A Survey of Retentive Network

Retentive Network (RetNet) represents a significant advancement in neural network architecture, offering an efficient alternative to the Transformer. While Transformers rely on self-attention to model dependencies, they suffer from high memory costs and limited scalability when handling long sequences due to their quadratic complexity. To mitigate these limitations, RetNet introduces a retention mechanism that unifies the inductive bias of recurrence with the global dependency modeling of attention. This mechanism enables linear-time inference, facilitates efficient modeling of extended contexts, and remains compatible with fully parallelizable training pipelines. RetNet has garnered significant research interest due to its consistently demonstrated cross-domain effectiveness, achieving robust performance across machine learning paradigms including natural language processing, speech recognition, and time-series analysis. However, a comprehensive review of RetNet is still missing from the current literature. This paper aims to fill that gap by offering the first detailed survey of the RetNet architecture, its key innovations, and its diverse applications. We also explore the main challenges associated with RetNet and propose future research directions to support its continued advancement in both academic research and practical deployment.

Don't Take the Premise for Granted: Evaluating the Premise Critique Ability of Large Language Models

Large language models (LLMs) have witnessed rapid advancements, demonstrating remarkable capabilities. However, a notable vulnerability persists: LLMs often uncritically accept flawed or contradictory premises, leading to inefficient reasoning and unreliable outputs. This emphasizes the significance of possessing the \textbf{Premise Critique Ability} for LLMs, defined as the capacity to proactively identify and articulate errors in input premises. Most existing studies assess LLMs' reasoning ability in ideal settings, largely ignoring their vulnerabilities when faced with flawed premises. Thus, we introduce the \textbf{Premise Critique Bench (PCBench)}, designed by incorporating four error types across three difficulty levels, paired with multi-faceted evaluation metrics. We conducted systematic evaluations of 15 representative LLMs. Our findings reveal: (1) Most models rely heavily on explicit prompts to detect errors, with limited autonomous critique; (2) Premise critique ability depends on question difficulty and error type, with direct contradictions being easier to detect than complex or procedural errors; (3) Reasoning ability does not consistently correlate with the premise critique ability; (4) Flawed premises trigger overthinking in reasoning models, markedly lengthening responses due to repeated attempts at resolving conflicts. These insights underscore the urgent need to enhance LLMs' proactive evaluation of input validity, positioning premise critique as a foundational capability for developing reliable, human-centric systems. The code is available at https://github.com/MLGroupJLU/Premise_Critique.

THINK-Bench: Evaluating Thinking Efficiency and Chain-of-Thought Quality of Large Reasoning Models

Large reasoning models (LRMs) have achieved impressive performance in complex tasks, often outperforming conventional large language models (LLMs). However, the prevalent issue of overthinking severely limits their computational efficiency. Overthinking occurs when models generate excessive and redundant tokens that contribute little to accurate outcomes, especially in simple tasks, resulting in a significant waste of computational resources. To systematically investigate this issue, we introduce Think-Bench, a benchmark designed to evaluate the reasoning efficiency of LRMs. We also propose novel efficiency metrics and conduct a comprehensive evaluation of various LRMs across multiple dimensions, including the reasoning process, outcome quality, and chain-of-thought (CoT) characteristics. Our analysis reveals that most LRMs exhibit overthinking in handling easy questions, generating unnecessarily lengthy reasoning chains. While many LRMs demonstrate high CoT quality, several suffer from low efficiency. We hope that Think-Bench can serve as a robust foundation for advancing research into LRMs.

See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction

Occupancy prediction aims to estimate the 3D spatial distribution of occupied regions along with their corresponding semantic labels. Existing vision-based methods perform well on daytime benchmarks but struggle in nighttime scenarios due to limited visibility and challenging lighting conditions. To address these challenges, we propose \textbf{LIAR}, a novel framework that learns illumination-affined representations. LIAR first introduces Selective Low-light Image Enhancement (SLLIE), which leverages the illumination priors from daytime scenes to adaptively determine whether a nighttime image is genuinely dark or sufficiently well-lit, enabling more targeted global enhancement. Building on the illumination maps generated by SLLIE, LIAR further incorporates two illumination-aware components: 2D Illumination-guided Sampling (2D-IGS) and 3D Illumination-driven Projection (3D-IDP), to respectively tackle local underexposure and overexposure. Specifically, 2D-IGS modulates feature sampling positions according to illumination maps, assigning larger offsets to darker regions and smaller ones to brighter regions, thereby alleviating feature degradation in underexposed areas. Subsequently, 3D-IDP enhances semantic understanding in overexposed regions by constructing illumination intensity fields and supplying refined residual queries to the BEV context refinement process. Extensive experiments on both real and synthetic datasets demonstrate the superior performance of LIAR under challenging nighttime scenarios. The source code and pretrained models are available \href{https://github.com/yanzq95/LIAR}{here}.

