PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models

Safety alignment of large language models (LLMs) typically depends on high-quality supervision data, such as safe demonstrations or preference pairs. However, in real-world deployment, emerging safety requirements are often specified as natural-language policies, while corresponding supervision data may be costly, delayed, or unavailable. This creates a mismatch between rapidly evolving safety policies and conventional data-driven alignment methods. To address this, we propose PolicyAlign, a simple yet effective framework for directly aligning LLMs with safety policies. Given a safety policy, PolicyAlign first synthesizes policy-violating instructions and then performs on-policy self-distillation to internalize policy-guided behavior. To improve training stability and data efficiency, we further introduce Policy-Sensitive Filtering, which selects instructions where the policy induces the largest behavioral shift. Experiments across multiple models show that PolicyAlign consistently improves safety while maintaining low over-refusal and preserving general capabilities. PolicyAlign also generalizes to medical, legal, and financial safety scenarios, highlighting its potential as a scalable and maintainable approach to policy-based LLM safety alignment. The code is released at https://github.com/Qwen-Applications/PolicyAlign.

DriveStack-VLA: Render-Teacher Alignment for BEV-Based DeepStack Vision-Language-Action Model

Vision-Language-Action driving models convert a pretrained Vision-Language Model into a driving policy, allowing them to use world knowledge and follow language guidances. However, existing VLA driving models still lack driving-oriented spatial intelligence: their policies are mainly grounded on perspective image tokens and language priors, while precise motion planning requires metric geometry, top-down scene structure, and attention to safety-critical perceptual cues. This limitation makes current models vulnerable to weak visual geometry modeling and perceptual coverage in expert demonstrations. In this paper, we present DriveStack-VLA, a framework built upon a large VLM backbone. To strengthen the spatial grounding of VLA driving, we develop dual visual modeling components. We inject a Bird-Eye-View representation into the Large Language Model decoder through a DeepStack-style connection, and propose Render-Teacher Alignment to align the perceptual focus of real images with that of rasterized images. Furthermore, to bridge the gap in multimodal trajectory selection, we introduce a head-based self-critique module that ranks sampled trajectories and conditionally refines the best one. DriveStack-VLA achieves 91.6 PDMS on NAVSIMv1, 91.0 EPDMS on NAVSIMv2 (with the human penalty filter enabled), and a driving score of 79.49 with a success rate of 56.36\% on the closed-loop Bench2Drive. More visualizations are available on our project page: https://anonymous.4open.science/w/drivestack-vla/.

RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer

Visual Geometry Grounded Transformer (VGGT) recovers dense 3D scene structure from multi-view images in one forward pass, but quadratic cross-frame attention limits its scalability. Existing training-free accelerators reduce computation uniformly along one axis, missing layer heterogeneity. Our spectral, probing, and causal analyses reveal three regimes: shallow layers lack cross-view structure, middle layers drive cross-view alignment, and deep layers are redundant for dense geometry yet their cross-frame attention remains essential for pose. RegimeVGGT applies layer-wise U-shaped compression along two axes: Saliency-Guided Banded Merging protects geometry- and edge-salient tokens, while Selectively Protected K/V Downsampling preserves cross-frame spatial coverage and the pose-critical path through a phase-shifted spatial grid, a reference-frame anchor, and uncompressed camera/register tokens. Training-free, RegimeVGGT achieves a 6.7x speedup over VGGT* at matched reconstruction quality.

Seam-to-Graph Reconstruction for Garment Configuration Alignment

Seams encode rich structural information about garments but are frequently partially observable in robotic manipulation scenarios. To robustly leverage seam information, we propose a Seam-to-Graph network based on graph neural networks and attention mechanisms. This network maps unstructured seam observations to a topology-encoded structural skeleton graph for real-time garment state estimation. Using this skeleton-graph-based state estimation, we design a deformation-aware, hierarchical visual servoing controller for garment configuration alignment. We implement this controller on a bimanual robot system to load a garment onto a screen printing platen and to align it to the desired configuration precisely. Real-robot experiments demonstrate that the robot using the proposed method not only achieves human-level alignment accuracy with reduced variance in alignment error but is also robust to different garments. These results demonstrate that the use of seam information is effective for garment manipulation.

