Modern video generators still struggle with complex physical dynamics, often falling short of physical realism. Existing approaches address this using external verifiers or additional training on augmented data, which is computationally expensive and still limited in capturing fine-grained motion. In this work, we present self-refining video sampling, a simple method that uses a pre-trained video generator trained on large-scale datasets as its own self-refiner. By interpreting the generator as a denoising autoencoder, we enable iterative inner-loop refinement at inference time without any external verifier or additional training. We further introduce an uncertainty-aware refinement strategy that selectively refines regions based on self-consistency, which prevents artifacts caused by over-refinement. Experiments on state-of-the-art video generators demonstrate significant improvements in motion coherence and physics alignment, achieving over 70% human preference compared to the default sampler and guidance-based sampler.

Recent work shows that large multimodal models (LMMs) can self-improve from unlabeled data via self-play and intrinsic feedback. Yet existing self-evolving frameworks mainly reward final outcomes, leaving intermediate reasoning weakly constrained despite its importance for visually grounded decision making. We propose iReasoner, a self-evolving framework that improves an LMM's implicit reasoning by explicitly eliciting chain-of-thought (CoT) and rewarding its internal agreement. In a Proposer--Solver loop over unlabeled images, iReasoner augments outcome-level intrinsic rewards with a trajectory-aware signal defined over intermediate reasoning steps, providing learning signals that distinguish reasoning paths leading to the same answer without ground-truth labels or external judges. Starting from Qwen2.5-VL-7B, iReasoner yields up to $+2.1$ points across diverse multimodal reasoning benchmarks under fully unsupervised post-training. We hope this work serves as a starting point for reasoning-aware self-improvement in LMMs in purely unsupervised settings. Our code is available at https://meghanaasunil.github.io/iReasoner.

Aligning text-to-video diffusion models with human preferences is crucial for generating high-quality videos. Existing Direct Preference Otimization (DPO) methods rely on multi-sample ranking and task-specific critic models, which is inefficient and often yields ambiguous global supervision. To address these limitations, we propose LocalDPO, a novel post-training framework that constructs localized preference pairs from real videos and optimizes alignment at the spatio-temporal region level. We design an automated pipeline to efficiently collect preference pair data that generates preference pairs with a single inference per prompt, eliminating the need for external critic models or manual annotation. Specifically, we treat high-quality real videos as positive samples and generate corresponding negatives by locally corrupting them with random spatio-temporal masks and restoring only the masked regions using the frozen base model. During training, we introduce a region-aware DPO loss that restricts preference learning to corrupted areas for rapid convergence. Experiments on Wan2.1 and CogVideoX demonstrate that LocalDPO consistently improves video fidelity, temporal coherence and human preference scores over other post-training approaches, establishing a more efficient and fine-grained paradigm for video generator alignment.The code is available at https://github.com/1170300714/Local-DPO.

Low-resource indigenous languages often lack the parallel corpora required for effective neural machine translation (NMT). Synthetic data generation offers a practical strategy for mitigating this limitation in data-scarce settings. In this work, we augment curated parallel datasets for indigenous languages of the Americas with synthetic sentence pairs generated using a high-capacity multilingual translation model. We fine-tune a multilingual mBART model on curated-only and synthetically augmented data and evaluate translation quality using chrF++, the primary metric used in recent AmericasNLP shared tasks for agglutinative languages. We further apply language-specific preprocessing, including orthographic normalization and noise-aware filtering, to reduce corpus artifacts. Experiments on Guarani-Spanish and Quechua-Spanish translation show consistent chrF++ improvements from synthetic data augmentation, while diagnostic experiments on Aymara highlight the limitations of generic preprocessing for highly agglutinative languages.

We revisit the analogy between feed-forward deep neural networks (DNNs) and discrete dynamical systems derived from neural integral equations and their corresponding partial differential equation (PDE) forms. A comparative analysis between the numerical/exact solutions of the Burgers' and Eikonal equations, and the same obtained via PINNs is presented. We show that PINN learning provides a different computational pathway compared to standard numerical discretization in approximating essentially the same underlying dynamics of the system. Within this framework, DNNs can be interpreted as discrete dynamical systems whose layer-wise evolution approaches attractors, and multiple parameter configurations may yield comparable solutions, reflecting the degeneracy of the inverse mapping. In contrast to the structured operators associated with finite-difference (FD) procedures, PINNs learn dense parameter representations that are not directly associated with classical discretization stencils. This distributed representation generally involves a larger number of parameters, leading to reduced interpretability and increased computational cost. However, the additional flexibility of such representations may offer advantages in high-dimensional settings where classical grid-based methods become impractical.

