Zihan Liu

Boxin Wang, Wei Ping, Peng Xu et al. · nvidia

Large decoder-only language models (LMs) can be largely improved in terms of perplexity by retrieval (e.g., RETRO), but its impact on text generation quality and downstream task accuracy is unclear. Thus, it is still an open question: shall we pretrain large autoregressive LMs with retrieval? To answer it, we perform a comprehensive study on a scalable pre-trained retrieval-augmented LM (i.e., RETRO) compared with standard GPT and retrieval-augmented GPT incorporated at fine-tuning or inference stages. We first provide the recipe to reproduce RETRO up to 9.5B parameters while retrieving a text corpus with 330B tokens. Based on that, we have the following novel findings: i) RETRO outperforms GPT on text generation with much less degeneration (i.e., repetition), moderately higher factual accuracy, and slightly lower toxicity with a nontoxic retrieval database. ii) On the LM Evaluation Harness benchmark, RETRO largely outperforms GPT on knowledge-intensive tasks, but is on par with GPT on other tasks. Furthermore, we introduce a simple variant of the model, RETRO++, which largely improves open-domain QA results of original RETRO (e.g., EM score +8.6 on Natural Question) and significantly outperforms retrieval-augmented GPT in both fine-tuning and zero-shot evaluation settings. Our findings highlight the promising direction of pretraining autoregressive LMs with retrieval as future foundation models. We release our code and model at: https://github.com/NVIDIA/Megatron-LM/blob/main/tools/retro/README.md

Yuqi Xu, Rizhen Hu, Zihan Liu et al.

Traditional Mixture-of-Experts (MoE) training typically proceeds without any structural priors, effectively requiring the model to simultaneously train expert weights while searching for an optimal routing policy within a vast combinatorial space. This entanglement often leads to sluggish convergence and training instabilities. This paper introduces Grouter, a preemptive routing method that by distilling high-quality structures from fully-trained MoE models and serving as a fixed router for target models. By decoupling structural optimization from weight updates, Grouter significantly accelerates both the speed and quality of model convergence. To ensure the framework's versatility, we also introduce expert folding to adapt Grouter across varying model configurations and expert tuning to rebalance workloads across different data distributions. Furthermore, by leveraging the structural priors provided by preemptive routing, we can implement targeted optimizations to further enhance training throughput. Experiments demonstrate that Grouter achieves superior performance and efficiency which boosts pre-training data utilization by 4.28x and achieves up to 33.5% throughput acceleration, establishing preemptive routing as a fundamental paradigm for scalable MoE training. We publicly release our code and pretrained Grouter checkpoints at https://github.com/JimmyAwoe/Grouter.

Wenliang Dai, Nayeon Lee, Boxin Wang et al. · nvidia

We introduce NVLM 1.0, a family of frontier-class multimodal large language models (LLMs) that achieve state-of-the-art results on vision-language tasks, rivaling the leading proprietary models (e.g., GPT-4o) and open-access models (e.g., Llama 3-V 405B and InternVL 2). Remarkably, NVLM 1.0 shows improved text-only performance over its LLM backbone after multimodal training. In terms of model design, we perform a comprehensive comparison between decoder-only multimodal LLMs (e.g., LLaVA) and cross-attention-based models (e.g., Flamingo). Based on the strengths and weaknesses of both approaches, we propose a novel architecture that enhances both training efficiency and multimodal reasoning capabilities. Furthermore, we introduce a 1-D tile-tagging design for tile-based dynamic high-resolution images, which significantly boosts performance on multimodal reasoning and OCR-related tasks. Regarding training data, we meticulously curate and provide detailed information on our multimodal pretraining and supervised fine-tuning datasets. Our findings indicate that dataset quality and task diversity are more important than scale, even during the pretraining phase, across all architectures. Notably, we develop production-grade multimodality for the NVLM-1.0 models, enabling them to excel in vision-language tasks while maintaining and even improving text-only performance compared to their LLM backbones. To achieve this, we craft and integrate a high-quality text-only dataset into multimodal training, alongside a substantial amount of multimodal math and reasoning data, leading to enhanced math and coding capabilities across modalities. To advance research in the field, we release the model weights at https://huggingface.co/nvidia/NVLM-D-72B and will open-source the training code for the community soon.

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Zhuolin Yang, Wei Ping, Zihan Liu et al. · stanford

Augmenting pretrained language models (LMs) with a vision encoder (e.g., Flamingo) has obtained the state-of-the-art results in image-to-text generation. However, these models store all the knowledge within their parameters, thus often requiring enormous model parameters to model the abundant visual concepts and very rich textual descriptions. Additionally, they are inefficient in incorporating new data, requiring a computational-expensive fine-tuning process. In this work, we introduce a Retrieval-augmented Visual Language Model, Re-ViLM, built upon the Flamingo, that supports retrieving the relevant knowledge from the external database for zero and in-context few-shot image-to-text generations. By storing certain knowledge explicitly in the external database, our approach reduces the number of model parameters and can easily accommodate new data during evaluation by simply updating the database. We also construct an interleaved image and text data that facilitates in-context few-shot learning capabilities. We demonstrate that Re-ViLM significantly boosts performance for image-to-text generation tasks, especially for zero-shot and few-shot generation in out-of-domain settings with 4 times less parameters compared with baseline methods.

Yue Yu, Wei Ping, Zihan Liu et al.

Large language models (LLMs) typically utilize the top-k contexts from a retriever in retrieval-augmented generation (RAG). In this work, we propose a novel instruction fine-tuning framework RankRAG, which instruction-tunes a single LLM for the dual purpose of context ranking and answer generation in RAG. In particular, the instruction-tuned LLMs work surprisingly well by adding a small fraction of ranking data into the training blend, and outperform existing expert ranking models, including the same LLM exclusively fine-tuned on a large amount of ranking data. For generation, we compare our model with many strong baselines, including GPT-4-0613, GPT-4-turbo-2024-0409, and ChatQA-1.5, an open-sourced model with the state-of-the-art performance on RAG benchmarks. Specifically, our Llama3-RankRAG significantly outperforms Llama3-ChatQA-1.5 and GPT-4 models on nine knowledge-intensive benchmarks. In addition, it also performs comparably to GPT-4 on five RAG benchmarks in the biomedical domain without instruction fine-tuning on biomedical data, demonstrating its superb capability for generalization to new domains.

Zihan Liu, Mostofa Patwary, Ryan Prenger et al.

