Why GRPO Needs Normalization: A Local-Curvature Perspective on Adaptive Gradients

Reinforcement learning (RL) has become a key driver of language model reasoning. Among RL algorithms, Group Relative Policy Optimization (GRPO) is the de facto standard, avoiding the need for a critic by using per-prompt baselines and variance normalization. Yet why and when this normalization helps remains unclear. In this work, we provide an explanation through the lens of local curvature of the sequence-level policy gradient: standard deviation normalization implements an adaptive gradient. Theoretically, under mild conditions, GRPO enjoys a strictly improved convergence rate over unnormalized REINFORCE, with gains characterized by the average within-prompt reward standard deviation across prompts and iterations. Empirically, our analysis on GSM8K and MATH benchmarks reveals three distinct training phases governed by the interplay between feature orthogonality and reward variance: (I) an early acceleration phase where high variance and orthogonality favor adaptive scaling; (II) a relatively stable transition phase; and (III) a late-stage regime where the loss of orthogonality limits further gains. Together, these results provide a principled account of when std normalization helps in GRPO, and offer broader insights into the design of critic-free RL algorithms.

TTSA3R: Training-Free Temporal-Spatial Adaptive Persistent State for Streaming 3D Reconstruction

Streaming recurrent models enable efficient 3D reconstruction by maintaining persistent state representations. However, they suffer from catastrophic memory forgetting over long sequences due to balancing historical information with new observations. Recent methods alleviate this by deriving adaptive signals from attention perspective, but they operate on single dimensions without considering temporal and spatial consistency. To this end, we propose a training-free framework termed TTSA3R that leverages both temporal state evolution and spatial observation quality for adaptive state updates in 3D reconstruction. In particular, we devise a Temporal Adaptive Update Module that regulates update magnitude by analyzing temporal state evolution patterns. Then, a Spatial Contextual Update Module is introduced to localize spatial regions that require updates through observation-state alignment and scene dynamics. These complementary signals are finally fused to determine the state updating strategies. Extensive experiments demonstrate the effectiveness of TTSA3R in diverse 3D tasks. Moreover, our method exhibits only 15% error increase compared to over 200% degradation in baseline models on extended sequences, significantly improving long-term reconstruction stability. Our codes will be available soon.

Smoothing the Score Function for Generalization in Diffusion Models: An Optimization-based Explanation Framework

Diffusion models achieve remarkable generation quality, yet face a fundamental challenge known as memorization, where generated samples can replicate training samples exactly. We develop a theoretical framework to explain this phenomenon by showing that the empirical score function (the score function corresponding to the empirical distribution) is a weighted sum of the score functions of Gaussian distributions, in which the weights are sharp softmax functions. This structure causes individual training samples to dominate the score function, resulting in sampling collapse. In practice, approximating the empirical score function with a neural network can partially alleviate this issue and improve generalization. Our theoretical framework explains why: In training, the neural network learns a smoother approximation of the weighted sum, allowing the sampling process to be influenced by local manifolds rather than single points. Leveraging this insight, we propose two novel methods to further enhance generalization: (1) Noise Unconditioning enables each training sample to adaptively determine its score function weight to increase the effect of more training samples, thereby preventing single-point dominance and mitigating collapse. (2) Temperature Smoothing introduces an explicit parameter to control the smoothness. By increasing the temperature in the softmax weights, we naturally reduce the dominance of any single training sample and mitigate memorization. Experiments across multiple datasets validate our theoretical analysis and demonstrate the effectiveness of the proposed methods in improving generalization while maintaining high generation quality.

