Thinking-Based Non-Thinking: Solving the Reward Hacking Problem in Training Hybrid Reasoning Models via Reinforcement Learning

Large reasoning models (LRMs) have attracted much attention due to their exceptional performance. However, their performance mainly stems from thinking, a long Chain of Thought (CoT), which significantly increase computational overhead. To address this overthinking problem, existing work focuses on using reinforcement learning (RL) to train hybrid reasoning models that automatically decide whether to engage in thinking or not based on the complexity of the query. Unfortunately, using RL will suffer the the reward hacking problem, e.g., the model engages in thinking but is judged as not doing so, resulting in incorrect rewards. To mitigate this problem, existing works either employ supervised fine-tuning (SFT), which incurs high computational costs, or enforce uniform token limits on non-thinking responses, which yields limited mitigation of the problem. In this paper, we propose Thinking-Based Non-Thinking (TNT). It does not employ SFT, and sets different maximum token usage for responses not using thinking across various queries by leveraging information from the solution component of the responses using thinking. Experiments on five mathematical benchmarks demonstrate that TNT reduces token usage by around 50% compared to DeepSeek-R1-Distill-Qwen-1.5B/7B and DeepScaleR-1.5B, while significantly improving accuracy. In fact, TNT achieves the optimal trade-off between accuracy and efficiency among all tested methods. Additionally, the probability of reward hacking problem in TNT's responses, which are classified as not using thinking, remains below 10% across all tested datasets.

ManiLong-Shot: Interaction-Aware One-Shot Imitation Learning for Long-Horizon Manipulation

One-shot imitation learning (OSIL) offers a promising way to teach robots new skills without large-scale data collection. However, current OSIL methods are primarily limited to short-horizon tasks, thus limiting their applicability to complex, long-horizon manipulations. To address this limitation, we propose ManiLong-Shot, a novel framework that enables effective OSIL for long-horizon prehensile manipulation tasks. ManiLong-Shot structures long-horizon tasks around physical interaction events, reframing the problem as sequencing interaction-aware primitives instead of directly imitating continuous trajectories. This primitive decomposition can be driven by high-level reasoning from a vision-language model (VLM) or by rule-based heuristics derived from robot state changes. For each primitive, ManiLong-Shot predicts invariant regions critical to the interaction, establishes correspondences between the demonstration and the current observation, and computes the target end-effector pose, enabling effective task execution. Extensive simulation experiments show that ManiLong-Shot, trained on only 10 short-horizon tasks, generalizes to 20 unseen long-horizon tasks across three difficulty levels via one-shot imitation, achieving a 22.8% relative improvement over the SOTA. Additionally, real-robot experiments validate ManiLong-Shot's ability to robustly execute three long-horizon manipulation tasks via OSIL, confirming its practical applicability.

VITA-VLA: Efficiently Teaching Vision-Language Models to Act via Action Expert Distillation

Vision-Language Action (VLA) models significantly advance robotic manipulation by leveraging the strong perception capabilities of pretrained vision-language models (VLMs). By integrating action modules into these pretrained models, VLA methods exhibit improved generalization. However, training them from scratch is costly. In this work, we propose a simple yet effective distillation-based framework that equips VLMs with action-execution capability by transferring knowledge from pretrained small action models. Our architecture retains the original VLM structure, adding only an action token and a state encoder to incorporate physical inputs. To distill action knowledge, we adopt a two-stage training strategy. First, we perform lightweight alignment by mapping VLM hidden states into the action space of the small action model, enabling effective reuse of its pretrained action decoder and avoiding expensive pretraining. Second, we selectively fine-tune the language model, state encoder, and action modules, enabling the system to integrate multimodal inputs with precise action generation. Specifically, the action token provides the VLM with a direct handle for predicting future actions, while the state encoder allows the model to incorporate robot dynamics not captured by vision alone. This design yields substantial efficiency gains over training large VLA models from scratch. Compared with previous state-of-the-art methods, our method achieves 97.3% average success rate on LIBERO (11.8% improvement) and 93.5% on LIBERO-LONG (24.5% improvement). In real-world experiments across five manipulation tasks, our method consistently outperforms the teacher model, achieving 82.0% success rate (17% improvement), which demonstrate that action distillation effectively enables VLMs to generate precise actions while substantially reducing training costs.

