Large language model (LLM)-based multi-agent systems (MASs) have shown impressive performance in solving a wide range of complex problems. However, previous studies mainly focus on designing customized MAS for specific tasks, while a critical research problem remains unclear: Do LLM agent groups exhibit a form of ``general intelligence'' that reflects their general ability across various tasks? Researchers have found a Collective Intelligence (CI) factor in human groups that captures their general capability. Inspired by this, in this study, we aim to investigate whether an analogous CI factor also exists in LLM agent groups, which is crucial for building generalizable MAS. Motivated by human cognitive psychology experiments, we construct 108 LLM agent groups with diverse group sizes, LLM compositions, and communication topologies. We systematically evaluate these groups across a wide range of tasks and analyze their performances. Our results demonstrate that an Artificial Collective Intelligence (ACI) factor can be extracted from LLM agent groups to predict the generalization performance on new tasks. Inspired by this, we train a model to predict the ACI based on the features of MAS, and show that it can be used as a plug-in to enhance the generalization ability of MAS optimization methods.

The reproducibility of artificial intelligence research is hindered by the complexity of configuring execution environments for "wild" academic repositories, which often suffer from missing dependencies, hardcoded paths, and non-standard structures. Existing automated tools fail to bridge the gap between standardized software engineering and the chaotic nature of research artifacts. In this paper, we introduce Auto-Reproduce, a novel multi-agent framework designed to automate the environment construction and reproduction of scientific code. Our approach introduces three key innovations. First, we propose a Context-Aware Standardization mechanism using a non-intrusive "Shim Architecture." By leveraging Large Language Models (LLMs) to parse paper methodologies and code ASTs, we synthesize external adapters—such as mocked data loaders and config shims—to standardize repositories without altering core logic. Second, to address the data scarcity in environment debugging, we develop Prog-Env, a trajectory generation strategy that injects topologically progressive errors—ranging from atomic mismatches to coupled dependency cycles—into functional environments, training agents to solve complex constraint satisfaction problems. Third, we implement a Structured Long-Term Memory that indexes hierarchical error fingerprints and stores counterfactual negative constraints to prevent recurring installation failures.

Dynamic origin-destination (OD) flow generation aims to synthesize plausible mobility patterns under given temporal contexts without access to historical OD observations. This setting requires translating semantic temporal cues into temporally coherent mobility patterns while preserving spatial heterogeneity. In this paper, we present DynaOD, a modular and scalable framework for dynamic OD generation. DynaOD represents temporal semantics from two complementary perspectives: discrete directional trends that capture qualitative changes in urban activities, and continuous temporal evolution that characterizes how such changes unfold over time. By integrating these temporal semantics, DynaOD constructs time-varying region features that drive dynamic OD generation using pretrained static OD models in a plug-and-play manner. DynaOD further supports cross-city generalization and scalable deployment through lightweight modular designs. Experiments on large-scale real-world datasets show that DynaOD consistently outperforms representative baselines, producing dynamic OD flows with improved accuracy and distributional fidelity.

The large-scale deployment of wind and solar resources has fundamentally reshaped the operating regime of power systems, giving rise to new electricity markets in which price volatility and system imbalance risks are significantly amplified. In this context, peak shaving and valley filling are no longer auxiliary tasks but central operational requirements. Virtual power plants (VPPs) have therefore emerged as a key intermediary in modern power systems, bridging market signals and the physical dispatch of distributed resources. Through a combination of market-based and administrative incentives, VPPs are authorized to participate in electricity markets at the aggregation level, commit to scheduled outputs, and coordinate heterogeneous controllable resources within their jurisdiction to meet pre-defined delivery targets. Concretely, a VPP must simultaneously satisfy system constraints and market rules while dispatching devices distributed across multiple nodes with highly heterogeneous flexibility, response speeds, and operational costs. Its objective is to maximize arbitrage revenue while controlling deviation risks under uncertain renewable generation and fluctuating prices. This leads to a fundamental coordination problem: how to effectively coordinate heterogeneous devices distributed across multiple nodes so that a VPP can fulfill its market commitments and optimize economic performance without incurring excessive deviation penalties. In this work, we present our recent progress in exploring the capacity of large language models to evolve coordination strategies in this dynamic and complex setting. We will introduce the model design and experimental setup in detail, and we warmly welcome feedback and suggestions.

Agents powered by large language models (LLMs) are becoming key tools for enhancing the user experience of smartphones. The latest advances have enabled agents to shift from reactively responding to user instructions to proactively offering support. However, existing proactive agents only utilize the immediate context observable in the current environment, neglecting the persistent context of users' long-term phone usage history. As a result, they are unable to extract user preferences and habits from history, making it difficult for proactive assistance to align with specific users' intents. To address these issues, we propose FingerTip-V2, the first proactive agent framework that fully leverages the historical intent sequence of users' phone usage. FingerTip-V2 retrieves the most relevant historical intents to the current context and dynamically maintains a semantically rich user persona, ultimately predicting users' current intents and proactively suggesting tasks without user instructions. Additionally, to verify the generalization ability of FingerTip-V2, we have created a diverse synthetic dataset containing 100 different users and 20,000 long-term intent sequences using persona-driven simulation, which enables LLMs to imitate human intents. We have further developed an Android app and conducted user experiments to evaluate the performance of FingerTip-V2 in real-world usage scenarios.

Inferring a network's evolutionary history from a single final snapshot with limited temporal annotations is fundamental yet challenging. Existing approaches predominantly rely on topology alone, which often provides insufficient and noisy cues. This paper leverages network steady-state dynamics—converged node states under a given dynamical process—as an additional and widely accessible observation for network evolution history inference. We propose CS², which explicitly models structure–state coupling to capture how topology modulates steady states and how the two signals jointly enhance edge discrimination for formation-order recovery. Evaluations on multiple real temporal networks under diverse dynamical processes indicate that CS² provides more accurate and robust edge-ordering estimates than competitive baselines across both local and global ordering criteria. Beyond edge-level ordering, CS² is able to recover salient macroscopic evolutionary patterns, including clustering formation, degree heterogeneity, and hub growth. Furthermore, a steady-state-only variant remains competitive when reliable topology information is limited, highlighting steady states as an independent and informative signal for network evolution inference.

