Secure Tug-of-War (SecTOW): Iterative Defense-Attack Training with Reinforcement Learning for Multimodal Model Security

The rapid advancement of multimodal large language models (MLLMs) has led to breakthroughs in various applications, yet their security remains a critical challenge. One pressing issue involves unsafe image-query pairs--jailbreak inputs specifically designed to bypass security constraints and elicit unintended responses from MLLMs. Compared to general multimodal data, such unsafe inputs are relatively sparse, which limits the diversity and richness of training samples available for developing robust defense models. Meanwhile, existing guardrail-type methods rely on external modules to enforce security constraints but fail to address intrinsic vulnerabilities within MLLMs. Traditional supervised fine-tuning (SFT), on the other hand, often over-refuses harmless inputs, compromising general performance. Given these challenges, we propose Secure Tug-of-War (SecTOW), an innovative iterative defense-attack training method to enhance the security of MLLMs. SecTOW consists of two modules: a defender and an auxiliary attacker, both trained iteratively using reinforcement learning (GRPO). During the iterative process, the attacker identifies security vulnerabilities in the defense model and expands jailbreak data. The expanded data are then used to train the defender, enabling it to address identified security vulnerabilities. We also design reward mechanisms used for GRPO to simplify the use of response labels, reducing dependence on complex generative labels and enabling the efficient use of synthetic data. Additionally, a quality monitoring mechanism is used to mitigate the defender's over-refusal of harmless inputs and ensure the diversity of the jailbreak data generated by the attacker. Experimental results on safety-specific and general benchmarks demonstrate that SecTOW significantly improves security while preserving general performance.

Test-Time-Matching: Decouple Personality, Memory, and Linguistic Style in LLM-based Role-Playing Language Agent

The rapid advancement of large language models (LLMs) has enabled role-playing language agents to demonstrate significant potential in various applications. However, relying solely on prompts and contextual inputs often proves insufficient for achieving deep immersion in specific roles, particularly well-known fictional or public figures. On the other hand, fine-tuning-based approaches face limitations due to the challenges associated with data collection and the computational resources required for training, thereby restricting their broader applicability. To address these issues, we propose Test-Time-Matching (TTM), a training-free role-playing framework through test-time scaling and context engineering. TTM uses LLM agents to automatically decouple a character's features into personality, memory, and linguistic style. Our framework involves a structured, three-stage generation pipeline that utilizes these features for controlled role-playing. It achieves high-fidelity role-playing performance, also enables seamless combinations across diverse linguistic styles and even variations in personality and memory. We evaluate our framework through human assessment, and the results demonstrate that our method achieves the outstanding performance in generating expressive and stylistically consistent character dialogues.

Inverse Reinforcement Learning Meets Large Language Model Post-Training: Basics, Advances, and Opportunities

In the era of Large Language Models (LLMs), alignment has emerged as a fundamental yet challenging problem in the pursuit of more reliable, controllable, and capable machine intelligence. The recent success of reasoning models and conversational AI systems has underscored the critical role of reinforcement learning (RL) in enhancing these systems, driving increased research interest at the intersection of RL and LLM alignment. This paper provides a comprehensive review of recent advances in LLM alignment through the lens of inverse reinforcement learning (IRL), emphasizing the distinctions between RL techniques employed in LLM alignment and those in conventional RL tasks. In particular, we highlight the necessity of constructing neural reward models from human data and discuss the formal and practical implications of this paradigm shift. We begin by introducing fundamental concepts in RL to provide a foundation for readers unfamiliar with the field. We then examine recent advances in this research agenda, discussing key challenges and opportunities in conducting IRL for LLM alignment. Beyond methodological considerations, we explore practical aspects, including datasets, benchmarks, evaluation metrics, infrastructure, and computationally efficient training and inference techniques. Finally, we draw insights from the literature on sparse-reward RL to identify open questions and potential research directions. By synthesizing findings from diverse studies, we aim to provide a structured and critical overview of the field, highlight unresolved challenges, and outline promising future directions for improving LLM alignment through RL and IRL techniques.