In our study, we evaluated large language model (LLM) performance on pharmacy licensure-style question-answering tasks and developed an external knowledge integration method to improve accuracy. We benchmarked ten LLMs with varying parameter sizes (8 billion to 70+ billion) using a 141-question pharmacy dataset, measuring baseline accuracy without modification. Baseline performance ranged from 46% to 92%, with GPT-5 (92%) and o3 (89%) achieving the highest scores, while smaller open-source models showed substantially lower performance. We then developed DrugRAG, a three-step retrieval-augmented generation (RAG) pipeline that retrieves structured, evidence-based drug information and augments model prompts with contextual pharmacological evidence, operating externally and requiring no changes to model architecture or parameters. DrugRAG increased accuracy across all five evaluated models, with gains ranging from 7 to 21 percentage points (e.g., Gemma 3 27B: 61.0% to 71%, Llama 3.1 8B: 46% to 67%). McNemar analyses demonstrated statistically significant paired improvements primarily in smaller and mid-sized open-source models. These findings demonstrate that integrating structured external drug knowledge via DrugRAG can improve LLM performance on pharmacy-focused question-answering tasks without modifying the underlying models, providing a practical pipeline for enhancing evidence-based pharmacy-focused AI applications.

Safety-critical learning requires policies that improve performance without leaving the safe operating regime. We study constrained policy learning where model parameters must satisfy rollout-based safety constraints that can be evaluated but not differentiated analytically. We propose SCPO, a sampling-based weight-space projection method that enforces safety directly in parameter space without requiring gradient access to the constraint functions. SCPO constructs a local safe region by combining rollout-based safety evaluations with smoothness bounds relating parameter perturbations to changes in safety metrics, and projects each gradient update via a convex QCQP. We establish a safe-by-induction guarantee: starting from any safe initialization, all intermediate policies remain safe given feasible projections. In constrained control settings with a stabilizing backup policy, SCPO further ensures closed-loop stability while enabling safe adaptation beyond the conservative backup. Experiments on constrained regression with harmful supervision and double-integrator imitation with a malicious expert show that SCPO rejects unsafe updates, maintains feasibility throughout training, and achieves meaningful objective improvement.

Intraoperative navigation in spine surgery demands millimeter-level accuracy. Currently, this is achieved through radiation-intensive intraoperative imaging and bone-anchored markers that are invasive and disrupt surgical workflow. Markerless RGB-D registration methods offer a promising alternative. However, existing approaches rely on weak segmentation labels to isolate relevant anatomical structures, potentially propagating errors through the registration process. We present End2Reg, an end-to-end deep learning framework that jointly optimizes segmentation and registration, eliminating the need for segmentation labels and manual steps. The network learns task-specific segmentation masks optimized for registration, guided solely by the registration objective without explicit segmentation supervision. End2Reg achieves state-of-the-art performance on ex- and in-vivo benchmarks, reducing median Target Registration Error by 32% and mean Root Mean Square Error by 61%, while maintaining robust performance under partial occlusions. Ablation results confirm that end-to-end optimization significantly improves registration accuracy. Overall, End2Reg advances towards fully automatic, markerless intraoperative navigation. Code and interactive visualizations are available at: https://lorenzopettinari.github.io/end-2-reg/.

Deep learning-based methods have recently achieved significant success in image reconstruction problems. However, challenges have emerged, as these methods may generate unrealistic artifacts or hallucinations, which can interfere with analysis in safety-critical scenarios. This paper introduces a framework for quantifying and characterizing hallucinated artifacts in image reconstruction models. The proposed method, termed the Conformal Hallucination Estimation Metric (CHEM), enables the identification of hallucination-prone regions in model predictions. It leverages wavelet and shearlet representations to localize such regions at the level of image features, and uses conformalized quantile regression to assess hallucination levels in a distribution-free manner. A theoretical analysis is provided, characterizing the sensitivity of CHEM to hallucinated artifacts and its relationship to the mean squared error. Building on these insights and adopting a viewpoint grounded in approximation theory, we investigate why U-shaped networks, widely used architectures for image reconstruction, tend to hallucination-prone predictions. We assess the effectiveness of the proposed approach on astronomical image deconvolution using the CANDELS dataset with architectures such as U-Net, SwinUNet, and Learnlets, and on natural image super-resolution using the DIV2K dataset with models such as DRUNet, Unfolded DRS, RAM, and DPS.

