Hui Su

Meituan LongCat Team, Anchun Gui, Bei Li et al.

We introduce LongCat-Flash-Thinking-2601, a 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model with superior agentic reasoning capability. LongCat-Flash-Thinking-2601 achieves state-of-the-art performance among open-source models on a wide range of agentic benchmarks, including agentic search, agentic tool use, and tool-integrated reasoning. Beyond benchmark performance, the model demonstrates strong generalization to complex tool interactions and robust behavior under noisy real-world environments. Its advanced capability stems from a unified training framework that combines domain-parallel expert training with subsequent fusion, together with an end-to-end co-design of data construction, environments, algorithms, and infrastructure spanning from pre-training to post-training. In particular, the model's strong generalization capability in complex tool-use are driven by our in-depth exploration of environment scaling and principled task construction. To optimize long-tailed, skewed generation and multi-turn agentic interactions, and to enable stable training across over 10,000 environments spanning more than 20 domains, we systematically extend our asynchronous reinforcement learning framework, DORA, for stable and efficient large-scale multi-environment training. Furthermore, recognizing that real-world tasks are inherently noisy, we conduct a systematic analysis and decomposition of real-world noise patterns, and design targeted training procedures to explicitly incorporate such imperfections into the training process, resulting in improved robustness for real-world applications. To further enhance performance on complex reasoning tasks, we introduce a Heavy Thinking mode that enables effective test-time scaling by jointly expanding reasoning depth and width through intensive parallel thinking.

Hui Su, Yue Ye, Zhiwei Chen et al.

Weakly supervised object localization is a challenging task which aims to localize objects with coarse annotations such as image categories. Existing deep network approaches are mainly based on class activation map, which focuses on highlighting discriminative local region while ignoring the full object. In addition, the emerging transformer-based techniques constantly put a lot of emphasis on the backdrop that impedes the ability to identify complete objects. To address these issues, we present a re-attention mechanism termed token refinement transformer (TRT) that captures the object-level semantics to guide the localization well. Specifically, TRT introduces a novel module named token priority scoring module (TPSM) to suppress the effects of background noise while focusing on the target object. Then, we incorporate the class activation map as the semantically aware input to restrain the attention map to the target object. Extensive experiments on two benchmarks showcase the superiority of our proposed method against existing methods with image category annotations. Source code is available in \url{https://github.com/su-hui-zz/ReAttentionTransformer}.

Meituan LongCat Team, Bin Xiao, Chao Wang et al.

The prevailing Next-Token Prediction (NTP) paradigm has driven the success of large language models through discrete autoregressive modeling. However, contemporary multimodal systems remain language-centric, often treating non-linguistic modalities as external attachments, leading to fragmented architectures and suboptimal integration. To transcend this limitation, we introduce Discrete Native Autoregressive (DiNA), a unified framework that represents multimodal information within a shared discrete space, enabling a consistent and principled autoregressive modeling across modalities. A key innovation is the Discrete Native Any-resolution Visual Transformer (dNaViT), which performs tokenization and de-tokenization at arbitrary resolutions, transforming continuous visual signals into hierarchical discrete tokens. Building on this foundation, we develop LongCat-Next, a native multimodal model that processes text, vision, and audio under a single autoregressive objective with minimal modality-specific design. As an industrial-strength foundation model, it excels at seeing, painting, and talking within a single framework, achieving strong performance across a wide range of multimodal benchmarks. In particular, LongCat-Next addresses the long-standing performance ceiling of discrete vision modeling on understanding tasks and provides a unified approach to effectively reconcile the conflict between understanding and generation. As an attempt toward native multimodality, we open-source the LongCat-Next and its tokenizers, hoping to foster further research and development in the community. GitHub: https://github.com/meituan-longcat/LongCat-Next

Hui Su, Yue Ye, Wei Hua et al.

Semantic segmentation based on sparse annotation has advanced in recent years. It labels only part of each object in the image, leaving the remainder unlabeled. Most of the existing approaches are time-consuming and often necessitate a multi-stage training strategy. In this work, we propose a simple yet effective sparse annotated semantic segmentation framework based on segformer, dubbed SASFormer, that achieves remarkable performance. Specifically, the framework first generates hierarchical patch attention maps, which are then multiplied by the network predictions to produce correlated regions separated by valid labels. Besides, we also introduce the affinity loss to ensure consistency between the features of correlation results and network predictions. Extensive experiments showcase that our proposed approach is superior to existing methods and achieves cutting-edge performance. The source code is available at \url{https://github.com/su-hui-zz/SASFormer}.

Hyoshin Park, Justice Darko, Niharika Deshpande et al.

We introduce temporal multimodal multivariate learning, a new family of decision making models that can indirectly learn and transfer online information from simultaneous observations of a probability distribution with more than one peak or more than one outcome variable from one time stage to another. We approximate the posterior by sequentially removing additional uncertainties across different variables and time, based on data-physics driven correlation, to address a broader class of challenging time-dependent decision-making problems under uncertainty. Extensive experiments on real-world datasets ( i.e., urban traffic data and hurricane ensemble forecasting data) demonstrate the superior performance of the proposed targeted decision-making over the state-of-the-art baseline prediction methods across various settings.

Hui Su, Xiao Zhou, Houjin Yu et al.

Large Language Models pre-trained with self-supervised learning have demonstrated impressive zero-shot generalization capabilities on a wide spectrum of tasks. In this work, we present WeLM: a well-read pre-trained language model for Chinese that is able to seamlessly perform different types of tasks with zero or few-shot demonstrations. WeLM is trained with 10B parameters by "reading" a curated high-quality corpus covering a wide range of topics. We show that WeLM is equipped with broad knowledge on various domains and languages. On 18 monolingual (Chinese) tasks, WeLM can significantly outperform existing pre-trained models with similar sizes and match the performance of models up to 25 times larger. WeLM also exhibits strong capabilities in multi-lingual and code-switching understanding, outperforming existing multilingual language models pre-trained on 30 languages. Furthermore, We collected human-written prompts for a large set of supervised datasets in Chinese and fine-tuned WeLM with multi-prompted training. The resulting model can attain strong generalization on unseen types of tasks and outperform the unsupervised WeLM in zero-shot learning. Finally, we demonstrate that WeLM has basic skills at explaining and calibrating the decisions from itself, which can be promising directions for future research. Our models can be applied from https://welm.weixin.qq.com/docs/api/.