Existing knowledge-grounded dialogue systems typically use finetuned versions of a pretrained language model (LM) and large-scale knowledge bases. These models typically fail to generalize on topics outside of the knowledge base, and require maintaining separate potentially large checkpoints each time finetuning is needed. In this paper, we aim to address these limitations by leveraging the inherent knowledge stored in the pretrained LM as well as its powerful generation ability. We propose a multi-stage prompting approach to generate knowledgeable responses from a single pretrained LM. We first prompt the LM to generate knowledge based on the dialogue context. Then, we further prompt it to generate responses based on the dialogue context and the previously generated knowledge. Results show that our knowledge generator outperforms the state-of-the-art retrieval-based model by 5.8% when combining knowledge relevance and correctness. In addition, our multi-stage prompting outperforms the finetuning-based dialogue model in terms of response knowledgeability and engagement by up to 10% and 5%, respectively. Furthermore, we scale our model up to 530 billion parameters and show that larger LMs improve the generation correctness score by up to 10%, and response relevance, knowledgeability and engagement by up to 10%. Our code is available at: https://github.com/NVIDIA/Megatron-LM.

Aditi, Niket Agarwal, Arslan Ali et al.

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI -- effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 https://openmdw.ai/license/1-1/ License at https://github.com/nvidia/cosmos}{github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3 . The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3 .

Cong Guo, Chen Zhang, Jingwen Leng et al. · microsoft-research

Quantization is a technique to reduce the computation and memory cost of DNN models, which are getting increasingly large. Existing quantization solutions use fixed-point integer or floating-point types, which have limited benefits, as both require more bits to maintain the accuracy of original models. On the other hand, variable-length quantization uses low-bit quantization for normal values and high-precision for a fraction of outlier values. Even though this line of work brings algorithmic benefits, it also introduces significant hardware overheads due to variable-length encoding and decoding. In this work, we propose a fixed-length adaptive numerical data type called ANT to achieve low-bit quantization with tiny hardware overheads. Our data type ANT leverages two key innovations to exploit the intra-tensor and inter-tensor adaptive opportunities in DNN models. First, we propose a particular data type, flint, that combines the advantages of float and int for adapting to the importance of different values within a tensor. Second, we propose an adaptive framework that selects the best type for each tensor according to its distribution characteristics. We design a unified processing element architecture for ANT and show its ease of integration with existing DNN accelerators. Our design results in 2.8$\times$ speedup and 2.5$\times$ energy efficiency improvement over the state-of-the-art quantization accelerators.

Samuel Cahyawijaya, Tiezheng Yu, Zihan Liu et al.

Self-supervised pre-training methods have brought remarkable breakthroughs in the understanding of text, image, and speech. Recent developments in genomics has also adopted these pre-training methods for genome understanding. However, they focus only on understanding haploid sequences, which hinders their applicability towards understanding genetic variations, also known as single nucleotide polymorphisms (SNPs), which is crucial for genome-wide association study. In this paper, we introduce SNP2Vec, a scalable self-supervised pre-training approach for understanding SNP. We apply SNP2Vec to perform long-sequence genomics modeling, and we evaluate the effectiveness of our approach on predicting Alzheimer's disease risk in a Chinese cohort. Our approach significantly outperforms existing polygenic risk score methods and all other baselines, including the model that is trained entirely with haploid sequences. We release our code and dataset on https://github.com/HLTCHKUST/snp2vec.

Peng Xu, Wei Ping, Xianchao Wu et al.

In this work, we introduce ChatQA 2, an Llama 3.0-based model with a 128K context window, designed to bridge the gap between open-source LLMs and leading proprietary models (e.g., GPT-4-Turbo-2024-04-09) in long context understanding and retrieval-augmented generation (RAG) capabilities. These two capabilities are complementary to each other and essential for LLMs to process large volumes of information that cannot fit into a single prompt. We present a detailed continued training recipe to extend the context window of Llama3-70B-base from 8K to 128K tokens, along with a three-stage instruction tuning process to enhance the model's instruction-following, RAG performance, and long-context understanding capabilities. Our results demonstrate that the Llama3-ChatQA-2-70B model outperforms most existing state-of-the-art models, including GPT-4-Turbo-2024-04-09, Qwen2-72B-Instruct, and Llama3.1-70B-Instruct, on ultra-long tasks beyond 100K tokens, as well as on the RAG benchmark using only a 4K context window, showing the strong long context capability across varying sequence lengths. We further provide extensive comparisons between direct long-context and RAG solutions using the same state-of-the-art long-context LLMs. Interestingly, we find that the performance of strong long-context LLMs using RAG improves when retrieving a larger number of chunks. With a large set of top-k chunks, RAG consistently outperforms direct long-context solution using the same state-of-the-art long-context models (e.g., Llama3-ChatQA-2-70B and Qwen2-72B-Instruct) on both 32K and 128K benchmarks. We open-source the model weights, training data, and the evaluation setup for the for the community: https://chatqa2-project.github.io/

Mengnan Qi, Yufan Huang, Maoquan Wang et al. · microsoft-research

Automatic Program translation has enormous application value and hence has been attracting significant interest from AI researchers. However, we observe that current program translation models still make elementary syntax errors, particularly, when the target language does not have syntax elements in the source language. Metrics like BLUE, CodeBLUE and computation accuracy may not expose these issues. In this paper we introduce a new metrics for programming language translation and these metrics address these basic syntax errors. We develop a novel active defects probing suite called Syntactic Unit Tests (SUT) which includes a highly interpretable evaluation harness for accuracy and test scoring. Experiments have shown that even powerful models like ChatGPT still make mistakes on these basic unit tests. Specifically, compared to previous program translation task evaluation dataset, its pass rate on our unit tests has decreased by 26.15%. Further our evaluation harness reveal syntactic element errors in which these models exhibit deficiencies.

Peng Xu, Wei Ping, Xianchao Wu et al.

Extending the context window of large language models (LLMs) is getting popular recently, while the solution of augmenting LLMs with retrieval has existed for years. The natural questions are: i) Retrieval-augmentation versus long context window, which one is better for downstream tasks? ii) Can both methods be combined to get the best of both worlds? In this work, we answer these questions by studying both solutions using two state-of-the-art pretrained LLMs, i.e., a proprietary 43B GPT and Llama2-70B. Perhaps surprisingly, we find that LLM with 4K context window using simple retrieval-augmentation at generation can achieve comparable performance to finetuned LLM with 16K context window via positional interpolation on long context tasks, while taking much less computation. More importantly, we demonstrate that retrieval can significantly improve the performance of LLMs regardless of their extended context window sizes. Our best model, retrieval-augmented Llama2-70B with 32K context window, outperforms GPT-3.5-turbo-16k and Davinci003 in terms of average score on nine long context tasks including question answering, query-based summarization, and in-context few-shot learning tasks. It also outperforms its non-retrieval Llama2-70B-32k baseline by a margin, while being much faster at generation. Our study provides general insights on the choice of retrieval-augmentation versus long context extension of LLM for practitioners.