Driven by the "Double Carbon" goals (2030/2060), China is witnessing an explosive expansion in wind and solar capacity. However, the stability of this transition is increasingly threatened by climate change, which induces frequent, high-intensity, and compound extreme weather events. Traditional evaluation methods—limited by coarse granularity and physical simplifications—fail to capture the extreme nonlinear volatility and micro-scale power fluctuations caused by these events. To address this gap, we propose a novel pipeline: Extreme Weather Risks to China’s Wind and Solar Energy Quantified by Machine Learning. Our approach leverages AI models to identify the spatial distribution of wind and solar installations and assesses their power generation potential on a national scale. By capturing the spatio-temporal coupling of extreme events, we provide finer-grained and more precise risk quantification across China. This framework offers critical insights for renewable site selection, energy integration, and resilient policy formulation.

Current scientific data infrastructures, such as OpenAlex and AMiner, serve as the cornerstone of "AI for Science." However, they suffer from a fundamental limitation: they are merely static snapshots of massive, noisy data. Existing maintenance methods, relying on heuristic rules or one-off model predictions, fail to handle complex ambiguities (e.g., author name conflicts) and lack the adaptability to process new information. This leads to high maintenance costs and performance degradation over time. To bridge the gap between static accumulation and dynamic maintenance, we propose SciDataAgent, a multi-agent framework designed as a self-evolving data operating system. Unlike traditional cleaning pipelines, SciDataAgent orchestrates multiple expert agents to mimic the workflow of human data scientists. Crucially, it integrates external tools (e.g., Web Search, full-text parsing) to ground decisions in real-world evidence, significantly mitigating hallucinations in Large Language Models (LLMs). This enables the system to autonomously resolve complex entity conflicts and continuously update the knowledge graph. Our experiments demonstrate that SciDataAgent achieves superior accuracy on challenging curation benchmarks. This is an ongoing work. In the future, we hope to optimize the system and improve its performance in downstream tasks (such as complex citation prediction and scientific question answering), providing a scalable paradigm for building a sustainable, high-quality scientific data infrastructure.

Probabilistic workload forecasting is crucial for efficient resource management. However, mapping job information to computing workloads involves high-dimensional uncertainty. This volatility is driven by both temporal dynamics, arising from periodic user behavior, and job-induced dynamics, caused by distributed tasks running across multiple machines. To address this, we propose a Dual-Branch Diffusion model. First, we identify implicit connections between machines by analyzing job co-occurrence. Second, we introduce a parallel dual-branch structure to separate the modeling process. One branch focuses on smooth temporal trends, while the other captures the sharp, job-induced structural fluctuations. By separating these dynamics, our approach effectively captures high-dimensional uncertainty and spatial correlations among machines, leading to precise predictions

As trajectory prediction models enter the scaling-law regime, both parameter counts and predictive accuracy have increased rapidly. However, the entropy-based predictability theory of Song et al. (2010), formulated for independent discrete sequences, is increasingly misaligned with modern deep learning trained on high-dimensional, population-scale mobility data. In this work, we introduce a predictability framework designed for representation learning and validate it on a large-scale pedestrian trajectory dataset. Across models and data scales, our metric more reliably anticipates attainable performance ceilings than classical entropy-based measures. Beyond model evaluation, the proposed predictability enables a quantitative analysis of how urban socioeconomic conditions—such as income, unemployment, and commuting patterns—relate to the regularity and predictability of human mobility.

Development in Artificial Intelligence (AI) has accelerated scientific discovery. Alongside recent AI-oriented Nobel prizes, these trends establish the role of AI tools in science. This advancement raises questions about the potential influences of AI tools on scientists and science. To evaluate, we used a pretrained language model to identify AI-augmented research, with an F1-score of 0.875 in validation against expert-labeled data, constructing a dataset of 41.3 million research papers across natural science and covering distinct eras of AI. Through comprehensive analyses, we uncover (1) a potential conflict between individual and collective science, where AI leads to an expansion of individual scientists' impact but a contraction in collective science's reach; (2) the emerging and infiltrating patterns of AI in conventional science fields, where AI first emerges in focused spots and then gradually spread the entire field, with the speed of such infiltration in different fields influenced by the data abundance, the collaboration willingness with AI researchers, and etc.; (3) the evolution of author collaboration modes in AI for Science, where more and more works are led by conventional science authors, and highly qualified scientists are willing to adopt AI in their research. These findings reveal the future potential and current limitations of AI for Science, providing enlightens to its development and elucidating on the paradigm shift from AI for Science to AI Scientist.

High-dimensional observational data typically lie on low-dimensional manifolds characterized by complex topological structures. Due to the continuity constraints inherent in global coordinate systems, Euclidean autoencoders struggle to compress such data to its intrinsic dimension without inducing topological distortions, such as tearing or overlapping. Inspired by the concept of "Atlas" in manifold theory, we propose a novel framework that integrates Diffusion Models with a Mixture of Experts (MoE) mechanism to losslessly learn the intrinsic geometry of data. Our method leverages the dynamical properties of diffusion models, effectively separating noise from manifold signals by analyzing symmetry breaking and entropy collapse phases during the denoising process. To overcome the topological limitations associated with a global latent space, we incorporate an MoE mechanism within the bottleneck layer of a diffusion autoencoder. In this architecture, distinct expert networks adaptively specialize in local manifold regions, collectively forming an atlas that covers the entire topology. This design allows the latent space dimension to strictly match the data's intrinsic dimension.

Since 2020, our laboratory has been dedicated to the research and development of urban simulators, achieving significant milestones. As we stand in 2025, amidst the emerging era of Large Models and World Models, it is imperative to chart the course for the next generation of urban simulators. This report aims to explore this direction from the perspective of constructing Urban World Models. By integrating Foundation Models, Intelligent Agents, and Physics Engines into a unified framework, we propose a new vision that we hope will offer fresh insights and inspiration to the community.

As Large Language Models evolve into specialized personal agents, aligning them with individual user preferences has become essential. This brief introduction categorizes the recent personalization landscape into three technical layers of intervention. At the Prompt Layer, we focus on few-shot learning and cognitive memory architectures that track evolving user beliefs. The States Layer examines the real-time manipulation of internal model representations during inference, utilizing activation steering and latent feature modeling. Finally, the Parameter Layer analyzes scalable adaptation through parameter-efficient fine-tuning and individualized preference alignment. By synthesizing these approaches, we identify key challenges and promising directions in LLM personalization area.

