and further organizes each vertex's edges using data structures adaptive to its degree. Lightweight arrays are used for low-degree vertices to maximize traversal locality, while learned indexes are applied to high-degree vertices to improve update throughput. Extensive experiments show that LHGstore achieves 5.9-28.2$\times$ higher throughput and significantly faster analytics than SOTA in-memory graph storage systems.

Xiaomin Li, Mingye Gao, Zhiwei Zhang et al.

Reinforcement Learning from Human Feedback (RLHF) is commonly employed to tailor models to human preferences, especially to improve the safety of outputs from large language models (LLMs). Traditionally, this method depends on selecting preferred responses from pairs. However, due to the variability in human opinions and the challenges in directly comparing two responses, there is an increasing trend towards fine-grained annotation approaches that evaluate responses using multiple targeted metrics or rules. The challenge lies in efficiently choosing and applying these rules to handle the diverse range of preference data. In this paper, we propose a dynamic method that adaptively selects the most important rules for each response pair. We introduce a mathematical framework that utilizes the maximum discrepancy across paired responses and demonstrate theoretically that this approach maximizes the mutual information between the rule-based annotations and the underlying true preferences. We then train an 8B reward model using this adaptively labeled preference dataset and assess its efficacy using RewardBench. As of January 25, 2025, our model achieved the highest safety performance on the leaderboard, surpassing various larger models.

Fali Wang, Hui Liu, Zhenwei Dai et al.

Test-time scaling (TTS) enhances the performance of large language models (LLMs) by allocating additional compute resources during inference. However, existing research primarily investigates TTS in single-stage tasks; while many real-world problems are multi-stage complex tasks, composed of a sequence of heterogeneous subtasks with each subtask requires LLM of specific capability. Therefore, we study a novel problem: the test-time compute-optimal scaling in multi-stage complex tasks, aiming to select suitable models and allocate budgets per subtask to maximize overall performance. TTS in multi-stage tasks introduces two fundamental challenges: (i) The combinatorial search space of model and budget allocations, combined with the high cost of inference, makes brute-force search impractical. (ii) The optimal model and budget allocations across subtasks are interdependent, increasing the complexity of the compute-optimal search. To address this gap, we conduct extensive pilot experiments on four tasks across six datasets, deriving three empirical insights characterizing the behavior of LLMs in multi-stage complex tasks. Informed by these insights, we propose AgentTTS, an LLM-agent-based framework that autonomously searches for compute-optimal allocations through iterative feedback-driven interactions with the execution environment. Experimental results demonstrate that AgentTTS significantly outperforms traditional and other LLM-based baselines in search efficiency, and shows improved robustness to varying training set sizes and enhanced interpretability.

Ali Al-Lawati, Jason Lucas, Zhiwei Zhang et al.

In-context learning (ICL) effectively conditions large language models (LLMs) for molecular tasks, such as property prediction and molecule captioning, by embedding carefully selected demonstration examples into the input prompt. This approach avoids the computational overhead of extensive pertaining and fine-tuning. However, current prompt retrieval methods for molecular tasks have relied on molecule feature similarity, such as Morgan fingerprints, which do not adequately capture the global molecular and atom-binding relationships. As a result, these methods fail to represent the full complexity of molecular structures during inference. Moreover, small-to-medium-sized LLMs, which offer simpler deployment requirements in specialized systems, have remained largely unexplored in the molecular ICL literature. To address these gaps, we propose a self-supervised learning technique, GAMIC (Graph-Aligned Molecular In-Context learning, which aligns global molecular structures, represented by graph neural networks (GNNs), with textual captions (descriptions) while leveraging local feature similarity through Morgan fingerprints. In addition, we introduce a Maximum Marginal Relevance (MMR) based diversity heuristic during retrieval to optimize input prompt demonstration samples. Our experimental findings using diverse benchmark datasets show GAMIC outperforms simple Morgan-based ICL retrieval methods across all tasks by up to 45%.

Qian Ma, Hongliang Chi, Hengrui Zhang et al.

The rise of self-supervised learning, which operates without the need for labeled data, has garnered significant interest within the graph learning community. This enthusiasm has led to the development of numerous Graph Contrastive Learning (GCL) techniques, all aiming to create a versatile graph encoder that leverages the wealth of unlabeled data for various downstream tasks. However, the current evaluation standards for GCL approaches are flawed due to the need for extensive hyper-parameter tuning during pre-training and the reliance on a single downstream task for assessment. These flaws can skew the evaluation away from the intended goals, potentially leading to misleading conclusions. In our paper, we thoroughly examine these shortcomings and offer fresh perspectives on how GCL methods are affected by hyper-parameter choices and the choice of downstream tasks for their evaluation. Additionally, we introduce an enhanced evaluation framework designed to more accurately gauge the effectiveness, consistency, and overall capability of GCL methods.

Yatong Wang, Fali Wang, Naibin Gu et al.

Cloud-hosted large language models (LLMs) commonly rely on LoRA for domain adaptation, yet domain data are distributed across multiple edge devices and cannot be uploaded due to privacy constraints. This raises a fundamental question: how can knowledge from multiple private edges be integrated into a cloud LLM for cross-domain problem solving? A natural solution is to train LoRA adapters locally and fuse them in the cloud; however, existing pipelines rely on unrealistic assumptions that edge devices can host cloud-scale LLMs and are evaluated mainly on single-domain tasks. To address these limitations, we propose a prune-train-recover framework that enables local LoRA training on pruned models and privacy-preserving cloud integration. We further introduce MMLU-CD, a cross-domain benchmark that composes multiple domain samples into a single instance, enabling explicit evaluation of cross-domain problem solving. This allows us to ask a concrete question: Can existing LoRA fusion methods support cross-domain tasks in cloud-edge collaboration? Our empirical answer is negative. Existing LoRA fusion methods perform poorly on MMLU-CD, often underperforming the base LLM, revealing their inability to support cross-domain problem solving. We attribute this failure to parameter conflicts among LoRA adapters and propose a simple conflict-resolution module, LoRA-CR, which mitigates conflicting updates and improves LoRA fusion performance by up to 3.8%. These results identify conflict mitigation as a critical yet largely overlooked factor in cloud-edge LoRA fusion, warranting further investigation in future research.