Yi-Kai Zhang, Zhiyuan Yao, Hongyan Hao et al.

Policy gradient methods rely on a baseline to measure the relative advantage of an action, ensuring the model reinforces behaviors that outperform its current average capability. In the training of Large Language Models (LLMs) using Actor-Critic methods (e.g., PPO), this baseline is typically estimated by a Value Model (Critic) often as large as the policy model itself. However, as the policy continuously evolves, the value model requires expensive, synchronous incremental training to accurately track the shifting capabilities of the policy. To avoid this overhead, Group Relative Policy Optimization (GRPO) eliminates the coupled value model by using the average reward of a group of rollouts as the baseline; yet, this approach necessitates extensive sampling to maintain estimation stability. In this paper, we propose $V_0$, a Generalist Value Model capable of estimating the expected performance of any model on unseen prompts without requiring parameter updates. We reframe value estimation by treating the policy's dynamic capability as an explicit context input; specifically, we leverage a history of instruction-performance pairs to dynamically profile the model, departing from the traditional paradigm that relies on parameter fitting to perceive capability shifts. Focusing on value estimation at State Zero (i.e., the initial prompt, hence $V_0$), our model serves as a critical resource scheduler. During GRPO training, $V_0$ predicts success rates prior to rollout, allowing for efficient sampling budget allocation; during deployment, it functions as a router, dispatching instructions to the most cost-effective and suitable model. Empirical results demonstrate that $V_0$ significantly outperforms heuristic budget allocation and achieves a Pareto-optimal trade-off between performance and cost in LLM routing tasks.

Wentao Shi, Yu Wang, Yuyang Zhao et al.

As reinforcement learning continues to scale the training of large language model-based agents, reliably verifying agent behaviors in complex environments has become increasingly challenging. Existing approaches rely on rule-based verifiers or LLM-as-a-Judge models, which struggle to generalize beyond narrow domains. Agent-as-a-Judge addresses this limitation by actively interacting with environments and tools to acquire verifiable evidence, yet its capabilities remain underexplored. We introduce a benchmark AJ-Bench to systematically evaluate Agent-as-a-Judge across three domains-search, data systems, and graphical user interfaces-comprising 155 tasks and 516 annotated trajectories. The benchmark comprehensively assesses judge agents' abilities in information acquisition, state verification, and process verification. Experiments demonstrate consistent performance gains over LLM-as-a-Judge baselines, while also revealing substantial open challenges in agent-based verification. Our data and code are available at https://aj-bench.github.io/.

Xinghao Chen, Zhijing Sun, Wenjin Guo et al.

Large Language Models (LLMs) excel in reasoning tasks through Chain-of-Thought (CoT) prompting. However, CoT prompting greatly increases computational demands, which has prompted growing interest in distilling CoT capabilities into Small Language Models (SLMs). This study systematically examines the factors influencing CoT distillation, including the choice of granularity, format and teacher model. Through experiments involving four teacher models and seven student models across seven mathematical and commonsense reasoning datasets, we uncover three key findings: (1) Unlike LLMs, SLMs exhibit a non-monotonic relationship with granularity, with stronger models benefiting from finer-grained reasoning and weaker models performing better with simpler CoT supervision; (2) CoT format significantly impacts LLMs but has minimal effect on SLMs, likely due to their reliance on supervised fine-tuning rather than pretraining preferences; (3) Stronger teacher models do NOT always produce better student models, as diversity and complexity in CoT supervision can outweigh accuracy alone. These findings emphasize the need to tailor CoT strategies to specific student model, offering actionable insights for optimizing CoT distillation in SLMs. The code and datasets are available at https://github.com/EIT-NLP/Distilling-CoT-Reasoning.

Meituan LongCat Team, Bayan, Bei Li et al.

We introduce LongCat-Flash, a 560-billion-parameter Mixture-of-Experts (MoE) language model designed for both computational efficiency and advanced agentic capabilities. Stemming from the need for scalable efficiency, LongCat-Flash adopts two novel designs: (a) Zero-computation Experts, which enables dynamic computational budget allocation and activates 18.6B-31.3B (27B on average) per token depending on contextual demands, optimizing resource usage. (b) Shortcut-connected MoE, which enlarges the computation-communication overlap window, demonstrating notable gains in inference efficiency and throughput compared to models of a comparable scale. We develop a comprehensive scaling framework for large models that combines hyperparameter transfer, model-growth initialization, a multi-pronged stability suite, and deterministic computation to achieve stable and reproducible training. Notably, leveraging the synergy among scalable architectural design and infrastructure efforts, we complete model training on more than 20 trillion tokens within 30 days, while achieving over 100 tokens per second (TPS) for inference at a cost of \$0.70 per million output tokens. To cultivate LongCat-Flash towards agentic intelligence, we conduct a large-scale pre-training on optimized mixtures, followed by targeted mid- and post-training on reasoning, code, and instructions, with further augmentation from synthetic data and tool use tasks. Comprehensive evaluations demonstrate that, as a non-thinking foundation model, LongCat-Flash delivers highly competitive performance among other leading models, with exceptional strengths in agentic tasks. The model checkpoint of LongCat-Flash is open-sourced to foster community research. LongCat Chat: https://longcat.ai Hugging Face: https://huggingface.co/meituan-longcat GitHub: https://github.com/meituan-longcat

Yaorui Shi, Shugui Liu, Yu Yang et al.