Event-aware spatio-temporal forecasting aims to predict system dynamics under the influence of internal and external events. However, existing methods are largely limited by their reliance on historical time series and by the lack of datasets with well-aligned event annotations, without explicitly modeling the impact of planned future events. To support rigorous evaluation, we first construct an event-aware metro flow dataset based on the New York City subway system. The dataset integrates station-level passenger flows with real-world traffic system-related news and further incorporates weather conditions and point-of-interest (POI) information. Importantly, events in our dataset are aligned in both time and space, and include future scheduled events consistent with real-world announcements, enabling systematic validation of event-aware forecasting under planned-event scenarios. Motivated by the effectiveness and generalization of spatio-temporal foundation models (STFMs), we propose an event-aware multimodal STFM post-training manner to improve forecasting accuracy with sparse events.

Crowd simulation holds crucial applications in various domains, such as urban planning, emergency response, and traffic management. However, existing approaches typically emphasize either macroscopic crowd counting or microscopic pedestrian simulation, and thus struggle to ensure consistency with real-world phenomena at both the level of macroscopic density evolution and that of microscopic individual trajectories. To address this challenge, we propose a density-guided crowd simulation framework that jointly enforces scene-level crowd dynamics and individual-level behavioral fidelity within a unified formulation. Specifically, we propose a crowd emitter that follows macroscopic density guidance to progressively generate and place new agents over time, while assigning each agent the corresponding attributes. We further introduce a crowd simulator for trajectory prediction that explicitly models intrinsic intention and extrinsic interaction, where a spatiotemporal aggregation module extracts interaction-aware representations and an intention module encodes goal-driven intention features. Under macroscopic density constraints, the proposed framework enables more realistic crowd simulation.

Urban mobility networks shape daily social encounters, yet their role in reinforcing segregation and inequality remains mechanistically opaque. This talk introduces an adaptive dynamics framework to model how community-place movement patterns co-evolve with socioeconomic structures. By integrating theoretical analysis and large-scale simulations, we recover empirical mobility distributions and reveal phase transitions where small changes in urban scale or heterogeneity trigger abrupt shifts in segregation regimes. Crucially, we demonstrate that segregation and mobility inequality are intrinsically coupled through feedback loops in human movement—not merely correlated. Leveraging this model, we simulate targeted interventions such as spatial reconfiguration and accessibility enhancements, showing they can significantly disrupt segregation cycles without requiring behavioral overhauls. Our work provides a foundation for designing equitable urban systems where reduced segregation aligns with practical planning, bridging network science and sustainable development.

Accurate and comprehensive evaluation of world models is important for building reliable embodied agents and decision-making systems. Existing evaluation practices mostly build on image and video-level metrics, sometimes extended into multi-dimensional benchmarks that score aspects such as visual fidelity, temporal consistency, and text alignment. More recent efforts also begin to probe whether models respect basic physical laws and commonsense. However, these approaches are typically designed in isolation, remain largely video-centric, and only partially capture the capabilities that matter for downstream control and planning. In this talk, we first review representative evaluation metrics and benchmarks for video generation, highlighting what aspects of model behavior they each capture and where they fall short for world-model use cases. Based on this review, we then outline a preliminary evaluation framework for world models that explicitly emphasizes three aspects: video generation quality, adherence to physical laws, and instruction controllability. Our goal is to connect video-level assessment with the needs of embodied settings, and to take a first step toward more holistic, comparable, and practically meaningful evaluation of world models.

World models are generative systems that learn to predict an environment’s evolution conditioned on actions, enabling agents to reason, plan, and simulate future trajectories without requiring direct interaction. Recent advances in video diffusion models have made them a powerful backbone for world modeling, driving rapid progress in domains such as model-based reinforcement learning, video prediction, gaming, and embodied AI.Despite this progress, current video generation models struggle from 3D consistency, object appearance drifts over time, and the scene shifts inconsistently across frames in the video. To address these challenges, we introduce explicit 3D GS features to guide video generation toward better spatial coherence and geometric fidelity. Specifically, we jointly train the video generation model with a 3D Gaussian Splatting reconstruction module to ensure that the diffusion-based video generator preserves the underlying 3D information.

Millions of urban residents suffer from unequal heat exposure due to intensifying climate change and static urban risk assessments, and precise exposure monitoring is critical for sustainable urban development. Existing satellite-only or residence-based methods are either too coarse-grained to capture street-level microclimates or ignore human mobility, yielding biased assessments that overlook the "heat traps" faced by vulnerable populations. In this presentation, we first conduct an analytical study: we quantify cross-sectional disparities in baseline heat exposure across demographic groups, and further investigate longitudinal inequalities in exposure intensification, determining whether vulnerable populations face a disproportionately steeper increase in heat burden as temperatures rise. Subsequently, we present a physics-aware multimodal learning approach to automatically map fine-grained dynamic heat exposure.We propose a novel fusion architecture to integrate macro-scale satellite imagery and micro-scale street-view imagery.Crucially, to capture human-centric dynamics, we devise a teacher-student distillation strategy: a teacher model utilizes hourly mobility profiles as "privileged information" to learn diurnal activity rhythms, and transfers this dynamic knowledge to an imagery-only student model, enabling scalable deployment without sensitive data. Extensive experiments on eight major U.S. metropolitan areas verify the effectiveness of our model, which significantly reduces mean absolute error by roughly 32% against state-of-the-art baselines and reveals distinct socioeconomic heat gradients.

Accurately forecasting chaotic systems, prevalent in domains such as weather prediction and fluid dynamics, remains a significant scientific challenge. The inherent sensitivity of these systems to initial conditions, coupled with a scarcity of observational data, severely constrains traditional modeling approaches. Since these models are typically trained for a specific system, they lack the generalization capacity necessary for real-world applications, which demand robust zero-shot or few-shot forecasting on novel or data-limited scenarios. To overcome this generalization barrier, we propose ChaosNexus, a foundation model pre-trained on a diverse corpus of chaotic dynamics. ChaosNexus employs a novel multi-scale architecture named ScaleFormer augmented with Mixture-of-Experts layers, to capture both universal patterns and system-specific behaviors. The model demonstrates state-of-the-art zero-shot generalization across both synthetic and real-world benchmarks. On a large-scale testbed comprising over 9,000 synthetic chaotic systems, it demonstrates notable improvements in the fidelity of long-term attractor statistics while achieving competitive short-term forecasting accuracy compared to the leading baseline. This robust performance extends to real-world applications with exceptional data efficiency. For instance, in 5-day global weather forecasting, ChaosNexus achieves a competitive zero-shot mean error below 1°C—a result that further improves with few-shot fine-tuning. Moreover, experiments on the scaling behavior of ChaosNexus provide a guiding principle for scientific foundation models: cross-system generalization stems from the diversity of training systems, rather than sheer data volume.