Zhiwei Zhang, Hui Liu, Xiaomin Li et al.

Reward models trained on human preference data have demonstrated strong effectiveness in aligning Large Language Models (LLMs) with human intent under the framework of Reinforcement Learning from Human Feedback (RLHF). However, RLHF remains vulnerable to reward hacking, where the policy exploits imperfections in the reward function rather than genuinely learning the intended behavior. Although significant efforts have been made to mitigate reward hacking, they predominantly focus on and evaluate in-distribution scenarios, where the training and testing data for the reward model share the same distribution. In this paper, we empirically show that state-of-the-art methods struggle in more challenging out-of-distribution (OOD) settings. We further demonstrate that incorporating fine-grained multi-attribute scores helps address this challenge. However, the limited availability of high-quality data often leads to weak performance of multi-objective reward functions, which can negatively impact overall performance and become the bottleneck. To address this issue, we propose a unified reward modeling framework that jointly trains Bradley--Terry (BT) single-objective and multi-objective regression-based reward functions using a shared embedding space. We theoretically establish a connection between the BT loss and the regression objective and highlight their complementary benefits. Specifically, the regression task enhances the single-objective reward function's ability to mitigate reward hacking in challenging OOD settings, while BT-based training improves the scoring capability of the multi-objective reward function, enabling a 7B model to outperform a 70B baseline. Extensive experimental results demonstrate that our framework significantly improves both the robustness and the scoring performance of reward models.

Ming Yang, Zhiwei Zhang, Jiahang Li et al.

Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize the artifact (a visually plausible deck) while under-optimizing the delivery process (pacing, narrative, and presentation preparation). We present DeepSlide, a human-in-the-loop multi-agent system that supports preparing the full presentation process, from requirement elicitation and time-budgeted narrative planning, to evidence-grounded slide--script generation, attention augmentation, and rehearsal support. DeepSlide integrates (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to ensure renderability. We further introduce a dual-scoreboard benchmark that cleanly separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while consistently achieving larger gains on delivery metrics, improving narrative flow, pacing precision, and slide--script synergy with clearer attention guidance.

Juepeng Zheng, Zi Ye, Yibin Wen et al.

Powered by advances in multiple remote sensing sensors, the production of high spatial resolution images provides great potential to achieve cost-efficient and high-accuracy agricultural inventory and analysis in an automated way. Lots of studies that aim at providing an inventory of the level of each agricultural parcel have generated many methods for Agricultural Parcel and Boundary Delineation (APBD). This review covers APBD methods for detecting and delineating agricultural parcels and systematically reviews the past and present of APBD-related research applied to remote sensing images. With the goal to provide a clear knowledge map of existing APBD efforts, we conduct a comprehensive review of recent APBD papers to build a meta-data analysis, including the algorithm, the study site, the crop type, the sensor type, the evaluation method, etc. We categorize the methods into three classes: (1) traditional image processing methods (including pixel-based, edge-based and region-based); (2) traditional machine learning methods (such as random forest, decision tree); and (3) deep learning-based methods. With deep learning-oriented approaches contributing to a majority, we further discuss deep learning-based methods like semantic segmentation-based, object detection-based and Transformer-based methods. In addition, we discuss five APBD-related issues to further comprehend the APBD domain using remote sensing data, such as multi-sensor data in APBD task, comparisons between single-task learning and multi-task learning in the APBD domain, comparisons among different algorithms and different APBD tasks, etc. Finally, this review proposes some APBD-related applications and a few exciting prospects and potential hot topics in future APBD research. We hope this review help researchers who involved in APBD domain to keep track of its development and tendency.

Zhiwei Zhang, Zi Ye, Yibin Wen et al.

Agricultural parcels serve as basic units for conducting agricultural practices and applications, which is vital for land ownership registration, food security assessment, soil erosion monitoring, etc. However, existing agriculture parcel extraction studies only focus on mid-resolution mapping or regular plain farmlands while lacking representation of complex terraced terrains due to the demands of precision agriculture.In this paper, we introduce a more fine-grained terraced parcel dataset named GTPBD (Global Terraced Parcel and Boundary Dataset), which is the first fine-grained dataset covering major worldwide terraced regions with more than 200,000 complex terraced parcels with manual annotation. GTPBD comprises 47,537 high-resolution images with three-level labels, including pixel-level boundary labels, mask labels, and parcel labels. It covers seven major geographic zones in China and transcontinental climatic regions around the world.Compared to the existing datasets, the GTPBD dataset brings considerable challenges due to the: (1) terrain diversity; (2) complex and irregular parcel objects; and (3) multiple domain styles. Our proposed GTPBD dataset is suitable for four different tasks, including semantic segmentation, edge detection, terraced parcel extraction, and unsupervised domain adaptation (UDA) tasks.Accordingly, we benchmark the GTPBD dataset on eight semantic segmentation methods, four edge extraction methods, three parcel extraction methods, and five UDA methods, along with a multi-dimensional evaluation framework integrating pixel-level and object-level metrics. GTPBD fills a critical gap in terraced remote sensing research, providing a basic infrastructure for fine-grained agricultural terrain analysis and cross-scenario knowledge transfer.