Tong Zhu

Yicheng Zou, Dongsheng Zhu, Lin Zhu et al.

We introduce Intern-S1-Pro, the first one-trillion-parameter scientific multimodal foundation model. Scaling to this unprecedented size, the model delivers a comprehensive enhancement across both general and scientific domains. Beyond stronger reasoning and image-text understanding capabilities, its intelligence is augmented with advanced agent capabilities. Simultaneously, its scientific expertise has been vastly expanded to master over 100 specialized tasks across critical science fields, including chemistry, materials, life sciences, and earth sciences. Achieving this massive scale is made possible by the robust infrastructure support of XTuner and LMDeploy, which facilitates highly efficient Reinforcement Learning (RL) training at the 1-trillion parameter level while ensuring strict precision consistency between training and inference. By seamlessly integrating these advancements, Intern-S1-Pro further fortifies the fusion of general and specialized intelligence, working as a Specializable Generalist, demonstrating its position in the top tier of open-source models for general capabilities, while outperforming proprietary models in the depth of specialized scientific tasks.

Tong Zhu, Junfei Ren, Zijian Yu et al.

Sharing knowledge between information extraction tasks has always been a challenge due to the diverse data formats and task variations. Meanwhile, this divergence leads to information waste and increases difficulties in building complex applications in real scenarios. Recent studies often formulate IE tasks as a triplet extraction problem. However, such a paradigm does not support multi-span and n-ary extraction, leading to weak versatility. To this end, we reorganize IE problems into unified multi-slot tuples and propose a universal framework for various IE tasks, namely Mirror. Specifically, we recast existing IE tasks as a multi-span cyclic graph extraction problem and devise a non-autoregressive graph decoding algorithm to extract all spans in a single step. It is worth noting that this graph structure is incredibly versatile, and it supports not only complex IE tasks, but also machine reading comprehension and classification tasks. We manually construct a corpus containing 57 datasets for model pretraining, and conduct experiments on 30 datasets across 8 downstream tasks. The experimental results demonstrate that our model has decent compatibility and outperforms or reaches competitive performance with SOTA systems under few-shot and zero-shot settings. The code, model weights, and pretraining corpus are available at https://github.com/Spico197/Mirror .

Yuanqing Wang, Kenichiro Takaba, Michael S. Chen et al.

A force field as accurate as quantum mechanics (QM) and as fast as molecular mechanics (MM), with which one can simulate a biomolecular system efficiently enough and meaningfully enough to get quantitative insights, is among the most ardent dreams of biophysicists -- a dream, nevertheless, not to be fulfilled any time soon. Machine learning force fields (MLFFs) represent a meaningful endeavor towards this direction, where differentiable neural functions are parametrized to fit ab initio energies, and furthermore forces through automatic differentiation. We argue that, as of now, the utility of the MLFF models is no longer bottlenecked by accuracy but primarily by their speed (as well as stability and generalizability), as many recent variants, on limited chemical spaces, have long surpassed the chemical accuracy of $1$ kcal/mol -- the empirical threshold beyond which realistic chemical predictions are possible -- though still magnitudes slower than MM. Hoping to kindle explorations and designs of faster, albeit perhaps slightly less accurate MLFFs, in this review, we focus our attention on the design space (the speed-accuracy tradeoff) between MM and ML force fields. After a brief review of the building blocks of force fields of either kind, we discuss the desired properties and challenges now faced by the force field development community, survey the efforts to make MM force fields more accurate and ML force fields faster, envision what the next generation of MLFF might look like.

Tong Zhu, Guoliang Zhang, Zechang Li et al.

Sentence-by-sentence information extraction from long documents is an exhausting and error-prone task. As the indicator of document skeleton, catalogs naturally chunk documents into segments and provide informative cascade semantics, which can help to reduce the search space. Despite their usefulness, catalogs are hard to be extracted without the assist from external knowledge. For documents that adhere to a specific template, regular expressions are practical to extract catalogs. However, handcrafted heuristics are not applicable when processing documents from different sources with diverse formats. To address this problem, we build a large manually annotated corpus, which is the first dataset for the Catalog Extraction from Documents (CED) task. Based on this corpus, we propose a transition-based framework for parsing documents into catalog trees. The experimental results demonstrate that our proposed method outperforms baseline systems and shows a good ability to transfer. We believe the CED task could fill the gap between raw text segments and information extraction tasks on extremely long documents. Data and code are available at \url{https://github.com/Spico197/CatalogExtraction}

Hitarth Choubisa, Md Azimul Haque, Tong Zhu et al.

The exploration of thermoelectric materials is challenging considering the large materials space, combined with added exponential degrees of freedom coming from doping and the diversity of synthetic pathways. Here we seek to incorporate historical data and update and refine it using experimental feedback by employing error-correction learning (ECL). We thus learn from prior datasets and then adapt the model to differences in synthesis and characterization that are otherwise difficult to parameterize. We then apply this strategy to discovering thermoelectric materials where we prioritize synthesis at temperatures < 300°C. We document a previously unreported chemical family of thermoelectric materials, PbSe:SnSb, finding that the best candidate in this chemical family, 2 wt% SnSb doped PbSe, exhibits a power factor more than 2x that of PbSe. Our investigations show that our closed-loop experimentation strategy reduces the required number of experiments to find an optimized material by as much as 3x compared to high-throughput searches powered by state-of-the-art machine learning models. We also observe that this improvement is dependent on the accuracy of prior in a manner that exhibits diminishing returns, and after a certain accuracy is reached, it is factors associated with experimental pathways that dictate the trends.

Zhaochen Su, Jun Zhang, Xiaoye Qu et al.