Accurate channel state information (CSI) is crucial for wireless communication systems, yet traditional pilot-based estimation becomes inefficient in 6G scenarios with massive antennas. Existing deep learning approaches alleviate pilot overhead but remain constrained, as they are often tailored to a single task or system configuration, limiting their scalability and adaptability. We propose a foundation model for CSI modeling. The proposed framework leverages diffusion models as the backbone to capture the probabilistic characteristics of signal propagation. Additionally, our model adopts a unified structure that incorporates multi-masking strategies, enabling the joint learning of CSI prediction and estimation, and thus allowing the model to generalize across diverse CSI modeling tasks. Furthermore, the model applies context-specific encoders to map contextual information into latent space, serving as conditions to guide the modeling process. The designs enhance the model’s adaptability in diverse scenarios of varying frequency bands, paving the way for enhanced wireless network arrangement and optimization.

Social stratification profoundly shapes individuals' life chances, access to resources, and overall well-being and has been as a central and enduring concern within sociology. Although abundant existing research has provided invaluable insights to social stratification, it often faces limitations in capturing the dynamics and mechanisms of social structures and fails to unveil the complex interplay between individual behaviors and broader social forces. We propose to unpack and inform social stratification by employing Large Language Model agent-based social simulations. Through equipping agents with diverse realistic socioeconomic profiles and the capacities for complex reasoning, dynamic decision-making, and nuanced interactions, we endeavor not only to replicate established patterns regarding social stratification, but also to delve into the underlying mechanisms, explore plausible counterfactuals, and derive actionable insights to understand and alleviate social stratification.

Recent work on diffusion models for reinforcement learning demonstrates that diffusion-based policies can represent rich, multimodal action distributions and generate stable long-horizon behaviors through iterative denoising. These properties yield improved robustness and flexibility compared with conventional policy parameterizations.Applying diffusion policies to wireless network optimization is a promising direction. Wireless control tasks—such as power control, beamforming, and scheduling—feature large action spaces, strong non-convexity, and dynamic uncertainty. Diffusion models can explore these spaces more effectively, offer diverse candidate solutions, and embed structural priors into the policy. Combined with RL-based resource allocation and multi-agent coordination, diffusion policies may enable more adaptive and efficient optimization for next-generation wireless systems.

The generation of high-quality 3D worlds represents a critical research frontier with profound implications for immersive media, large-scale simulation, and the development of embodied intelligence and world models. While recent works have made strides in scalability, quality, and reality-alignment, challenges remain, particularly regarding the structural consistency of buildings and road networks. Perception and 3D generation are identified as two pivotal factors in synthesizing reality-aligned 3D urban environments. Integrating multimodal information beyond standard street views and OSM shows significant potential in guiding 3D generation. Furthermore, the emerging paradigm of unified multimodal models—capable of image editing, understanding, and generation—offers new possibilities for urban information retrieval and curation. Regarding 3D generation, while general image-to-3D models exhibit promising performance, developing robust and efficient domain-specific models remains essential.

Large language models (LLMs) acquire extensive capabilities through pretraining and supervised fine tuning, yet reinforcement learning (RL) based post training often yields superior performance in reasoning tasks. Despite this success, the internal mechanistic shifts driving these enhancements remain underexplored. In this study, we employ Edge Attribution Patching (EAP) to investigate the internal circuitry of LLMs before and after RL training. Our analysis across multiple model families reveals that online RL methods, such as PPO and GRPO, systematically alter internal information flow by increasing Activation Intensity and Information Complexity while decreasing Distribution Kurtosis. These changes suggest that online RL recruits broader, more redundant, and diverse functional pathways, whereas Direct Preference Optimization (DPO) fails to consistently induce these structural shifts in some scenarios due to its relatively more static data support. To validate the causality between these internal changes and reasoning capability, we conduct controlled experiments on the Qwen2.5 series using DAPO with modulated sampling temperatures. We demonstrate that suppressing the development of activation intensity and complexity through higher temperature sampling directly correlates with degraded performance on mathematical benchmarks. We posit that the dynamic on policy sampling inherent to online RL expands the exploration state space, necessitating the activation of latent circuits that remain dormant under supervised or preference based optimization. Our findings bridge empirical performance gains with mechanistic interpretability, highlighting internal circuit diversity as a pivotal driver of robust generalization.

The growing demand for large language models (LLMs) with tunable reasoning capabilities in many real-world applications highlights a critical need for methods that can efficiently produce a spectrum of models balancing reasoning depth and computational cost. Model merging has emerged as a promising, training-free technique to address this challenge by arithmetically combining the weights of a general-purpose model with a specialized reasoning model. While various merging techniques exist, their potential to create a spectrum of models with fine-grained control over reasoning abilities remains largely unexplored. This work presents a large-scale empirical study evaluating a range of model merging techniques across multiple reasoning benchmarks. We systematically vary merging strengths to construct accuracy-efficiency curves, providing the first comprehensive view of the tunable performance landscape. Our findings reveal that model merging offers an effective and controllable method for calibrating the trade-off between reasoning accuracy and token efficiency, even when parent models have highly divergent weight spaces. Crucially, we identify instances of Pareto Improvement, where a merged model achieves both higher accuracy and lower token consumption than one of its parents. Our study provides the first comprehensive analysis of this tunable space, offering practical guidelines for creating LLMs with specific reasoning profiles to meet diverse application demands.