Wenliang Dai, Zihan Liu, Ziwei Ji et al. · nvidia

Large-scale vision-language pre-trained (VLP) models are prone to hallucinate non-existent visual objects when generating text based on visual information. In this paper, we systematically study the object hallucination problem from three aspects. First, we examine recent state-of-the-art VLP models, showing that they still hallucinate frequently, and models achieving better scores on standard metrics (e.g., CIDEr) could be more unfaithful. Second, we investigate how different types of image encoding in VLP influence hallucination, including region-based, grid-based, and patch-based. Surprisingly, we find that patch-based features perform the best and smaller patch resolution yields a non-trivial reduction in object hallucination. Third, we decouple various VLP objectives and demonstrate that token-level image-text alignment and controlled generation are crucial to reducing hallucination. Based on that, we propose a simple yet effective VLP loss named ObjMLM to further mitigate object hallucination. Results show that it reduces object hallucination by up to 17.4% when tested on two benchmarks (COCO Caption for in-domain and NoCaps for out-of-domain evaluation).

Zhuolin Yang, Zihan Liu, Yang Chen et al. · nvidia

We introduce Nemotron-Cascade 2, an open 30B MoE model with 3B activated parameters that delivers best-in-class reasoning and strong agentic capabilities. Despite its compact size, its mathematical and coding reasoning performance approaches that of frontier open models. It is the second open-weight LLM, after DeepSeekV3.2-Speciale-671B-A37B, to achieve Gold Medal-level performance in the 2025 International Mathematical Olympiad (IMO), the International Olympiad in Informatics (IOI), and the ICPC World Finals, demonstrating remarkably high intelligence density with 20x fewer parameters. In contrast to Nemotron-Cascade 1, the key technical advancements are as follows. After SFT on a meticulously curated dataset, we substantially expand Cascade RL to cover a much broader spectrum of reasoning and agentic domains. Furthermore, we introduce multi-domain on-policy distillation from the strongest intermediate teacher models for each domain throughout the Cascade RL process, allowing us to efficiently recover benchmark regressions and sustain strong performance gains along the way. We release the collection of model checkpoint and training data.

Junchen Ye, Zihan Liu, Bowen Du et al.

Recent studies have shown great promise in applying graph neural networks for multivariate time series forecasting, where the interactions of time series are described as a graph structure and the variables are represented as the graph nodes. Along this line, existing methods usually assume that the graph structure (or the adjacency matrix), which determines the aggregation manner of graph neural network, is fixed either by definition or self-learning. However, the interactions of variables can be dynamic and evolutionary in real-world scenarios. Furthermore, the interactions of time series are quite different if they are observed at different time scales. To equip the graph neural network with a flexible and practical graph structure, in this paper, we investigate how to model the evolutionary and multi-scale interactions of time series. In particular, we first provide a hierarchical graph structure cooperated with the dilated convolution to capture the scale-specific correlations among time series. Then, a series of adjacency matrices are constructed under a recurrent manner to represent the evolving correlations at each layer. Moreover, a unified neural network is provided to integrate the components above to get the final prediction. In this way, we can capture the pair-wise correlations and temporal dependency simultaneously. Finally, experiments on both single-step and multi-step forecasting tasks demonstrate the superiority of our method over the state-of-the-art approaches.

Wanshuang Gou, Zihan Liu

Large language model-driven multi-agent systems enhance the reliability of complex reasoning tasks through multi-round deliberation, role specialization, and cross-validation. However, existing multi-agent debate and collaboration frameworks typically adopt fully connected communication, causing the number of messages, token costs, and end-to-end latency to grow approximately quadratically with the number of agents; although fixed sparse topologies reduce overhead, they cannot adapt communication relationships to different task instances or intermediate reasoning states, making them prone either to preserving low-value interactions or to losing critical error-correction information. To address this problem, this paper proposes DySCo (Dynamic Sparse Consensus), a dynamic trust-aware sparse consensus mechanism. In each round of reasoning, DySCo estimates the value of communication edges based on agent reliability, answer divergence, and task relevance, and selects a small number of high-value edges for message exchange under budget constraints; it then aggregates the answers of different agents through dynamic trust weights and terminates the discussion early once consensus stabilizes. This mechanism replaces universal broadcasting with on-demand communication, thereby reducing communication overhead while preserving essential cross-validation information. We further present analyses of communication complexity and consensus stability, and evaluate the performance of DySCo on mathematical reasoning, logical reasoning, and factual question-answering tasks.

Zexin Wang, Changhua Pei, Yuanhao Liu et al.

As the reasoning capabilities of Large Language Models (LLMs) continue to advance, LLM-based agent systems offer advantages in flexibility and interpretability over traditional systems, garnering increasing attention. However, despite the widespread research interest and industrial application of agent systems, these systems, like their traditional counterparts, frequently encounter anomalies. These anomalies lead to instability and insecurity, hindering their further development. Therefore, a comprehensive and systematic approach to the operation and maintenance of agent systems is urgently needed. Unfortunately, current research on the operations of agent systems is sparse. To address this gap, we have undertaken a survey on agent system operations with the aim of establishing a clear framework for the field, defining the challenges, and facilitating further development. Specifically, this paper begins by systematically defining anomalies within agent systems, categorizing them into intra-agent anomalies and inter-agent anomalies. Next, we introduce a novel and comprehensive operational framework for agent systems, dubbed Agent System Operations (AgentOps). We provide detailed definitions and explanations of its four key stages: monitoring, anomaly detection, root cause localization, and resolution.

Yun Luo, Zihan Liu, Stan Z. Li et al.

(Dis)agreement detection aims to identify the authors' attitudes or positions (\textit{agree, disagree, neutral}) towards a specific text. It is limited for existing methods merely using textual information for identifying (dis)agreements, especially for cross-domain settings. Social relation information can play an assistant role in the (dis)agreement task besides textual information. We propose a novel method to extract such relation information from (dis)agreement data into an inductive social relation graph, merely using the comment-reply pairs without any additional platform-specific information. The inductive social relation globally considers the historical discussion and the relation between authors. Textual information based on a pre-trained language model and social relation information encoded by pre-trained RGCN are jointly considered for (dis)agreement detection. Experimental results show that our model achieves state-of-the-art performance for both the in-domain and cross-domain tasks on the benchmark -- DEBAGREEMENT. We find social relations can boost the performance of the (dis)agreement detection model, especially for the long-token comment-reply pairs, demonstrating the effectiveness of the social relation graph. We also explore the effect of the knowledge graph embedding methods, the information fusing method, and the time interval in constructing the social relation graph, which shows the effectiveness of our model.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We present Nemotron 3 Nano 30B-A3B, a Mixture-of-Experts hybrid Mamba-Transformer language model. Nemotron 3 Nano was pretrained on 25 trillion text tokens, including more than 3 trillion new unique tokens over Nemotron 2, followed by supervised fine tuning and large-scale RL on diverse environments. Nemotron 3 Nano achieves better accuracy than our previous generation Nemotron 2 Nano while activating less than half of the parameters per forward pass. It achieves up to 3.3x higher inference throughput than similarly-sized open models like GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, while also being more accurate on popular benchmarks. Nemotron 3 Nano demonstrates enhanced agentic, reasoning, and chat abilities and supports context lengths up to 1M tokens. We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

Samuel Cahyawijaya, Holy Lovenia, Willy Chung et al.