Large language models (LLMs) have achieved impressive advancements across numerous disciplines, yet the critical issue of knowledge conflicts, a major source of hallucinations, has rarely been studied. Only a few research explored the conflicts between the inherent knowledge of LLMs and the retrieved contextual knowledge. However, a thorough assessment of knowledge conflict in LLMs is still missing. Motivated by this research gap, we present ConflictBank, the first comprehensive benchmark developed to systematically evaluate knowledge conflicts from three aspects: (i) conflicts encountered in retrieved knowledge, (ii) conflicts within the models' encoded knowledge, and (iii) the interplay between these conflict forms. Our investigation delves into four model families and twelve LLM instances, meticulously analyzing conflicts stemming from misinformation, temporal discrepancies, and semantic divergences. Based on our proposed novel construction framework, we create 7,453,853 claim-evidence pairs and 553,117 QA pairs. We present numerous findings on model scale, conflict causes, and conflict types. We hope our ConflictBank benchmark will help the community better understand model behavior in conflicts and develop more reliable LLMs.

Jihai Zhang, Xiaoye Qu, Tong Zhu et al.

Contrastive Language-Image Pre-training (CLIP) has become a cornerstone in multimodal intelligence. However, recent studies discovered that CLIP can only encode one aspect of the feature space, leading to substantial information loss and indistinctive features. To mitigate this issue, this paper introduces a novel strategy that fine-tunes a series of complementary CLIP models and transforms them into a CLIP-MoE. Specifically, we propose a model-agnostic Diversified Multiplet Upcycling (DMU) framework for CLIP. Instead of training multiple CLIP models from scratch, DMU leverages a pre-trained CLIP and fine-tunes it into a diverse set with highly cost-effective multistage contrastive learning, thus capturing distinct feature subspaces efficiently. To fully exploit these fine-tuned models while minimizing computational overhead, we transform them into a CLIP-MoE, which dynamically activates a subset of CLIP experts, achieving an effective balance between model capacity and computational cost. Comprehensive experiments demonstrate the superior performance of CLIP-MoE across various zero-shot retrieval, zero-shot image classification tasks, and downstream Multimodal Large Language Model (MLLM) benchmarks when used as a vision encoder.

Zhenhua Liu, Tong Zhu, Chuanyuan Tan et al.

Large language models (LLMs) exhibit remarkable capabilities in understanding and generating natural language. However, these models can inadvertently memorize private information, posing significant privacy risks. This study addresses the challenge of enabling LLMs to protect specific individuals' private data without the need for complete retraining. We propose \return, a Real-world pErsonal daTa UnleaRNing dataset, comprising 2,492 individuals from Wikipedia with associated QA pairs, to evaluate machine unlearning (MU) methods for protecting personal data in a realistic scenario. Additionally, we introduce the Name-Aware Unlearning Framework (NAUF) for Privacy Protection, which enables the model to learn which individuals' information should be protected without affecting its ability to answer questions related to other unrelated individuals. Our extensive experiments demonstrate that NAUF achieves a state-of-the-art average unlearning score, surpassing the best baseline method by 5.65 points, effectively protecting target individuals' personal data while maintaining the model's general capabilities.

Zefeng He, Siyuan Huang, Xiaoye Qu et al.

Recent multimodal generation models have achieved remarkable progress on general-purpose generation tasks, yet continue to struggle with complex instructions and specialized downstream tasks. Inspired by the success of advanced agent frameworks such as Claude Code, we propose \textbf{GEMS} (Agent-Native Multimodal \textbf{GE}neration with \textbf{M}emory and \textbf{S}kills), a framework that pushes beyond the inherent limitations of foundational models on both general and downstream tasks. GEMS is built upon three core components. Agent Loop introduces a structured multi-agent framework that iteratively improves generation quality through closed-loop optimization. Agent Memory provides a persistent, trajectory-level memory that hierarchically stores both factual states and compressed experiential summaries, enabling a global view of the optimization process while reducing redundancy. Agent Skill offers an extensible collection of domain-specific expertise with on-demand loading, allowing the system to effectively handle diverse downstream applications. Across five mainstream tasks and four downstream tasks, evaluated on multiple generative backends, GEMS consistently achieves significant performance gains. Most notably, it enables the lightweight 6B model Z-Image-Turbo to surpass the state-of-the-art Nano Banana 2 on GenEval2, demonstrating the effectiveness of agent harness in extending model capabilities beyond their original limits.

Xiaofang Yang, Lijun Li, Heng Zhou et al.

Recent years have witnessed increasing interest in extending large language models into agentic systems. While the effectiveness of agents has continued to improve, efficiency, which is crucial for real-world deployment, has often been overlooked. This paper therefore investigates efficiency from three core components of agents: memory, tool learning, and planning, considering costs such as latency, tokens, steps, etc. Aimed at conducting comprehensive research addressing the efficiency of the agentic system itself, we review a broad range of recent approaches that differ in implementation yet frequently converge on shared high-level principles including but not limited to bounding context via compression and management, designing reinforcement learning rewards to minimize tool invocation, and employing controlled search mechanisms to enhance efficiency, which we discuss in detail. Accordingly, we characterize efficiency in two complementary ways: comparing effectiveness under a fixed cost budget, and comparing cost at a comparable level of effectiveness. This trade-off can also be viewed through the Pareto frontier between effectiveness and cost. From this perspective, we also examine efficiency oriented benchmarks by summarizing evaluation protocols for these components and consolidating commonly reported efficiency metrics from both benchmark and methodological studies. Moreover, we discuss the key challenges and future directions, with the goal of providing promising insights.

Mengsong Wu, Tong Zhu, Han Han et al.

This paper presents a new tool learning dataset Seal-Tools, which contains self-instruct API-like tools. Seal-Tools not only offers a large number of tools, but also includes instances which demonstrate the practical application of tools. Seeking to generate data on a large scale while ensuring reliability, we propose a self-instruct method to generate tools and instances, allowing precise control over the process. Moreover, our Seal-Tools contains hard instances that call multiple tools to complete the job, among which some are nested tool callings. For precise and comprehensive evaluation, we use strict format control and design three metrics from different dimensions. Therefore, Seal-Tools can serve as a new benchmark to evaluate the tool-calling ability of LLMs. Finally, we evaluate several prevalent LLMs and our finetuned model on Seal-Tools. The results show that current systems are far from perfect. The code, data and experiment results are available at https://github.com/fairyshine/Seal-Tools .