Structurally repetitive environments produce visually plausible but aliased LiDAR loop candidates that can destabilize pose-graph optimization when admitted as loop factors. We propose Query-Calibrated Segmental Admission (QCSA), a descriptor-agnostic sparse loop-admission policy for graph-stability-oriented insertion. The policy scores short descriptor segments against hard negatives, calibrates which query-level segment hypotheses reach geometry, and inserts representative pairs validated by Generalized Iterative Closest Point (G-ICP). We evaluate it on the SNU Library Dataset (SNULib) and HeLiPR overlap routes. Aggregated over seven LiDAR descriptor families on SNULib, QCSA reduces inserted loop factors by 3.8 times, raises factor precision from 0.542 to 0.717, and sharply lowers false admissions per query group. With this sparser graph, it maintains comparable mean absolute trajectory error (ATE) and substantially reduces worst-sequence ATE versus dense Top1+G-ICP, from 1.064 to 0.778 m. The aggregate mean and worst-sequence ATE remain lower than the odometry-only reference. Under a matched factor budget, QCSA also attains lower trajectory error than SeqSLAM and sparse Top1+G-ICP selections. Fixed-transfer validation on HeLiPR, with no route-specific tuning, likewise suppresses hard-negative admissions. These results support the proposed admission layer for aliasing-heavy simultaneous localization and mapping (SLAM). Our implementation and dataset will be released at: https://github.com/wanderingcar/snu_library_dataset.

Discovering discrete algebraic rules from data is a fundamental challenge in machine learning. We formalize this problem through Cayley-table completion -- an algebraic counterpart to classical matrix completion -- where the degree of associativity violation replaces linear rank as the intrinsic measure of complexity. We provide a rigorous landscape analysis of HyperCube, an operator-valued tensor factorization, on the fully observed target table $δ$, proving that its global infimum $H_{\inf}(δ) := \inf_{Θ\in F_δ} H(Θ)$ implicitly defines an exact differentiable measure for this complexity. We show that HyperCube's native objective $H(Θ)$ decomposes into two components: geometric alignment (collinearity) and an inverse $\ell_2$ penalty. We establish that these continuous variational pressures induce core discrete properties: collinearity enforces associativity (Collinearity--Associativity Equivalence), and the inverse $\ell_2$ penalty reduces to an exact inverse rank penalty within the collinear manifold, driving the parameters toward full-rank unitarity. Consequently, we derive an absolute lower bound $H(Θ) \ge H_{\inf}(δ) \ge 3 \, |δ|$, where $|δ|$ is the target table size. We prove this absolute floor is attained if and only if the target is isotopic to a group, and characterize the global minimizer as the regular representation of the underlying group (up to unitary gauge), resolving the central open conjecture of Huh (2025). This work serves as an existence proof that certain discrete algebraic structures can be exactly characterized by differentiable measures, enabling gradient-based discovery without the need for combinatorial search. All theoretical results are mechanically verified in Lean 4 and confirmed via small-scale experiments.

Text-based automated Cognitive Distortion detection is a challenging task due to its subjective nature, with low agreement scores observed even among expert human annotators, leading to unreliable annotations. We explore the use of Large Language Models (LLMs) as consistent and reliable annotators, and propose that multiple independent LLM runs can reveal stable labeling patterns despite the inherent subjectivity of the task. Furthermore, to fairly compare models trained on datasets with different characteristics, we introduce a dataset-agnostic evaluation framework using Cohen's kappa as an effect size measure. This methodology allows for fair cross-dataset and cross-study comparisons where traditional metrics like F1 score fall short. Our results show that GPT-4 can produce consistent annotations (Fleiss's Kappa = 0.78), resulting in improved test set performance for models trained on these annotations compared to those trained on human-labeled data. Our findings suggest that LLMs can offer a scalable and internally consistent alternative for generating training data that supports strong downstream performance in subjective NLP tasks.