EPIC: Efficient Prompt Interaction for Text-Image Classification

In recent years, large-scale pre-trained multimodal models (LMMs) generally emerge to integrate the vision and language modalities, achieving considerable success in multimodal tasks, such as text-image classification. The growing size of LMMs, however, results in a significant computational cost for fine-tuning these models for downstream tasks. Hence, prompt-based interaction strategy is studied to align modalities more efficiently. In this context, we propose a novel efficient prompt-based multimodal interaction strategy, namely Efficient Prompt Interaction for text-image Classification (EPIC). Specifically, we utilize temporal prompts on intermediate layers, and integrate different modalities with similarity-based prompt interaction, to leverage sufficient information exchange between modalities. Utilizing this approach, our method achieves reduced computational resource consumption and fewer trainable parameters (about 1\% of the foundation model) compared to other fine-tuning strategies. Furthermore, it demonstrates superior performance on the UPMC-Food101 and SNLI-VE datasets, while achieving comparable performance on the MM-IMDB dataset.

DynamicBench: Evaluating Real-Time Report Generation in Large Language Models

Traditional benchmarks for large language models (LLMs) typically rely on static evaluations through storytelling or opinion expression, which fail to capture the dynamic requirements of real-time information processing in contemporary applications. To address this limitation, we present DynamicBench, a benchmark designed to evaluate the proficiency of LLMs in storing and processing up-to-the-minute data. DynamicBench utilizes a dual-path retrieval pipeline, integrating web searches with local report databases. It necessitates domain-specific knowledge, ensuring accurate responses report generation within specialized fields. By evaluating models in scenarios that either provide or withhold external documents, DynamicBench effectively measures their capability to independently process recent information or leverage contextual enhancements. Additionally, we introduce an advanced report generation system adept at managing dynamic information synthesis. Our experimental results confirm the efficacy of our approach, with our method achieving state-of-the-art performance, surpassing GPT4o in document-free and document-assisted scenarios by 7.0% and 5.8%, respectively. The code and data will be made publicly available.

OmniFluids: Unified Physics Pre-trained Modeling of Fluid Dynamics

High-fidelity and efficient simulation of fluid dynamics drive progress in various scientific and engineering applications. Traditional computational fluid dynamics methods offer strong interpretability and guaranteed convergence, but rely on fine spatial and temporal meshes, incurring prohibitive computational costs. Physics-informed neural networks (PINNs) and neural operators aim to accelerate PDE solvers using deep learning techniques. However, PINNs require extensive retraining and careful tuning, and purely data-driven operators demand large labeled datasets. Hybrid physics-aware methods embed numerical discretizations into network architectures or loss functions, but achieve marginal speed gains and become unstable when balancing coarse priors against high-fidelity measurements. To this end, we introduce OmniFluids, a unified physics pre-trained operator learning framework that integrates physics-only pre-training, coarse-grid operator distillation, and few-shot fine-tuning, which enables fast inference and accurate prediction under limited or zero data supervision. For architectural design, the key components of OmniFluids include a mixture of operators, a multi-frame decoder, and factorized Fourier layers, which enable efficient and scalable modeling of diverse physical tasks while maintaining seamless integration with physics-based supervision. Across a broad range of two- and three-dimensional benchmarks, OmniFluids significantly outperforms state-of-the-art AI-driven methods in flow field reconstruction and turbulence statistics accuracy, delivering 10-100x speedups compared to classical solvers, and accurately recovers unknown physical parameters from sparse, noisy data. This work establishes a new paradigm for efficient and generalizable surrogate modeling in complex fluid systems under limited data availability.

SlotPi: Physics-informed Object-centric Reasoning Models

Understanding and reasoning about dynamics governed by physical laws through visual observation, akin to human capabilities in the real world, poses significant challenges. Currently, object-centric dynamic simulation methods, which emulate human behavior, have achieved notable progress but overlook two critical aspects: 1) the integration of physical knowledge into models. Humans gain physical insights by observing the world and apply this knowledge to accurately reason about various dynamic scenarios; 2) the validation of model adaptability across diverse scenarios. Real-world dynamics, especially those involving fluids and objects, demand models that not only capture object interactions but also simulate fluid flow characteristics. To address these gaps, we introduce SlotPi, a slot-based physics-informed object-centric reasoning model. SlotPi integrates a physical module based on Hamiltonian principles with a spatio-temporal prediction module for dynamic forecasting. Our experiments highlight the model's strengths in tasks such as prediction and Visual Question Answering (VQA) on benchmark and fluid datasets. Furthermore, we have created a real-world dataset encompassing object interactions, fluid dynamics, and fluid-object interactions, on which we validated our model's capabilities. The model's robust performance across all datasets underscores its strong adaptability, laying a foundation for developing more advanced world models.

