Nonconvex Optimization

Recently, GRPO-based reinforcement learning has shown remarkable progress in optimizing flow-matching models, effectively improving their alignment with task-specific rewards. Within these frameworks, the policy update relies on importance-ratio clipping to constrain overconfident positive and negative gradients. However, in practice, we observe a systematic shift in the importance-ratio distribution--its mean falls below 1 and its variance differs substantially across timesteps. This left-shifted and inconsistent distribution prevents positive-advantage samples from entering the clipped region, causing the mechanism to fail in constraining overconfident positive updates. As a result, the policy model inevitably enters an **implicit over-optimization stage**--while the proxy reward continues to increase, essential metrics such as image quality and text-prompt alignment deteriorate sharply, ultimately making the learned policy impractical for real-world use. To address this issue, we introduce **GRPO-Guard**, a simple yet effective enhancement to existing GRPO frameworks. Our method incorporates ratio normalization, which restores a balanced and step-consistent importance ratio, ensuring that PPO clipping properly constrains harmful updates across denoising timesteps. In addition, a gradient reweighting strategy equalizes policy gradients over noise conditions, preventing excessive updates from particular timestep regions. Together, these designs act as a regulated clipping mechanism, stabilizing optimization and substantially mitigating implicit over-optimization without relying on heavy KL regularization. Extensive experiments on multiple diffusion backbones (e.g., SD3.5M, Flux.1-dev) and diverse proxy tasks demonstrate that GRPO-Guard significantly reduces over-optimization while maintaining or even improving generation quality. We provide detailed demonstrations of the over-optimization process and corresponding visualizations in https://jingw193.github.io/GRPO-Guard/.

Reinforcement learning (RL) has demonstrated significant potential for post-training language models and autoregressive visual generative models, but adapting RL to masked generative models (MGMs) remains challenging. The core factor is that policy optimization requires the probability likelihood of each step due to its multi-step iteration process. This reliance on entire sampling trajectories introduces high computational cost, whereas natively optimizing random steps often yields suboptimal results. In this paper, we present MaskFocus, a novel RL framework that achieves effective policy optimization for MGMs by focusing on critical steps. Specifically, we determine the step-level information gain by measuring the similarity between the intermediate images at each sampling step and the final generated image. Crucially, we leverage this to identify the most critical and valuable steps and execute policy optimization on them. Furthermore, we design a dynamic routing sampling based on entropy to encourage the model to explore more valuable masking strategies. Extensive experiments on multiple Text-to-Image benchmarks validate the effectiveness of our method.

Text-to-Image (T2I) generation has achieved remarkable progress in recent years. Meanwhile, reinforcement learning methods, particularly those based on Group Relative Policy Optimization (GRPO), have attracted widespread attention and been successfully applied to T2I tasks. However, the uniform sampling strategy commonly used during training often ignores the match between sample difficulty and the model's current learning capability, leading to low training efficiency. We argue that improving training efficiency requires continuously prioritizing prompts that match the model's evolving capability and remain actively learnable. To this end, we propose Curriculum Group Policy Optimization (CGPO), an adaptive curriculum training framework. During training, each prompt produces a group of images scored by a reward model. We use the variance of group rewards as an online proxy for prompt inconsistency. A higher variance suggests that the model has partially captured the prompt requirements but has not yet achieved stable mastery. Such prompts are more likely to provide useful learning signals, so we increase their sampling probabilities accordingly. Additionally, to address data imbalance in multi-category datasets, we design a category calibration method based on proportional fairness optimization, which balances training difficulty across categories. Experiments on GenEval, T2I-CompBench++, and DPG Bench demonstrate that our framework effectively improves generation performance.

Rehearsal-based methods are the cornerstone of modern online class-incremental learning (OCIL), yet they face a fundamental challenge: the gradient of the current task often conflicts with that of the rehearsal data from the memory buffer, leading to catastrophic forgetting. Recent works have implicitly addressed this by using hypergradients, but the underlying mechanism has remained poorly understood. In this paper, we provide a formal analysis revealing that hypergradients mitigate forgetting by aligning task-specific gradients towards a common meta-objective, thereby reducing their conflict. However, we argue that this conflict-reducing alignment is inherently myopic--it only considers the immediate gradient directions, failing to account for the loss landscape geometry one step ahead. To overcome this limitation, we introduce a novel framework: Lookahead Optimization for Rehearsal (LOR). LOR explores a set of future model states by taking lookahead steps along different directions that balance plasticity and stability and optimizes a first-order Log-Sum-Exp (LSE) surrogate to emphasize the worst-performing sampled lookahead directions. Theoretical analysis from both optimization and statistical perspectives corroborates the robustness of our approach. Extensive experiments on Seq-CIFAR10, Seq-CIFAR100, and Seq-TinyImageNet demonstrate that LOR significantly outperforms state-of-the-art methods, establishing a new and more robust paradigm for rehearsal-based OCIL.