Speech emotion recognition plays a crucial role in human-computer interactions. However, most speech emotion recognition research is biased toward English-speaking adults, which hinders its applicability to other demographic groups in different languages and age groups. In this work, we analyze the transferability of emotion recognition across three different languages--English, Mandarin Chinese, and Cantonese; and 2 different age groups--adults and the elderly. To conduct the experiment, we develop an English-Mandarin speech emotion benchmark for adults and the elderly, BiMotion, and a Cantonese speech emotion dataset, YueMotion. This study concludes that different language and age groups require specific speech features, thus making cross-lingual inference an unsuitable method. However, cross-group data augmentation is still beneficial to regularize the model, with linguistic distance being a significant influence on cross-lingual transferability. We release publicly release our code at https://github.com/HLTCHKUST/elderly_ser.

Zihan Liu, Zewei Sun, Shanbo Cheng et al. · bytedance

Document-level Neural Machine Translation (DocNMT) has been proven crucial for handling discourse phenomena by introducing document-level context information. One of the most important directions is to input the whole document directly to the standard Transformer model. In this case, efficiency becomes a critical concern due to the quadratic complexity of the attention module. Existing studies either focus on the encoder part, which cannot be deployed on sequence-to-sequence generation tasks, e.g., Machine Translation (MT), or suffer from a significant performance drop. In this work, we keep the translation performance while gaining 20\% speed up by introducing extra selection layer based on lightweight attention that selects a small portion of tokens to be attended. It takes advantage of the original attention to ensure performance and dimension reduction to accelerate inference. Experimental results show that our method could achieve up to 95\% sparsity (only 5\% tokens attended) approximately, and save 93\% computation cost on the attention module compared with the original Transformer, while maintaining the performance.

Jiale Xu, Rui Zhang, Yi Xiong et al.

Large Language Models are increasingly being deployed in datacenters. Serving these models requires careful memory management, as their memory usage includes static weights, dynamic activations, and key-value caches. While static weights are constant and predictable, dynamic components such as activations and KV caches change frequently during runtime, presenting significant challenges for efficient memory management. Modern LLM serving systems typically handle runtime memory and KV caches at distinct abstraction levels: runtime memory management relies on static tensor abstractions, whereas KV caches utilize a page table-based virtualization layer built on top of the tensor abstraction. This virtualization dynamically manages KV caches to mitigate memory fragmentation. However, this dual-level approach fundamentally isolates runtime memory and KV cache management, resulting in suboptimal memory utilization under dynamic workloads, which can lead to a nearly 20% drop in throughput. To address these limitations, we propose eLLM, an elastic memory management framework inspired by the classical memory ballooning mechanism in operating systems. The core components of eLLM include: (1) Virtual Tensor Abstraction, which decouples the virtual address space of tensors from the physical GPU memory, creating a unified and flexible memory pool; (2) an Elastic Memory Mechanism that dynamically adjusts memory allocation through runtime memory inflation and deflation, leveraging CPU memory as an extensible buffer; and (3) a Lightweight Scheduling Strategy employing SLO-aware policies to optimize memory utilization and effectively balance performance trade-offs under stringent SLO constraints. Comprehensive evaluations demonstrate that eLLM significantly outperforms state-of-the-art systems, 2.32x higher decoding throughput, and supporting 3x larger batch sizes for 128K-token inputs.

Zihan Liu, Yun Luo, Lirong Wu et al.

It has become cognitive inertia to employ cross-entropy loss function in classification related tasks. In the untargeted attacks on graph structure, the gradients derived from the attack objective are the attacker's basis for evaluating a perturbation scheme. Previous methods use negative cross-entropy loss as the attack objective in attacking node-level classification models. However, the suitability of the cross-entropy function for constructing the untargeted attack objective has yet been discussed in previous works. This paper argues about the previous unreasonable attack objective from the perspective of budget allocation. We demonstrate theoretically and empirically that negative cross-entropy tends to produce more significant gradients from nodes with lower confidence in the labeled classes, even if the predicted classes of these nodes have been misled. To free up these inefficient attack budgets, we propose a simple attack model for untargeted attacks on graph structure based on a novel attack objective which generates unweighted gradients on graph structures that are not affected by the node confidence. By conducting experiments in gray-box poisoning attack scenarios, we demonstrate that a reasonable budget allocation can significantly improve the effectiveness of gradient-based edge perturbations without any extra hyper-parameter.

Zihan Liu, Jiaqi Wang, Yun Luo et al.

In recent years, there has been an explosion of research on the application of deep learning to the prediction of various peptide properties, due to the significant development and market potential of peptides. Molecular dynamics has enabled the efficient collection of large peptide datasets, providing reliable training data for deep learning. However, the lack of systematic analysis of the peptide encoding, which is essential for AI-assisted peptide-related tasks, makes it an urgent problem to be solved for the improvement of prediction accuracy. To address this issue, we first collect a high-quality, colossal simulation dataset of peptide self-assembly containing over 62,000 samples generated by coarse-grained molecular dynamics (CGMD). Then, we systematically investigate the effect of peptide encoding of amino acids into sequences and molecular graphs using state-of-the-art sequential (i.e., RNN, LSTM, and Transformer) and structural deep learning models (i.e., GCN, GAT, and GraphSAGE), on the accuracy of peptide self-assembly prediction, an essential physiochemical process prior to any peptide-related applications. Extensive benchmarking studies have proven Transformer to be the most powerful sequence-encoding-based deep learning model, pushing the limit of peptide self-assembly prediction to decapeptides. In summary, this work provides a comprehensive benchmark analysis of peptide encoding with advanced deep learning models, serving as a guide for a wide range of peptide-related predictions such as isoelectric points, hydration free energy, etc.

Ziwei Ji, Zihan Liu, Nayeon Lee et al.

Dialogue systems can leverage large pre-trained language models and knowledge to generate fluent and informative responses. However, these models are still prone to produce hallucinated responses not supported by the input source, which greatly hinders their application. The heterogeneity between external knowledge and dialogue context challenges representation learning and source integration, and further contributes to unfaithfulness. To handle this challenge and generate more faithful responses, this paper presents RHO ($ρ$) utilizing the representations of linked entities and relation predicates from a knowledge graph (KG). We propose (1) local knowledge grounding to combine textual embeddings with the corresponding KG embeddings; and (2) global knowledge grounding to equip RHO with multi-hop reasoning abilities via the attention mechanism. In addition, we devise a response re-ranking technique based on walks over KG sub-graphs for better conversational reasoning. Experimental results on OpenDialKG show that our approach significantly outperforms state-of-the-art methods on both automatic and human evaluation by a large margin, especially in hallucination reduction (17.54% in FeQA).

Zihan Liu, Ge Wang, Jiaqi Wang et al.

