Yu Zhao

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Tianyi Zhang, Zhiling Yan, Chunhui Li et al.

In pathology image analysis, obtaining and maintaining high-quality annotated samples is an extremely labor-intensive task. To overcome this challenge, mixing-based methods have emerged as effective alternatives to traditional preprocessing data augmentation techniques. Nonetheless, these methods fail to fully consider the unique features of pathology images, such as local specificity, global distribution, and inner/outer-sample instance relationships. To better comprehend these characteristics and create valuable pseudo samples, we propose the CellMix framework, which employs a novel distribution-oriented in-place shuffle approach. By dividing images into patches based on the granularity of pathology instances and shuffling them within the same batch, the absolute relationships between instances can be effectively preserved when generating new samples. Moreover, we develop a curriculum learning-inspired, loss-driven strategy to handle perturbations and distribution-related noise during training, enabling the model to adaptively fit the augmented data. Our experiments in pathology image classification tasks demonstrate state-of-the-art (SOTA) performance on 7 distinct datasets. This innovative instance relationship-centered method has the potential to inform general data augmentation approaches for pathology image classification. The associated codes are available at https://github.com/sagizty/CellMix.

Xiang Li, Yiqun Yao, Xin Jiang et al.

Large Language Models (LLMs) represent a significant stride toward Artificial General Intelligence. As scaling laws underscore the potential of increasing model sizes, the academic community has intensified its investigations into LLMs with capacities exceeding 50 billion parameters. This technical report builds on our prior work with Tele-FLM (also known as FLM-2), a publicly available 52-billion-parameter model. We delve into two primary areas: we first discuss our observation of Supervised Fine-tuning (SFT) on Tele-FLM-52B, which supports the "less is more" approach for SFT data construction; second, we demonstrate our experiments and analyses on the best practices for progressively growing a model from 52 billion to 102 billion, and subsequently to 1 trillion parameters. We will open-source a 1T model checkpoint, namely Tele-FLM-1T, to advance further training and research.

Yu Zhao, Hao Guan, Yongcheng Jing et al.

Large Language Models (LLMs) have shown strong potential in complex medical reasoning yet face diminishing gains under inference scaling laws. While existing studies augment LLMs with various knowledge types, it remains unclear how effectively the additional costs translate into accuracy. In this paper, we explore how meta-cognition of LLMs, i.e., their self-assessment of their own cognitive states, can regulate the reasoning process. Specifically, we propose MedCoG, a Medical Meta-Cognition Agent with Knowledge Graph, where the meta-cognitive assessments of task complexity, familiarity, and knowledge density dynamically regulate utilization of procedural, episodic, and factual knowledge. The LLM-centric on-demand reasoning aims to mitigate the diminishing returns under scaling law by (1) reducing costs via avoiding indiscriminate scaling, (2) improving accuracy via filtering out distractive knowledge. To validate this, we empirically characterize the scaling curve and introduce inference density to quantify inference efficiency. Experiments demonstrate the effectiveness and efficiency of MedCoG on five hard sets of medical benchmarks, yielding 6.2x inference density. Furthermore, the Oracle study highlights the significant potential of meta-cognitive regulation.

Jiawei Yang, Hanbo Chen, Yu Zhao et al.

Whole slide image (WSI) classification often relies on deep weakly supervised multiple instance learning (MIL) methods to handle gigapixel resolution images and slide-level labels. Yet the decent performance of deep learning comes from harnessing massive datasets and diverse samples, urging the need for efficient training pipelines for scaling to large datasets and data augmentation techniques for diversifying samples. However, current MIL-based WSI classification pipelines are memory-expensive and computation-inefficient since they usually assemble tens of thousands of patches as bags for computation. On the other hand, despite their popularity in other tasks, data augmentations are unexplored for WSI MIL frameworks. To address them, we propose ReMix, a general and efficient framework for MIL based WSI classification. It comprises two steps: reduce and mix. First, it reduces the number of instances in WSI bags by substituting instances with instance prototypes, i.e., patch cluster centroids. Then, we propose a ``Mix-the-bag'' augmentation that contains four online, stochastic and flexible latent space augmentations. It brings diverse and reliable class-identity-preserving semantic changes in the latent space while enforcing semantic-perturbation invariance. We evaluate ReMix on two public datasets with two state-of-the-art MIL methods. In our experiments, consistent improvements in precision, accuracy, and recall have been achieved but with orders of magnitude reduced training time and memory consumption, demonstrating ReMix's effectiveness and efficiency. Code is available.

Yu Zhao, Xiangrui Cai, Yike Wu et al.

Multimodal knowledge graph completion (MKGC) aims to predict missing entities in MKGs. Previous works usually share relation representation across modalities. This results in mutual interference between modalities during training, since for a pair of entities, the relation from one modality probably contradicts that from another modality. Furthermore, making a unified prediction based on the shared relation representation treats the input in different modalities equally, while their importance to the MKGC task should be different. In this paper, we propose MoSE, a Modality Split representation learning and Ensemble inference framework for MKGC. Specifically, in the training phase, we learn modality-split relation embeddings for each modality instead of a single modality-shared one, which alleviates the modality interference. Based on these embeddings, in the inference phase, we first make modality-split predictions and then exploit various ensemble methods to combine the predictions with different weights, which models the modality importance dynamically. Experimental results on three KG datasets show that MoSE outperforms state-of-the-art MKGC methods. Codes are available at https://github.com/OreOZhao/MoSE4MKGC.

Yike Wu, Yu Zhao, Shiwan Zhao et al.

Despite the great progress of Visual Question Answering (VQA), current VQA models heavily rely on the superficial correlation between the question type and its corresponding frequent answers (i.e., language priors) to make predictions, without really understanding the input. In this work, we define the training instances with the same question type but different answers as \textit{superficially similar instances}, and attribute the language priors to the confusion of VQA model on such instances. To solve this problem, we propose a novel training framework that explicitly encourages the VQA model to distinguish between the superficially similar instances. Specifically, for each training instance, we first construct a set that contains its superficially similar counterparts. Then we exploit the proposed distinguishing module to increase the distance between the instance and its counterparts in the answer space. In this way, the VQA model is forced to further focus on the other parts of the input beyond the question type, which helps to overcome the language priors. Experimental results show that our method achieves the state-of-the-art performance on VQA-CP v2. Codes are available at \href{https://github.com/wyk-nku/Distinguishing-VQA.git}{Distinguishing-VQA}.

Qiong Wu, Yu Zhao, Qiang Fan et al.