MedSentry: Understanding and Mitigating Safety Risks in Medical LLM Multi-Agent Systems

As large language models (LLMs) are increasingly deployed in healthcare, ensuring their safety, particularly within collaborative multi-agent configurations, is paramount. In this paper we introduce MedSentry, a benchmark comprising 5 000 adversarial medical prompts spanning 25 threat categories with 100 subthemes. Coupled with this dataset, we develop an end-to-end attack-defense evaluation pipeline to systematically analyze how four representative multi-agent topologies (Layers, SharedPool, Centralized, and Decentralized) withstand attacks from 'dark-personality' agents. Our findings reveal critical differences in how these architectures handle information contamination and maintain robust decision-making, exposing their underlying vulnerability mechanisms. For instance, SharedPool's open information sharing makes it highly susceptible, whereas Decentralized architectures exhibit greater resilience thanks to inherent redundancy and isolation. To mitigate these risks, we propose a personality-scale detection and correction mechanism that identifies and rehabilitates malicious agents, restoring system safety to near-baseline levels. MedSentry thus furnishes both a rigorous evaluation framework and practical defense strategies that guide the design of safer LLM-based multi-agent systems in medical domains.

SEPT: Standard-Definition Map Enhanced Scene Perception and Topology Reasoning for Autonomous Driving

Online scene perception and topology reasoning are critical for autonomous vehicles to understand their driving environments, particularly for mapless driving systems that endeavor to reduce reliance on costly High-Definition (HD) maps. However, recent advances in online scene understanding still face limitations, especially in long-range or occluded scenarios, due to the inherent constraints of onboard sensors. To address this challenge, we propose a Standard-Definition (SD) Map Enhanced scene Perception and Topology reasoning (SEPT) framework, which explores how to effectively incorporate the SD map as prior knowledge into existing perception and reasoning pipelines. Specifically, we introduce a novel hybrid feature fusion strategy that combines SD maps with Bird's-Eye-View (BEV) features, considering both rasterized and vectorized representations, while mitigating potential misalignment between SD maps and BEV feature spaces. Additionally, we leverage the SD map characteristics to design an auxiliary intersection-aware keypoint detection task, which further enhances the overall scene understanding performance. Experimental results on the large-scale OpenLane-V2 dataset demonstrate that by effectively integrating SD map priors, our framework significantly improves both scene perception and topology reasoning, outperforming existing methods by a substantial margin.

DeCo: Task Decomposition and Skill Composition for Zero-Shot Generalization in Long-Horizon 3D Manipulation

Generalizing language-conditioned multi-task imitation learning (IL) models to novel long-horizon 3D manipulation tasks remains a significant challenge. To address this, we propose DeCo (Task Decomposition and Skill Composition), a model-agnostic framework compatible with various multi-task IL models, designed to enhance their zero-shot generalization to novel, compositional, long-horizon 3D manipulation tasks. DeCo first decomposes IL demonstrations into a set of modular atomic tasks based on the physical interaction between the gripper and objects, and constructs an atomic training dataset that enables models to learn a diverse set of reusable atomic skills during imitation learning. At inference time, DeCo leverages a vision-language model (VLM) to parse high-level instructions for novel long-horizon tasks, retrieve the relevant atomic skills, and dynamically schedule their execution; a spatially-aware skill-chaining module then ensures smooth, collision-free transitions between sequential skills. We evaluate DeCo in simulation using DeCoBench, a benchmark specifically designed to assess zero-shot generalization of multi-task IL models in compositional long-horizon 3D manipulation. Across three representative multi-task IL models (RVT-2, 3DDA, and ARP), DeCo achieves success rate improvements of 66.67%, 21.53%, and 57.92%, respectively, on 12 novel compositional tasks. Moreover, in real-world experiments, a DeCo-enhanced model trained on only 6 atomic tasks successfully completes 9 novel long-horizon tasks, yielding an average success rate improvement of 53.33% over the base multi-task IL model. Video demonstrations are available at: https://deco226.github.io.