Peptides are formed by the dehydration condensation of multiple amino acids. The primary structure of a peptide can be represented either as an amino acid sequence or as a molecular graph consisting of atoms and chemical bonds. Previous studies have indicated that deep learning routes specific to sequential and graphical peptide forms exhibit comparable performance on downstream tasks. Despite the fact that these models learn representations of the same modality of peptides, we find that they explain their predictions differently. Considering sequential and graphical models as two experts making inferences from different perspectives, we work on fusing expert knowledge to enrich the learned representations for improving the discriminative performance. To achieve this, we propose a peptide co-modeling method, RepCon, which employs a contrastive learning-based framework to enhance the mutual information of representations from decoupled sequential and graphical end-to-end models. It considers representations from the sequential encoder and the graphical encoder for the same peptide sample as a positive pair and learns to enhance the consistency of representations between positive sample pairs and to repel representations between negative pairs. Empirical studies of RepCon and other co-modeling methods are conducted on open-source discriminative datasets, including aggregation propensity, retention time, antimicrobial peptide prediction, and family classification from Peptide Database. Our results demonstrate the superiority of the co-modeling approach over independent modeling, as well as the superiority of RepCon over other methods under the co-modeling framework. In addition, the attribution on RepCon further corroborates the validity of the approach at the level of model explanation.

Zihan Liu, Yun Luo, Lirong Wu et al.

Graph edge perturbations are dedicated to damaging the prediction of graph neural networks by modifying the graph structure. Previous gray-box attackers employ gradients from the surrogate model to locate the vulnerable edges to perturb the graph structure. However, unreliability exists in gradients on graph structures, which is rarely studied by previous works. In this paper, we discuss and analyze the errors caused by the unreliability of the structural gradients. These errors arise from rough gradient usage due to the discreteness of the graph structure and from the unreliability in the meta-gradient on the graph structure. In order to address these problems, we propose a novel attack model with methods to reduce the errors inside the structural gradients. We propose edge discrete sampling to select the edge perturbations associated with hierarchical candidate selection to ensure computational efficiency. In addition, semantic invariance and momentum gradient ensemble are proposed to address the gradient fluctuation on semantic-augmented graphs and the instability of the surrogate model. Experiments are conducted in untargeted gray-box poisoning scenarios and demonstrate the improvement in the performance of our approach.

Jiale Xu, Rui Zhang, Cong Guo et al.

Large Language Models (LLMs) are widely used across various domains, processing millions of daily requests. This surge in demand poses significant challenges in optimizing throughput and latency while keeping costs manageable. The Key-Value (KV) cache, a standard method for retaining previous computations, makes LLM inference highly bounded by memory. While batching strategies can enhance performance, they frequently lead to significant memory fragmentation. Even though cutting-edge systems like vLLM mitigate KV cache fragmentation using paged Attention mechanisms, they still suffer from inefficient memory and computational operations due to the tightly coupled page management and computation kernels. This study introduces the vTensor, an innovative tensor structure for LLM inference based on GPU virtual memory management (VMM). vTensor addresses existing limitations by decoupling computation from memory defragmentation and offering dynamic extensibility. Our framework employs a CPU-GPU heterogeneous approach, ensuring efficient, fragmentation-free memory management while accommodating various computation kernels across different LLM architectures. Experimental results indicate that vTensor achieves an average speedup of 1.86x across different models, with up to 2.42x in multi-turn chat scenarios. Additionally, vTensor provides average speedups of 2.12x and 3.15x in kernel evaluation, reaching up to 3.92x and 3.27x compared to SGLang Triton prefix-prefilling kernels and vLLM paged Attention kernel, respectively. Furthermore, it frees approximately 71.25% (57GB) of memory on the NVIDIA A100 GPU compared to vLLM, enabling more memory-intensive workloads.

Yufei Huang, Siyuan Li, Jin Su et al.

Protein structure-based property prediction has emerged as a promising approach for various biological tasks, such as protein function prediction and sub-cellular location estimation. The existing methods highly rely on experimental protein structure data and fail in scenarios where these data are unavailable. Predicted protein structures from AI tools (e.g., AlphaFold2) were utilized as alternatives. However, we observed that current practices, which simply employ accurately predicted structures during inference, suffer from notable degradation in prediction accuracy. While similar phenomena have been extensively studied in general fields (e.g., Computer Vision) as model robustness, their impact on protein property prediction remains unexplored. In this paper, we first investigate the reason behind the performance decrease when utilizing predicted structures, attributing it to the structure embedding bias from the perspective of structure representation learning. To study this problem, we identify a Protein 3D Graph Structure Learning Problem for Robust Protein Property Prediction (PGSL-RP3), collect benchmark datasets, and present a protein Structure embedding Alignment Optimization framework (SAO) to mitigate the problem of structure embedding bias between the predicted and experimental protein structures. Extensive experiments have shown that our framework is model-agnostic and effective in improving the property prediction of both predicted structures and experimental structures. The benchmark datasets and codes will be released to benefit the community.

Zihan Liu, Ge Wang, Yun Luo et al.

Recent research has revealed that Graph Neural Networks (GNNs) are susceptible to adversarial attacks targeting the graph structure. A malicious attacker can manipulate a limited number of edges, given the training labels, to impair the victim model's performance. Previous empirical studies indicate that gradient-based attackers tend to add edges rather than remove them. In this paper, we present a theoretical demonstration revealing that attackers tend to increase inter-class edges due to the message passing mechanism of GNNs, which explains some previous empirical observations. By connecting dissimilar nodes, attackers can more effectively corrupt node features, making such attacks more advantageous. However, we demonstrate that the inherent smoothness of GNN's message passing tends to blur node dissimilarity in the feature space, leading to the loss of crucial information during the forward process. To address this issue, we propose a novel surrogate model with multi-level propagation that preserves the node dissimilarity information. This model parallelizes the propagation of unaggregated raw features and multi-hop aggregated features, while introducing batch normalization to enhance the dissimilarity in node representations and counteract the smoothness resulting from topological aggregation. Our experiments show significant improvement with our approach.Furthermore, both theoretical and experimental evidence suggest that adding inter-class edges constitutes an easily observable attack pattern. We propose an innovative attack loss that balances attack effectiveness and imperceptibility, sacrificing some attack effectiveness to attain greater imperceptibility. We also provide experiments to validate the compromise performance achieved through this attack loss.

PAN Team, Qiyue Gao, Kun Zhou et al.