Long-horizon agentic reasoning necessitates effectively compressing growing interaction histories into a limited context window. Most existing memory systems serialize history as text, where token-level cost is uniform and scales linearly with length, often spending scarce budget on low-value details. To this end, we introduce MemOCR, a multimodal memory agent that improves long-horizon reasoning under tight context budgets by allocating memory space with adaptive information density through visual layout. Concretely, MemOCR maintains a structured rich-text memory (e.g., headings, highlights) and renders it into an image that the agent consults for memory access, visually prioritizing crucial evidence while aggressively compressing auxiliary details. To ensure robustness across varying memory budgets, we train MemOCR with reinforcement learning under budget-aware objectives that expose the agent to diverse compression levels. Across long-context multi-hop and single-hop question-answering benchmarks, MemOCR outperforms strong text-based baselines and achieves more effective context utilization under extreme budgets.

Junyan Lin, Haoran Chen, Yue Fan et al.

Multimodal Large Language Models (MLLMs) have made significant advancements in recent years, with visual features playing an increasingly critical role in enhancing model performance. However, the integration of multi-layer visual features in MLLMs remains underexplored, particularly with regard to optimal layer selection and fusion strategies. Existing methods often rely on arbitrary design choices, leading to suboptimal outcomes. In this paper, we systematically investigate two core aspects of multi-layer visual feature fusion: (1) selecting the most effective visual layers and (2) identifying the best fusion approach with the language model. Our experiments reveal that while combining visual features from multiple stages improves generalization, incorporating additional features from the same stage typically leads to diminished performance. Furthermore, we find that direct fusion of multi-layer visual features at the input stage consistently yields superior and more stable performance across various configurations. We make all our code publicly available: https://github.com/EIT-NLP/Layer_Select_Fuse_for_MLLM.

Anhao Zhao, Fanghua Ye, Yingqi Fan et al.

Large language models (LLMs) achieve remarkable performance across tasks but incur substantial computational costs due to their deep, multi-layered architectures. Layer pruning has emerged as a strategy to alleviate these inefficiencies, but conventional static pruning methods overlook two critical dynamics inherent to LLM inference: (1) horizontal dynamics, where token-level heterogeneity demands context-aware pruning decisions, and (2) vertical dynamics, where the distinct functional roles of MLP and self-attention layers necessitate component-specific pruning policies. We introduce SkipGPT, a dynamic layer pruning framework designed to optimize computational resource allocation through two core innovations: (1) global token-aware routing to prioritize critical tokens, and (2) decoupled pruning policies for MLP and self-attention components. To mitigate training instability, we propose a two-stage optimization paradigm: first, a disentangled training phase that learns routing strategies via soft parameterization to avoid premature pruning decisions, followed by parameter-efficient LoRA fine-tuning to restore performance impacted by layer removal. Extensive experiments demonstrate that SkipGPT reduces over 40% of model parameters while matching or exceeding the performance of the original dense model across benchmarks. By harmonizing dynamic efficiency with preserved expressivity, SkipGPT advances the practical deployment of scalable, resource-aware LLMs. Our code is publicly available at: https://github.com/EIT-NLP/SkipGPT.

Junlong Tong, Jinlan Fu, Zixuan Lin et al.

Large Language Models (LLMs) are primarily designed for batch processing. Existing methods for adapting LLMs to streaming rely either on expensive re-encoding or specialized architectures with limited scalability. This work identifies three key mismatches in adapting batch-oriented LLMs to streaming: (1) input-attention, (2) output-attention, and (3) position-ID mismatches. While it is commonly assumed that the latter two mismatches require frequent re-encoding, our analysis reveals that only the input-attention mismatch significantly impacts performance, indicating re-encoding outputs is largely unnecessary. To better understand this discrepancy with the common assumption, we provide the first comprehensive analysis of the impact of position encoding on LLMs in streaming, showing that preserving relative positions within source and target contexts is more critical than maintaining absolute order. Motivated by the above analysis, we introduce a group position encoding paradigm built on batch architectures to enhance consistency between streaming and batch modes. Extensive experiments on cross-lingual and cross-modal tasks demonstrate that our method outperforms existing approaches. Our method requires no architectural modifications, exhibits strong generalization in both streaming and batch modes. The code is available at repository https://github.com/EIT-NLP/StreamingLLM.

Yingqi Fan, Anhao Zhao, Jinlan Fu et al.

Multimodal Large Language Models (MLLMs) have achieved strong performance across vision-language tasks, but suffer from significant computational overhead due to the quadratic growth of attention computations with the number of multimodal tokens. Though efforts have been made to prune tokens in MLLMs, \textit{they lack a fundamental understanding of how MLLMs process and fuse multimodal information.} Through systematic analysis, we uncover a \textbf{three-stage} cross-modal interaction process: (1) Shallow layers recognize task intent, with visual tokens acting as passive attention sinks; (2) Cross-modal fusion occurs abruptly in middle layers, driven by a few critical visual tokens; (3) Deep layers discard vision tokens, focusing solely on linguistic refinement. Based on these findings, we propose \emph{VisiPruner}, a training-free pruning framework that reduces up to 99\% of vision-related attention computations and 53.9\% of FLOPs on LLaVA-v1.5 7B. It significantly outperforms existing token pruning methods and generalizes across diverse MLLMs. Beyond pruning, our insights further provide actionable guidelines for training efficient MLLMs by aligning model architecture with its intrinsic layer-wise processing dynamics. Our code is available at: https://github.com/EIT-NLP/VisiPruner.

Meituan LongCat Team, Anchun Gui, Bei Li et al.