The efficiency of hyperparameter optimization (HPO) is critical for deep learning, yet state-of-the-art methods share a fundamental flaw: they are difficulty-agnostic, treating all hyperparameter configurations homogeneously. This approach leads to inefficient resource allocation, wasting budget in simple regions while under-exploring complex, rugged landscapes, and thereby critically undermining both search efficiency and final performance. To address this universal challenge, we introduce DABO, a framework that pioneers difficulty-aware tuning within the efficient context of Freeze-Thaw Bayesian Optimization. We first model optimization difficulty hierarchically. Then, departing from hand-crafted priors, we train a conditional diffusion model on 120,000 real learning curves, generating synthetic data with 2.3xhigher fidelity. This data trains our difficulty-aware surrogate model and acquisition function to dynamically adapt the search strategy. Across 75 tasks, DABO reduces regret by 11-18% compared to the leading difficulty-agnostic method, ifBO. Our work establishes a new paradigm for HPO, shifting the focus from configuration-centric to difficulty-aware resource allocation to enable more robust and efficient optimization.

Diffusion models have achieved outstanding success in image generation, yet their objectives are often limited to reconstruction, making it difficult to align with human preferences directly. Reinforcement learning (RL) offers a promising approach to address this by optimizing models using explicit reward signals. However, most studies apply RL across the entire denoising process, which is both computationally expensive and tends to weaken preference alignment, i.e., doing more but achieving less. We observe that the impact of RL fine-tuning varies significantly across denoising stages. In the early stage, image structures are unstable and distant from the final reward signal. Applying RL at this stage leads to delayed rewards and action-reward mismatching, resulting in high variance and inefficient updates. Conversely, in the later stage, reward gains saturate, and continued training tends to overfit local details, intensifying reward hacking. To tackle these challenges, we propose \ourmethod , an RL-enhanced plug-in that improves generation quality while reducing computational cost. Specifically, \ourmethod adaptively identifies the optimal timing for RL by perceiving the structural evolution and semantic consistency during denoising, and dynamically terminates training once the denoising converges and reward gains saturate. As a result, it achieves a rare `dual benefit': a reduction in computational costs alongside a significant performance improvement. Theoretical analysis from an entropy perspective and extensive experiments verify our claims: compared with state-of-the-art methods, \ourmethod improves performance by xx% while cutting computational cost by xx%.

Fairness in Continual Learning for Large Multimodal Models (LMMs) is an emerging yet underexplored challenge, particularly in the presence of imbalanced data distributions that can lead to biased model updates and suboptimal performance across tasks. While recent continual learning studies have made progress in addressing catastrophic forgetting, the problem of fairness caused by imbalanced data remains largely underexplored. This paper presents a novel Fairness Direct Preference Optimization (FaiDPO or \phi-DPO) framework for continual learning in LMMs. In particular, we first propose a new continual learning paradigm based on Direct Preference Optimization (DPO) to mitigate catastrophic forgetting by aligning learning with pairwise preference signals. Then, we identify the limitations of conventional DPO in imbalanced data and present a new \phi-DPO loss that explicitly addresses distributional biases. We provide a comprehensive theoretical analysis demonstrating that our approach addresses both forgetting and data imbalance. Additionally, to enable \phi-DPO-based continual learning, we construct pairwise preference annotations for existing benchmarks in the context of continual learning. Extensive experiments and ablation studies show the proposed \phi-DPO achieves State-of-the-Art performance across multiple benchmarks, outperforming prior continual learning methods of LMMs.