LMHLD: A Large-scale Multi-source High-resolution Landslide Dataset for Landslide Detection based on Deep Learning

Landslides are among the most common natural disasters globally, posing significant threats to human society. Deep learning (DL) has proven to be an effective method for rapidly generating landslide inventories in large-scale disaster areas. However, DL models rely heavily on high-quality labeled landslide data for strong feature extraction capabilities. And landslide detection using DL urgently needs a benchmark dataset to evaluate the generalization ability of the latest models. To solve the above problems, we construct a Large-scale Multi-source High-resolution Landslide Dataset (LMHLD) for Landslide Detection based on DL. LMHLD collects remote sensing images from five different satellite sensors across seven study areas worldwide: Wenchuan, China (2008); Rio de Janeiro, Brazil (2011); Gorkha, Nepal (2015); Jiuzhaigou, China (2015); Taiwan, China (2018); Hokkaido, Japan (2018); Emilia-Romagna, Italy (2023). The dataset includes a total of 25,365 patches, with different patch sizes to accommodate different landslide scales. Additionally, a training module, LMHLDpart, is designed to accommodate landslide detection tasks at varying scales and to alleviate the issue of catastrophic forgetting in multi-task learning. Furthermore, the models trained by LMHLD is applied in other datasets to highlight the robustness of LMHLD. Five dataset quality evaluation experiments designed by using seven DL models from the U-Net family demonstrate that LMHLD has the potential to become a benchmark dataset for landslide detection. LMHLD is open access and can be accessed through the link: https://doi.org/10.5281/zenodo.11424988. This dataset provides a strong foundation for DL models, accelerates the development of DL in landslide detection, and serves as a valuable resource for landslide prevention and mitigation efforts.

Mixup Model Merge: Enhancing Model Merging Performance through Randomized Linear Interpolation

Model merging integrates the parameters of multiple models into a unified model, combining their diverse capabilities. Existing model merging methods are often constrained by fixed parameter merging ratios. In this study, we propose Mixup Model Merge (M$^3$), an innovative approach inspired by the Mixup data augmentation technique. This method merges the parameters of two large language models (LLMs) by randomly generating linear interpolation ratios, allowing for a more flexible and comprehensive exploration of the parameter space. Extensive experiments demonstrate the superiority of our proposed M$^3$ method in merging fine-tuned LLMs: (1) it significantly improves performance across multiple tasks, (2) it enhances LLMs' out-of-distribution (OOD) robustness and adversarial robustness, (3) it achieves superior results when combined with sparsification techniques such as DARE, and (4) it offers a simple yet efficient solution that does not require additional computational resources. In conclusion, M$^3$ is a simple yet effective model merging method that significantly enhances the performance of the merged model by randomly generating contribution ratios for two fine-tuned LLMs. The code is available at https://github.com/MLGroupJLU/MixupModelMerge.

R-LoRA: Random Initialization of Multi-Head LoRA for Multi-Task Learning

Fine-tuning large language models (LLMs) is prohibitively expensive in terms of computational and memory costs. Low-rank Adaptation (LoRA), as one of the most popular parameter-efficient fine-tuning (PEFT) methods, offers a cost-effective alternative by approximating the model changes $\Delta W \in \mathbb{R}^{m \times n}$ through the product of down-projection matrix $A \in \mathbb{R}^{m \times r}$ and head matrix $B \in \mathbb{R}^{r \times n}$, where $r \ll \min(m, n)$. In real-world scenarios, LLMs are fine-tuned on data from multiple domains to perform tasks across various fields, embodying multi-task learning (MTL). LoRA often underperforms in such complex scenarios. To enhance LoRA's capability in multi-task learning, we propose R-LoRA, which incorporates Multi-Head Randomization. Multi-Head Randomization diversifies the head matrices through Multi-Head Random Initialization and Multi-Head Dropout, enabling more efficient learning of task-specific features while maintaining shared knowledge representation. Extensive experiments demonstrate that R-LoRA is better at capturing task-specific knowledge, thereby improving performance in multi-task scenarios. The code is available at https://github.com/jinda-liu/R-LoRA.