Zefeng He, Xiaoye Qu, Yafu Li et al.

While recent Multimodal Large Language Models (MLLMs) have attained significant strides in multimodal reasoning, their reasoning processes remain predominantly text-centric, leading to suboptimal performance in complex long-horizon, vision-centric tasks. In this paper, we establish a novel Generative Multimodal Reasoning paradigm and introduce DiffThinker, a diffusion-based reasoning framework. Conceptually, DiffThinker reformulates multimodal reasoning as a native generative image-to-image task, achieving superior logical consistency and spatial precision in vision-centric tasks. We perform a systematic comparison between DiffThinker and MLLMs, providing the first in-depth investigation into the intrinsic characteristics of this paradigm, revealing four core properties: efficiency, controllability, native parallelism, and collaboration. Extensive experiments across four domains (sequential planning, combinatorial optimization, constraint satisfaction, and spatial configuration) demonstrate that DiffThinker significantly outperforms leading closed source models including GPT-5 (+314.2\%) and Gemini-3-Flash (+111.6\%), as well as the fine-tuned Qwen3-VL-32B baseline (+39.0\%), highlighting generative multimodal reasoning as a promising approach for vision-centric reasoning.

Xiaoye Qu, Daize Dong, Xuyang Hu et al.

Recently, inspired by the concept of sparsity, Mixture-of-Experts (MoE) models have gained increasing popularity for scaling model size while keeping the number of activated parameters constant. In this study, we thoroughly investigate the sparsity of the dense LLaMA model by constructing MoE for both the attention (i.e., Attention MoE) and MLP (i.e., MLP MoE) modules in the transformer blocks. Specifically, we investigate different expert construction methods and granularities under the same activation conditions to analyze the impact of sparsifying the model. Additionally, to comprehensively evaluate the model's capabilities across various domains (e.g., conversation, code, math) after sparsification, we apply sparsity to the instructed large language models (LLMs) and construct instructed MoE models. To counteract the performance degradation resulting from increased sparsity, we design a two-stage post-training strategy to enhance model performance. Experiments on the LLaMA3 model demonstrate the potential effectiveness of this approach for future developments of instructed MoE models. The source codes and models are available at: \url{https://github.com/OpenSparseLLMs/LLaMA-MoE-v2}.

Linfeng Zhang, Siheng Chen, Yuzhu Cai et al.

AI agents are emerging as a practical way to run multi-step scientific workflows that interleave reasoning with tool use and verification, pointing to a shift from isolated AI-assisted steps toward \emph{agentic science at scale}. This shift is increasingly feasible, as scientific tools and models can be invoked through stable interfaces and verified with recorded execution traces, and increasingly necessary, as AI accelerates scientific output and stresses the peer-review and publication pipeline, raising the bar for traceability and credible evaluation. However, scaling agentic science remains difficult: workflows are hard to observe and reproduce; many tools and laboratory systems are not agent-ready; execution is hard to trace and govern; and prototype AI Scientist systems are often bespoke, limiting reuse and systematic improvement from real workflow signals. We argue that scaling agentic science requires an infrastructure-and-ecosystem approach, instantiated in Bohrium+SciMaster. Bohrium acts as a managed, traceable hub for AI4S assets -- akin to a HuggingFace of AI for Science -- that turns diverse scientific data, software, compute, and laboratory systems into agent-ready capabilities. SciMaster orchestrates these capabilities into long-horizon scientific workflows, on which scientific agents can be composed and executed. Between infrastructure and orchestration, a \emph{scientific intelligence substrate} organizes reusable models, knowledge, and components into executable building blocks for workflow reasoning and action, enabling composition, auditability, and improvement through use. We demonstrate this stack with eleven representative master agents in real workflows, achieving orders-of-magnitude reductions in end-to-end scientific cycle time and generating execution-grounded signals from real workloads at multi-million scale.

Jiacheng Ruan, Daize Dong, Xiaoye Qu et al.

Mixture-of-Experts (MoE) models substantially improve performance by increasing the capacity of dense architectures. However, directly training MoE models requires considerable computational resources and introduces extra overhead in parameter storage and deployment. Therefore, it is critical to develop an approach that leverages the multi-expert capability of MoE to enhance performance while incurring minimal additional cost. To this end, we propose a novel pre-training approach, termed ExFusion, which improves the efficiency of Transformer training through multi-expert fusion. Specifically, during the initialization phase, ExFusion upcycles the feed-forward network (FFN) of the Transformer into a multi-expert configuration, where each expert is assigned a weight for later parameter fusion. During training, these weights allow multiple experts to be fused into a single unified expert equivalent to the original FFN, which is subsequently used for forward computation. As a result, ExFusion introduces multi-expert characteristics into the training process while incurring only marginal computational cost compared to standard dense training. After training, the learned weights are used to integrate multi-experts into a single unified expert, thereby eliminating additional overhead in storage and deployment. Extensive experiments on a variety of computer vision and natural language processing tasks demonstrate the effectiveness of the proposed method.

Zihan Zhang, Haohui Que, Junhan Chang et al.

The emergence of Self-Driving Laboratories (SDLs) transforms scientific discovery methodology by integrating AI with robotic automation to create closed-loop experimental systems capable of autonomous hypothesis generation, experimentation, and analysis. While promising to compress research timelines from years to weeks, their deployment introduces unprecedented safety challenges differing from traditional laboratories or purely digital AI. This paper presents Safe-SDL, a comprehensive framework for establishing robust safety boundaries and control mechanisms in AI-driven autonomous laboratories. We identify and analyze the critical ``Syntax-to-Safety Gap'' -- the disconnect between AI-generated syntactically correct commands and their physical safety implications -- as the central challenge in SDL deployment. Our framework addresses this gap through three synergistic components: (1) formally defined Operational Design Domains (ODDs) that constrain system behavior within mathematically verified boundaries, (2) Control Barrier Functions (CBFs) that provide real-time safety guarantees through continuous state-space monitoring, and (3) a novel Transactional Safety Protocol (CRUTD) that ensures atomic consistency between digital planning and physical execution. We ground our theoretical contributions through analysis of existing implementations including UniLabOS and the Osprey architecture, demonstrating how these systems instantiate key safety principles. Evaluation against the LabSafety Bench reveals that current foundation models exhibit significant safety failures, demonstrating that architectural safety mechanisms are essential rather than optional. Our framework provides both theoretical foundations and practical implementation guidance for safe deployment of autonomous scientific systems, establishing the groundwork for responsible acceleration of AI-driven discovery.