World models (WMs) are intended to serve as internal simulators of the real world that enable agents to understand, anticipate, and act upon complex environments. Existing WM benchmarks remain narrowly focused on next-state prediction and visual fidelity, overlooking the richer simulation capabilities required for intelligent behavior. To address this gap, we introduce WR-Arena, a comprehensive benchmark for evaluating WMs along three fundamental dimensions of next world simulation: (i) Action Simulation Fidelity, the ability to interpret and follow semantically meaningful, multi-step instructions and generate diverse counterfactual rollouts; (ii) Long-horizon Forecast, the ability to sustain accurate, coherent, and physically plausible simulations across extended interactions; and (iii) Simulative Reasoning and Planning, the ability to support goal-directed reasoning by simulating, comparing, and selecting among alternative futures in both structured and open-ended environments. We build a task taxonomy and curate diverse datasets designed to probe these capabilities, moving beyond single-turn and perceptual evaluations. Through extensive experiments with state-of-the-art WMs, our results expose a substantial gap between current models and human-level hypothetical reasoning, and establish WR-Arena as both a diagnostic tool and a guideline for advancing next-generation world models capable of robust understanding, forecasting, and purposeful action. The code is available at https://github.com/MBZUAI-IFM/WR-Arena.

Yun Luo, Zihan Liu, Yuefeng Shi et al.

The stance detection task aims to classify the stance toward given documents and topics. Since the topics can be implicit in documents and unseen in training data for zero-shot settings, we propose to boost the transferability of the stance detection model by using sentiment and commonsense knowledge, which are seldom considered in previous studies. Our model includes a graph autoencoder module to obtain commonsense knowledge and a stance detection module with sentiment and commonsense. Experimental results show that our model outperforms the state-of-the-art methods on the zero-shot and few-shot benchmark dataset--VAST. Meanwhile, ablation studies prove the significance of each module in our model. Analysis of the relations between sentiment, common sense, and stance indicates the effectiveness of sentiment and common sense.

Yun Luo, Fang Guo, Zihan Liu et al.

Cross-domain sentiment analysis aims to predict the sentiment of texts in the target domain using the model trained on the source domain to cope with the scarcity of labeled data. Previous studies are mostly cross-entropy-based methods for the task, which suffer from instability and poor generalization. In this paper, we explore contrastive learning on the cross-domain sentiment analysis task. We propose a modified contrastive objective with in-batch negative samples so that the sentence representations from the same class will be pushed close while those from the different classes become further apart in the latent space. Experiments on two widely used datasets show that our model can achieve state-of-the-art performance in both cross-domain and multi-domain sentiment analysis tasks. Meanwhile, visualizations demonstrate the effectiveness of transferring knowledge learned in the source domain to the target domain and the adversarial test verifies the robustness of our model.

Boxin Wang, Chankyu Lee, Nayeon Lee et al.

Building general-purpose reasoning models with reinforcement learning (RL) entails substantial cross-domain heterogeneity, including large variation in inference-time response lengths and verification latency. Such variability complicates the RL infrastructure, slows training, and makes training curriculum (e.g., response length extension) and hyperparameter selection challenging. In this work, we propose cascaded domain-wise reinforcement learning (Cascade RL) to develop general-purpose reasoning models, Nemotron-Cascade, capable of operating in both instruct and deep thinking modes. Departing from conventional approaches that blend heterogeneous prompts from different domains, Cascade RL orchestrates sequential, domain-wise RL, reducing engineering complexity and delivering state-of-the-art performance across a wide range of benchmarks. Notably, RLHF for alignment, when used as a pre-step, boosts the model's reasoning ability far beyond mere preference optimization, and subsequent domain-wise RLVR stages rarely degrade the benchmark performance attained in earlier domains and may even improve it (see an illustration in Figure 1). Our 14B model, after RL, outperforms its SFT teacher, DeepSeek-R1-0528, on LiveCodeBench v5/v6/Pro and achieves silver-medal performance in the 2025 International Olympiad in Informatics (IOI). We transparently share our training and data recipes.

Zihan Liu

Natural language understanding (NLU) is the task of semantic decoding of human languages by machines. NLU models rely heavily on large training data to ensure good performance. However, substantial languages and domains have very few data resources and domain experts. It is necessary to overcome the data scarcity challenge, when very few or even zero training samples are available. In this thesis, we focus on developing cross-lingual and cross-domain methods to tackle the low-resource issues. First, we propose to improve the model's cross-lingual ability by focusing on the task-related keywords, enhancing the model's robustness and regularizing the representations. We find that the representations for low-resource languages can be easily and greatly improved by focusing on just the keywords. Second, we present Order-Reduced Modeling methods for the cross-lingual adaptation, and find that modeling partial word orders instead of the whole sequence can improve the robustness of the model against word order differences between languages and task knowledge transfer to low-resource languages. Third, we propose to leverage different levels of domain-related corpora and additional masking of data in the pre-training for the cross-domain adaptation, and discover that more challenging pre-training can better address the domain discrepancy issue in the task knowledge transfer. Finally, we introduce a coarse-to-fine framework, Coach, and a cross-lingual and cross-domain parsing framework, X2Parser. Coach decomposes the representation learning process into a coarse-grained and a fine-grained feature learning, and X2Parser simplifies the hierarchical task structures into flattened ones. We observe that simplifying task structures makes the representation learning more effective for low-resource languages and domains.

Yue Guan, Changming Yu, Shihan Fang et al.

Speculative decoding improves LLM inference by generating and verifying multiple tokens in parallel, but existing systems suffer from suboptimal performance due to a mismatch between dynamic speculation and static runtime assumptions. We present Yggdrasil, a co-designed system that enables latency-optimal speculative decoding through context-aware tree drafting and compiler-friendly execution. Yggdrasil introduces an equal-growth tree structure for static graph compatibility, a latency-aware optimization objective for draft selection, and stage-based scheduling to reduce overhead. Yggdrasil supports unmodified LLMs and achieves up to $3.98\times$ speedup over state-of-the-art baselines across multiple hardware setups.

Mengchen Zhang, Qi Chen, Tong Wu et al.

Despite progress in video-to-audio generation, the field focuses predominantly on mono output, lacking spatial immersion. Existing binaural approaches remain constrained by a two-stage pipeline that first generates mono audio and then performs spatialization, often resulting in error accumulation and spatio-temporal inconsistencies. To address this limitation, we introduce the task of end-to-end binaural spatial audio generation directly from silent video. To support this task, we present the BiAudio dataset, comprising approximately 97K video-binaural audio pairs spanning diverse real-world scenes and camera rotation trajectories, constructed through a semi-automated pipeline. Furthermore, we propose ViSAudio, an end-to-end framework that employs conditional flow matching with a dual-branch audio generation architecture, where two dedicated branches model the audio latent flows. Integrated with a conditional spacetime module, it balances consistency between channels while preserving distinctive spatial characteristics, ensuring precise spatio-temporal alignment between audio and the input video. Comprehensive experiments demonstrate that ViSAudio outperforms existing state-of-the-art methods across both objective metrics and subjective evaluations, generating high-quality binaural audio with spatial immersion that adapts effectively to viewpoint changes, sound-source motion, and diverse acoustic environments. Project website: https://kszpxxzmc.github.io/ViSAudio-project.

Zihan Liu, Zhuolin Yang, Yang Chen et al.