The vehicular edge computing (VEC) can cache contents in different RSUs at the network edge to support the real-time vehicular applications. In VEC, owing to the high-mobility characteristics of vehicles, it is necessary to cache the user data in advance and learn the most popular and interesting contents for vehicular users. Since user data usually contains privacy information, users are reluctant to share their data with others. To solve this problem, traditional federated learning (FL) needs to update the global model synchronously through aggregating all users' local models to protect users' privacy. However, vehicles may frequently drive out of the coverage area of the VEC before they achieve their local model trainings and thus the local models cannot be uploaded as expected, which would reduce the accuracy of the global model. In addition, the caching capacity of the local RSU is limited and the popular contents are diverse, thus the size of the predicted popular contents usually exceeds the cache capacity of the local RSU. Hence, the VEC should cache the predicted popular contents in different RSUs while considering the content transmission delay. In this paper, we consider the mobility of vehicles and propose a cooperative Caching scheme in the VEC based on Asynchronous Federated and deep Reinforcement learning (CAFR). We first consider the mobility of vehicles and propose an asynchronous FL algorithm to obtain an accurate global model, and then propose an algorithm to predict the popular contents based on the global model. In addition, we consider the mobility of vehicles and propose a deep reinforcement learning algorithm to obtain the optimal cooperative caching location for the predicted popular contents in order to optimize the content transmission delay. Extensive experimental results have demonstrated that the CAFR scheme outperforms other baseline caching schemes.

Yu Zhao, Yongfang Liu, Zhongkui Li et al.

This paper studies the distributed average tracking problem for multiple time-varying signals generated by linear dynamics, whose reference inputs are nonzero and not available to any agent in the network. In the edge-based framework, a pair of continuous algorithms with, respectively, static and adaptive coupling strengths are designed. Based on the boundary layer concept, the proposed continuous algorithm with static coupling strengths can asymptotically track the average of multiple reference signals without the chattering phenomenon. Furthermore, for the case of algorithms with adaptive coupling strengths, average tracking errors are uniformly ultimately bounded and exponentially converge to a small adjustable bounded set. Finally, a simulation example is presented to show the validity of theoretical results.

Tianyi Zhang, Youdan Feng, Yunlu Feng et al.

The rapid on-site evaluation (ROSE) technique can signifi-cantly accelerate the diagnosis of pancreatic cancer by im-mediately analyzing the fast-stained cytopathological images. Computer-aided diagnosis (CAD) can potentially address the shortage of pathologists in ROSE. However, the cancerous patterns vary significantly between different samples, making the CAD task extremely challenging. Besides, the ROSE images have complicated perturbations regarding color distribution, brightness, and contrast due to different staining qualities and various acquisition device types. To address these challenges, we proposed a shuffle instances-based Vision Transformer (SI-ViT) approach, which can reduce the perturbations and enhance the modeling among the instances. With the regrouped bags of shuffle instances and their bag-level soft labels, the approach utilizes a regression head to make the model focus on the cells rather than various perturbations. Simultaneously, combined with a classification head, the model can effectively identify the general distributive patterns among different instances. The results demonstrate significant improvements in the classification accuracy with more accurate attention regions, indicating that the diverse patterns of ROSE images are effectively extracted, and the complicated perturbations are significantly reduced. It also suggests that the SI-ViT has excellent potential in analyzing cytopathological images. The code and experimental results are available at https://github.com/sagizty/MIL-SI.

Yu Zhao, Yunxin Li, Yuxiang Wu et al.

Medical dialogue generation is an important yet challenging task. Most previous works rely on the attention mechanism and large-scale pretrained language models. However, these methods often fail to acquire pivotal information from the long dialogue history to yield an accurate and informative response, due to the fact that the medical entities usually scatters throughout multiple utterances along with the complex relationships between them. To mitigate this problem, we propose a medical response generation model with Pivotal Information Recalling (MedPIR), which is built on two components, i.e., knowledge-aware dialogue graph encoder and recall-enhanced generator. The knowledge-aware dialogue graph encoder constructs a dialogue graph by exploiting the knowledge relationships between entities in the utterances, and encodes it with a graph attention network. Then, the recall-enhanced generator strengthens the usage of these pivotal information by generating a summary of the dialogue before producing the actual response. Experimental results on two large-scale medical dialogue datasets show that MedPIR outperforms the strong baselines in BLEU scores and medical entities F1 measure.

Yuxiang Wu, Yu Zhao, Baotian Hu et al.

Access to external knowledge is essential for many natural language processing tasks, such as question answering and dialogue. Existing methods often rely on a parametric model that stores knowledge in its parameters, or use a retrieval-augmented model that has access to an external knowledge source. Parametric and retrieval-augmented models have complementary strengths in terms of computational efficiency and predictive accuracy. To combine the strength of both approaches, we propose the Efficient Memory-Augmented Transformer (EMAT) -- it encodes external knowledge into a key-value memory and exploits the fast maximum inner product search for memory querying. We also introduce pre-training tasks that allow EMAT to encode informative key-value representations, and to learn an implicit strategy to integrate multiple memory slots into the transformer. Experiments on various knowledge-intensive tasks such as question answering and dialogue datasets show that, simply augmenting parametric models (T5-base) using our method produces more accurate results (e.g., 25.8 -> 44.3 EM on NQ) while retaining a high throughput (e.g., 1000 queries/s on NQ). Compared to retrieval-augmented models, EMAT runs substantially faster across the board and produces more accurate results on WoW and ELI5. Our code and datasets are available at https://github. com/uclnlp/EMAT.

Yu Zhao, Hao Fei, Yixin Cao et al.

Video Semantic Role Labeling (VidSRL) aims to detect the salient events from given videos, by recognizing the predict-argument event structures and the interrelationships between events. While recent endeavors have put forth methods for VidSRL, they can be mostly subject to two key drawbacks, including the lack of fine-grained spatial scene perception and the insufficiently modeling of video temporality. Towards this end, this work explores a novel holistic spatio-temporal scene graph (namely HostSG) representation based on the existing dynamic scene graph structures, which well model both the fine-grained spatial semantics and temporal dynamics of videos for VidSRL. Built upon the HostSG, we present a nichetargeting VidSRL framework. A scene-event mapping mechanism is first designed to bridge the gap between the underlying scene structure and the high-level event semantic structure, resulting in an overall hierarchical scene-event (termed ICE) graph structure. We further perform iterative structure refinement to optimize the ICE graph, such that the overall structure representation can best coincide with end task demand. Finally, three subtask predictions of VidSRL are jointly decoded, where the end-to-end paradigm effectively avoids error propagation. On the benchmark dataset, our framework boosts significantly over the current best-performing model. Further analyses are shown for a better understanding of the advances of our methods.

Yu Zhao, Jianguo Wei, Zhichao Lin et al.

Image-to-text tasks, such as open-ended image captioning and controllable image description, have received extensive attention for decades. Here, we further advance this line of work by presenting Visual Spatial Description (VSD), a new perspective for image-to-text toward spatial semantics. Given an image and two objects inside it, VSD aims to produce one description focusing on the spatial perspective between the two objects. Accordingly, we manually annotate a dataset to facilitate the investigation of the newly-introduced task and build several benchmark encoder-decoder models by using VL-BART and VL-T5 as backbones. In addition, we investigate pipeline and joint end-to-end architectures for incorporating visual spatial relationship classification (VSRC) information into our model. Finally, we conduct experiments on our benchmark dataset to evaluate all our models. Results show that our models are impressive, providing accurate and human-like spatial-oriented text descriptions. Meanwhile, VSRC has great potential for VSD, and the joint end-to-end architecture is the better choice for their integration. We make the dataset and codes public for research purposes.