Weigao Sun, Disen Lan, Tong Zhu et al.

Linear Sequence Modeling (LSM) like linear attention, state space models and linear RNNs, and Mixture-of-Experts (MoE) have recently emerged as significant architectural improvements. In this paper, we introduce Linear-MoE, a production-level system for modeling and training large-scale models that integrate LSM with MoE. Linear-MoE leverages the advantages of both LSM modules for linear-complexity sequence modeling and MoE layers for sparsely activation, aiming to offer high performance with efficient training. The Linear-MoE system comprises: 1) Modeling subsystem, which provides a unified framework supporting all instances of LSM. and 2) Training subsystem, which facilitates efficient training by incorporating various advanced parallelism technologies, particularly Sequence Parallelism designed for Linear-MoE models. Additionally, we explore hybrid models that combine Linear-MoE layers with standard Transformer-MoE layers with its Sequence Parallelism to further enhance model flexibility and performance. Evaluations on two model series, A0.3B-2B and A1B-7B, demonstrate Linear-MoE achieves efficiency gains while maintaining competitive performance on various benchmarks, showcasing its potential as a next-generation foundational model architecture. Code: https://github.com/OpenSparseLLMs/Linear-MoE.

Haoran Zhang, Luxin Xu, Zhilin Wang et al.

The rise of personal assistant agents, e.g., OpenClaw, highlights the growing potential of large language models to support users across everyday life and work. A core challenge in these settings is proactive assistance, since users often begin with underspecified requests and leave important needs, constraints, or preferences unstated. However, existing benchmarks rarely evaluate whether agents can identify and act on such hidden intents before they are explicitly stated, especially in sustained multi-turn interactions where user needs emerge gradually. To address this gap, we introduce $π$-Bench, a benchmark for proactive assistance comprising 100 multi-turn tasks across 5 domain-specific user personas. By incorporating hidden user intents, inter-task dependencies, and cross-session continuity, $π$-Bench evaluates agents' ability to anticipate and address user needs over extended interactions, jointly measuring proactivity and task completion in long-horizon trajectories that better reflect real-world use. Experiments show (1) proactive assistance remains challenging, (2) a clear distinction between task completion and proactivity, and (3) the value of prior interaction for proactive intent resolution in later tasks.

Quanlong Guan, Tong Zhu, Liangda Fang et al.

Belief revision and update, two significant types of belief change, both focus on how an agent modify her beliefs in presence of new information. The most striking difference between them is that the former studies the change of beliefs in a static world while the latter concentrates on a dynamically-changing world. The famous AGM and KM postulates were proposed to capture rational belief revision and update, respectively. However, both of them are too permissive to exclude some unreasonable changes in the iteration. In response to this weakness, the DP postulates and its extensions for iterated belief revision were presented. Furthermore, Rodrigues integrated these postulates in belief update. Unfortunately, his approach does not meet the basic requirement of iterated belief update. This paper is intended to solve this problem of Rodrigues's approach. Firstly, we present a modification of the original KM postulates based on belief states. Subsequently, we migrate several well-known postulates for iterated belief revision to iterated belief update. Moreover, we provide the exact semantic characterizations based on partial preorders for each of the proposed postulates. Finally, we analyze the compatibility between the above iterated postulates and the KM postulates for belief update.

Mengsong Wu, Tong Zhu, Han Han et al.

Tool learning can further broaden the usage scenarios of large language models (LLMs). However most of the existing methods either need to finetune that the model can only use tools seen in the training data, or add tool demonstrations into the prompt with lower efficiency. In this paper, we present a new Tool Learning method Chain-of-Tools. It makes full use of the powerful semantic representation capability of frozen LLMs to finish tool calling in CoT reasoning with a huge and flexible tool pool which may contain unseen tools. Especially, to validate the effectiveness of our approach in the massive unseen tool scenario, we construct a new dataset SimpleToolQuestions. We conduct experiments on two numerical reasoning benchmarks (GSM8K-XL and FuncQA) and two knowledge-based question answering benchmarks (KAMEL and SimpleToolQuestions). Experimental results show that our approach performs better than the baseline. We also identify dimensions of the model output that are critical in tool selection, enhancing the model interpretability. Our code and data are available at: https://github.com/fairyshine/Chain-of-Tools .

Ziyu Zhao, Tong Zhu, Zhi Zhang et al.

Mixture-of-Experts (MoE) effectively scales model capacity while preserving computational efficiency through sparse expert activation. However, training high-quality MoEs from scratch is prohibitively expensive. A promising alternative is to convert pretrained dense models into sparse MoEs. Existing dense-to-MoE methods fall into two categories: \textbf{dynamic structural pruning} that converts dense models into MoE architectures with moderate sparsity to balance performance and inference efficiency, and \textbf{downcycling} approaches that use pretrained dense models to initialize highly sparse MoE architectures. However, existing methods break the intrinsic activation patterns within dense models, leading to suboptimal expert construction. In this work, we argue that the Gated Linear Unit (GLU) mechanism provides a natural blueprint for dense-to-MoE conversion. We show that the fine-grained neural-wise activation patterns of GLU reveal a coarse-grained structure, uncovering an inherent MoE architecture composed of consistently activated universal neurons and dynamically activated specialized neurons. Leveraging this discovery, we introduce ExpertWeaver, a training-free framework that partitions neurons according to their activation patterns and constructs shared experts and specialized routed experts with layer-adaptive configurations. Our experiments demonstrate that ExpertWeaver significantly outperforms existing methods, both as a training-free dynamic structural pruning technique and as a downcycling strategy for superior MoE initialization.