Chain-of-thought has been proven essential for enhancing the complex reasoning abilities of Large Language Models (LLMs), but it also leads to high computational costs. Recent advances have explored the method to route queries among multiple models and proved it as a promising approach. However, previous works directly operate at the task level, i.e., assigning user queries to suitable LLMs, which does not allow hybrid LLMs to truly collaborate on finer-grained sub-tasks. Collaboration at the level of intermediate reasoning steps (thoughts) could enable more efficient coordination, but it also poses significant challenges for router scheduling, placing immense demands on the quality of task decomposition and the precision of the router. To address this, we propose R2-Reasoner, a novel framework centered around a Reinforced Model Router designed to efficiently scale LLM reasoning. This router orchestrates collaboration across 9 heterogeneous models, of whom the parameter scale ranges from less than 1B to hundreds of billions, by first breaking down a complex query into subtasks with a decomposer, and then assigning each subtask to the optimal model with a subtask allocator, balancing performance with cost. To train this router involves a two-stage alternating process for the decomposer and the allocator, integrating supervised fine-tuning with reinforcement learning to enable effective self-supervised refinement. Extensive experiments across six challenging reasoning benchmarks demonstrate that R2-Reasoner reduces API costs by 84.46% compared with state-of-the-art baselines while maintaining competitive reasoning accuracy. Our framework paves the way for the development of more scalable and efficient reasoning systems.

Building realistic and generalizable 3D urban environments is essential for world modeling and embodied AI. We introduce CITYRAISE, a Reality-Aligned Intelligent Synthesis Engine that automatically constructs detailed 3D urban environments from real-world geographic data. The framework combines procedural modeling and generative refinement to reconstruct buildings, road networks, and fine-grained urban objects with both geometric precision and visual realism. Unlike conventional handcrafted pipelines, CITYRAISE generalizes across different cities with minimal human intervention, enabling simulation systems and embodied agents to migrate seamlessly between regions. Experiments on multiple urban areas demonstrate its scalability, realism, and potential to accelerate city-level world modeling.

We propose a scalable multi-agent paradigm that generates agentic trajectories by asynchronously sampling task descriptions and environment parameters, embedding feedback loops without locking step-wise synchronization. The framework decouples trajectory production from runtime control, enabling arbitrary parallel expansion and continual re-mixing of experience fragments. The resulting architecture supports open-ended growth in agent count, task diversity and environmental complexity. We will first synthesize a medium-scale dataset without intermediate intervention and then conduct validation experiments on BFCL and Smithery.

Integrating memory systems with large language models (LLMs) is a key frontier in building persistent, personalized agents. However, existing frameworks, often focused on conversational recall or general long-context processing, struggle to capture the complex, domain-specific reasoning required for specialized fields like scientific research. We aim to develop a system that enables AuthorQA, a query mode designed to match human scientists in both factual accuracy and cognitive style. However, current role-playing agents lack deep knowledge grounding, while retrieval-based systems like PaperQA capture facts but overlook higher-level conceptual links. To address this, we propose ScientistMind, a hierarchical memory system that mimics cognitive consolidation. It fuses public academic data with private materials, progressively forming both factual and stylistic memories. Using a benchmark built on OpenAlex, ScientistMind demonstrates superior performance in both factual accuracy and stylistic alignment. Furthermore, inspired by the Society-of-Mind theory, we extend this framework into a multi-agent system. By leveraging a hierarchical science network and the memory system, it aims to create an “omniscient scientist” agent—enhancing knowledge integration, reducing hallucinations in cross-disciplinary Q&A, and fostering novel scientific collaboration.

We begin by contextualizing the application of RL in traditional visual tasks, elucidating its mechanisms for optimizing performance in foundational computer vision paradigms. Subsequently, the discourse shifts to the role of RL in advancing Large Language Models (LLMs), with a focus on how RL often synergized with human feedback (RLHF) facilitates alignment with human preferences, enhances factual robustness, and refines generative coherence. The discussion then extends to RL-driven advancements in Vision-Language Models (VLMs), investigating its efficacy in mitigating cross-modal misalignment and improving the models' capacity for semantic correspondence between visual and linguistic modalities. Finally, the talk delineates ongoing research endeavors centered on infusing reasoning capabilities into VLMs, exploring methodologies to enable complex logical inference, multi-step problem-solving, and context-aware decision-making in multi-modal environments.

Designing high-performing robot morphologies is a grand challenge for developing specialized autonomous agents. However, the vast, combinatorial, and non-differentiable nature of the morphological design space has been a primary obstacle. Existing methods tackle this problem indirectly, relying on either semantically-blind genetic operators or reinforcement learning with predefined modification actions, both of which constrain exploration. In this work, we introduce MorphoGen, a novel framework that reframes morphological design as a code generation problem. MorphoGen leverages large language models (LLMs) to directly iterate the XML files as codes that define an agent’s morphology, solving the original open problem without being limited by any prior constraints or fixed action spaces. Gradient-like textual guidance is provided to steer the evolution of robot morphologies through prompted mutations and crossovers. Our approach allows the LLMs to apply its understanding of structure and syntax to generate complex and semantically coherent design variations, enabling an unconstrained and efficient exploration of the design space. On a suite of challenging locomotion benchmarks, MorphoGen discovers novel and high-performing morphologies, significantly outperforming strong baselines by over 57.5% in downstream motoring evaluation.

The ability of Large Language Models (LLMs) to adopt specific user personas is crucial for personalized web services and user interactions.However, endowing LLMs with stable and distinct user personas remains a significant challenge. Existing methods like fine-tuning create a rigid, blended model that struggles to adapt to a wide range of distinct personas, while simple prompting lacks consistency and depth.To address these limitations, we propose PersonaSteer, a novel and lightweight side-network approach that enables controllable persona emulation via inference-time activation steering. We train a small side-network which learns to generate a persona-specific vector that is injected into the backbone LLM's hidden states, effectively `steering' the output style without altering the base model's weights.Experiments show that PersonaSteer achieves superior persona consistency depth. Its modular design allows a single model to flexibly adopt a wide range of distinct and even composable personas.

Mobile network traffic generation serves as a critical approach for proactively anticipating mobile user demands and capturing the dynamics of network load. Network traffic at different resolutions, including Base Station (BS)-level traffic, cell-level traffic, and gridlevel traffic, can reflect the operational states of mobile networks across various service levels, supporting diverse types of network planning and resource allocation. However, existing methods focus on generating network traffic with a single spatiotemporal resolution, making it difficult to achieve joint generation of multiresolution traffic. In this paper, we propose ZoomDiff, a denoising refinement diffusion model for multi-resolution mobile traffic generation. Aligned with the hierarchical structure of BSs, cells, and small grids in mobile networks, ZoomDiff introduces a multi-scale decomposition of noise data during the denoising process. Furthermore, it incorporates a prior noise guidance mechanism to model the correlations of mobile traffic across different spatial and temporal resolutions. This design enables the model to progressively generate traffic data at multiple granularities within a single denoising trajectory. We also incorporate a spatio-temporal positional encoding to enhance the model’s ability to understand and capture the intrinsic temporal and spatial correlations within high-dimensional traffic data. Evaluations on real-world cellular mobile network traffic datasets from Nanning City demonstrate that ZoomDiff achieves average performance improvements of at least 18.4% over state-ofthe-art baselines on generation tasks at BS traffic, 100m grid traffic, and 50m grid traffic, showcasing its strong capability in generating multi-resolution mobile traffic.