We present LongCat-Flash-Thinking, an efficient 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model. Its advanced capabilities are cultivated through a meticulously crafted training process, beginning with long Chain-of-Thought (CoT) data cold-start and culminating in large-scale Reinforcement Learning (RL). We first employ a well-designed cold-start training strategy, which significantly enhances the reasoning potential and equips the model with specialized skills in both formal and agentic reasoning. Then, a core innovation is our domain-parallel training scheme, which decouples optimization across distinct domains (e.g., STEM, Code, Agentic) and subsequently fuses the resulting expert models into a single, nearly Pareto-optimal model. This entire process is powered by our Dynamic ORchestration for Asynchronous rollout (DORA) system, a large-scale RL framework that delivers a greater than threefold training speedup over synchronous methods on tens of thousands of accelerators. As a result, LongCat-Flash-Thinking achieves state-of-the-art performance among open-source models on a suite of complex reasoning tasks. The model exhibits exceptional efficiency in agentic reasoning, reducing average token consumption by 64.5% (from 19, 653 to 6, 965) on AIME-25, without degrading task accuracy. We release LongCat-Flash-Thinking to promote further advances in reasoning systems and agentic AI research.

Jinluan Yang, Yuxin Liu, Zhengyu Chen et al.

Training tool-use agents typically relies on outcome-based filtering: Supervised Fine-Tuning (SFT) on successful trajectories and Reinforcement Learning (RL) on pass-rate-selected tasks. However, this paradigm ignores interaction dynamics: successful trajectories may lack error recovery or exhibit redundancy, while pass rates fail to distinguish structurally informative tasks from trivial ones. We propose \textbf{TopoCurate}, an interaction-aware framework that projects multi-trial rollouts from the same task into a unified semantic quotient topology. By merging equivalent action-observation states, this projection transforms scattered linear trajectories into a structured manifold that explicitly captures how tool invocations and environmental responses drive the divergence between effective strategies and failure modes. Leveraging this representation, we introduce a dual-selection mechanism: for SFT, we prioritize trajectories demonstrating reflective recovery, semantic efficiency, and strategic diversity to mitigate covariate shift and mode collapse; for RL, we select tasks with high error branch ratios and strategic heterogeneity, maximizing gradient Signal-to-Noise Ratio to address vanishing signals in sparse-reward settings. Evaluations on BFCLv3 and Tau2 Bench show that TopoCurate achieves consistent gains of 4.2\% (SFT) and 6.9\% (RL) over state-of-the-art baselines. We will release the code and data soon for further investigations.

Dunwei Tu, Hongyan Hao, Hansi Yang et al.

Training generalist agents capable of adapting to diverse scenarios requires interactive environments for self-exploration. However, interactive environments remain critically scarce, and existing synthesis methods suffer from significant limitations regarding environmental diversity and scalability. To address these challenges, we introduce ScaleEnv, a framework that constructs fully interactive environments and verifiable tasks entirely from scratch. Specifically, ScaleEnv ensures environment reliability through procedural testing, and guarantees task completeness and solvability via tool dependency graph expansion and executable action verification. By enabling agents to learn through exploration within ScaleEnv, we demonstrate significant performance improvements on unseen, multi-turn tool-use benchmarks such as $τ^2$-Bench and VitaBench, highlighting strong generalization capabilities. Furthermore, we investigate the relationship between increasing number of domains and model generalization performance, providing empirical evidence that scaling environmental diversity is critical for robust agent learning.

Jiarui Fang, Zilin Zhu, Shenggui Li et al.

The pre-trained model (PTM) is revolutionizing Artificial Intelligence (AI) technology. However, the hardware requirement of PTM training is prohibitively high, making it a game for a small proportion of people. Therefore, we proposed PatrickStar system to lower the hardware requirements of PTMs and make them accessible to everyone. PatrickStar uses the CPU-GPU heterogeneous memory space to store the model data. Different from existing works, we organize the model data in memory chunks and dynamically distribute them in the heterogeneous memory. Guided by the runtime memory statistics collected in a warm-up iteration, chunks are orchestrated efficiently in heterogeneous memory and generate lower CPU-GPU data transmission volume and higher bandwidth utilization. Symbiosis with the Zero Redundancy Optimizer, PatrickStar scales to multiple GPUs on multiple nodes. % using data parallelism. The system can train tasks on bigger models and larger batch sizes, which cannot be accomplished by existing works. Experimental results show that PatrickStar extends model scales 2.27 and 2.5 times of DeepSpeed, and consistently exhibits significantly higher execution speed. PatricStar also successfully runs the 175B GPT3 training task on a 32 GPU cluster. Our code is publicly available at https://github.com/Tencent/PatrickStar.

Xiangyang Mou, Chenghao Yang, Mo Yu et al.

Recent advancements in open-domain question answering (ODQA), i.e., finding answers from large open-domain corpus like Wikipedia, have led to human-level performance on many datasets. However, progress in QA over book stories (Book QA) lags behind despite its similar task formulation to ODQA. This work provides a comprehensive and quantitative analysis about the difficulty of Book QA: (1) We benchmark the research on the NarrativeQA dataset with extensive experiments with cutting-edge ODQA techniques. This quantifies the challenges Book QA poses, as well as advances the published state-of-the-art with a $\sim$7\% absolute improvement on Rouge-L. (2) We further analyze the detailed challenges in Book QA through human studies.\footnote{\url{https://github.com/gorov/BookQA}.} Our findings indicate that the event-centric questions dominate this task, which exemplifies the inability of existing QA models to handle event-oriented scenarios.

Hui Su, Zhi Tian, Xiaoyu Shen et al.