Yafu Li, Runzhe Zhan, Haoran Zhang et al.

Recent progress in reasoning models has substantially advanced long-horizon mathematical and scientific problem solving, with several systems now reaching gold-medal-level performance on International Mathematical Olympiad (IMO) and International Physics Olympiad (IPhO) problems. In this paper, we introduce a simple and unified recipe for converting a post-trained reasoning backbone into a rigorous olympiad-level solver. The recipe first uses a reverse-perplexity curriculum for SFT to instill rigorous proof-search and self-checking behaviors, then scales these behaviors through a two-stage RL pipeline that progresses from RL with verifiable rewards to more delicate proof-level RL, and finally boosts solving performance with test-time scaling. Applying this recipe, we train a 30B-A3B backbone with SFT on around 340K sub-8K-token trajectories followed by 200 RL steps. The resulting model, SU-01, supports stable reasoning on difficult problems with trajectories exceeding 100K tokens, while achieving gold-medal-level performance on mathematical and physical olympiad competitions, including IMO 2025/USAMO 2026 and IPhO 2024/2025. It also demonstrates strong generalization of scientific reasoning to domains beyond mathematics and physics.

Weige Cai, Tong Zhu, Jinyi Niu et al.

With the rapid advancement of large language models (LLMs), Multi-agent Systems (MAS) have achieved significant progress in various application scenarios. However, substantial challenges remain in designing versatile, robust, and efficient platforms for agent deployment. To address these limitations, we propose \textbf{LightAgent}, a lightweight yet powerful agentic framework, effectively resolving the trade-off between flexibility and simplicity found in existing frameworks. LightAgent integrates core functionalities such as Memory (mem0), Tools, and Tree of Thought (ToT), while maintaining an extremely lightweight structure. As a fully open-source solution, it seamlessly integrates with mainstream chat platforms, enabling developers to easily build self-learning agents. We have released LightAgent at \href{https://github.com/wxai-space/LightAgent}{https://github.com/wxai-space/LightAgent}

Tong Zhu, Xiaoye Qu, Daize Dong et al.

Mixture-of-Experts (MoE) has gained increasing popularity as a promising framework for scaling up large language models (LLMs). However, training MoE from scratch in a large-scale setting still suffers from data-hungry and instability problems. Motivated by this limit, we investigate building MoE models from existing dense large language models. Specifically, based on the well-known LLaMA-2 7B model, we obtain an MoE model by: (1) Expert Construction, which partitions the parameters of original Feed-Forward Networks (FFNs) into multiple experts; (2) Continual Pre-training, which further trains the transformed MoE model and additional gate networks. In this paper, we comprehensively explore different methods for expert construction and various data sampling strategies for continual pre-training. After these stages, our LLaMA-MoE models could maintain language abilities and route the input tokens to specific experts with part of the parameters activated. Empirically, by training 200B tokens, LLaMA-MoE-3.5B models significantly outperform dense models that contain similar activation parameters. The source codes and models are available at https://github.com/pjlab-sys4nlp/llama-moe .

Zhaochen Su, Jun Zhang, Tong Zhu et al.

Reasoning about time is essential for Large Language Models (LLMs) to understand the world. Previous works focus on solving specific tasks, primarily on time-sensitive question answering. While these methods have proven effective, they cannot generalize to a wider spectrum of temporal reasoning tasks. Therefore, we propose a crucial question: Can we build a universal framework to handle a variety of temporal reasoning tasks? To that end, we systematically study 38 temporal reasoning tasks. Based on the observation that 19 tasks are directly related to mathematics, we first leverage the available mathematical dataset to set a solid foundation for temporal reasoning. However, the in-depth study indicates that focusing solely on mathematical enhancement falls short of addressing pure temporal reasoning tasks. To mitigate this limitation, we propose a simple but effective self-critic temporal optimization method to enhance the model's temporal reasoning capabilities without sacrificing general task abilities. Finally, we develop Timo, a model designed to excel in temporal reasoning at the 7B and 13B scales. Notably, Timo outperforms the counterpart LLMs by 10.0 and 7.6 in average accuracy scores and achieves the new state-of-the-art (SOTA) performance of comparable size. Extensive experiments further validate our framework's effectiveness and its generalization across diverse temporal tasks. The code is available at https://github.com/zhaochen0110/Timo.

Tong Zhu, Daize Dong, Xiaoye Qu et al.

Mixture-of-Experts (MoE) models have shown remarkable capability in instruction tuning, especially when the number of tasks scales. However, previous methods simply merge all training tasks (e.g. creative writing, coding, and mathematics) and apply fixed sampling weights, without considering the importance of different tasks as the model training state changes. In this way, the most helpful data cannot be effectively distinguished, leading to suboptimal model performance. To reduce the potential redundancies of datasets, we make the first attempt and propose a novel dynamic data mixture for MoE instruction tuning. Specifically, inspired by MoE's token routing preference, we build dataset-level representations and then capture the subtle differences among datasets. Finally, we propose to dynamically adjust the sampling weight of datasets by their inter-redundancies, thus maximizing global performance under a limited training budget. The experimental results on two MoE models demonstrate the effectiveness of our approach on both downstream knowledge \& reasoning tasks and open-ended queries. Code and models are available at https://github.com/Spico197/MoE-SFT .

Zhaochen Su, Juntao Li, Jun Zhang et al.