Are Video Generation Models Geographically Fair? An Attraction-Centric Evaluation of Global Visual Knowledge

Recent advances in text-to-video generation have produced visually compelling results, yet it remains unclear whether these models encode geographically equitable visual knowledge. In this work, we investigate the geo-equity and geographically grounded visual knowledge of text-to-video models through an attraction-centric evaluation. We introduce Geo-Attraction Landmark Probing (GAP), a systematic framework for assessing how faithfully models synthesize tourist attractions from diverse regions, and construct GEOATTRACTION-500, a benchmark of 500 globally distributed attractions spanning varied regions and popularity levels. GAP integrates complementary metrics that disentangle overall video quality from attraction-specific knowledge, including global structural alignment, fine-grained keypoint-based alignment, and vision-language model judgments, all validated against human evaluation. Applying GAP to the state-of-the-art text-to-video model Sora 2, we find that, contrary to common assumptions of strong geographic bias, the model exhibits a relatively uniform level of geographically grounded visual knowledge across regions, development levels, and cultural groupings, with only weak dependence on attraction popularity. These results suggest that current text-to-video models express global visual knowledge more evenly than expected, highlighting both their promise for globally deployed applications and the need for continued evaluation as such systems evolve.

Mosaic: Unlocking Long-Context Inference for Diffusion LLMs via Global Memory Planning and Dynamic Peak Taming

Diffusion-based large language models (dLLMs) have emerged as a promising paradigm, utilizing simultaneous denoising to enable global planning and iterative refinement. While these capabilities are particularly advantageous for long-context generation, deploying such models faces a prohibitive memory capacity barrier stemming from severe system inefficiencies. We identify that existing inference systems are ill-suited for this paradigm: unlike autoregressive models constrained by the cumulative KV-cache, dLLMs are bottlenecked by transient activations recomputed at every step. Furthermore, general-purpose memory reuse mechanisms lack the global visibility to adapt to dLLMs' dynamic memory peaks, which toggle between logits and FFNs. To address these mismatches, we propose Mosaic, a memory-efficient inference system that shifts from local, static management to a global, dynamic paradigm. Mosaic integrates a mask-only logits kernel to eliminate redundancy, a lazy chunking optimizer driven by an online heuristic search to adaptively mitigate dynamic peaks, and a global memory manager to resolve fragmentation via virtual addressing. Extensive evaluations demonstrate that Mosaic achieves an average 2.71$\times$ reduction in the memory peak-to-average ratio and increases the maximum inference sequence length supportable on identical hardware by 15.89-32.98$\times$. This scalability is achieved without compromising accuracy and speed, and in fact reducing latency by 4.12%-23.26%.

Local Path Optimization in The Latent Space Using Learned Distance Gradient

Constrained motion planning is a common but challenging problem in robotic manipulation. In recent years, data-driven constrained motion planning algorithms have shown impressive planning speed and success rate. Among them, the latent motion method based on manifold approximation is the most efficient planning algorithm. Due to errors in manifold approximation and the difficulty in accurately identifying collision conflicts within the latent space, time-consuming path validity checks and path replanning are required. In this paper, we propose a method that trains a neural network to predict the minimum distance between the robot and obstacles using latent vectors as inputs. The learned distance gradient is then used to calculate the direction of movement in the latent space to move the robot away from obstacles. Based on this, a local path optimization algorithm in the latent space is proposed, and it is integrated with the path validity checking process to reduce the time of replanning. The proposed method is compared with state-of-the-art algorithms in multiple planning scenarios, demonstrating the fastest planning speed

SparseSurf: Sparse-View 3D Gaussian Splatting for Surface Reconstruction

Recent advances in optimizing Gaussian Splatting for scene geometry have enabled efficient reconstruction of detailed surfaces from images. However, when input views are sparse, such optimization is prone to overfitting, leading to suboptimal reconstruction quality. Existing approaches address this challenge by employing flattened Gaussian primitives to better fit surface geometry, combined with depth regularization to alleviate geometric ambiguities under limited viewpoints. Nevertheless, the increased anisotropy inherent in flattened Gaussians exacerbates overfitting in sparse-view scenarios, hindering accurate surface fitting and degrading novel view synthesis performance. In this paper, we propose \net{}, a method that reconstructs more accurate and detailed surfaces while preserving high-quality novel view rendering. Our key insight is to introduce Stereo Geometry-Texture Alignment, which bridges rendering quality and geometry estimation, thereby jointly enhancing both surface reconstruction and view synthesis. In addition, we present a Pseudo-Feature Enhanced Geometry Consistency that enforces multi-view geometric consistency by incorporating both training and unseen views, effectively mitigating overfitting caused by sparse supervision. Extensive experiments on the DTU, BlendedMVS, and Mip-NeRF360 datasets demonstrate that our method achieves the state-of-the-art performance.