Bobo Li, Hao Fei, Lizi Liao et al.

It has been a hot research topic to enable machines to understand human emotions in multimodal contexts under dialogue scenarios, which is tasked with multimodal emotion analysis in conversation (MM-ERC). MM-ERC has received consistent attention in recent years, where a diverse range of methods has been proposed for securing better task performance. Most existing works treat MM-ERC as a standard multimodal classification problem and perform multimodal feature disentanglement and fusion for maximizing feature utility. Yet after revisiting the characteristic of MM-ERC, we argue that both the feature multimodality and conversational contextualization should be properly modeled simultaneously during the feature disentanglement and fusion steps. In this work, we target further pushing the task performance by taking full consideration of the above insights. On the one hand, during feature disentanglement, based on the contrastive learning technique, we devise a Dual-level Disentanglement Mechanism (DDM) to decouple the features into both the modality space and utterance space. On the other hand, during the feature fusion stage, we propose a Contribution-aware Fusion Mechanism (CFM) and a Context Refusion Mechanism (CRM) for multimodal and context integration, respectively. They together schedule the proper integrations of multimodal and context features. Specifically, CFM explicitly manages the multimodal feature contributions dynamically, while CRM flexibly coordinates the introduction of dialogue contexts. On two public MM-ERC datasets, our system achieves new state-of-the-art performance consistently. Further analyses demonstrate that all our proposed mechanisms greatly facilitate the MM-ERC task by making full use of the multimodal and context features adaptively. Note that our proposed methods have the great potential to facilitate a broader range of other conversational multimodal tasks.

Yu Zhao, Yongfang Liu, Guanrong Chen

This technical note addresses the distributed fixed-time consensus protocol design problem for multi-agent systems with general linear dynamics over directed communication graphs. By using motion planning approaches, a class of distributed fixed-time consensus algorithms are developed, which rely only on the sampling information at some sampling instants. For linear multi-agent systems, the proposed algorithms solve the fixed-time consensus problem for any directed graph containing a directed spanning tree. In particular, the settling time can be off-line pre-assigned according to task requirements. Compared with the existing results for multi-agent systems, to our best knowledge, it is the first-time to solve fixed-time consensus problems for general linear multi-agent systems over directed graphs having a directed spanning tree. Extensions to the fixed-time formation flying are further studied for multiple satellites described by Hill equations.

Yu Zhao, Yongfang Liu

In this paper, a distributed average tracking problem is studied for Lipschitz-type nonlinear dynamical systems. The objective is to design distributed average tracking algorithms for locally interactive agents to track the average of multiple reference signals. Here, in both the agents' and the reference signals' dynamics, there is a nonlinear term satisfying the Lipschitz-type condition. Three types of distributed average tracking algorithms are designed. First, based on state-dependent-gain designing approaches, a robust distributed average tracking algorithm is developed to solve distributed average tracking problems without requiring the same initial condition. Second, by using a gain adaption scheme, an adaptive distributed average tracking algorithm is proposed in this paper to remove the requirement that the Lipschitz constant is known for agents. Third, to reduce chattering and make the algorithms easier to implement, a continuous distributed average tracking algorithm based on a time-varying boundary layer is further designed as a continuous approximation of the previous discontinuous distributed average tracking algorithms.

Yu Zhao, Ying Zhang, Baohang Zhou et al.

A large number of studies have emerged for Multimodal Knowledge Graph Completion (MKGC) to predict the missing links in MKGs. However, fewer studies have been proposed to study the inductive MKGC (IMKGC) involving emerging entities unseen during training. Existing inductive approaches focus on learning textual entity representations, which neglect rich semantic information in visual modality. Moreover, they focus on aggregating structural neighbors from existing KGs, which of emerging entities are usually limited. However, the semantic neighbors are decoupled from the topology linkage and usually imply the true target entity. In this paper, we propose the IMKGC task and a semantic neighbor retrieval-enhanced IMKGC framework CMR, where the contrast brings the helpful semantic neighbors close, and then the memorize supports semantic neighbor retrieval to enhance inference. Specifically, we first propose a unified cross-modal contrastive learning to simultaneously capture the textual-visual and textual-textual correlations of query-entity pairs in a unified representation space. The contrastive learning increases the similarity of positive query-entity pairs, therefore making the representations of helpful semantic neighbors close. Then, we explicitly memorize the knowledge representations to support the semantic neighbor retrieval. At test time, we retrieve the nearest semantic neighbors and interpolate them to the query-entity similarity distribution to augment the final prediction. Extensive experiments validate the effectiveness of CMR on three inductive MKGC datasets. Codes are available at https://github.com/OreOZhao/CMR.

Yongfang Liu, Yu Zhao, Wei Ren et al.

This paper investigates the fixed-time consensus problem under directed topologies. By using a motion-planning approach, a class of distributed fixed-time algorithms are developed for a multi-agent system with double-integrator dynamics. In the context of the fixed-time consensus, we focus on both directed fixed and switching topologies. Under the directed fixed topology, a novel class of distributed algorithms are designed, which guarantee the consensus of the multi-agent system with a fixed settling time if the topology has a directed spanning tree. Under the directed periodically switching topologies, the fixedtime consensus is solved via the proposed algorithms if the topologies jointly have a directed spanning tree. In particular, the fixed settling time can be off-line pre-assigned according to task requirements. Compared with the existing results, to our best knowledge, it is the first time to solve the fixed-time consensus problem for double-integrator systems under directed topologies. Finally, a numerical example is given to illustrate the effectiveness of the analytical results.

Zeyu Liu, Tianyi Zhang, Yufang He et al.

Deep learning is widely applied in computer-aided pathological diagnosis, which alleviates the pathologist workload and provide timely clinical analysis. However, most models generally require large-scale annotated data for training, which faces challenges due to the sampling and annotation scarcity in pathological images. The rapid developing generative models shows potential to generate more training samples from recent studies. However, they also struggle in generalization diversity with limited training data, incapable of generating effective samples. Inspired by the pathological transitions between different stages, we propose an adaptive depth-controlled diffusion (ADD) network to generate pathological progressive images for effective data augmentation. This novel approach roots in domain migration, where a hybrid attention strategy guides the bidirectional diffusion, blending local and global attention priorities. With feature measuring, the adaptive depth-controlled strategy ensures the migration and maintains locational similarity in simulating the pathological feature transition. Based on tiny training set (samples less than 500), the ADD yields cross-domain progressive images with corresponding soft-labels. Experiments on two datasets suggest significant improvements in generation diversity, and the effectiveness with generated progressive samples are highlighted in downstream classifications. The code is available at https://github.com/Rowerliu/ADD.

Shaopeng Wei, Manzhen Wei, Haoyu Wang et al.