Temporal reasoning is fundamental for large language models (LLMs) to comprehend the world. Current temporal reasoning datasets are limited to questions about single or isolated events, falling short in mirroring the realistic temporal characteristics involving concurrent nature and intricate temporal interconnections. In this paper, we introduce CoTempQA, a comprehensive co-temporal Question Answering (QA) benchmark containing four co-temporal scenarios (Equal, Overlap, During, Mix) with 4,748 samples for evaluating the co-temporal comprehension and reasoning abilities of LLMs. Our extensive experiments reveal a significant gap between the performance of current LLMs and human-level reasoning on CoTempQA tasks. Even when enhanced with Chain of Thought (CoT) methodologies, models consistently struggle with our task. In our preliminary exploration, we discovered that mathematical reasoning plays a significant role in handling co-temporal events and proposed a strategy to boost LLMs' co-temporal reasoning from a mathematical perspective. We hope that our CoTempQA datasets will encourage further advancements in improving the co-temporal reasoning capabilities of LLMs. Our code is available at https://github.com/zhaochen0110/Cotempqa.

Zhenhua Liu, Tong Zhu, Chuanyuan Tan et al.

Large Language Models (LLMs) have shown their impressive capabilities, while also raising concerns about the data contamination problems due to privacy issues and leakage of benchmark datasets in the pre-training phase. Therefore, it is vital to detect the contamination by checking whether an LLM has been pre-trained on the target texts. Recent studies focus on the generated texts and compute perplexities, which are superficial features and not reliable. In this study, we propose to utilize the probing technique for pre-training data detection by examining the model's internal activations. Our method is simple and effective and leads to more trustworthy pre-training data detection. Additionally, we propose ArxivMIA, a new challenging benchmark comprising arxiv abstracts from Computer Science and Mathematics categories. Our experiments demonstrate that our method outperforms all baselines, and achieves state-of-the-art performance on both WikiMIA and ArxivMIA, with additional experiments confirming its efficacy (Our code and dataset are available at https://github.com/zhliu0106/probing-lm-data).

Tong Zhu, Xiaoye Qu, Wenliang Chen et al.

Most previous studies of document-level event extraction mainly focus on building argument chains in an autoregressive way, which achieves a certain success but is inefficient in both training and inference. In contrast to the previous studies, we propose a fast and lightweight model named as PTPCG. In our model, we design a novel strategy for event argument combination together with a non-autoregressive decoding algorithm via pruned complete graphs, which are constructed under the guidance of the automatically selected pseudo triggers. Compared to the previous systems, our system achieves competitive results with 19.8\% of parameters and much lower resource consumption, taking only 3.8\% GPU hours for training and up to 8.5 times faster for inference. Besides, our model shows superior compatibility for the datasets with (or without) triggers and the pseudo triggers can be the supplements for annotated triggers to make further improvements. Codes are available at https://github.com/Spico197/DocEE .

Baiting Chen, Tong Zhu, Rui Yu et al.

Aligning large language models (LLMs) with human preferences is inherently multi-objective: different users and evaluation criteria impose heterogeneous and often conflicting requirements on model outputs. We propose CAGE (Common-Agency Games for Alignment), a training-free, game-theoretic framework for multi-objective test-time alignment. CAGE models alignment objectives as strategic principals that allocate token-level incentives to a shared LLM, inducing an equilibrium policy that captures the joint effect of competing objectives. We develop an efficient algorithm based on equilibrium problems with equilibrium constraints (EPEC) to compute this equilibrium, and establish theoretical guarantees including existence and uniqueness of the equilibrium policy, convergence and stability of the algorithm, and no-regret learning dynamics. Empirically, CAGE enables flexible and fine-grained trade-offs across objectives at inference time, consistently outperforming existing test-time alignment methods while requiring no retraining. It further supports weak-to-strong generalization, making multi-objective alignment practical in resource-constrained settings.

Siyuan Huang, Xiaoye Qu, Yafu Li et al.

While autoregressive Large Vision-Language Models (LVLMs) demonstrate remarkable proficiency in multimodal tasks, they face a "Visual Signal Dilution" phenomenon, where the accumulation of textual history expands the attention partition function, causing visual attention to decay inversely with generated sequence length. To counteract this, we propose Persistent Visual Memory (PVM), a lightweight learnable module designed to ensure sustained, on-demand visual perception. Integrated as a parallel branch alongside the Feed-Forward Network (FFN) in LVLMs, PVM establishes a distance-agnostic retrieval pathway that directly provides visual embeddings for precise visual perception, thereby structurally mitigating the signal suppression inherent to deep generation. Extensive experiments on Qwen3-VL models demonstrate that PVM brings notable improvements with negligible parameter overhead, delivering consistent average accuracy gains across both 4B and 8B scales, particularly in complex reasoning tasks that demand persistent visual perception. Furthermore, in-depth analysis reveals that PVM can resist length-induced signal decay and accelerate internal prediction convergence.

Weigao Sun, Jiaxi Hu, Yucheng Zhou et al.

Large Language Models (LLMs) have delivered impressive results in language understanding, generation, reasoning, and pushes the ability boundary of multimodal models. Transformer models, as the foundation of modern LLMs, offer a strong baseline with excellent scaling properties. However, the traditional transformer architecture requires substantial computations and poses significant obstacles for large-scale training and practical deployment. In this survey, we offer a systematic examination of innovative LLM architectures that address the inherent limitations of transformers and boost the efficiency. Starting from language modeling, this survey covers the background and technical details of linear and sparse sequence modeling methods, efficient full attention variants, sparse mixture-of-experts, hybrid model architectures incorporating the above techniques, and emerging diffusion LLMs. Additionally, we discuss applications of these techniques to other modalities and consider their wider implications for developing scalable, resource-aware foundation models. By grouping recent studies into the above category, this survey presents a blueprint of modern efficient LLM architectures, and we hope this could help motivate future research toward more efficient, versatile AI systems.

Zhenhua Liu, Lijun Li, Ruizhe Chen et al.

While guided decoding, especially value-guided methods, has emerged as a cost-effective alternative for controlling language model outputs without re-training models, its effectiveness is limited by the accuracy of the value function. We identify that this inaccuracy stems from a core distributional gap: existing methods train static value functions on trajectories sampled exclusively from the base policy, which inherently confines their training to a narrow and suboptimal view of the potential output space. We propose Iterative Value Refinement, a novel framework designed to bridge this gap. It employs Value Exploration to provide a more comprehensive and robust training signal, complemented by Iterative Self-Refinement, which uses the improved value function from one iteration to guide the generation of higher-quality data for the next. Extensive experiments on text summarization, multi-turn dialogue, and instruction following demonstrate the effectiveness of our framework in aligning language models. Our approach not only achieves alignment but also significantly reduces computational costs by leveraging principled value function optimization for efficient and effective control.