Robust Dataset Distillation by Matching Adversarial Trajectories

Dataset distillation synthesizes compact datasets that enable models to achieve performance comparable to training on the original large-scale datasets. However, existing distillation methods overlook the robustness of the model, resulting in models that are vulnerable to adversarial attacks when trained on distilled data. To address this limitation, we introduce the task of ``robust dataset distillation", a novel paradigm that embeds adversarial robustness into the synthetic datasets during the distillation process. We propose Matching Adversarial Trajectories (MAT), a method that integrates adversarial training into trajectory-based dataset distillation. MAT incorporates adversarial samples during trajectory generation to obtain robust training trajectories, which are then used to guide the distillation process. As experimentally demonstrated, even through natural training on our distilled dataset, models can achieve enhanced adversarial robustness while maintaining competitive accuracy compared to existing distillation methods. Our work highlights robust dataset distillation as a new and important research direction and provides a strong baseline for future research to bridge the gap between efficient training and adversarial robustness.

RoboHorizon: An LLM-Assisted Multi-View World Model for Long-Horizon Robotic Manipulation

Efficient control in long-horizon robotic manipulation is challenging due to complex representation and policy learning requirements. Model-based visual reinforcement learning (RL) has shown great potential in addressing these challenges but still faces notable limitations, particularly in handling sparse rewards and complex visual features in long-horizon environments. To address these limitations, we propose the Recognize-Sense-Plan-Act (RSPA) pipeline for long-horizon tasks and further introduce RoboHorizon, an LLM-assisted multi-view world model tailored for long-horizon robotic manipulation. In RoboHorizon, pre-trained LLMs generate dense reward structures for multi-stage sub-tasks based on task language instructions, enabling robots to better recognize long-horizon tasks. Keyframe discovery is then integrated into the multi-view masked autoencoder (MAE) architecture to enhance the robot's ability to sense critical task sequences, strengthening its multi-stage perception of long-horizon processes. Leveraging these dense rewards and multi-view representations, a robotic world model is constructed to efficiently plan long-horizon tasks, enabling the robot to reliably act through RL algorithms. Experiments on two representative benchmarks, RLBench and FurnitureBench, show that RoboHorizon outperforms state-of-the-art visual model-based RL methods, achieving a 23.35% improvement in task success rates on RLBench's 4 short-horizon tasks and a 29.23% improvement on 6 long-horizon tasks from RLBench and 3 furniture assembly tasks from FurnitureBench.

Distractor-free Generalizable 3D Gaussian Splatting

We present DGGS, a novel framework addressing the previously unexplored challenge of Distractor-free Generalizable 3D Gaussian Splatting (3DGS). It accomplishes two key objectives: fortifying generalizable 3DGS against distractor-laden data during both training and inference phases, while successfully extending cross-scene adaptation capabilities to conventional distractor-free approaches. To achieve these objectives, DGGS introduces a scene-agnostic reference-based mask prediction and refinement methodology during training phase, coupled with a training view selection strategy, effectively improving distractor prediction accuracy and training stability. Moreover, to address distractor-induced voids and artifacts during inference stage, we propose a two-stage inference framework for better reference selection based on the predicted distractor masks, complemented by a distractor pruning module to eliminate residual distractor effects. Extensive generalization experiments demonstrate DGGS's advantages under distractor-laden conditions. Additionally, experimental results show that our scene-agnostic mask inference achieves accuracy comparable to scene-specific trained methods. Homepage is \url{https://github.com/bbbbby-99/DGGS}.

Magnetic Preference Optimization: Achieving Last-iterate Convergence for Language Models Alignment

Self-play methods have demonstrated remarkable success in enhancing model capabilities across various domains. In the context of Reinforcement Learning from Human Feedback (RLHF), self-play not only boosts Large Language Model (LLM) performance but also overcomes the limitations of traditional Bradley-Terry (BT) model assumptions by finding the Nash equilibrium (NE) of a preference-based, two-player constant-sum game. However, existing methods either guarantee only average-iterate convergence, incurring high storage and inference costs, or converge to the NE of a regularized game, failing to accurately reflect true human preferences. In this paper, we introduce Magnetic Preference Optimization (MPO), a novel approach capable of achieving last-iterate convergence to the NE of the original game, effectively overcoming the limitations of existing methods. Building upon Magnetic Mirror Descent (MMD), MPO attains a linear convergence rate, making it particularly suitable for fine-tuning LLMs. To ensure our algorithm is both theoretically sound and practically viable, we present a simple yet effective implementation that adapts the theoretical insights to the RLHF setting. Empirical results demonstrate that MPO can significantly enhance the performance of LLMs, highlighting the potential of self-play methods in alignment.