We present the Melody-Guided Music Generation (MG2) model, a novel approach using melody to guide the text-to-music generation that, despite a simple method and limited resources, achieves excellent performance. Specifically, we first align the text with audio waveforms and their associated melodies using the newly proposed Contrastive Language-Music Pretraining, enabling the learned text representation fused with implicit melody information. Subsequently, we condition the retrieval-augmented diffusion module on both text prompt and retrieved melody. This allows MG2 to generate music that reflects the content of the given text description, meantime keeping the intrinsic harmony under the guidance of explicit melody information. We conducted extensive experiments on two public datasets: MusicCaps and MusicBench. Surprisingly, the experimental results demonstrate that the proposed MG2 model surpasses current open-source text-to-music generation models, achieving this with fewer than 1/3 of the parameters or less than 1/200 of the training data compared to state-of-the-art counterparts. Furthermore, we conducted comprehensive human evaluations involving three types of users and five perspectives, using newly designed questionnaires to explore the potential real-world applications of MG2.

Xiaofeng Cong, Jing Zhang, Yeying Jin et al.

Underwater images often suffer from quality degradation due to absorption and scattering effects. Most existing underwater image enhancement algorithms produce a single, fixed-color image, limiting user flexibility and application. To address this limitation, we propose a method called \textit{ColorCode}, which enhances underwater images while offering a range of controllable color outputs. Our approach involves recovering an underwater image to a reference enhanced image through supervised training and decomposing it into color and content codes via self-reconstruction and cross-reconstruction. The color code is explicitly constrained to follow a Gaussian distribution, allowing for efficient sampling and interpolation during inference. ColorCode offers three key features: 1) color enhancement, producing an enhanced image with a fixed color; 2) color adaptation, enabling controllable adjustments of long-wavelength color components using guidance images; and 3) color interpolation, allowing for the smooth generation of multiple colors through continuous sampling of the color code. Quantitative and visual evaluations on popular and challenging benchmark datasets demonstrate the superiority of ColorCode over existing methods in providing diverse, controllable, and color-realistic enhancement results. The source code is available at https://github.com/Xiaofeng-life/ColorCode.

Edoardo D'Amico, Marco De Nadai, Praveen Chandar et al.

Podcast listening is often grounded in a set of favorite shows, while listener intent can evolve over time. This combination of stable preferences and changing intent motivates recommendation approaches that support both familiarity and exploration. Traditional recommender systems typically emphasize long-term interaction patterns, and are less explicitly designed to incorporate rich contextual signals or flexible, intent-aware discovery objectives. In this setting, models that can jointly reason over semantics, context, and user state offer a promising direction. Large Language Models (LLMs) provide strong semantic reasoning and contextual conditioning for discovery-oriented recommendation, but deploying them in production introduces challenges in catalog grounding, user-level personalization, and latency-critical serving. We address these challenges with GLIDE, a production-scale generative recommender for podcast discovery at Spotify. GLIDE formulates recommendation as an instruction-following task over a discretized catalog using Semantic IDs, enabling grounded generation over a large inventory. The model conditions on recent listening history and lightweight user context, while injecting long-term user embeddings as soft prompts to capture stable preferences under strict inference constraints. We evaluate GLIDE using offline retrieval metrics, human judgments, and LLM-based evaluation, and validate its impact through large-scale online A/B testing. Across experiments involving millions of users, GLIDE increases non-habitual podcast streaming on Spotify home surface by up to 5.4% and new-show discovery by up to 14.3%, while meeting production cost and latency constraints.

Zihao Cheng, Weixin Wang, Yu Zhao et al.

Long-term memory is fundamental for personalized agents capable of accumulating knowledge, reasoning over user experiences, and adapting across time. However, existing memory benchmarks primarily target declarative memory, specifically semantic and episodic types, where all information is explicitly presented in dialogues. In contrast, real-world actions are also governed by non-declarative memory, including habitual and procedural types, and need to be inferred from diverse digital traces. To bridge this gap, we introduce Lifebench, which features densely connected, long-horizon event simulation. It pushes AI agents beyond simple recall, requiring the integration of declarative and non-declarative memory reasoning across diverse and temporally extended contexts. Building such a benchmark presents two key challenges: ensuring data quality and scalability. We maintain data quality by employing real-world priors, including anonymized social surveys, map APIs, and holiday-integrated calendars, thus enforcing fidelity, diversity and behavioral rationality within the dataset. Towards scalability, we draw inspiration from cognitive science and structure events according to their partonomic hierarchy; enabling efficient parallel generation while maintaining global coherence. Performance results show that top-tier, state-of-the-art memory systems reach just 55.2\% accuracy, highlighting the inherent difficulty of long-horizon retrieval and multi-source integration within our proposed benchmark. The dataset and data synthesis code are available at https://github.com/1754955896/LifeBench.

Aaron Branson Cigres Li, Zhaowei Wang, Yu Zhao et al.

Large vision-language models increasingly rely on long-context modeling to reason over documents, hour-level videos, and long-horizon agent trajectories, requiring them to locate relevant evidence across interleaved text and images. Prior work has studied this behavior using retrieval heads in large language models, but its copy-based criterion does not directly apply when evidence appears in images. We introduce a multimodal retrieval head detection method that scores attention from question tokens to textual or visual evidence. With this method, we show that multimodal retrieval heads are sparse, intrinsic, and causally important: only 4.4-10.2% of attention heads account for 50% of the positive retrieval-score mass, and masking the top-5% selected heads drops MMLongBench-Doc from 48.2% to 5.7% and SlideVQA from 71.2% to 8.9%, while random-head masking is far less damaging. Further analysis shows that these heads are partly shared across modalities yet remain dynamic within each modality, with image retrieval heads changing more than text retrieval heads as context length and haystack modality change. Without further training, we find that these heads can also be used directly to rank visually rich documents: on MMDocIR, Qwen3-VL-8B selected-head scoring improves Recall@1 by 7.7/7.4 macro/micro points for page retrieval and 6.3/6.8 points for layout retrieval over the strongest reported baseline.

Yu Zhao, Zhenzhou Jin, Jinke Tang et al.

With the widespread application of location-based services, fingerprint-based localization has demonstrated advantages in environments with complex signal propagation. Deep learning has significantly improved the efficiency of both offline training and online matching in localization processes. However, existing fingerprints only contain terminal position information without capturing motion states, and neural network designs have not fully incorporated structural features such as fingerprint sparsity. In this paper, we propose a triple-beam fingerprint (TBF) incorporating Doppler information and design a Transformer-based localization and orientation awareness network (LOA-Net) to simultaneously estimate user position and motion direction in massive multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) systems. We first show the correlation between TBF and multipath information, and investigate the collinearity of different TBFs, demonstrating that TBF is an effective small-size sparse fingerprint. Then, we propose LOA-Net containing a mask-augmented detection Transformer for regression (MaskDETR-Reg) module and a fusion-enhanced Transformer for direction classification (Fusion-TDC) module to process angle-delay domain information and Doppler domain information, respectively. Finally, in the simulation of indoor scenarios defined in 3GPP 38.901, the proposed method achieves significantly better localization accuracy than weighted $K$-nearest neighbors (WKNN), 2D and 3D convolutional neural networks (CNNs), and achieves satisfactory motion direction estimation accuracy.