GeoMin: Data-Efficient Semi-Supervised RLVR via Geometric Distribution Modeling

Reinforcement learning with verifiable rewards (RLVR) significantly advances LLM reasoning, yet it faces a dilemma: standard supervised scaling is throttled by high annotation costs, while unsupervised alternatives suffer from severe model collapse. Recent semi-supervised RLVR methods address this by using a small labeled set to guide unlabeled data, achieving a promising trade-off between training efficacy and annotation cost. However, they suffer from a severe data-efficiency bottleneck due to the reliance on coarse performance heuristics, leaving a vast majority of valuable instances underutilized. To this end, we propose GeoMin, which models global feature distributions on labeled data to decode the structural discrepancy between correct and incorrect rollouts, thereby establishing a robust prior to assess the reliability of self-reward signals and fully unleash the potential of unlabeled data. Empirically, GeoMin outperforms the strongest baselines by +4.1% and even surpasses fully supervised models with only 10% of the annotations, demonstrating remarkable data efficiency.

Skill-RM: Unifying Heterogeneous Evaluation Criteria via Agent Skill

Reward models (RMs) provide critical feedback signals for LLM post-training, notably in reinforced fine-tuning (RFT) and reinforcement learning (RL) pipelines. However, current reward evaluation relies on heterogeneous criteria such as rule-based verifiers, ground-truth references, procedural checklists, and complex rubrics, where a unified mechanism to integrate all types of evidence remains unexplored. To this end, we propose Skill Reward Model (Skill-RM), a unified framework that reformulates reward modeling as the execution of a reusable Reward-Evaluation Skill. By treating reward computation as a structured agentic task, Skill-RM provides a consistent interface to orchestrate heterogeneous resources, dynamically selecting and aggregating evidence tailored to the specific requirements of each input. This approach enables the reward model to move beyond static evaluation, ensuring consistency and transparency across diverse tasks. Extensive experiments on reward benchmarks and downstream applications, including best-of-N selection and reinforcement learning, demonstrate that Skill-RM consistently outperforms traditional judge baselines. Our findings suggest that Skill-RM not only provides a unified solution for reward modeling but also achieves superior performance through the strategic and dynamic orchestration of evidence. The code is at https://github.com/Qwen-Applications/Skill-RM.

Phase-Conditioned Imitation Learning with Autonomous Failure Recovery for Robust Deformable Object Manipulation

This paper presents a phase-conditioned, force-aware framework for robust deformable object manipulation. Standard imitation learning policies such as Action Chunking with Transformers (ACT) rely on a Markovian assumption at inference, causing state aliasing when visually similar observations require contradictory actions and preventing autonomous recovery from execution failures. We address this with a closed-loop hierarchical architecture. A FiLM-conditioned ACT encoder modulates feature extraction based on the current task phase, enabling a single unified policy to produce phase-specific behaviors while sharing action dynamics across phases. A multi-modal phase predictor fusing visual, force, and pose feedback estimates the phase in real time, detecting contact failures that are invisible to vision alone and autonomously triggering recovery trajectories. The system is completed by a hybrid impedance controller for compliant execution and a haptic teleoperation interface for force-aware data collection. Ablation studies show that FiLM-based modulation significantly outperforms both unconditioned and token-level conditioned baselines, and t-SNE analysis confirms that FiLM induces well-separated, phase-specific feature representations. Validated on hanging and removing a T-shirt with dual arms, the closed-loop system improves the hanging success rate from 56\% to 87\% through autonomous error recovery. Code and videos: https://leledeyuan00.github.io/phaser/

expo: Exploration-prioritized policy optimization via adaptive kl regulation and gaussian curriculum sampling