Scaling law principles indicate a power-law correlation between loss and variables such as model size, dataset size, and computational resources utilized during training. These principles play a vital role in optimizing various aspects of model pre-training, ultimately contributing to the success of large language models such as GPT-4, Llama and Gemini. However, the original scaling law paper by OpenAI did not disclose the complete details necessary to derive the precise scaling law formulas, and their conclusions are only based on models containing up to 1.5 billion parameters. Though some subsequent works attempt to unveil these details and scale to larger models, they often neglect the training dependency of important factors such as the learning rate, context length and batch size, leading to their failure to establish a reliable formula for predicting the test loss trajectory. In this technical report, we confirm that the scaling law formulations proposed in the original OpenAI paper remain valid when scaling the model size up to 33 billion, but the constant coefficients in these formulas vary significantly with the experiment setup. We meticulously identify influential factors and provide transparent, step-by-step instructions to estimate all constant terms in scaling-law formulas by training on models with only 1M~60M parameters. Using these estimated formulas, we showcase the capability to accurately predict various attributes for models with up to 33B parameters before their training, including (1) the minimum possible test loss; (2) the minimum required training steps and processed tokens to achieve a specific loss; (3) the critical batch size with an optimal time/computation trade-off at any loss value; and (4) the complete test loss trajectory with arbitrary batch size.

Zhiyuan Yao, Yi-Kai Zhang, Yuxin Chen et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a key approach for enhancing LLM reasoning. However, standard frameworks like Group Relative Policy Optimization (GRPO) typically employ a uniform rollout budget, leading to resource inefficiency. Moreover, existing adaptive methods often rely on instance-level metrics, such as task pass rates, failing to capture the model's dynamic learning state. To address these limitations, we propose CoBA-RL, a reinforcement learning algorithm designed to adaptively allocate rollout budgets based on the model's evolving capability. Specifically, CoBA-RL utilizes a Capability-Oriented Value function to map tasks to their potential training gains and employs a heap-based greedy strategy to efficiently self-calibrate the distribution of computational resources to samples with high training value. Extensive experiments demonstrate that our approach effectively orchestrates the trade-off between exploration and exploitation, delivering consistent generalization improvements across multiple challenging benchmarks. These findings underscore that quantifying sample training value and optimizing budget allocation are pivotal for advancing LLM post-training efficiency.

Haoran Chen, Junyan Lin, Xinghao Chen et al.

Multimodal large language models (MLLMs) typically extract visual features from the final layers of a pretrained Vision Transformer (ViT). This widespread deep-layer bias, however, is largely driven by empirical convention rather than principled analysis. While prior studies suggest that different ViT layers capture different types of information, with shallower layers focusing on fine visual details and deeper layers aligning more closely with textual semantics, the impact of this variation on MLLM performance remains underexplored. We present the first comprehensive study of visual layer selection for MLLMs, analyzing representation similarity across ViT layers to establish shallow, middle, and deep layer groupings. Through extensive evaluation of MLLMs (1.4B-7B parameters) across 10 benchmarks encompassing 60+ tasks, we find that while deep layers excel in semantic-rich tasks like OCR, shallow and middle layers significantly outperform them on fine-grained visual tasks including counting, positioning, and object localization. Building on these insights, we propose a lightweight feature fusion method that strategically incorporates shallower layers, achieving consistent improvements over both single-layer and specialized fusion baselines. Our work offers the first principled study of visual layer selection in MLLMs, showing that MLLMs can often see better when they look shallower.

Kuai Dai, Xutao Li, Junying Fang et al.

Convection (thunderstorm) develops rapidly within hours and is highly destructive, posing a significant challenge for nowcasting and resulting in substantial losses to nature and society. After the emergence of artificial intelligence (AI)-based methods, convection nowcasting has experienced rapid advancements, with its performance surpassing that of physics-based numerical weather prediction and other conventional approaches. However, the lead time and coverage of it still leave much to be desired and hardly meet the needs of disaster emergency response. Here, we propose deep diffusion models of satellite (DDMS) to establish an AI-based convection nowcasting system. Specifically, DDMS employs diffusion processes to effectively simulate complicated spatiotemporal evolution patterns of convective clouds, significantly improving the forecast lead time. Additionally, it combines geostationary satellite brightness temperature data and domain knowledge from meteorological experts, thereby achieving planetary-scale forecast coverage. During long-term tests and objective validation based on the FengYun-4A satellite, our system achieves, for the first time, effective convection nowcasting up to 4 hours, with broad coverage (about 20,000,000 km$^2$), remarkable accuracy, and high resolution (15 minutes; 4 km). Its performance reaches a new height in convection nowcasting compared to the existing models. In terms of application, our system is highly transferable with the potential to collaborate with multiple satellites for global convection nowcasting. Furthermore, our results highlight the remarkable capabilities of diffusion models in convective clouds forecasting, as well as the significant value of geostationary satellite data when empowered by AI technologies.

Ruipeng Wang, Yuxin Chen, Yukai Wang et al.

Recent advances in large language models have enabled LLM-based agents to achieve strong performance on a variety of benchmarks. However, their performance in real-world deployments often that observed on benchmark settings, especially in complex and imperfect environments. This discrepancy largely arises because prevailing training and evaluation paradigms are typically built on idealized assumptions, overlooking the inherent stochasticity and noise present in real-world interactions. To bridge this gap, we introduce AgentNoiseBench, a framework for systematically evaluating the robustness of agentic models under noisy environments. We first conduct an in-depth analysis of biases and uncertainties in real-world scenarios and categorize environmental noise into two primary types: user-noise and tool-noise. Building on this analysis, we develop an automated pipeline that injects controllable noise into existing agent-centric benchmarks while preserving task solvability. Leveraging this pipeline, we perform extensive evaluations across a wide range of models with diverse architectures and parameter scales. Our results reveal consistent performance variations under different noise conditions, highlighting the sensitivity of current agentic models to realistic environmental perturbations.

Jiafeng Liang, Shixin Jiang, Xuan Dong et al.

Large Multimodal Models (LMMs) have recently demonstrated impressive performance on general video comprehension benchmarks. Nevertheless, for broader applications, the robustness of their temporal analysis capability needs to be thoroughly investigated yet predominantly ignored. Motivated by this, we propose a novel temporal robustness benchmark (TemRobBench), which introduces temporal inconsistency perturbations separately at the visual and textual modalities to assess the robustness of models. We evaluate 16 mainstream LMMs and find that they exhibit over-reliance on prior knowledge and textual context in adversarial environments, while ignoring the actual temporal dynamics in the video. To mitigate this issue, we design panoramic direct preference optimization (PanoDPO), which encourages LMMs to incorporate both visual and linguistic feature preferences simultaneously. Experimental results show that PanoDPO can effectively enhance the model's robustness and reliability in temporal analysis.