Closed-set object detection in remote sensing imagery has made significant progress, but achieving high detection accuracy remains challenging. Vision-Language Models (VLMs), which possess rich prior knowledge, offer a promising solution to this challenge. However, most existing VLMs are designed for open-vocabulary tasks and exhibit inherent limitations when directly applied to closed-set scenarios, such as notable accuracy degradation and high deployment costs. To address these issues, we propose VLM4RSDet, a novel collaborative training framework that leverages vision-language model to enhance the performance of conventional closed-set remote sensing object detectors. Notably, during inference, VLM4RSDet only retains the standard object detection architecture, thus avoiding any additional deployment overhead. Furthermore, we introduce a Global-Local Cross-Attention (GLCA) module and a Learnable Hierarchical Prediction Strategy (LHPS) to further improve collaborative training performance. Extensive experiments on five benchmark datasets demonstrate the effectiveness and robustness of our approach. In particular, our method outperforms the state-of-the-art by 7.5% in mAP_ 0.5:0.95 on the VisDrone2019 dataset. Our code is available at \href https://github.com/cszzshi/VLM4RSDet https://github.com/cszzshi/VLM4RSDet .

Test-Time Adaptation (TTA) is essential for enabling deep learning models to handle real-world data distribution shifts. However, current approaches face significant limitations: backpropagation-based methods are not suitable for low-end deployment devices, due to their high computation and memory requirements, as well as their tendency to modify model weights during adaptation; while traditional backpropagation-free techniques exhibit constrained adaptation capabilities. In this work, we propose Forward-Only Zeroth-Order Optimization (FOZO), a novel and practical backpropagation-free paradigm for TTA. FOZO leverages a memory-efficient zeroth-order prompt optimization, which is led by objectives optimizing both intermediate feature statistics and prediction entropy. To ensure efficient and stable adaptation over the out-of-distribution data stream, we introduce a dynamically decaying perturbation scale during zeroth-order gradient estimation and theoretically prove its convergence under the TTA data stream assumption. Extensive continual adaptation experiments on ImageNet-C, ImageNet-R, and ImageNet-Sketch demonstrate FOZO's superior performance, achieving 59.52% Top-1 accuracy on ImageNet-C (5K, level 5) and outperforming main gradient-based methods and SOTA forward-only FOA (58.13%). Furthermore, FOZO exhibits strong generalization on quantized (INT8) models. These findings demonstrate that FOZO is a highly competitive solution for TTA deployment in resource-limited scenarios. Code: https://github.com/eVI-group-SCU/FOZO

Group Relative Policy Optimization (GRPO) has shown promise in aligning image and video generative models with human preferences. However, applying it to modern flow matching models is challenging because of its deterministic sampling paradigm. Current methods address this issue by converting Ordinary Differential Equations (ODEs) to Stochastic Differential Equations (SDEs), which introduce stochasticity. However, this SDE-based GRPO suffers from issues of inefficient credit assignment and incompatibility with high-order solvers for fewer-step sampling. In this paper, we first reinterpret existing SDE-based GRPO methods from a distance optimization perspective, revealing their underlying mechanism as a form of contrastive learning. Based on this insight, we propose Neighbor GRPO, a novel alignment algorithm that completely bypasses the need for SDEs. Neighbor GRPO generates a diverse set of candidate trajectories by perturbing the initial noise conditions of the ODE and optimizes the model using a softmax distance-based surrogate leaping policy. We establish a theoretical connection between this distance-based objective and policy gradient optimization, rigorously integrating our approach into the GRPO framework. Our method fully preserves the advantages of deterministic ODE sampling, including efficiency and compatibility with high-order solvers. We further introduce symmetric anchor sampling for computational efficiency and group-wise quasi-norm reweighting to address reward flattening. Extensive experiments demonstrate that Neighbor GRPO significantly outperforms SDE-based counterparts in terms of training cost, convergence speed, and generation quality.

What makes a good viewpoint? The quality of the data used to learn 3D reconstructions is crucial for enabling efficient and accurate scene modeling. We study the active view selection problem and develop a principled analysis that yields a simple and interpretable criterion for selecting informative camera poses. Our key insight is that informative views can be obtained by minimizing a tractable approximation of the Fisher Information Gain, which reduces to favoring viewpoints that cover geometry that has been insufficiently observed by past cameras. This leads to a lightweight coverage-based view selection metric that avoids expensive transmittance estimation and is robust to noise and training dynamics. We call this metric CONVERGE (Coverage Optimized Novel Viewset Estimation for Reconstructing Geometry Efficiently). We integrate our method into the Nerfstudio framework and evaluate it on real datasets within fixed and embodied data acquisition scenarios. Across multiple datasets and radiance-field baselines, our method consistently improves reconstruction quality compared to state-of-the-art active view selection methods.