Reinforcement Learning with Verifiable Rewards (RLVR) has become the standard paradigm for LLM mathematical reasoning, where Group Relative Policy Optimization (GRPO) serves as the mainstream algorithm. We point out two understudied inefficiencies existing in GRPO. First, the fixed KL penalty coefficient overly restricts policy exploration at stages where the model requires significant deviation from the reference policy. Second, uniform sampling of training questions ignores that moderately difficult problems provide the most informative gradient signals for optimization. We propose Exploration-Prioritized Policy Optimization (EXPO) with two lightweight plug-in modules. The Accuracy-Conditioned KL Scaling (AKL) dynamically adjusts KL regularization strength through a smooth nonlinear function of batch average accuracy, relaxing the penalty when the model underperforms and strengthening it when the model achieves good results. The Gaussian Curriculum Sampling (GCS) assigns sampling weights to questions following a Gaussian distribution centered at moderate accuracy around 0.5, focusing training on the model's learning frontier. We conduct extensive experiments on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-8B-Base over six mathematical reasoning benchmarks. The results show EXPO steadily surpasses vanilla GRPO. It obtains an absolute gain of 13.34 on AIME 2025 pass@32, rising from 63.33 percent to 76.67 percent, and achieves an average pass@32 improvement of 2.66 on the 8B model. The much larger performance gains on pass@32 compared with pass@1 demonstrate that EXPO effectively enlarges the model's exploration boundary under a fixed inference cost budget.

Reinforced Curriculum Pre-Alignment for Domain-Adaptive VLMs

Vision-Language Models (VLMs) demonstrate remarkable general-purpose capabilities but often fall short in specialized domains such as medical imaging or geometric problem-solving. Supervised Fine-Tuning (SFT) can enhance performance within a target domain, but it typically causes catastrophic forgetting, limiting its generalization. The central challenge, therefore, is to adapt VLMs to new domains while preserving their general-purpose capabilities. Continual pretraining is effective for expanding knowledge in Large Language Models (LLMs), but it is less feasible for VLMs due to prohibitive computational costs and the unavailability of pretraining data for most open-source models. This necessitates efficient post-training adaptation methods. Reinforcement learning (RL)-based approaches such as Group Relative Policy Optimization (GRPO) have shown promise in preserving general abilities, yet they often fail in domain adaptation scenarios where the model initially lacks sufficient domain knowledge, leading to optimization collapse. To bridge this gap, we propose Reinforced Curriculum Pre-Alignment (RCPA), a novel post-training paradigm that introduces a curriculum-aware progressive modulation mechanism. In the early phase, RCPA applies partial output constraints to safely expose the model to new domain concepts. As the model's domain familiarity increases, training gradually transitions to full generation optimization, refining responses and aligning them with domain-specific preferences. This staged adaptation balances domain knowledge acquisition with the preservation of general multimodal capabilities. Extensive experiments across specialized domains and general benchmarks validate the effectiveness of RCPA, establishing a practical pathway toward building high-performing and domain-adaptive VLMs.

TC-IDM: Grounding Video Generation for Executable Zero-shot Robot Motion

The vision-language-action (VLA) paradigm has enabled powerful robotic control by leveraging vision-language models, but its reliance on large-scale, high-quality robot data limits its generalization. Generative world models offer a promising alternative for general-purpose embodied AI, yet a critical gap remains between their pixel-level plans and physically executable actions. To this end, we propose the Tool-Centric Inverse Dynamics Model (TC-IDM). By focusing on the tool's imagined trajectory as synthesized by the world model, TC-IDM establishes a robust intermediate representation that bridges the gap between visual planning and physical control. TC-IDM extracts the tool's point cloud trajectories via segmentation and 3D motion estimation from generated videos. Considering diverse tool attributes, our architecture employs decoupled action heads to project these planned trajectories into 6-DoF end-effector motions and corresponding control signals. This plan-and-translate paradigm not only supports a wide range of end-effectors but also significantly improves viewpoint invariance. Furthermore, it exhibits strong generalization capabilities across long-horizon and out-of-distribution tasks, including interacting with deformable objects. In real-world evaluations, the world model with TC-IDM achieves an average success rate of 61.11 percent, with 77.7 percent on simple tasks and 38.46 percent on zero-shot deformable object tasks. It substantially outperforms end-to-end VLA-style baselines and other inverse dynamics models.