Zhenhua Liu, Tong Zhu, Jianxiang Xiang et al.

Data augmentation (DA) is crucial to mitigate model training instability and over-fitting problems in low-resource open-domain dialogue generation. However, traditional DA methods often neglect semantic data diversity, restricting the overall quality. Recently, large language models (LLM) have been used for DA to generate diversified dialogues. However, they have limited controllability and tend to generate dialogues with a distribution shift compared to the seed dialogues. To maximize the augmentation diversity and address the controllability problem, we propose \textbf{S}ummary-based \textbf{D}ialogue \textbf{A}ugmentation with LLM (SDA). Our approach enhances the controllability of LLM by using dialogue summaries as a planning tool. Based on summaries, SDA can generate high-quality and diverse dialogue data even with a small seed dataset. To evaluate the efficacy of data augmentation methods for open-domain dialogue, we designed a clustering-based metric to characterize the semantic diversity of the augmented dialogue data. The experimental results show that SDA can augment high-quality and semantically diverse dialogues given a small seed dataset and an LLM, and the augmented data can boost the performance of open-domain dialogue models.

Chuanyuan Tan, Wenbiao Shao, Hao Xiong et al.

Handling unanswerable questions (UAQ) is crucial for LLMs, as it helps prevent misleading responses in complex situations. While previous studies have built several datasets to assess LLMs' performance on UAQ, these datasets lack factual knowledge support, which limits the evaluation of LLMs' ability to utilize their factual knowledge when handling UAQ. To address the limitation, we introduce a new unanswerable question dataset UAQFact, a bilingual dataset with auxiliary factual knowledge created from a Knowledge Graph. Based on UAQFact, we further define two new tasks to measure LLMs' ability to utilize internal and external factual knowledge, respectively. Our experimental results across multiple LLM series show that UAQFact presents significant challenges, as LLMs do not consistently perform well even when they have factual knowledge stored. Additionally, we find that incorporating external knowledge may enhance performance, but LLMs still cannot make full use of the knowledge which may result in incorrect responses.

Baiting Chen, Tong Zhu, Jiale Han et al.

Large Language Models (LLMs) have demonstrated strong generative capabilities but remain prone to inconsistencies and hallucinations. We introduce Peer Elicitation Games (PEG), a training-free, game-theoretic framework for aligning LLMs through a peer elicitation mechanism involving a generator and multiple discriminators instantiated from distinct base models. Discriminators interact in a peer evaluation setting, where utilities are computed using a determinant-based mutual information score that provably incentivizes truthful reporting without requiring ground-truth labels. We establish theoretical guarantees showing that each agent, via online learning, achieves sublinear regret in the sense their cumulative performance approaches that of the best fixed truthful strategy in hindsight. Moreover, we prove last-iterate convergence to a truthful Nash equilibrium, ensuring that the actual policies used by agents converge to stable and truthful behavior over time. Empirical evaluations across multiple benchmarks demonstrate significant improvements in factual accuracy. These results position PEG as a practical approach for eliciting truthful behavior from LLMs without supervision or fine-tuning.

Ningning Yao, Huan Xi, Lang Chen et al.

Despite policymakers deploying various tools to mitigate emissions of ozone (O\textsubscript{3}) precursors, such as nitrogen oxides (NO\textsubscript{x}), carbon monoxide (CO), and volatile organic compounds (VOCs), the effectiveness of policy combinations remains uncertain. We employ an integrated framework that couples structural break detection with machine learning to pinpoint effective interventions across the building, electricity, industrial, and transport sectors, identifying treatment effects as abrupt changes without prior assumptions about policy treatment assignment and timing. Applied to two decades of global O\textsubscript{3} precursor emissions data, we detect 78, 77, and 78 structural breaks for NO\textsubscript{x}, CO, and VOCs, corresponding to cumulative emission reductions of 0.96-0.97 Gt, 2.84-2.88 Gt, and 0.47-0.48 Gt, respectively. Sector-level analysis shows that electricity sector structural policies cut NO\textsubscript{x} by up to 32.4\%, while in buildings, developed countries combined adoption subsidies with carbon taxes to achieve 42.7\% CO reductions and developing countries used financing plus fuel taxes to secure 52.3\%. VOCs abatement peaked at 38.5\% when fossil-fuel subsidy reforms were paired with financial incentives. Finally, hybrid strategies merging non-price measures (subsidies, bans, mandates) with pricing instruments delivered up to an additional 10\% co-benefit. These findings guide the sequencing and complementarity of context-specific policy portfolios for O\textsubscript{3} precursor mitigation.

Han Han, Tong Zhu, Xiang Zhang et al.

Large language models (LLMs) combined with tool learning have gained impressive results in real-world applications. During tool learning, LLMs may call multiple tools in nested orders, where the latter tool call may take the former response as its input parameters. However, current research on the nested tool learning capabilities is still under-explored, since the existing benchmarks lack relevant data instances. To address this problem, we introduce NesTools to bridge the current gap in comprehensive nested tool learning evaluations. NesTools comprises a novel automatic data generation method to construct large-scale nested tool calls with different nesting structures. With manual review and refinement, the dataset is in high quality and closely aligned with real-world scenarios. Therefore, NesTools can serve as a new benchmark to evaluate the nested tool learning abilities of LLMs. We conduct extensive experiments on 22 LLMs, and provide in-depth analyses with NesTools, which shows that current LLMs still suffer from the complex nested tool learning task.

Tianwen Tang, Tong Zhu, Haodong Liu et al.