MingCai Chen, Yu Zhao, Bing He et al.

Learning with Noisy Labels (LNL) has attracted significant attention from the research community. Many recent LNL methods rely on the assumption that clean samples tend to have "small loss". However, this assumption always fails to generalize to some real-world cases with imbalanced subpopulations, i.e., training subpopulations varying in sample size or recognition difficulty. Therefore, recent LNL methods face the risk of misclassifying those "informative" samples (e.g., hard samples or samples in the tail subpopulations) into noisy samples, leading to poor generalization performance. To address the above issue, we propose a novel LNL method to simultaneously deal with noisy labels and imbalanced subpopulations. It first leverages sample correlation to estimate samples' clean probabilities for label correction and then utilizes corrected labels for Distributionally Robust Optimization (DRO) to further improve the robustness. Specifically, in contrast to previous works using classification loss as the selection criterion, we introduce a feature-based metric that takes the sample correlation into account for estimating samples' clean probabilities. Then, we refurbish the noisy labels using the estimated clean probabilities and the pseudo-labels from the model's predictions. With refurbished labels, we use DRO to train the model to be robust to subpopulation imbalance. Extensive experiments on a wide range of benchmarks demonstrate that our technique can consistently improve current state-of-the-art robust learning paradigms against noisy labels, especially when encountering imbalanced subpopulations.

Pan Deng, Yu Zhao, Junting Liu et al.

As a representative of public transportation, the fundamental issue of managing bike-sharing systems is bike flow prediction. Recent methods overemphasize the spatio-temporal correlations in the data, ignoring the effects of contextual conditions on the transportation system and the inter-regional timevarying causality. In addition, due to the disturbance of incomplete observations in the data, random contextual conditions lead to spurious correlations between data and features, making the prediction of the model ineffective in special scenarios. To overcome this issue, we propose a Spatio-temporal Neural Structure Causal Model(STNSCM) from the perspective of causality. First, we build a causal graph to describe the traffic prediction, and further analyze the causal relationship between the input data, contextual conditions, spatiotemporal states, and prediction results. Second, we propose to apply the frontdoor criterion to eliminate confounding biases in the feature extraction process. Finally, we propose a counterfactual representation reasoning module to extrapolate the spatio-temporal state under the factual scenario to future counterfactual scenarios to improve the prediction performance. Experiments on real-world datasets demonstrate the superior performance of our model, especially its resistance to fluctuations caused by the external environment. The source code and data will be released.

Changzai Pan, Jie Zhang, Kaiwen Wei et al.

Recent advancements in Large Language Models (LLMs) have significantly catalyzed table-based question answering (TableQA). However, existing TableQA benchmarks often overlook the intricacies of industrial scenarios, which are characterized by multi-table structures, nested headers, and massive scales. These environments demand robust table reasoning through deep structured inference, presenting a significant challenge that remains inadequately addressed by current methodologies. To bridge this gap, we present ReasonTabQA, a large-scale bilingual benchmark encompassing 1,932 tables across 30 industry domains such as energy and automotive. ReasonTabQA provides high-quality annotations for both final answers and explicit reasoning chains, supporting both thinking and no-thinking paradigms. Furthermore, we introduce TabCodeRL, a reinforcement learning method that leverages table-aware verifiable rewards to guide the generation of logical reasoning paths. Extensive experiments on ReasonTabQA and 4 TableQA datasets demonstrate that while TabCodeRL yields substantial performance gains on open-source LLMs, the persistent performance gap on ReasonTabQA underscores the inherent complexity of real-world industrial TableQA.

Yu Guo, Zhilong Xie, Xingyan Chen et al.

Natural language understanding (NLU) has two core tasks: intent classification and slot filling. The success of pre-training language models resulted in a significant breakthrough in the two tasks. One of the promising solutions called BERT can jointly optimize the two tasks. We note that BERT-based models convert each complex token into multiple sub-tokens by wordpiece algorithm, which generates a mismatch between the lengths of the tokens and the labels. This leads to BERT-based models do not do well in label prediction which limits model performance improvement. Many existing models can be compatible with this issue but some hidden semantic information is discarded in the fine-tuning process. We address the problem by introducing a novel joint method on top of BERT which explicitly models the multiple sub-tokens features after wordpiece tokenization, thereby contributing to the two tasks. Our method can well extract the contextual features from complex tokens by the proposed sub-words attention adapter (SAA), which preserves overall utterance information. Additionally, we propose an intent attention adapter (IAA) to obtain the full sentence features to aid users to predict intent. Experimental results confirm that our proposed model is significantly improved on two public benchmark datasets. In particular, the slot filling F1 score is improved from 96.1 to 98.2 (2.1% absolute) on the Airline Travel Information Systems (ATIS) dataset.

Donya Khaledyan, Thomas J. Marini, Avice OConnell et al.

Objective. Limited access to breast cancer diagnosis globally leads to delayed treatment. Ultrasound, an effective yet underutilized method, requires specialized training for sonographers, which hinders its widespread use. Approach. Volume sweep imaging (VSI) is an innovative approach that enables untrained operators to capture high-quality ultrasound images. Combined with deep learning, like convolutional neural networks (CNNs), it can potentially transform breast cancer diagnosis, enhancing accuracy, saving time and costs, and improving patient outcomes. The widely used UNet architecture, known for medical image segmentation, has limitations, such as vanishing gradients and a lack of multi-scale feature extraction and selective region attention. In this study, we present a novel segmentation model known as Wavelet_Attention_UNet (WATUNet). In this model, we incorporate wavelet gates (WGs) and attention gates (AGs) between the encoder and decoder instead of a simple connection to overcome the limitations mentioned, thereby improving model performance. Main results. Two datasets are utilized for the analysis. The public "Breast Ultrasound Images" (BUSI) dataset of 780 images and a VSI dataset of 3818 images. Both datasets contained segmented lesions categorized into three types: no mass, benign mass, and malignant mass. Our segmentation results show superior performance compared to other deep networks. The proposed algorithm attained a Dice coefficient of 0.94 and an F1 score of 0.94 on the VSI dataset and scored 0.93 and 0.94 on the public dataset, respectively.

Huangsen Cao, Hongkang Chu, Yuxi Li et al.