Novel research ideas play a critical role in advancing scientific inquiries. Recent advancements in Large Language Models (LLMs) have demonstrated their potential to generate novel research ideas by leveraging large-scale scientific literature. However, previous work in research ideation has primarily relied on simplistic methods, such as keyword co-occurrence or semantic similarity. These approaches focus on identifying statistical associations in the literature but overlook the complex, contextual relationships between scientific concepts, which are essential to effectively leverage knowledge embedded in human literature. For instance, papers that simultaneously mention "keyword A" and "keyword B" often present research ideas that integrate both concepts. Additionally, some LLM-driven methods propose and iteratively enhance research ideas using the model's vast internal knowledge, but they fail to effectively leverage the valuable scientific concept network, limiting the grounding of these ideas in established research. To address these challenges, we propose the \textbf{Deep Ideation} framework, which integrates a scientific network that not only captures keyword co-occurrence but also incorporates contextual relationships between keywords, providing a richer scientific foundation for LLM-driven ideation. Our framework introduces an explore-expand-evolve workflow for Deep-Ideation which integrates several key components to iteratively refine research ideas. Throughout this workflow, we maintain an Idea Stack to track research progress across iterations. To guide this search and evolution process, we integrate a critic engine trained on real-world reviewer feedback, providing continuous signals on the novelty and feasibility of generated ideas. Experimental results across multiple AI domains show that our approach significantly improves the overall quality of generated ideas by \textbf{10.67\%} compared to other methods, with the generated ideas exceeding the acceptance level of top conferences. Human evaluation highlights the practical value of the generated ideas in supporting scientific research while ablation studies further confirm the effectiveness of each component of the workflow.

Stock markets are one of the most complex systems in the modern world, where prices emerge from billions of decentralized interactions among heterogeneous participants in an ever-evolving information landscape. Building a high-fidelity stock market simulator is not only a cornerstone for understanding such complexity, but also offers a valuable testbed for anticipating and mitigating crises and disruptions. Despite decades of efforts, existing methods remain confined to an unresolved dilemma: structural fidelity often comes at the cost of non-intelligent agents, while large language model (LLM) agents can only participate in oversimplified market environments. To this end, we propose MarketSim, a large-scale stock market simulation framework with generative agents. Specifically, we first design a hierarchical multi-agent architecture. By decoupling agents’ strategic reasoning from their high-frequency actions, this architecture enables LLM agents to participate in a nanosecond-resolution, NASDAQ-like continuous double auction market. Building on this, we simulate over 15k diverse market participant agents, whose billions of interactions collectively create an evolving market environment in which agents learn from feedback and adapt their strategies accordingly. Furthermore, we ground these agents in a rich informational landscape that covers over 12k real-world news articles, policy documents, and earnings reports. To evaluate our proposed MarketSim, we develop a comprehensive benchmark that includes stocks from 8 GICS sectors and 3 representative real-world scenarios, along with 5 stylized facts for market complexity and 5 price-related statistical metrics. Extensive experiments demonstrate that MarketSim not only captures the complexity characterizing real-world markets, but also accurately tracks real-world high-frequency price dynamics with an average MAPE of 3.48%.

Harnessing publicly available, large-scale web data, such as street view and satellite imagery, urban socio-economic sensing is of paramount importance for achieving global sustainable development goals. With the emergence of Large Vision-Language Models (LVLMs), new opportunities have arisen to solve this task by treating it as a multimodal perception and understanding problem. However, recent studies reveal that LVLMs still struggle with accurate and interpretable socio-economic predictions from visual data. To address these limitations and maximize the potential of LVLMs, we introduce CityRiSE, a novel framework for Reasoning urban Socio-Economic status in LVLMs through pure reinforcement learning (RL). With carefully curated multi-modal data and verifiable reward design, our approach guides the LVLM to focus on semantically meaningful visual cues, enabling structured and goal-oriented reasoning for generalist socio-economic status prediction. Experiments demonstrate that CityRiSE with emergent reasoning process significantly outperforms existing baselines, improving both prediction accuracy and generalization across diverse urban contexts, particularly for prediction on unseen cities and unseen indicators. This work highlights the promise of combining RL and LVLMs for interpretable and generalist urban socio-economic sensing.

Large language models (LLMs) have gained significant attention for their poten-tial to replicate human participants in social science simulations. However, pre-vious works on LLM reasoning focus on enhancing the capabilities for math and logical problems, overlooking the reasoning process behind social behavior, such as controversial social attitudes, moral dilemmas, and economic games. In this study, we explore the limitations of current models and propose a new approach to improve their human-likeness in social behavioral reasoning tasks. We introduce the Social-Behavioral-Reasoning (SBR) dataset, comprising 1,560 quadruples of human profiles, social questions, reasoning processes, and final choices. Utilizing this dataset, we evaluate large reasoning models (LRMs), revealing a contradic-tion: while LRMs increase society-level diversity, they fail to maintain individual-level accuracy. Our findings further indicate that the observed increase in diversity is primarily attributed to random variation introduced by longer reasoning dura-tions, rather than improved understanding of human diversity. To address these issues, we propose the Reasoning-Enhanced-SFT method, which explicitly aligns both the reasoning and final choices with human data. Our experimental results demonstrate that our method significantly improves both in-domain and out-of-domain performance, enhancing the generalization ability across diverse social contexts. Our user study results confirm the model’s ability to produce a reasoning process more closely aligned with specific human reasoning patterns. Our work offers a new pathway to overcome the challenges that limit the use of LLMs in social simulations. Aligning model outputs with human reasoning boosts LLMs’ credibility and applicability in social science, enabling more precise and insightful simulations of human behavior.