Zero-shot dialogue state tracking (DST) transfers knowledge to unseen domains, reducing the cost of annotating new datasets. Previous zero-shot DST models mainly suffer from domain transferring and partial prediction problems. To address these challenges, we propose Mixture of Prefix Experts (MoPE) to establish connections between similar slots in different domains, which strengthens the model transfer performance in unseen domains. Empirical results demonstrate that MoPE-DST achieves the joint goal accuracy of 57.13% on MultiWOZ2.1 and 55.40% on SGD.

Hongwei Liu, Junnan Liu, Shudong Liu et al.

The rapid advancement of Large Language Models (LLMs) has led to performance saturation on many established benchmarks, questioning their ability to distinguish frontier models. Concurrently, existing high-difficulty benchmarks often suffer from narrow disciplinary focus, oversimplified answer formats, and vulnerability to data contamination, creating a fidelity gap with real-world scientific inquiry. To address these challenges, we introduce ATLAS (AGI-Oriented Testbed for Logical Application in Science), a large-scale, high-difficulty, and cross-disciplinary evaluation suite composed of approximately 800 original problems. Developed by domain experts (PhD-level and above), ATLAS spans seven core scientific fields: mathematics, physics, chemistry, biology, computer science, earth science, and materials science. Its key features include: (1) High Originality and Contamination Resistance, with all questions newly created or substantially adapted to prevent test data leakage; (2) Cross-Disciplinary Focus, designed to assess models' ability to integrate knowledge and reason across scientific domains; (3) High-Fidelity Answers, prioritizing complex, open-ended answers involving multi-step reasoning and LaTeX-formatted expressions over simple multiple-choice questions; and (4) Rigorous Quality Control, employing a multi-stage process of expert peer review and adversarial testing to ensure question difficulty, scientific value, and correctness. We also propose a robust evaluation paradigm using a panel of LLM judges for automated, nuanced assessment of complex answers. Preliminary results on leading models demonstrate ATLAS's effectiveness in differentiating their advanced scientific reasoning capabilities. We plan to develop ATLAS into a long-term, open, community-driven platform to provide a reliable "ruler" for progress toward Artificial General Intelligence.

Tong Zhu, Yan Meng, Haotian Hu et al.

Although the use of pay-per-click mechanisms stimulates the prosperity of the mobile advertisement network, fraudulent ad clicks result in huge financial losses for advertisers. Extensive studies identify click fraud according to click/traffic patterns based on dynamic analysis. However, in this study, we identify a novel click fraud, named humanoid attack, which can circumvent existing detection schemes by generating fraudulent clicks with similar patterns to normal clicks. We implement the first tool ClickScanner to detect humanoid attacks on Android apps based on static analysis and variational AutoEncoder (VAE) with limited knowledge of fraudulent examples. We define novel features to characterize the patterns of humanoid attacks in the apps' bytecode level. ClickScanner builds a data dependency graph (DDG) based on static analysis to extract these key features and form a feature vector. We then propose a classification model only trained on benign datasets to overcome the limited knowledge of humanoid attacks. We leverage ClickScanner to conduct the first large-scale measurement on app markets (i.e.,120,000 apps from Google Play and Huawei AppGallery) and reveal several unprecedented phenomena. First, even for the top-rated 20,000 apps, ClickScanner still identifies 157 apps as fraudulent, which shows the prevalence of humanoid attacks. Second, it is observed that the ad SDK-based attack (i.e., the fraudulent codes are in the third-party ad SDKs) is now a dominant attack approach. Third, the manner of attack is notably different across apps of various categories and popularities. Finally, we notice there are several existing variants of the humanoid attack. Additionally, our measurements demonstrate the proposed ClickScanner is accurate and time-efficient (i.e., the detection overhead is only 15.35% of those of existing schemes).

Tong Zhu, Haitao Wang, Junjie Yu et al.

In recent years, distantly-supervised relation extraction has achieved a certain success by using deep neural networks. Distant Supervision (DS) can automatically generate large-scale annotated data by aligning entity pairs from Knowledge Bases (KB) to sentences. However, these DS-generated datasets inevitably have wrong labels that result in incorrect evaluation scores during testing, which may mislead the researchers. To solve this problem, we build a new dataset NYTH, where we use the DS-generated data as training data and hire annotators to label test data. Compared with the previous datasets, NYT-H has a much larger test set and then we can perform more accurate and consistent evaluation. Finally, we present the experimental results of several widely used systems on NYT-H. The experimental results show that the ranking lists of the comparison systems on the DS-labelled test data and human-annotated test data are different. This indicates that our human-annotated data is necessary for evaluation of distantly-supervised relation extraction.

Jinzhe Zeng, Liqun Cao, Mingyuan Xu et al.

Understanding and prediction of the chemical reactions are fundamental demanding in the study of many complex chemical systems. Reactive molecular dynamics (MD) simulation has been widely used for this purpose as it can offer atomic details and can help us better interpret chemical reaction mechanisms. In this study, two reference datasets were constructed and corresponding neural network (NN) potentials were trained based on them. For given large-scale reaction systems, the NN potentials can predict the potential energy and atomic forces of DFT precision, while it is orders of magnitude faster than the conventional DFT calculation. With these two models, reactive MD simulations were performed to explore the combustion mechanisms of hydrogen and methane. Benefit from the high efficiency of the NN model, nanosecond MD trajectories for large-scale systems containing hundreds of atoms were produced and detailed combustion mechanism was obtained. Through further development, the algorithms in this study can be used to explore and discovery reaction mechanisms of many complex reaction systems, such as combustion, synthesis, and heterogeneous catalysis without any predefined reaction coordinates and elementary reaction steps.

Haitao Wang, Zhengqiu He, Tong Zhu et al.

The CCKS2019 shared task was devoted to inter-personal relationship extraction. Given two person entities and at least one sentence containing these two entities, participating teams are asked to predict the relationship between the entities according to a given relation list. This year, 358 teams from various universities and organizations participated in this task. In this paper, we present the task definition, the description of data and the evaluation methodology used during this shared task. We also present a brief overview of the various methods adopted by the participating teams. Finally, we present the evaluation results.