In this work, we investigate the synergy between supervised fine-tuning (SFT) and reinforcement learning (RL) in developing strong reasoning models. We begin by curating the SFT training data through two scaling strategies: increasing the number of collected prompts and the number of generated responses per prompt. Both approaches yield notable improvements in reasoning performance, with scaling the number of prompts resulting in more substantial gains. We then explore the following questions regarding the synergy between SFT and RL: (i) Does a stronger SFT model consistently lead to better final performance after large-scale RL training? (ii) How can we determine an appropriate sampling temperature during RL training to effectively balance exploration and exploitation for a given SFT initialization? Our findings suggest that (i) holds true, provided effective RL training is conducted, particularly when the sampling temperature is carefully chosen to maintain the temperature-adjusted entropy around 0.3, a setting that strikes a good balance between exploration and exploitation. Notably, the performance gap between initial SFT models narrows significantly throughout the RL process. Leveraging a strong SFT foundation and insights into the synergistic interplay between SFT and RL, our AceReason-Nemotron-1.1 7B model significantly outperforms AceReason-Nemotron-1.0 and achieves new state-of-the-art performance among Qwen2.5-7B-based reasoning models on challenging math and code benchmarks, thereby demonstrating the effectiveness of our post-training recipe. We release the model and data at: https://huggingface.co/nvidia/AceReason-Nemotron-1.1-7B

Zihan Liu, Hanyi Wang, Yaoyu Kang et al.

Generative models have shown a giant leap in synthesizing photo-realistic images with minimal expertise, sparking concerns about the authenticity of online information. This study aims to develop a universal AI-generated image detector capable of identifying images from diverse sources. Existing methods struggle to generalize across unseen generative models when provided with limited sample sources. Inspired by the zero-shot transferability of pre-trained vision-language models, we seek to harness the nontrivial visual-world knowledge and descriptive proficiency of CLIP-ViT to generalize over unknown domains. This paper presents a novel parameter-efficient fine-tuning approach, mixture of low-rank experts, to fully exploit CLIP-ViT's potential while preserving knowledge and expanding capacity for transferable detection. We adapt only the MLP layers of deeper ViT blocks via an integration of shared and separate LoRAs within an MoE-based structure. Extensive experiments on public benchmarks show that our method achieves superiority over state-of-the-art approaches in cross-generator generalization and robustness to perturbations. Remarkably, our best-performing ViT-L/14 variant requires training only 0.08% of its parameters to surpass the leading baseline by +3.64% mAP and +12.72% avg.Acc across unseen diffusion and autoregressive models. This even outperforms the baseline with just 0.28% of the training data. Our code and pre-trained models will be available at https://github.com/zhliuworks/CLIPMoLE.

Zihan Liu, Shuangrui Ding, Zhixiong Zhang et al.

Text-to-song generation, the task of creating vocals and accompaniment from textual inputs, poses significant challenges due to domain complexity and data scarcity. Existing approaches often employ multi-stage generation procedures, leading to cumbersome training and inference pipelines, as well as suboptimal overall generation quality due to error accumulation across stages. In this paper, we propose SongGen, a fully open-source, single-stage auto-regressive transformer designed for controllable song generation. The proposed model facilitates fine-grained control over diverse musical attributes, including lyrics and textual descriptions of instrumentation, genre, mood, and timbre, while also offering an optional three-second reference clip for voice cloning. Within a unified auto-regressive framework, SongGen supports two output modes: mixed mode, which generates a mixture of vocals and accompaniment directly, and dual-track mode, which synthesizes them separately for greater flexibility in downstream applications. We explore diverse token pattern strategies for each mode, leading to notable improvements and valuable insights. Furthermore, we design an automated data preprocessing pipeline with effective quality control. To foster community engagement and future research, we will release our model weights, training code, annotated data, and preprocessing pipeline. The code is available at https://github.com/LiuZH-19/SongGen.

Zihan Liu, Hanyi Wang, Shilin Wang

As ultra-realistic face forgery techniques emerge, deepfake detection has attracted increasing attention due to security concerns. Many detectors cannot achieve accurate results when detecting unseen manipulations despite excellent performance on known forgeries. In this paper, we are motivated by the observation that the discrepancies between real and fake videos are extremely subtle and localized, and inconsistencies or irregularities can exist in some critical facial regions across various information domains. To this end, we propose a novel pipeline, Cross-Domain Local Forensics (XDLF), for more general deepfake video detection. In the proposed pipeline, a specialized framework is presented to simultaneously exploit local forgery patterns from space, frequency, and time domains, thus learning cross-domain features to detect forgeries. Moreover, the framework leverages four high-level forgery-sensitive local regions of a human face to guide the model to enhance subtle artifacts and localize potential anomalies. Extensive experiments on several benchmark datasets demonstrate the impressive performance of our method, and we achieve superiority over several state-of-the-art methods on cross-dataset generalization. We also examined the factors that contribute to its performance through ablations, which suggests that exploiting cross-domain local characteristics is a noteworthy direction for developing more general deepfake detectors.

PAN Team, Jiannan Xiang, Yi Gu et al.

A world model enables an intelligent agent to imagine, predict, and reason about how the world evolves in response to its actions, and accordingly to plan and strategize. While recent video generation models produce realistic visual sequences, they typically operate in the prompt-to-full-video manner without causal control, interactivity, or long-horizon consistency required for purposeful reasoning. Existing world modeling efforts, on the other hand, often focus on restricted domains (e.g., physical, game, or 3D-scene dynamics) with limited depth and controllability, and struggle to generalize across diverse environments and interaction formats. In this work, we introduce PAN, a general, interactable, and long-horizon world model that predicts future world states through high-quality video simulation conditioned on history and natural language actions. PAN employs the Generative Latent Prediction (GLP) architecture that combines an autoregressive latent dynamics backbone based on a large language model (LLM), which grounds simulation in extensive text-based knowledge and enables conditioning on language-specified actions, with a video diffusion decoder that reconstructs perceptually detailed and temporally coherent visual observations, to achieve a unification between latent space reasoning (imagination) and realizable world dynamics (reality). Trained on large-scale video-action pairs spanning diverse domains, PAN supports open-domain, action-conditioned simulation with coherent, long-term dynamics. Extensive experiments show that PAN achieves strong performance in action-conditioned world simulation, long-horizon forecasting, and simulative reasoning compared to other video generators and world models, taking a step towards general world models that enable predictive simulation of future world states for reasoning and acting.

Diwei Sheng, Vijayraj Gohil, Satyam Gaba et al.