Wei He, Yueqing Sun, Hongyan Hao et al.

As LLM-based agents are increasingly deployed in real-life scenarios, existing benchmarks fail to capture their inherent complexity of handling extensive information, leveraging diverse resources, and managing dynamic user interactions. To address this gap, we introduce VitaBench, a challenging benchmark that evaluates agents on versatile interactive tasks grounded in real-world settings. Drawing from daily applications in food delivery, in-store consumption, and online travel services, VitaBench presents agents with the most complex life-serving simulation environment to date, comprising 66 tools. Through a framework that eliminates domain-specific policies, we enable flexible composition of these scenarios and tools, yielding 100 cross-scenario tasks (main results) and 300 single-scenario tasks. Each task is derived from multiple real user requests and requires agents to reason across temporal and spatial dimensions, utilize complex tool sets, proactively clarify ambiguous instructions, and track shifting user intent throughout multi-turn conversations. Moreover, we propose a rubric-based sliding window evaluator, enabling robust assessment of diverse solution pathways in complex environments and stochastic interactions. Our comprehensive evaluation reveals that even the most advanced models achieve only 30% success rate on cross-scenario tasks, and less than 50% success rate on others. Overall, we believe VitaBench will serve as a valuable resource for advancing the development of AI agents in practical real-world applications. The code, dataset, and leaderboard are available at https://vitabench.github.io/

Alex Chandler, Devesh Surve, Hui Su

Accurate text summarization is one of the most common and important tasks performed by Large Language Models, where the costs of human review for an entire document may be high, but the costs of errors in summarization may be even greater. We propose Detecting Errors through Ensembling Prompts (DEEP) - an end-to-end large language model framework for detecting factual errors in text summarization. Our framework uses a diverse set of LLM prompts to identify factual inconsistencies, treating their outputs as binary features, which are then fed into ensembling models. We then calibrate the ensembled models to produce empirically accurate probabilities that a text is factually consistent or free of hallucination. We demonstrate that prior models for detecting factual errors in summaries perform significantly worse without optimizing the thresholds on subsets of the evaluated dataset. Our framework achieves state-of-the-art (SOTA) balanced accuracy on the AggreFact-XSUM FTSOTA, TofuEval Summary-Level, and HaluEval Summarization benchmarks in detecting factual errors within transformer-generated text summaries. It does so without any fine-tuning of the language model or reliance on thresholding techniques not available in practical settings.

Rongzhi Zhang, Yulong Gu, Xiaoyu Shen et al.

Recent research has achieved impressive progress in the session-based recommendation. However, information such as item knowledge and click time interval, which could be potentially utilized to improve the performance, remains largely unexploited. In this paper, we propose a framework called Knowledge-enhanced Session-based Recommendation with Temporal Transformer (KSTT) to incorporate such information when learning the item and session embeddings. Specifically, a knowledge graph, which models contexts among items within a session and their corresponding attributes, is proposed to obtain item embeddings through graph representation learning. We introduce time interval embedding to represent the time pattern between the item that needs to be predicted and historical click, and use it to replace the position embedding in the original transformer (called temporal transformer). The item embeddings in a session are passed through the temporal transformer network to get the session embedding, based on which the final recommendation is made. Extensive experiments demonstrate that our model outperforms state-of-the-art baselines on four benchmark datasets.

Xiangyang Mou, Mo Yu, Shiyu Chang et al.

This paper proposes a new problem of complementary evidence identification for open-domain question answering (QA). The problem aims to efficiently find a small set of passages that covers full evidence from multiple aspects as to answer a complex question. To this end, we proposes a method that learns vector representations of passages and models the sufficiency and diversity within the selected set, in addition to the relevance between the question and passages. Our experiments demonstrate that our method considers the dependence within the supporting evidence and significantly improves the accuracy of complementary evidence selection in QA domain.

Ernie Chang, Xiaoyu Shen, Dawei Zhu et al.

For many new application domains for data-to-text generation, the main obstacle in training neural models consists of a lack of training data. While usually large numbers of instances are available on the data side, often only very few text samples are available. To address this problem, we here propose a novel few-shot approach for this setting. Our approach automatically augments the data available for training by (i) generating new text samples based on replacing specific values by alternative ones from the same category, (ii) generating new text samples based on GPT-2, and (iii) proposing an automatic method for pairing the new text samples with data samples. As the text augmentation can introduce noise to the training data, we use cycle consistency as an objective, in order to make sure that a given data sample can be correctly reconstructed after having been formulated as text (and that text samples can be reconstructed from data). On both the E2E and WebNLG benchmarks, we show that this weakly supervised training paradigm is able to outperform fully supervised seq2seq models with less than 10% annotations. By utilizing all annotated data, our model can boost the performance of a standard seq2seq model by over 5 BLEU points, establishing a new state-of-the-art on both datasets.

Zaid Bin Tariq, Arun Iyengar, Lara Marcuse et al.

Electroencephalogram (EEG) is a prominent way to measure the brain activity for studying epilepsy, thereby helping in predicting seizures. Seizure prediction is an active research area with many deep learning based approaches dominating the recent literature for solving this problem. But these models require a considerable number of patient-specific seizures to be recorded for extracting the preictal and interictal EEG data for training a classifier. The increase in sensitivity and specificity for seizure prediction using the machine learning models is noteworthy. However, the need for a significant number of patient-specific seizures and periodic retraining of the model because of non-stationary EEG creates difficulties for designing practical device for a patient. To mitigate this process, we propose a Siamese neural network based seizure prediction method that takes a wavelet transformed EEG tensor as an input with convolutional neural network (CNN) as the base network for detecting change-points in EEG. Compared to the solutions in the literature, which utilize days of EEG recordings, our method only needs one seizure for training which translates to less than ten minutes of preictal and interictal data while still getting comparable results to models which utilize multiple seizures for seizure prediction.