This paper addresses the Kidnapped Robot Problem (KRP), a core localization challenge of relocalizing a robot in a known map without prior pose estimate upon localization loss or at SLAM initialization. For this purpose, a passive 2-D global relocalization framework is proposed. It estimates the global pose efficiently and reliably from a single LiDAR scan and an occupancy grid map while the robot remains stationary, thereby enhancing the long-term autonomy of mobile robots. The proposed framework casts global relocalization as a non-convex problem and solves it via the multi-hypothesis scheme with batched multi-stage inference and early termination, balancing completeness and efficiency. The Rapidly-exploring Random Tree (RRT), under traversability constraints, asymptotically covers the reachable space to generate sparse, uniformly distributed feasible positional hypotheses, fundamentally reducing the sampling space. The hypotheses are preliminarily ordered by the proposed Scan Mean Absolute Difference (SMAD), a coarse beam-error level metric that facilitates the early termination by prioritizing high-likelihood candidates. The SMAD computation is optimized for limited scan measurements. The Translation-Affinity Scan-to-Map Alignment Metric (TAM) is proposed for reliable orientation selection at hypothesized positions and accurate final global pose evaluation to mitigate degradation in conventional likelihood-field metrics under translational uncertainty induced by sparse hypotheses, as well as non-panoramic LiDAR scan and environmental changes. Real-world experiments on a resource-constrained mobile robot with non-panoramic LiDAR scans show that the proposed framework achieves competitive performance in success rate, robustness under measurement uncertainty, and computational efficiency.

The scope of neural code intelligence is rapidly expanding beyond text-based source code to encompass the rich visual outputs that programs generate. This visual dimension is critical for advanced applications like flexible content generation and precise, program-driven editing of visualizations. However, progress has been impeded by the scarcity of high-quality multimodal code data, a bottleneck stemming from challenges in synthesis and quality assessment. To address these challenges, we make contributions from both a data and modeling perspective. We first introduce a complete synthesis toolkit that leverages reciprocal synergies between data modalities to efficiently produce a large-scale, high-quality corpus spanning from standard charts to complex interactive web UIs and code-driven animations. Leveraging this toolkit, we construct JanusCode-800K, the largest multimodal code corpus to date. This powers the training of our models, JanusCoder and JanusCoderV, which establish a visual-programmatic interface for generating code from textual instructions, visual inputs, or a combination of both. Our unified model is a departure from existing approaches that build specialized models for isolated tasks. Extensive experiments on both text-centric and vision-centric coding tasks demonstrate the superior performance of the JanusCoder series, with our 7B to 14B scale models approaching or even exceeding the performance of commercial models. Furthermore, extensive analysis provides key insights into harmonizing programmatic logic with its visual expression. Our code and checkpoints are available at https://github.com/InternLM/JanusCoder.

Recent Progress in post-training flow matching for text-to-image (T2I) generation with Group Relative Policy Optimization (GRPO) has demonstrated strong potential. However, it is hindered by a critical limitation: inaccurate advantage attribution. In this work, we argue that aggregating consecutive steps into a coherent `chunk' and shifting the policy optimization paradigm from GRPO's step level to the chunk level can effectively mitigate the negative impact of this issue. Building on this insight, we propose Group Chunking Policy Optimization (GCPO), the first chunk-level reinforcement learning approach for post-training flow matching. Extensive experiments demonstrate that GCPO achieves superior performance on both standard T2I benchmarks and preference alignment, with up to 43% relative gains over GRPO, highlighting the promise of chunk-level policy optimization. The code is available on https://github.com/xingzhejun/GCPO.