The rapid advancement of generative models has made synthetic images increasingly realistic, challenging reliable detection. Existing methods are often limited to end-to-end classification or monolithic reasoning, and thus fail to model structured forensic reasoning and heterogeneous visual evidence. We revisit synthetic image detection from a cognitive perspective and propose a \textit{Heuristic-to-Reasoning} cognitive skill learning framework for evidence-based forensic analysis. Given an input image, our framework first extracts heuristic perceptual clues, selects the optimal forensic skill, and then performs skill-conditioned reasoning for evidence extraction and decision making. To support this paradigm, we introduce \textbf{ClueAegis-Bench}, which decomposes synthetic image detection into explicitly annotated forensic cognitive skills for structured evaluation beyond binary classification. Based on this benchmark, we propose \textbf{ClueAegis} (\underline{C}ognitive-skill \underline{L}earning for \underline{U}nified \underline{E}vidence-based Synthetic Image Detection), a two-stage agentic framework that conducts heuristic skill selection followed by evidence-guided reasoning through skill-conditioned toolchains. This design reformulates synthetic image detection as a configurable multi-skill reasoning process that bridges perception, skill selection, and forensic reasoning. Extensive experiments show that ClueAegis achieves state-of-the-art performance while improving cross-domain generalization and robustness. It also provides transparent reasoning trajectories and structured forensic evidence, offering a more explainable alternative to conventional end-to-end detectors.

Yu Zhao, Pan Deng, Junting Liu et al.

Multimodal traffic flow can reflect the health of the transportation system, and its prediction is crucial to urban traffic management. Recent works overemphasize spatio-temporal correlations of traffic flow, ignoring the physical concepts that lead to the generation of observations and their causal relationship. Spatio-temporal correlations are considered unstable under the influence of different conditions, and spurious correlations may exist in observations. In this paper, we analyze the physical concepts affecting the generation of multimode traffic flow from the perspective of the observation generation principle and propose a Causal Conditional Hidden Markov Model (CCHMM) to predict multimodal traffic flow. In the latent variables inference stage, a posterior network disentangles the causal representations of the concepts of interest from conditional information and observations, and a causal propagation module mines their causal relationship. In the data generation stage, a prior network samples the causal latent variables from the prior distribution and feeds them into the generator to generate multimodal traffic flow. We use a mutually supervised training method for the prior and posterior to enhance the identifiability of the model. Experiments on real-world datasets show that CCHMM can effectively disentangle causal representations of concepts of interest and identify causality, and accurately predict multimodal traffic flow.

Lucas Maystre, Daniel Russo, Yu Zhao

We present a novel podcast recommender system deployed at industrial scale. This system successfully optimizes personal listening journeys that unfold over months for hundreds of millions of listeners. In deviating from the pervasive industry practice of optimizing machine learning algorithms for short-term proxy metrics, the system substantially improves long-term performance in A/B tests. The paper offers insights into how our methods cope with attribution, coordination, and measurement challenges that usually hinder such long-term optimization. To contextualize these practical insights within a broader academic framework, we turn to reinforcement learning (RL). Using the language of RL, we formulate a comprehensive model of users' recurring relationships with a recommender system. Then, within this model, we identify our approach as a policy improvement update to a component of the existing recommender system, enhanced by tailored modeling of value functions and user-state representations. Illustrative offline experiments suggest this specialized modeling reduces data requirements by as much as a factor of 120,000 compared to black-box approaches.

Sijie Mai, Yu Zhao, Ying Zeng et al.

Multimodal sentiment analysis aims to effectively integrate information from various sources to infer sentiment, where in many cases there are no annotations for unimodal labels. Therefore, most works rely on multimodal labels for training. However, there exists the noisy label problem for the learning of unimodal signals as multimodal annotations are not always the ideal substitutes for the unimodal ones, failing to achieve finer optimization for individual modalities. In this paper, we explore the learning of unimodal labels under the weak supervision from the annotated multimodal labels. Specifically, we propose a novel meta uni-label generation (MUG) framework to address the above problem, which leverages the available multimodal labels to learn the corresponding unimodal labels by the meta uni-label correction network (MUCN). We first design a contrastive-based projection module to bridge the gap between unimodal and multimodal representations, so as to use multimodal annotations to guide the learning of MUCN. Afterwards, we propose unimodal and multimodal denoising tasks to train MUCN with explicit supervision via a bi-level optimization strategy. We then jointly train unimodal and multimodal learning tasks to extract discriminative unimodal features for multimodal inference. Experimental results suggest that MUG outperforms competitive baselines and can learn accurate unimodal labels.

Mingcai Chen, Yu Zhao, Zhonghuang Wang et al.

Immune repertoire classification, a typical multiple instance learning (MIL) problem, is a frontier research topic in computational biology that makes transformative contributions to new vaccines and immune therapies. However, the traditional instance-space MIL, directly assigning bag-level labels to instances, suffers from the massive amount of noisy labels and extremely low witness rate. In this work, we propose a noisy-label-learning formulation to solve the immune repertoire classification task. To remedy the inaccurate supervision of repertoire-level labels for a sequence-level classifier, we design a robust training strategy: The initial labels are smoothed to be asymmetric and are progressively corrected using the model's predictions throughout the training process. Furthermore, two models with the same architecture but different parameter initialization are co-trained simultaneously to remedy the known "confirmation bias" problem in the self-training-like schema. As a result, we obtain accurate sequence-level classification and, subsequently, repertoire-level classification. Experiments on the Cytomegalovirus (CMV) and Cancer datasets demonstrate our method's effectiveness and superior performance on sequence-level and repertoire-level tasks.

Yu Zhao, Ying Zhang, Xuhui Sui et al.

Knowledge-based Visual Question Answering (KBVQA) necessitates external knowledge incorporation beyond cross-modal understanding. Existing KBVQA methods either utilize implicit knowledge in multimodal large language models (MLLMs) via in-context learning or explicit knowledge via retrieval augmented generation. However, their reasoning processes remain implicit, without explicit multi-step trajectories from MLLMs. To address this gap, we provide a Hindsight Distilled Reasoning (HinD) framework with Knowledge Encouragement Preference Optimization (KEPO), designed to elicit and harness internal knowledge reasoning ability in MLLMs. First, to tackle the reasoning supervision problem, we propose to emphasize the hindsight wisdom of MLLM by prompting a frozen 7B-size MLLM to complete the reasoning process between the question and its ground truth answer, constructing Hindsight-Zero training data. Then we self-distill Hindsight-Zero into Chain-of-Thought (CoT) Generator and Knowledge Generator, enabling the generation of sequential steps and discrete facts. Secondly, to tackle the misalignment between knowledge correctness and confidence, we optimize the Knowledge Generator with KEPO, preferring under-confident but helpful knowledge over the over-confident but unhelpful one. The generated CoT and sampled knowledge are then exploited for answer prediction. Experiments on OK-VQA and A-OKVQA validate the effectiveness of HinD, showing that HinD with elicited reasoning from 7B-size MLLM achieves superior performance without commercial model APIs or outside knowledge.

Shaopeng Wei, Jun Wang, Yu Zhao et al.