Lisha Chen, Hui Su, Qiang Ji

Existing deep learning based facial landmark detection methods have achieved excellent performance. These methods, however, do not explicitly embed the structural dependencies among landmark points. They hence cannot preserve the geometric relationships between landmark points or generalize well to challenging conditions or unseen data. This paper proposes a method for deep structured facial landmark detection based on combining a deep Convolutional Network with a Conditional Random Field. We demonstrate its superior performance to existing state-of-the-art techniques in facial landmark detection, especially a better generalization ability on challenging datasets that include large pose and occlusion.

Xiangyang Mou, Brandyn Sigouin, Ian Steenstra et al.

Recently, a more challenging state tracking task, Audio-Video Scene-Aware Dialogue (AVSD), is catching an increasing amount of attention among researchers. Different from purely text-based dialogue state tracking, the dialogue in AVSD contains a sequence of question-answer pairs about a video and the final answer to the given question requires additional understanding of the video. This paper interprets the AVSD task from an open-domain Question Answering (QA) point of view and proposes a multimodal open-domain QA system to deal with the problem. The proposed QA system uses common encoder-decoder framework with multimodal fusion and attention. Teacher forcing is applied to train a natural language generator. We also propose a new data augmentation approach specifically under QA assumption. Our experiments show that our model and techniques bring significant improvements over the baseline model on the DSTC7-AVSD dataset and demonstrate the potentials of our data augmentation techniques.

Xiangyang Mou, Mo Yu, Bingsheng Yao et al.

A lot of progress has been made to improve question answering (QA) in recent years, but the special problem of QA over narrative book stories has not been explored in-depth. We formulate BookQA as an open-domain QA task given its similar dependency on evidence retrieval. We further investigate how state-of-the-art open-domain QA approaches can help BookQA. Besides achieving state-of-the-art on the NarrativeQA benchmark, our study also reveals the difficulty of evidence retrieval in books with a wealth of experiments and analysis - which necessitates future effort on novel solutions for evidence retrieval in BookQA.

Hui Su, Xiaoyu Shen, Sanqiang Zhao et al.

Neural network-based sequence-to-sequence (seq2seq) models strongly suffer from the low-diversity problem when it comes to open-domain dialogue generation. As bland and generic utterances usually dominate the frequency distribution in our daily chitchat, avoiding them to generate more interesting responses requires complex data filtering, sampling techniques or modifying the training objective. In this paper, we propose a new perspective to diversify dialogue generation by leveraging non-conversational text. Compared with bilateral conversations, non-conversational text are easier to obtain, more diverse and cover a much broader range of topics. We collect a large-scale non-conversational corpus from multi sources including forum comments, idioms and book snippets. We further present a training paradigm to effectively incorporate these text via iterative back translation. The resulting model is tested on two conversational datasets and is shown to produce significantly more diverse responses without sacrificing the relevance with context.

Xiaoyu Shen, Ernie Chang, Hui Su et al.

The neural attention model has achieved great success in data-to-text generation tasks. Though usually excelling at producing fluent text, it suffers from the problem of information missing, repetition and "hallucination". Due to the black-box nature of the neural attention architecture, avoiding these problems in a systematic way is non-trivial. To address this concern, we propose to explicitly segment target text into fragment units and align them with their data correspondences. The segmentation and correspondence are jointly learned as latent variables without any human annotations. We further impose a soft statistical constraint to regularize the segmental granularity. The resulting architecture maintains the same expressive power as neural attention models, while being able to generate fully interpretable outputs with several times less computational cost. On both E2E and WebNLG benchmarks, we show the proposed model consistently outperforms its neural attention counterparts.

Jaimie Drozdal, Justin Weisz, Dakuo Wang et al.

We explore trust in a relatively new area of data science: Automated Machine Learning (AutoML). In AutoML, AI methods are used to generate and optimize machine learning models by automatically engineering features, selecting models, and optimizing hyperparameters. In this paper, we seek to understand what kinds of information influence data scientists' trust in the models produced by AutoML? We operationalize trust as a willingness to deploy a model produced using automated methods. We report results from three studies -- qualitative interviews, a controlled experiment, and a card-sorting task -- to understand the information needs of data scientists for establishing trust in AutoML systems. We find that including transparency features in an AutoML tool increased user trust and understandability in the tool; and out of all proposed features, model performance metrics and visualizations are the most important information to data scientists when establishing their trust with an AutoML tool.

Xiaoyu Shen, Jun Suzuki, Kentaro Inui et al.

Many text generation tasks naturally contain two steps: content selection and surface realization. Current neural encoder-decoder models conflate both steps into a black-box architecture. As a result, the content to be described in the text cannot be explicitly controlled. This paper tackles this problem by decoupling content selection from the decoder. The decoupled content selection is human interpretable, whose value can be manually manipulated to control the content of generated text. The model can be trained end-to-end without human annotations by maximizing a lower bound of the marginal likelihood. We further propose an effective way to trade-off between performance and controllability with a single adjustable hyperparameter. In both data-to-text and headline generation tasks, our model achieves promising results, paving the way for controllable content selection in text generation.

Hui Su, Xiaoyu Shen, Rongzhi Zhang et al.

Recent research has made impressive progress in single-turn dialogue modelling. In the multi-turn setting, however, current models are still far from satisfactory. One major challenge is the frequently occurred coreference and information omission in our daily conversation, making it hard for machines to understand the real intention. In this paper, we propose rewriting the human utterance as a pre-process to help multi-turn dialgoue modelling. Each utterance is first rewritten to recover all coreferred and omitted information. The next processing steps are then performed based on the rewritten utterance. To properly train the utterance rewriter, we collect a new dataset with human annotations and introduce a Transformer-based utterance rewriting architecture using the pointer network. We show the proposed architecture achieves remarkably good performance on the utterance rewriting task. The trained utterance rewriter can be easily integrated into online chatbots and brings general improvement over different domains.