Lost in Tokenization: Context as the Key to Unlocking Biomolecular Understanding in Scientific LLMs

Scientific Large Language Models (Sci-LLMs) have emerged as a promising frontier for accelerating biological discovery. However, these models face a fundamental challenge when processing raw biomolecular sequences: the tokenization dilemma. Whether treating sequences as a specialized language, risking the loss of functional motif information, or as a separate modality, introducing formidable alignment challenges, current strategies fundamentally limit their reasoning capacity. We challenge this sequence-centric paradigm by positing that a more effective strategy is to provide Sci-LLMs with high-level structured context derived from established bioinformatics tools, thereby bypassing the need to interpret low-level noisy sequence data directly. Through a systematic comparison of leading Sci-LLMs on biological reasoning tasks, we tested three input modes: sequence-only, context-only, and a combination of both. Our findings are striking: the context-only approach consistently and substantially outperforms all other modes. Even more revealing, the inclusion of the raw sequence alongside its high-level context consistently degrades performance, indicating that raw sequences act as informational noise, even for models with specialized tokenization schemes. These results suggest that the primary strength of existing Sci-LLMs lies not in their nascent ability to interpret biomolecular syntax from scratch, but in their profound capacity for reasoning over structured, human-readable knowledge. Therefore, we argue for reframing Sci-LLMs not as sequence decoders, but as powerful reasoning engines over expert knowledge. This work lays the foundation for a new class of hybrid scientific AI agents, repositioning the developmental focus from direct sequence interpretation towards high-level knowledge synthesis. The code is available at github.com/opendatalab-raise-dev/CoKE.

ReTrack: Data Unlearning in Diffusion Models through Redirecting the Denoising Trajectory

Diffusion models excel at generating high-quality, diverse images but suffer from training data memorization, raising critical privacy and safety concerns. Data unlearning has emerged to mitigate this issue by removing the influence of specific data without retraining from scratch. We propose ReTrack, a fast and effective data unlearning method for diffusion models. ReTrack employs importance sampling to construct a more efficient fine-tuning loss, which we approximate by retaining only dominant terms. This yields an interpretable objective that redirects denoising trajectories toward the $k$-nearest neighbors, enabling efficient unlearning while preserving generative quality. Experiments on MNIST T-Shirt, CelebA-HQ, CIFAR-10, and Stable Diffusion show that ReTrack achieves state-of-the-art performance, striking the best trade-off between unlearning strength and generation quality preservation.

Discovery Learning accelerates battery design evaluation

Fast and reliable validation of novel designs in complex physical systems such as batteries is critical to accelerating technological innovation. However, battery research and development remain bottlenecked by the prohibitively high time and energy costs required to evaluate numerous new design candidates, particularly in battery prototyping and life testing. Despite recent progress in data-driven battery lifetime prediction, existing methods require labeled data of target designs to improve accuracy and cannot make reliable predictions until after prototyping, thus falling far short of the efficiency needed to enable rapid feedback for battery design. Here, we introduce Discovery Learning (DL), a scientific machine-learning paradigm that integrates active learning, physics-guided learning, and zero-shot learning into a human-like reasoning loop, drawing inspiration from learning theories in educational psychology. DL can learn from historical battery designs and actively reduce the need for prototyping, thus enabling rapid lifetime evaluation for unobserved material-design combinations without requiring additional data labeling. To test DL, we present 123 industrial-grade large-format lithium-ion pouch cells, spanning eight material-design combinations and diverse cycling protocols. Trained solely on public datasets of small-capacity cylindrical cells, DL achieves 7.2% test error in predicting the average cycle life under unknown device variability. This results in savings of 98% in time and 95% in energy compared to industrial practices. This work highlights the potential of uncovering insights from historical designs to inform and accelerate the development of next-generation battery technologies. DL represents a key advance toward efficient data-driven modeling and helps realize the promise of machine learning for accelerating scientific discovery and engineering innovation.