LightSAM: Parameter-Agnostic Sharpness-Aware Minimization

Sharpness-Aware Minimization (SAM) optimizer enhances the generalization ability of the machine learning model by exploring the flat minima landscape through weight perturbations. Despite its empirical success, SAM introduces an additional hyper-parameter, the perturbation radius, which causes the sensitivity of SAM to it. Moreover, it has been proved that the perturbation radius and learning rate of SAM are constrained by problem-dependent parameters to guarantee convergence. These limitations indicate the requirement of parameter-tuning in practical applications. In this paper, we propose the algorithm LightSAM which sets the perturbation radius and learning rate of SAM adaptively, thus extending the application scope of SAM. LightSAM employs three popular adaptive optimizers, including AdaGrad-Norm, AdaGrad and Adam, to replace the SGD optimizer for weight perturbation and model updating, reducing sensitivity to parameters. Theoretical results show that under weak assumptions, LightSAM could converge ideally with any choices of perturbation radius and learning rate, thus achieving parameter-agnostic. We conduct preliminary experiments on several deep learning tasks, which together with the theoretical findings validate the the effectiveness of LightSAM.

Impromptu VLA: Open Weights and Open Data for Driving Vision-Language-Action Models

Vision-Language-Action (VLA) models for autonomous driving show promise but falter in unstructured corner case scenarios, largely due to a scarcity of targeted benchmarks. To address this, we introduce Impromptu VLA. Our core contribution is the Impromptu VLA Dataset: over 80,000 meticulously curated video clips, distilled from over 2M source clips sourced from 8 open-source large-scale datasets. This dataset is built upon our novel taxonomy of four challenging unstructured categories and features rich, planning-oriented question-answering annotations and action trajectories. Crucially, experiments demonstrate that VLAs trained with our dataset achieve substantial performance gains on established benchmarks--improving closed-loop NeuroNCAP scores and collision rates, and reaching near state-of-the-art L2 accuracy in open-loop nuScenes trajectory prediction. Furthermore, our Q&A suite serves as an effective diagnostic, revealing clear VLM improvements in perception, prediction, and planning. Our code, data and models are available at https://github.com/ahydchh/Impromptu-VLA.

Spectral-inspired Neural Operator for Data-efficient PDE Simulation in Physics-agnostic Regimes

Partial differential equations (PDEs) govern the spatiotemporal evolution of various physical systems. Classical numerical solvers, while accurate, require fine discretization and full knowledge of the governing PDEs, limiting their applicability when the physics is unknown or fast inference is required. Data-driven neural PDE solvers alleviate these constraints by learning from data but demand large training datasets and perform poorly in data-scarce regimes. Physics-aware methods mitigate data requirements by incorporating physical knowledge yet rely on known PDE terms or local numerical schemes, restricting their ability to handle unknown or globally coupled systems. In this work, we propose the Spectral-inspired Neural Operator (SINO), a novel framework that learns PDE operators from limited trajectories (as few as 2-5), without any known PDE terms. SINO operates in the frequency domain and introduces a Frequency-to-Vector module to learn spectral representations analogous to derivative multipliers. To model nonlinear physical interactions, we design a nonlinear operator block that includes a $\Pi$-Block with low-pass filtering to prevent aliasing. Finally, we introduce an operator distillation technique to distill the trained model for efficient inference. SINO achieves state-of-the-art results across multiple PDE benchmarks, demonstrating strong discretization invariance and robust generalization to out-of-distribution initial conditions. To our knowledge, SINO is the first physics-aware method capable of accurately simulating globally coupled systems (e.g., the Navier-Stokes equations) from limited data without any explicit PDE terms.