StructFlowBench: A Structured Flow Benchmark for Multi-turn Instruction Following

Multi-turn instruction following capability constitutes a core competency of large language models (LLMs) in real-world applications. Existing evaluation benchmarks predominantly focus on fine-grained constraint satisfaction and domain-specific capability assessment, yet overlook the crucial structural dependency between dialogue turns that distinguishes multi-turn from single-turn interactions. This structural dependency not only reflects user intent but also establishes a second dimension for instruction following evaluation beyond constraint satisfaction. To address this gap, we propose StructFlowBench, a multi-turn instruction following benchmark with structural flow modeling. The benchmark innovatively defines a structural flow framework comprising six fundamental inter-turn relationships, which not only introduces novel structural constraints for model evaluation but also serves as generation parameters for creating customized dialogue flows tailored to specific scenarios. Adopting established LLM-based automatic evaluation methodologies, we conduct systematic evaluations of 13 leading open-source and closed-source LLMs. Experimental results reveal significant deficiencies in current models' comprehension of multi-turn dialogue structures. The code is available at \url{https://github.com/MLGroupJLU/StructFlowBench}.

Length-Controlled Margin-Based Preference Optimization without Reference Model

Direct Preference Optimization (DPO) is a widely adopted offline algorithm for preference-based reinforcement learning from human feedback (RLHF), designed to improve training simplicity and stability by redefining reward functions. However, DPO is hindered by several limitations, including length bias, memory inefficiency, and probability degradation. To address these challenges, we propose Length-Controlled Margin-Based Preference Optimization (LMPO), a more efficient and robust alternative. LMPO introduces a uniform reference model as an upper bound for the DPO loss, enabling a more accurate approximation of the original optimization objective. Additionally, an average log-probability optimization strategy is employed to minimize discrepancies between training and inference phases. A key innovation of LMPO lies in its Length-Controlled Margin-Based loss function, integrated within the Bradley-Terry framework. This loss function regulates response length while simultaneously widening the margin between preferred and rejected outputs. By doing so, it mitigates probability degradation for both accepted and discarded responses, addressing a significant limitation of existing methods. We evaluate LMPO against state-of-the-art preference optimization techniques on two open-ended large language models, Mistral and LLaMA3, across six conditional benchmarks. Our experimental results demonstrate that LMPO effectively controls response length, reduces probability degradation, and outperforms existing approaches. The code is available at \url{https://github.com/gengxuli/LMPO}.

Asymmetric Co-Training for Source-Free Few-Shot Domain Adaptation

Source-free unsupervised domain adaptation (SFUDA) has gained significant attention as an alternative to traditional unsupervised domain adaptation (UDA), which relies on the constant availability of labeled source data. However, SFUDA approaches come with inherent limitations that are frequently overlooked. These challenges include performance degradation when the unlabeled target data fails to meet critical assumptions, such as having a closed-set label distribution identical to that of the source domain, or when sufficient unlabeled target data is unavailable-a common situation in real-world applications. To address these issues, we propose an asymmetric co-training (ACT) method specifically designed for the SFFSDA scenario. SFFSDA presents a more practical alternative to SFUDA, as gathering a few labeled target instances is more feasible than acquiring large volumes of unlabeled target data in many real-world contexts. Our ACT method begins by employing a weak-strong augmentation to enhance data diversity. Then we use a two-step optimization process to train the target model. In the first step, we optimize the label smoothing cross-entropy loss, the entropy of the class-conditional distribution, and the reverse-entropy loss to bolster the model's discriminative ability while mitigating overfitting. The second step focuses on reducing redundancy in the output space by minimizing classifier determinacy disparity. Extensive experiments across four benchmarks demonstrate the superiority of our ACT approach, which outperforms state-of-the-art SFUDA methods and transfer learning techniques. Our findings suggest that adapting a source pre-trained model using only a small amount of labeled target data offers a practical and dependable solution. The code is available at https://github.com/gengxuli/ACT.