AgentSociety shows us the feasibility of social simulation and conducting social science research through LLM agent simulation. However, over time, despite the success of AgentSociety in terms of challenge organization, research promotion, etc., we have identified the main issue that constrains AgentSociety from moving further forward: complexity. The complexity of AgentSociety comes from its adherence to the traditional programming paradigm, which tries to impose deterministic constraints on most of the aspects, especially the environment interaction and program initialization parameters, making it difficult to be understood or further extended by users. Considering the magic of natural language and LLM, we plan to design new system oriented to LLM natively, named AgentSociety V2. This system will include natural language-driven agent-environment interactions, LLM-driven invocation of environment modules, generic “social beings” capable of adapting to a wide range of different environments and tasks, and LLM Agent-driven AI social scientists. We hope it can use LLM as the soul and glue to bond different data, different environments, and different agent designs, so that it can be more flexible, scalable and easy to use when facing social science problems.

Traffic simulation is important for transportation optimization and policy making. While existing simulators such as SUMO and MAT- Sim offer fully-featured platforms and utilities, users without too much knowledge about these platforms often face significant chal- lenges when conducting experiments from scratch and applying them to their daily work. To solve this challenge, we propose Traf- ficSimAgent, an LLM-based agent framework that serves as an expert in experiment design and decision optimization for general- purpose traffic simulation tasks. The framework facilitates execu- tion through cross-level collaboration among expert agents: high- level expert agents comprehend natural language instructions with high flexibility, plan the overall experiment workflow, and invoke corresponding MCP-compatible tools on demand; meanwhile, low- level expert agents select optimal action plans for fundamental ele- ments based on real-time traffic conditions. Extensive experiments across multiple scenarios show that TrafficSimAgent effectively executes simulations under various conditions and consistently produces reasonable outcomes even when user instructions are am- biguous. Besides, the carefully designed expert-level autonomous decision-driven optimization in TrafficSimAgent yields superior performance when compared with other systems and SOTA LLM based methods.

LLMs have recently been employed in large-scale social simulations to reproduce macro-phenomena such as economic and political outcomes. In our latest series of works, we follow along with these works by adapting LLM social simulation to two macro scenarios: perceptions of Americans towards China over 20 years from 2005 to 2025, and the responses by countries to the US's imposed tariffs in April 2025. For each scenario, the simulation framework involves robust citizen profiles, real-world data to influence agent opinion, and ground truth to validate simulation accuracy. In order to give agents a human-like mechanism for opinion changes, we design a cognitive mechanism while adding a strategic layer to their thinking using international relations theory. Experimental results show that our simulation results exhibit high fidelity with respect to the ground truth, demonstrating the strengths of our framework and shedding light on the potential of LLMs to reproduce a wide array of global trends if designed correctly. The results also provide some social insights, such as the types of news that contribute most to biased opinions and how media could be used to convey more objective views. We also aim to shed light on predictive capabilities of LLM for complex international relationships.

LLM agents have advanced from reactive responders to interactive entities capable of dynamic social interaction. As these systems expand to simulate complex societies, memory stands as a backbone for continuity, identity, and socially grounded behavior. Drawing on human cognition, from attention to short-term and long-term memory, computational memory systems allow agents to adapt, learn, and evolve within dynamic environments. Recent research highlights both the progress and the limitations of current approaches. Building on this foundation, the presentation will review the literature on LLM agent memory systems, explore their applications in social simulations, identify key design challenges, and propose improvements to LLM agent memory design within the AgentSociety platform.

World models are powerful generative frameworks that simulate future environment states, yet their inference speed remains a major bottleneck. This limitation is particularly prohibitive for real-time tasks like robotic control, where computational efficiency is crucial. This inefficiency stems from the per-frame generation paradigm, where models redundantly generate every frame from scratch despite only minor changes occurring in consecutive frames. To address this inefficiency, we propose KeyWorld, a framework that accelerates text-conditioned robotic world models by concentrating expensive computation on a few semantic key frames. Specifically, KeyWorld first identifies critical frames by simplifying the robot's motion trajectory to retain significant transitions. A dedicated diffusion-based model then generates these temporally distant key frames. Finally, a lightweight interpolator efficiently reconstructs the full video by inpainting all intermediate frames. Evaluations on the Libero benchmark demonstrate that KeyWorld achieves a 4× acceleration with minimal quality loss and the motion-aware key frame selection further contributes to improving generation quality, especially on complex tasks. Our approach highlights a practical path toward deploying world models in real-time robotic control and other efficiency-sensitive domains.

Symbolic regression discovers accurate and interpretable formulas to describe data. While accuracy is straightforward to measure, there is no established metric for interpretability. Existing approaches either depend on expert evaluation or approximate interpretability using complexity measures such as formula length. However, expert evaluation is time-consuming, and complexity ignores the essence of formulas as white-box models --- formulas with the same complexity can differ substantially in structure and interpretability. In this work, we propose a new evaluation metric based on valid digits loss, inspired by the observation that physical formulas are typically numerically stable under limited precision. By computing the loss of valid digits for each part of a formula, our method can identify unreasonable substructures such as catastrophic cancellation. We find that this metric distinguishes physical formulas in the Feynman problem set from those produced by existing symbolic regression algorithms, achieving an F1-score of 90\%. Extensive experiments further demonstrate its effectiveness: it improves the accuracy of a genetic programming algorithm by 15\% by excluding unreasonable formulas during the search, and boosts the performance of Transformer-based generative methods by 45\% by filtering implausible formulas during pre-training. Finally, for formulas with comparable R2 and complexity, our metric achieves 75\% agreement with interpretability evaluations provided by state-of-the-art dialogue models leveraging domain knowledge, underscoring its reliability.

Training a robust and generalizable policy in embodied scenarios remains a significant challenge. Imitation Learning (IL) acquires behavioral patterns from human expert demonstration data, yet it is susceptible to demonstration biases and distribution shifts, which limits its generalization capability in unseen scenarios. Reinforcement Learning (RL) enables agents to explore the environment autonomously; however, its generalization heavily relies on the setup of the initial environment. As the requirement for generalization increases, the cost of initializing scenarios also rises substantially. Nevertheless, world models inherently exhibit strong generalization across multiple scenarios, allowing them to provide a wide range of scenario options at a low cost for policy training. Therefore, we explore the use of world models as environments and adopt reinforcement learning methods to train policies, with the aim of enhancing their generalization performance.