Graph learning is a prevalent domain that endeavors to learn the intricate relationships among nodes and the topological structure of graphs. Over the years, graph learning has transcended from graph theory to graph data mining. With the advent of representation learning, it has attained remarkable performance in diverse scenarios. Owing to its extensive application prospects, graph learning attracts copious attention. While some researchers have accomplished impressive surveys on graph learning, they failed to connect related objectives, methods, and applications in a more coherent way. As a result, they did not encompass current ample scenarios and challenging problems due to the rapid expansion of graph learning. Particularly, large language models have recently had a disruptive effect on human life, but they also show relative weakness in structured scenarios. The question of how to make these models more powerful with graph learning remains open. Our survey focuses on the most recent advancements in integrating graph learning with pre-trained language models, specifically emphasizing their application within the domain of large language models. Different from previous surveys on graph learning, we provide a holistic review that analyzes current works from the perspective of graph structure, and discusses the latest applications, trends, and challenges in graph learning. Specifically, we commence by proposing a taxonomy and then summarize the methods employed in graph learning. We then provide a detailed elucidation of mainstream applications. Finally, we propose future directions.

Yuansong Zhu, Yu Zhao

Diffusion models have become a powerful family of deep generative models, with record-breaking performance in many applications. This paper first gives an overview and derivation of the basic theory of diffusion models, then reviews the research results of diffusion models in the field of natural language processing, from text generation, text-driven image generation and other four aspects, and analyzes and summarizes the relevant literature materials sorted out, and finally records the experience and feelings of this topic literature review research.

Andreas Grivas, Lorenzo Loconte, Emile van Krieken et al.

Multi-token prediction (MTP) is a prominent strategy to significantly speed up generation in large language models (LLMs), including byte-level LLMs, which are tokeniser-free but prohibitively slow. However, existing MTP methods often sacrifice expressiveness by assuming independence between future tokens. In this work, we investigate the trade-off between expressiveness and latency in MTP within the framework of probabilistic circuits (PCs). Our framework, named MTPC, allows one to explore different ways to encode the joint distributions over future tokens by selecting different circuit architectures, generalising classical models such as (hierarchical) mixture models, hidden Markov models and tensor networks. We show the efficacy of MTPC by retrofitting existing byte-level LLMs, such as EvaByte. Our experiments show that, when combined with speculative decoding, MTPC significantly speeds up generation compared to MTP with independence assumptions, while guaranteeing to retain the performance of the original verifier LLM. We also rigorously study the optimal trade-off between expressiveness and latency when exploring the possible parameterisations of MTPC, such as PC architectures and partial layer sharing between the verifier and draft LLMs.

Shaopeng Wei, Beni Egressy, Xingyan Chen et al.

Enterprise credit assessment is critical for evaluating financial risk, and Graph Neural Networks (GNNs), with their advanced capability to model inter-entity relationships, are a natural tool to get a deeper understanding of these financial networks. However, existing GNN-based methodologies predominantly emphasize entity-level attention mechanisms for contagion risk aggregation, often overlooking the heterogeneous importance of different feature dimensions, thus falling short in adequately modeling credit risk levels. To address this issue, we propose a novel architecture named Graph Dimension Attention Network (GDAN), which incorporates a dimension-level attention mechanism to capture fine-grained risk-related characteristics. Furthermore, we explore the interpretability of the GNN-based method in financial scenarios and propose a simple but effective data-centric explainer for GDAN, called GDAN-DistShift. DistShift provides edge-level interpretability by quantifying distribution shifts during the message-passing process. Moreover, we collected a real-world, multi-source Enterprise Credit Assessment Dataset (ECAD) and have made it accessible to the research community since high-quality datasets are lacking in this field. Extensive experiments conducted on ECAD demonstrate the effectiveness of our methods. In addition, we ran GDAN on the well-known datasets SMEsD and DBLP, also with excellent results.

Huaming Du, Xingyan Chen, Yu Zhao et al.

Enterprise financial risk analysis aims at predicting the future financial risk of enterprises. Due to its wide and significant application, enterprise financial risk analysis has always been the core research topic in the fields of Finance and Management. Based on advanced computer science and artificial intelligence technologies, enterprise risk analysis research is experiencing rapid developments and making significant progress. Therefore, it is both necessary and challenging to comprehensively review the relevant studies. Although there are already some valuable and impressive surveys on enterprise risk analysis from the perspective of Finance and Management, these surveys introduce approaches in a relatively isolated way and lack recent advances in enterprise financial risk analysis. In contrast, this paper attempts to provide a systematic literature survey of enterprise risk analysis approaches from Big Data perspective, which reviews more than 250 representative articles in the past almost 50 years (from 1968 to 2023). To the best of our knowledge, this is the first and only survey work on enterprise financial risk from Big Data perspective. Specifically, this survey connects and systematizes the existing enterprise financial risk studies, i.e. to summarize and interpret the problems, methods, and spotlights in a comprehensive way. In particular, we first introduce the issues of enterprise financial risks in terms of their types,granularity, intelligence, and evaluation metrics, and summarize the corresponding representative works. Then, we compare the analysis methods used to learn enterprise financial risk, and finally summarize the spotlights of the most representative works. Our goal is to clarify current cutting-edge research and its possible future directions to model enterprise risk, aiming to fully understand the mechanisms of enterprise risk generation and contagion.

Jiahuan Yan, Zhouyang Hong, Yu Zhao et al.

Imitation based robot learning has recently gained significant attention in the robotics field due to its theoretical potential for transferability and generalizability. However, it remains notoriously costly, both in terms of hardware and data collection, and deploying it in real-world environments demands meticulous setup of robots and precise experimental conditions. In this paper, we present a low-cost robot learning framework that is both easily reproducible and transferable to various robots and environments. We demonstrate that deployable imitation learning can be successfully applied even to industrial-grade robots, not just expensive collaborative robotic arms. Furthermore, our results show that multi-task robot learning is achievable with simple network architectures and fewer demonstrations than previously thought necessary. As the current evaluating method is almost subjective when it comes to real-world manipulation tasks, we propose Voting Positive Rate (VPR) - a novel evaluation strategy that provides a more objective assessment of performance. We conduct an extensive comparison of success rates across various self-designed tasks to validate our approach. To foster collaboration and support the robot learning community, we have open-sourced all relevant datasets and model checkpoints, available at huggingface.co/ZhiChengAI.

Che Chen, Lanhua Li, Shimin Gong et al.

In this paper, we employ multiple UAVs to accelerate data transmissions from ground users (GUs) to a remote base station (BS) via the UAVs' relay communications. The UAVs' intermittent information exchanges typically result in delays in acquiring the complete system state and hinder their effective collaboration. To maximize the overall throughput, we first propose a delay-tolerant multi-agent deep reinforcement learning (MADRL) algorithm that integrates a delay-penalized reward to encourage information sharing among UAVs, while jointly optimizing the UAVs' trajectory planning, network formation, and transmission control strategies. Additionally, considering information loss due to unreliable channel conditions, we further propose a spatio-temporal attention based prediction approach to recover the lost information and enhance each UAV's awareness of the network state. These two designs are envisioned to enhance the network capacity in UAV-assisted wireless networks with limited communications. The simulation results reveal that our new approach achieves over 50\% reduction in information delay and 75% throughput gain compared to the conventional MADRL. Interestingly, it is shown that improving the UAVs' information sharing will not sacrifice the network capacity. Instead, it significantly improves the learning performance and throughput simultaneously. It is also effective in reducing the need for UAVs' information exchange and thus fostering practical deployment of MADRL in UAV-assisted wireless networks.