Mathew Peveler, Jeffery Kephart, Hui Su

We introduce Reagent, a technology that readily converts ordinary webpages containing structured data into software agents with which one can interact naturally, via a combination of speech and pointing. Previous efforts to make webpage content manipulable by third-party software components in browsers or desktop applications have generally relied upon specialized instrumentation included in the webpages -- a practice that neither scales well nor applies to pre-existing webpages. In contrast, Reagent automatically captures semantic details and semantically-meaningful mouse events from arbitrary webpages that contain no pre-existing special instrumentation. Reagent combines these events with text transcriptions of user speech to derive and execute parameterized commands representing human intent. Thus, users may request various visualization or analytic operations to be performed on data displayed on a page by speaking to it and/or pointing to elements within it. When unable to infer translations between event labels and human terminology, Reagent proactively asks users for definitions and adds them to its dictionary. We demonstrate Reagent in the context of a collection of pre-existing webpages that contain football team and player statistics.

Hui Su, Xiaoyu Shen, Wenjie Li et al.

Sequence-to-Sequence (seq2seq) models have become overwhelmingly popular in building end-to-end trainable dialogue systems. Though highly efficient in learning the backbone of human-computer communications, they suffer from the problem of strongly favoring short generic responses. In this paper, we argue that a good response should smoothly connect both the preceding dialogue history and the following conversations. We strengthen this connection through mutual information maximization. To sidestep the non-differentiability of discrete natural language tokens, we introduce an auxiliary continuous code space and map such code space to a learnable prior distribution for generation purpose. Experiments on two dialogue datasets validate the effectiveness of our model, where the generated responses are closely related to the dialogue context and lead to more interactive conversations.

Zhibing Zhao, Haoming Li, Junming Wang et al.

We propose a cost-effective framework for preference elicitation and aggregation under the Plackett-Luce model with features. Given a budget, our framework iteratively computes the most cost-effective elicitation questions in order to help the agents make a better group decision. We illustrate the viability of the framework with experiments on Amazon Mechanical Turk, which we use to estimate the cost of answering different types of elicitation questions. We compare the prediction accuracy of our framework when adopting various information criteria that evaluate the expected information gain from a question. Our experiments show carefully designed information criteria are much more efficient, i.e., they arrive at the correct answer using fewer queries, than randomly asking questions given the budget constraint.

Xiaoyu Shen, Hui Su, Shuzi Niu et al.

Variational encoder-decoders (VEDs) have shown promising results in dialogue generation. However, the latent variable distributions are usually approximated by a much simpler model than the powerful RNN structure used for encoding and decoding, yielding the KL-vanishing problem and inconsistent training objective. In this paper, we separate the training step into two phases: The first phase learns to autoencode discrete texts into continuous embeddings, from which the second phase learns to generalize latent representations by reconstructing the encoded embedding. In this case, latent variables are sampled by transforming Gaussian noise through multi-layer perceptrons and are trained with a separate VED model, which has the potential of realizing a much more flexible distribution. We compare our model with current popular models and the experiment demonstrates substantial improvement in both metric-based and human evaluations.

Yanran Li, Hui Su, Xiaoyu Shen et al.

We develop a high-quality multi-turn dialog dataset, DailyDialog, which is intriguing in several aspects. The language is human-written and less noisy. The dialogues in the dataset reflect our daily communication way and cover various topics about our daily life. We also manually label the developed dataset with communication intention and emotion information. Then, we evaluate existing approaches on DailyDialog dataset and hope it benefit the research field of dialog systems.

Matthew Peveler, Naveen Sundar Govindarajulu, Selmer Bringsjord et al.

As computational power has continued to increase, and sensors have become more accurate, the corresponding advent of systems that are at once cognitive and immersive has arrived. These \textit{cognitive and immersive systems} (CAISs) fall squarely into the intersection of AI with HCI/HRI: such systems interact with and assist the human agents that enter them, in no small part because such systems are infused with AI able to understand and reason about these humans and their knowledge, beliefs, goals, communications, plans, etc. We herein explain our approach to engineering CAISs. We emphasize the capacity of a CAIS to develop and reason over a `theory of the mind' of its human partners. This capacity entails that the AI in question has a sophisticated model of the beliefs, knowledge, goals, desires, emotions, etc.\ of these humans. To accomplish this engineering, a formal framework of very high expressivity is needed. In our case, this framework is a \textit{cognitive event calculus}, a particular kind of quantified multi-operator modal logic, and a matching high-expressivity automated reasoner and planner. To explain, advance, and to a degree validate our approach, we show that a calculus of this type satisfies a set of formal requirements, and can enable a CAIS to understand a psychologically tricky scenario couched in what we call the \textit{cognitive polysolid framework} (CPF). We also formally show that a room that satisfies these requirements can have a useful property we term \emph{expectation of usefulness}. CPF, a sub-class of \textit{cognitive microworlds}, includes machinery able to represent and plan over not merely blocks and actions (such as seen in the primitive `blocks worlds' of old), but also over agents and their mental attitudes about both other agents and inanimate objects.

Xiaoyu Shen, Hui Su, Yanran Li et al.

Deep latent variable models have been shown to facilitate the response generation for open-domain dialog systems. However, these latent variables are highly randomized, leading to uncontrollable generated responses. In this paper, we propose a framework allowing conditional response generation based on specific attributes. These attributes can be either manually assigned or automatically detected. Moreover, the dialog states for both speakers are modeled separately in order to reflect personal features. We validate this framework on two different scenarios, where the attribute refers to genericness and sentiment states respectively. The experiment result testified the potential of our model, where meaningful responses can be generated in accordance with the specified attributes.