Trajectory modeling plays a crucial role in understanding and predicting movement patterns. With rapid technological advancements and increasing data availability, trajectory modeling methods have evolved from rule-based approaches to those leveraging recurrent neural networks (RNNs) and Transformers. Despite these advances, key challenges remain, including generalization to new environments and generating interactive, conditional trajectories.We introduce an innovative solution inspired by multimodal fusion techniques, combining a large language model (LLM) with traditional trajectory sequence models to address these challenges. By aligning position and trajectory representations, the LLM leverages multimodal trajectory features, including semantic information from text and user behavioral insights, as well as trajectory regularities learned from numerical sequences in traditional models. This approach enhances user trajectory modeling capabilities and improves its generalization to new scenarios. Furthermore, we propose a self-reflective iterative strategy based on reinforcement learning (RL) to enable interactive and flexible trajectory generation for diverse scenarios. Our approach achieves state-of-the-art performance on multiple real-world datasets, demonstrating its effectiveness and potential to advance the field of trajectory modeling.

AI Scientists are emerging as a transformative paradigm in research. Achieving this vision requires accurate and comprehensive literature retrieval that effectively positions new work within the broader scholarly context, establishing a solid foundation for each research stage. However, existing keyword-based and embedding-based retrieval approaches fail to capture the rich relations in scientific literature, such as citation links and supportive or contradictory stances between papers. This fundamental limitation often leads to incomplete or inaccurate literature reviews, mischaracterization of prior work, and unreliable downstream scientific outputs. To better understand and evaluate these limitations, we introduce the first benchmark for assessing relational retrieval capabilities in scientific literature. Constructed from a corpus of over 18 million AI papers, our benchmark systematically evaluates three levels of relational reasoning: ego-centric retrieval of papers with distinctive intrinsic properties, pair-wise identification of scholarly relationships such as support or contradiction, and path-wise reconstruction of scientific evolutionary trajectories. Through extensive evaluation of four major categories of retrieval systems, we demonstrate that current methods consistently fail across all relational tasks. Notably, our experiments show that providing systems with relational ground truth substantially improves performance on downstream applications such as survey generation, highlighting the significant potential for relation-aware retrieval. We publicly release our benchmark to support future research toward developing more sophisticated, context-aware retrieval systems for scientific AI.

Model Predictive Control (MPC) is a powerful strategy for complex systems, yet its performance critically depends on an accurate dynamics model. Integrating learned models to handle unknown dynamics introduces the following challenges: high online optimization cost, difficulty in robust multi-step prediction for time-varying systems, and poor generalization to novel parameters. We introduces Generative Koopman Learning, a framework designed to address these issues. Our approach first leverages Koopman operator theory to linearize the nonlinear dynamics, recasting the control problem as a computationally efficient Quadratic Program (QP) to resolve the optimization bottleneck. The core novelty lies in shifting from deterministic prediction to a generative paradigm: we employ a conditional diffusion model to learn the joint distribution over the entire future sequence of time-varying Koopman operators. This generative modeling captures complex temporal structures for robust forecasting and enhances generalization by accounting for the underlying dynamics distribution. To ensure real-time viability, we introduce a pipelined sampling technique that reduces online inference latency. By unifying the efficiency of Koopman linearization with the expressive power and distributional awareness of generative models, our framework aims to create a controller that is fast, robust, and highly generalizable.

Network orchestration is critical for cellular networks, underpinning tasks such as coverage and capacity optimization, interference mitigation, energy savings, and mobility robustness. Expert-driven workflows and hand-tuned heuristics struggle to keep pace with dynamic traffic patterns and operational complexity. In this ongoing work, we present an initial LLM-agent workflow that targets coverage optimization as the first use case. The workflow leverages LLM capabilities in structured planning and tool-augmented reasoning to automatically decompose, generate, and execute optimization algorithms. The system iteratively refines solutions via two cooperating agents: an Organizer Agent, which integrates planning, reasoning, and aggregation modules, with the main task of designing stepwise plans, selecting algorithmic approaches, and determining execution strategies; and an Evaluation Agent, which executes the generated code and measures key performance indicators. Results are fed back to the Organizer to guide the next iteration.

Since the introduction of Sora, video generation has garnered substantial attention from both academia and industry, rapidly advancing the field. The emergence of models capable of generating photorealistic videos with physically realistic dynamics has brought new hope for real-world modeling. This talk will first provide a comprehensive review of recent high-performance foundational video generation models, with a focus on open-source contributions. We will detail their data pipelines, model architectures, training methodologies, and their highlighted features. Following this, we will discuss the evolution from large-scale video generation to world models, identifying the core challenges to be addressed. Finally, we will briefly introduce our preliminary ideas for a general world model's roadmap.

In social science and economics research, households are often regarded as the smallest behavioral unit. Modeling household behavior is of great significance in the fields of social simulation and economic simulation. However, household behavior is influenced by multiple factors, including the bounded rationality of decision-making by members, communication, division of labor, cooperation among family members, and household values and beliefs. Traditional mathematical models or rule-based agent models often struggle to accurately capture these intricate internal household factors. On the other hand, large language models (LLMs) demonstrate strong human-like reasoning and decision-making capabilities, making it promising to use LLMs to model household members and potentially achieve better results than traditional models. We aim to design an LLM-based multi-agent simulation approach to redefine the paradigm of household modeling.

Large Language Models (LLMs) are typically fine-tuned for reasoning tasks through a two-stage pipeline of Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL), a process fraught with catastrophic forgetting and suboptimal trade-offs between imitation and exploration. Recent single-stage methods attempt to unify SFT and RL using heuristics, but lack a principled mechanism for dynamically balancing the two paradigms. In this paper, we reframe this challenge through the theoretical lens of implicit rewards, viewing SFT and RL not as distinct methods but as complementary reward signals. We introduce Adaptive Meta Fine-Tuning (AMFT), a novel single-stage algorithm that learns the optimal balance between SFT’s implicit, pathlevel reward and RL’s explicit, outcome-based reward. The core of AMFT is a meta-gradient adaptive weight controller that treats the SFT-RL balance as a learnable parameter, dynamically optimizing it to maximize long-term task performance. This forward-looking approach, regularized by policy entropy for stability, autonomously discovers an effective training curriculum. We conduct a comprehensive evaluation on challenging benchmarks spanning mathematical reasoning, abstract visual reasoning (General Points), and vision-language navigation (V-IRL). AMFT consistently establishes a new state-of-the-art and demonstrats superior generalization on out-of-distribution (OOD) tasks. Ablation studies and training dynamic analysis confirm that the metalearning controller is crucial for AMFT’s stability, sample efficiency, and performance, offering a more principled and effective paradigm for LLM alignment.