Scene change detection (SCD) is crucial for urban monitoring and navigation but remains challenging in real-world environments due to lighting variations, seasonal shifts, viewpoint differences, and complex urban layouts. Existing methods rely primarily on low-level visual features, limiting their ability to accurately identify changed objects amid the visual complexity of urban scenes. In this paper, we propose LangSCD, a vision-language framework for scene change detection that overcomes this single-modal limitation by incorporating semantic reasoning through language. Our approach introduces a modular language component that leverages vision-language models (VLMs) to generate textual descriptions of scene changes, which are fused with visual features through a cross-modal feature enhancer. We further introduce a geometric-semantic matching module that refines the predicted masks by enforcing semantic consistency and spatial completeness. Existing real-world scene change detection benchmarks provide only binary change annotations, which are insufficient for downstream applications requiring fine-grained understanding of scene dynamics. To address this limitation, we introduce NYC-CD, a large-scale dataset of 8,122 real-world image pairs collected in New York City with multiclass change annotations generated through a semi-automatic pipeline. Extensive experiments across multiple street-view benchmarks demonstrate that our language and matching modules consistently improve existing change-detection architectures, achieving state-of-the-art performance and highlighting the value of integrating linguistic reasoning with visual representations for robust scene change detection.

Zekai Zhao, Qi Liu, Kun Zhou et al.

Despite the remarkable reasoning performance, eliciting the long chain-of-thought (CoT) ability in large language models (LLMs) typically requires costly reinforcement learning or supervised fine-tuning on high-quality distilled data. We investigate the internal mechanisms behind this capability and show that a small set of high-impact activations in the last few layers largely governs long-form reasoning attributes, such as output length and self-reflection. By simply amplifying these activations and inserting "wait" tokens, we can invoke the long CoT ability without any training, resulting in significantly increased self-reflection rates and accuracy. Moreover, we find that the activation dynamics follow predictable trajectories, with a sharp rise after special tokens and a subsequent exponential decay. Building on these insights, we introduce a general training-free activation control technique. It leverages a few contrastive examples to identify key activations, and employs simple analytic functions to modulate their values at inference time to elicit long CoTs. Extensive experiments confirm the effectiveness of our method in efficiently eliciting long CoT reasoning in LLMs and improving their performance. Additionally, we propose a parameter-efficient fine-tuning method that trains only a last-layer activation amplification module and a few LoRA layers, outperforming full LoRA fine-tuning on reasoning benchmarks with significantly fewer parameters. Our code and data are publicly released.

Zhilun Zhou, Zihan Liu, Jiahe Liu et al.

Large Language Model-based Multi-Agent Systems (LLM-based MAS), where multiple LLM agents collaborate to solve complex tasks, have shown impressive performance in many areas. However, MAS are typically distributed across different devices or environments, making them vulnerable to perturbations such as agent failures. While existing works have studied the adversarial attacks and corresponding defense strategies, they mainly focus on reactively detecting and mitigating attacks after they occur rather than proactively designing inherently resilient systems. In this work, we study the resilience of LLM-based MAS under perturbations and find that both the communication topology and prompt design significantly influence system resilience. Motivated by these findings, we propose ResMAS: a two-stage framework for enhancing MAS resilience. First, we train a reward model to predict the MAS's resilience, based on which we train a topology generator to automatically design resilient topology for specific tasks through reinforcement learning. Second, we introduce a topology-aware prompt optimization method that refines each agent's prompt based on its connections and interactions with other agents. Extensive experiments across a range of tasks show that our approach substantially improves MAS resilience under various constraints. Moreover, our framework demonstrates strong generalization ability to new tasks and models, highlighting its potential for building resilient MASs.

Zihan Liu, Zhikang Niu, Qiuyang Xiao et al.

Despite rapid progress in Multi-modal Large Language Models and Large Audio-Language Models, existing audio benchmarks largely test semantics that can be recovered from text captions, masking deficits in fine-grained perceptual reasoning. We formalize audio 4D intelligence that is defined as reasoning over sound dynamics in time and 3D space, and introduce STAR-Bench to measure it. STAR-Bench combines a Foundational Acoustic Perception setting (six attributes under absolute and relative regimes) with a Holistic Spatio-Temporal Reasoning setting that includes segment reordering for continuous and discrete processes and spatial tasks spanning static localization, multi-source relations, and dynamic trajectories. Our data curation pipeline uses two methods to ensure high-quality samples. For foundational tasks, we use procedurally synthesized and physics-simulated audio. For holistic data, we follow a four-stage process that includes human annotation and final selection based on human performance. Unlike prior benchmarks where caption-only answering reduces accuracy slightly, STAR-Bench induces far larger drops (-31.5\% temporal, -35.2\% spatial), evidencing its focus on linguistically hard-to-describe cues. Evaluating 19 models reveals substantial gaps compared with humans and a capability hierarchy: closed-source models are bottlenecked by fine-grained perception, while open-source models lag across perception, knowledge, and reasoning. Our STAR-Bench provides critical insights and a clear path forward for developing future models with a more robust understanding of the physical world.

Xinhao Luo, Zihan Liu, Yangjie Zhou et al.

Large language model (LLM) decoding suffers from high latency due to fragmented execution across operators and heavy reliance on off-chip memory for data exchange and reduction. This execution model limits opportunities for fusion and incurs significant memory traffic and kernel launch overhead. While modern architectures such as NVIDIA Hopper provide distributed shared memory and low-latency intra-cluster interconnects, they expose only low-level data movement instructions, lacking structured abstractions for collective on-chip communication. To bridge this software-hardware gap, we introduce two cluster-level communication primitives, ClusterReduce and ClusterGather, which abstract common communication patterns and enable structured, high-speed data exchange and reduction between thread blocks within a cluster, allowing intermediate results to be on-chip without involving off-chip memory. Building on these abstractions, we design ClusterFusion, an execution framework that schedules communication and computation jointly to expand operator fusion scope by composing decoding stages such as QKV Projection, Attention, and Output Projection into a single fused kernels. Evaluations on H100 GPUs show that ClusterFusion outperforms state-of-the-art inference frameworks by 1.61x on average in end-to-end latency across different models and configurations. The source code is available at https://github.com/xinhao-luo/ClusterFusion.

Zihan Liu, Xiangning Ren, Dezhang Kong et al.

Infrared small object detection urgently requires semi-supervised paradigms due to the high cost of annotation. However, existing methods like SAM face significant challenges of domain gaps, inability of encoding physical priors, and inherent architectural complexity. To address this, we designed a Hierarchical MoE Adapter consisting of four white-box neural operators. Building upon this core component, we propose a two-stage paradigm for knowledge distillation and transfer: (1) Prior-Guided Knowledge Distillation, where we use our MoE adapter and 10% of available fully supervised data to distill SAM into an expert teacher (Scalpel-SAM); and (2) Deployment-Oriented Knowledge Transfer, where we use Scalpel-SAM to generate pseudo labels for training lightweight and efficient downstream models. Experiments demonstrate that with minimal annotations, our paradigm enables downstream models to achieve performance comparable to, or even surpassing, their fully supervised counterparts. To our knowledge, this is the first semi-supervised paradigm that systematically addresses the data scarcity issue in IR-SOT using SAM as the teacher model.