Xin Ge, Ke Wang, Jiayi Wang et al.

This paper describes the joint submission of Alibaba and Soochow University, TSMind, to the WMT 2022 Shared Task on Translation Suggestion (TS). We participate in the English-German and English-Chinese tasks. Basically, we utilize the model paradigm fine-tuning on the downstream tasks based on large-scale pre-trained models, which has recently achieved great success. We choose FAIR's WMT19 English-German news translation system and MBART50 for English-Chinese as our pre-trained models. Considering the task's condition of limited use of training data, we follow the data augmentation strategies proposed by WeTS to boost our TS model performance. The difference is that we further involve the dual conditional cross-entropy model and GPT-2 language model to filter augmented data. The leader board finally shows that our submissions are ranked first in three of four language directions in the Naive TS task of the WMT22 Translation Suggestion task.

Yu Zhao, Fan Jiang, Tianle Liu et al.

Recent advancements in Large Reasoning Models (LRMs), exemplified by DeepSeek-R1, have underscored the potential of scaling inference-time compute through Group Relative Policy Optimization (GRPO). However, GRPO frequently suffers from gradient signal attenuation when encountering problems that are either too trivial or overly complex. In these scenarios, the disappearance of inter-group advantages makes the gradient signal susceptible to noise, thereby jeopardizing convergence stability. While variants like DAPO attempt to rectify gradient vanishing, they do not alleviate the substantial computational overhead incurred by exhaustive rollouts on low-utility samples. In this paper, we propose Difficulty-Estimated Policy Optimization (DEPO), a novel framework designed to optimize the efficiency and robustness of reasoning alignment. DEPO integrates an online Difficulty Estimator that dynamically assesses and filters training data before the rollout phase. This mechanism ensures that computational resources are prioritized for samples with high learning potential. Empirical results demonstrate that DEPO achieves up to a 2x reduction in rollout costs without compromising model performance. Our approach significantly lowers the computational barrier for training high-performance reasoning models, offering a more sustainable path for reasoning scaling. Code and data will be released upon acceptance.

Yu Zhao, Xinshuo Hu, Yunxin Li et al.

Dialog systems have achieved significant progress and have been widely used in various scenarios. The previous researches mainly focused on designing dialog generation models in a single scenario, while comprehensive abilities are required to handle tasks under various scenarios in the real world. In this paper, we propose a general Multi-Skill Dialog Framework, namely MSDF, which can be applied in different dialog tasks (e.g. knowledge grounded dialog and persona based dialog). Specifically, we propose a transferable response generator pre-trained on diverse large-scale dialog corpora as the backbone of MSDF, consisting of BERT-based encoders and a GPT-based decoder. To select the response consistent with dialog history, we propose a consistency selector trained through negative sampling. Moreover, the flexible copy mechanism of external knowledge is also employed to enhance the utilization of multiform knowledge in various scenarios. We conduct experiments on knowledge grounded dialog, recommendation dialog, and persona based dialog tasks. The experimental results indicate that our MSDF outperforms the baseline models with a large margin. In the Multi-skill Dialog of 2021 Language and Intelligence Challenge, our general MSDF won the 3rd prize, which proves our MSDF is effective and competitive.

Guining Cao, Jiaxin Peng, Chu Zeng et al.

Current reinforcement learning(RL) methods are broadly applicable and powerful in verifiable settings where scalar rewards can be provided. However, in open-ended generation tasks, verifying the correctness of responses remains challenging, and training reward models incurs substantial computational and annotation costs. Moreover, reinforcement learning (RLVR) often leads to diversity collapse and produces stereotypical or rigid outputs, outcomes that are particularly undesirable in open-domain scenarios. We propose Pairwise Preference Reward and Group-based Diversity Enhancement (PPR-GDE), a RL method that is more suitable for open-ended generation. PPR-GDE does not require scalar rewards and incorporates group-level diversity into the reward signal, it preserves the comparative structure of subjective evaluation through a pairwise preference reward, mitigates judge position bias via repeated comparisons with swapped response order, and introduces a group-based diversity reward that explicitly encourages semantic dispersion within a response group, all of these reward signals are integrated into a unified group-relative policy optimization objective. We instantiate PPR-GDE on role-playing task, experiments show that PPR-GDE achieves a better alignment quality as well as expressive diversity than strong RL baselines. Further analysis shows that pairwise preference is critical for preference alignment in subjective perspective, while the diversity metric plays an essential role in achieving superior expressive diversity and broader semantic coverage.

Huifeng Yin, Yu Zhao, Minghao Wu et al.

Large Reasoning Models(LRMs) such as OpenAI o1 and DeepSeek-R1 have shown remarkable reasoning capabilities by scaling test-time compute and generating long Chain-of-Thought(CoT). Distillation--post-training on LRMs-generated data--is a straightforward yet effective method to enhance the reasoning abilities of smaller models, but faces a critical bottleneck: we found that distilled long CoT data poses learning difficulty for small models and leads to the inheritance of biases (i.e. over-thinking) when using Supervised Fine-tuning (SFT) and Reinforcement Learning (RL) methods. To alleviate this bottleneck, we propose constructing tree-based CoT data from scratch via Monte Carlo Tree Search(MCTS). We then exploit a set of CoT-aware approaches, including Thoughts Length Balance, Fine-grained DPO, and Joint Post-training Objective, to enhance SFT and RL on the constructed data. We conduct evaluation on various benchmarks such as math (GSM8K, MATH, AIME). instruction-following (Multi-IF) and planning (Blocksworld), results demonstrate our approaches substantially improve the reasoning performance of distilled models compared to standard distilled models via reducing the hallucinations in long-time thinking. The project homepage is https://github.com/AIDC-AI/Marco-o1.

Xiang Li, Yiqun Yao, Xin Jiang et al.

Large language models (LLMs) have showcased profound capabilities in language understanding and generation, facilitating a wide array of applications. However, there is a notable paucity of detailed, open-sourced methodologies on efficiently scaling LLMs beyond 50 billion parameters with minimum trial-and-error cost and computational resources. In this report, we introduce Tele-FLM (aka FLM-2), a 52B open-sourced multilingual large language model that features a stable, efficient pre-training paradigm and enhanced factual judgment capabilities. Tele-FLM demonstrates superior multilingual language modeling abilities, measured by BPB on textual corpus. Besides, in both English and Chinese foundation model evaluation, it is comparable to strong open-sourced models that involve larger pre-training FLOPs, such as Llama2-70B and DeepSeek-67B. In addition to the model weights, we share the core designs, engineering practices, and training details, which we expect to benefit both the academic and industrial communities.