Active 3D Gaussian reconstruction achieves superior completeness and rendering quality by intelligently selecting viewpoints. However, existing methods suffer from two critical limitations: information gain metrics that prioritize geometric coverage while ignoring rendering quality, and overfitting to sparse view configurations that degrades novel view synthesis. We introduce ActivePolicy, a novel framework addressing both challenges through principled NBV selection and regularization. We propose GL-Graph, a graph-theoretic strategy that unifies geometric consistency, rendering quality, and observation redundancy into a single stability criterion. To counteract overfitting, we introduce 4D-Reg, which identifies floaters through manifold discrepancies among three depth types (R-Depth, alpha-Depth, C-Depth) and suppresses them via adaptive dropout. Extensive experiments demonstrate state-of-the-art reconstruction completeness and rendering fidelity on standard benchmarks.

3D Gaussian Splatting (3DGS) has revolutionized radiance field reconstruction by achieving high-quality novel view synthesis with fast rendering speed, introducing 3D Gaussian primitives to represent the scene. However, 3DGS encounters blurring and floaters when applied to complex scenes, caused by the reconstruction of redundant and ambiguous geometric structures. We attribute this issue to the unstable optimization of the Gaussians. To address this limitation, we present a plug-and-play PDE-based optimization method that overcomes the optimization constraints of 3DGS-based approaches in various tasks, such as novel view synthesis and surface reconstruction. Firstly, we theoretically derive that the 3DGS optimization procedure can be modeled as a PDE, and introduce a viscous term to ensure stable optimization. Secondly, we use the Material Point Method (MPM) to obtain a stable numerical solution of the PDE, which enhances both global and local constraints. Additionally, an effective Gaussian densification strategy and particle constraints are introduced to ensure fine-grained details. Extensive qualitative and quantitative experiments confirm that our method achieves state-of-the-art rendering and reconstruction quality.

Maintaining long-term accuracy of stereo camera calibration parameters is important for autonomous systems' perception. This work proposes Online Tracking of Essential Matrix by Stochastic Optimization (TESO). The core mechanisms of TESO are: 1) a robust loss function based on kernel correlation over tentative correspondences, 2) an adaptive online stochastic optimization on the essential manifold. TESO has low CPU and memory requirements, relies on a few hyperparameters, and eliminates the need for data-driven training, enabling the usage in resource-constrained online perception systems. We evaluated the influence of TESO on geometric precision, rectification quality, and stereo depth consistency. On the large-scale MAN TruckScenes dataset, TESO tracks rotational calibration drift with 0.12 deg precision in the Y-axis-critical for stereo accuracy-while the X- and Z-axes are five times more precise. Tracking applied to sequences with simulated drift shows similar precision with respect to the reference as tracking applied to no-drift sequences, indicating the tracker is unbiased. On the KITTI dataset, TESO revealed systematic inconsistencies in extrinsic parameters across stereo pairs, confirming previous published findings. We verified that intrinsic decalibration affected these errors, as evidenced by the conflicting behavior of the rectification and depth metrics. After correcting the reference calibration, TESO improved its rotation precision around the Y-axis 20 times to 0.025 deg and its depth accuracy 50 times. Despite its lightweight design, direct optimization of the proposed TESO loss function alone achieves accuracy comparable to that of neural network-based single-frame methods.

Video Large Language Models (VLLMs) demonstrate strong video understanding but suffer from inefficiency due to redundant visual tokens. Existing pruning primary targets intra-frame spatial redundancy or prunes inside the LLM with shallow-layer overhead, yielding suboptimal spatiotemporal reduction and underutilizing long-context compressibility. All of them often discard subtle yet informative context from merged or pruned tokens. In this paper, we propose a new perspective that elaborates token Anchors within intra-frame and inter-frame to comprehensively aggregate the informative contexts via local-global Optimal Transport (AOT). Specifically, we first establish local- and global-aware token anchors within each frame under the attention guidance, which then optimal transport aggregates the informative contexts from pruned tokens, constructing intra-frame token anchors. Then, building on the temporal frame clips, the first frame within each clip will be considered as the keyframe anchors to ensemble similar information from consecutive frames through optimal transport, while keeping distinct tokens to represent temporal dynamics, leading to efficient token reduction in a training-free manner. Extensive evaluations show that our proposed AOT obtains competitive performances across various short- and long-video benchmarks on leading video LLMs, obtaining substantial computational efficiency while preserving temporal and visual fidelity.

Although the temporal spike dynamics of spiking neural networks (SNNs) enable low-power temporal capture capabilities, they also incur inherent inconsistencies that severely compromise representation. In this paper, we perform dual consistency optimization via Stable Spike to mitigate this problem, thereby improving the recognition performance of SNNs. With the hardware-friendly "AND" bit operation, we efficiently decouple the stable spike skeleton from the multi-timestep spike maps, which captures critical semantics and reduces the inconsistency from variable noise spikes. Enforcing the unstable spike maps to converge to the stable spike skeleton significantly improves the inherent consistency across timesteps. Furthermore, we inject amplitude-aware spike noise into the stable spike skeleton to diversify the representations while preserving consistent semantics. The SNN is encouraged to produce perturbation-consistent predictions, thereby contributing to generalization. Extensive experiments across multiple architectures and datasets validate the effectiveness and versatility of our method. In particular, our method significantly advances neuromorphic object recognition under ultra-low latency, improving accuracy by up to 8.33%. This will help unlock the full power consumption and speed potential of SNNs.

Dynamic driving scene modeling is critical for autonomous driving simulation and closed-loop learning. While recent feed-forward methods offer fast inference through data-driven priors, they struggle with long-range driving sequences due to quadratic complexity in sequence length and restrictive assumptions for dynamic objects. We propose UFO, a novel recurrent paradigm that combines the strengths of optimization-based and feed-forward methods for efficient long-range 4D modeling. Our approach maintains a persistent 4D scene representation composed of scene tokens that are progressively refined as new frames arrive, enabling future observations to correct earlier uncertain predictions. A visibility-based filtering mechanism exploits the locality of 3D-to-pixel correspondence, reducing complexity from quadratic to near-linear in sequence length. For dynamic objects, we introduce object pose-guided modeling with learned lifespans, enabling complex long-range motion without kinematic assumptions. Experiments on the Waymo Open Dataset demonstrate that UFO significantly outperforms both per-scene optimization and feed-forward baselines across 2s, 8s, and zero-shot 16s sequences.

Accurate camera viewpoint estimation under sparse-view conditions remains challenging, particularly in two-view scenarios. Recent approaches leverage diffusion models such as Zero123 to synthesize novel views conditioned on relative viewpoint, showing promising results when repurposed for viewpoint estimation via optimization with MSE loss. However, existing methods often suffer from non-convex loss landscape with numerous local minima, making them sensitive to initialization and reliant on naive multi-start strategies. We analyze these optimization challenges and visualize failure cases, showing that geometric ambiguities, such as symmetry and self-similarity, can mislead gradient-based updates toward incorrect viewpoints. To address these limitations, we propose a score-based method that reshapes the optimization landscape to guide updates toward the ground-truth viewpoint, followed by a refinement stage using a viewpoint-conditioned diffusion model. Experiments show that our method improves convergence, reduces reliance on brute-force sampling, and achieves competitive accuracy with higher sample-efficiency.

We propose Fresco, a unified optimization pipeline designed to mitigate early over-sharpening, and cross-view drifting in head avatar reconstruction. Fresco combines a Laplacian-pyramid-based frequency curriculum with UV-space consistency regularization to progressively enhance reconstruction quality. The optimization begins by stabilizing low-frequency appearance in the image domain, which suppresses spurious details and promotes reliable convergence. As learning proceeds, consistency across different viewpoints is reinforced through pixel-level alignment on shared UV texture coordinates. Finally, high-frequency components are refined under explicit frequency-band constraints, and seam boundary regularization is applied to preserve local continuity. By optimizing in a frequency- and UV-aligned space, Fresco achieves robust convergence without pseudo high-frequency artifacts and yields consistent, high-fidelity results across views. Experiments on the NeRSemble dataset validate the effectiveness of our design, outperforming previous state-of-the-art methods.

Building reconstruction aims to extract compact wireframes from point clouds. Recent edge-based methods achieve impressive results but suffer from sparse supervision from one-to-one matching, which leaves most edge proposals under-optimized. In this paper, we present Group Relative Edge Optimization (GREO), the first attempt to incentivize dense supervision across edges proposals through reinforcement learning-style optimization in wireframe reconstruction. Specifically, GREO computes edge-level rewards based on geometric alignment quality and transforms them into target confidence distributions via group-wise normalization. In addition, we incorporate entropy regularization to maintain distributional stability and prevent confidence collapse. This joint optimization enables dense and discriminative supervision across all edge proposals through cross-entropy minimization. Experiments on the large-scale Building3D dataset demonstrate that our powerful and versatile GREO integrates seamlessly into existing edge-based methods as a plug-and-play training strategy, achieving state-of-the-art performance on both the Entry-level and Tallinn benchmarks while adding no inference overhead.