Autonomous driving systems remain critically vulnerable to the long-tail of rare, out-of-distribution semantic anomalies. While VLMs have emerged as promising tools for perception, their application in anomaly detection remains largely restricted to prompting proprietary models - limiting reliability, reproducibility, and deployment feasibility. To address this gap, we introduce SAVANT (Semantic Anomaly Verification/Analysis Toolkit), a novel model-agnostic reasoning framework that reformulates anomaly detection as a layered semantic consistency verification. By applying SAVANT's two-phase pipeline - structured scene description extraction and multi-modal evaluation - existing VLMs improve their scores in detecting anomalous driving scenarios from input images. Our approach replaces ad hoc prompting with semantic-aware reasoning, transforming VLM-based detection into a principled decomposition across four semantic domains. We show that across a balanced set of real-world driving scenarios, applying SAVANT improves VLM's absolute recall by approximately 18.5% compared to prompting baselines. Moreover, this gain enables reliable large-scale annotation: leveraging the best proprietary model within our framework, we automatically labeled around 10,000 real-world images with high confidence. We use the resulting high-quality dataset to fine-tune a 7B open-source model (Qwen2.5-VL) to perform single-shot anomaly detection, achieving 90.8% recall and 93.8% accuracy - surpassing all models evaluated while enabling local deployment at near-zero cost. By coupling structured semantic reasoning with scalable data curation, we provide a practical solution to data scarcity in semantic anomaly detection for autonomous systems. Supplementary material: https://TUM-AVS.github.io/SAVANT/.

The advancement of vision-language models (VLMs) is hampered by a fragmented landscape of inconsistent and contaminated public datasets. We introduce FineVision, a meticulously collected, curated, and unified corpus of 24 million samples - the largest open resource of its kind. We unify more than 200 sources into 185 subsets via a semi-automated, human-in-the-loop pipeline: automation performs bulk ingestion and schema mapping, while reviewers audit mappings and spot-check outputs to verify faithful consumption of annotations, appropriate formatting and diversity, and safety; issues trigger targeted fixes and re-runs. The workflow further applies rigorous de-duplication within and across sources and decontamination against 66 public benchmarks. FineVision also encompasses agentic/GUI tasks with a unified action space; reviewers validate schemas and inspect a sample of trajectories to confirm executable fidelity. Models trained on FineVision consistently outperform those trained on existing open mixtures across a broad evaluation suite, underscoring the benefits of scale, data hygiene, and balanced automation with human oversight. We release the corpus and curation tools to accelerate data-centric VLM research.

Post-training pruning can substantially reduce LLM inference costs, but it often degrades quality unless the remaining weights are adapted. Since global retraining is expensive at LLM scale, recent work has largely focused on increasingly sophisticated pruning criteria that aim to select better sparsity patterns without adaptation. We revisit this trade-off through local reconstruction: after pruning, we adapt one subset of the model parameters at a time on a calibration set, training it to match the corresponding intermediate activations of the dense model. We evaluate local reconstruction across model families and scales, up to 72B parameters, and establish three main findings. First, local reconstruction is an effective adaptation mechanism for LLMs: it matches post-pruning retraining while using over an order of magnitude less data and compute, even when using PEFT techniques. Second, reconstruction exhibits a broad "free-lunch" regime in granularity, i.e., the reconstruction parameter window: as long as the reconstructed region contains at least a nonlinear submodule, final quality is largely insensitive to the window size, allowing granularity to be chosen primarily based on memory constraints. In contrast, reconstructing individual matrices, despite being the natural approach often proposed in the literature, consistently underperforms, as small matrix-level errors accumulate into larger activation drift. Lastly, reconstruction reduces the relative importance of the pruning criterion: performance gaps between sophisticated criteria and simple baselines shrink with model scale, making simple methods competitive again. Overall, our results challenge the prevailing view that post-pruning adaptation is impractical for LLMs.

The output of image the segmentation process is usually not very clear due to low quality features of Satellite images. The purpose of this study is to find a suitable Conditional Random Field (CRF) to achieve better clarity in a segmented image. We started with different types of CRFs and studied them as to why they are or are not suitable for our purpose. We evaluated our approach on two different datasets - Satellite imagery having low quality features and high quality Aerial photographs. During the study we experimented with various CRFs to find which CRF gives the best results on images and compared our results on these datasets to show the pitfalls and potentials of different approaches.

Iconicity, the resemblance between linguistic form and meaning, is pervasive in signed languages, offering a natural testbed for visual grounding. For vision-language models (VLMs), the challenge is to recover such essential mappings from dynamic human motion rather than static context. We introduce the Visual Iconicity Challenge, a novel video-based benchmark that adapts psycholinguistic measures to evaluate VLMs on three tasks: (i) phonological sign-form prediction (e.g., handshape, location), (ii) transparency (inferring meaning from visual form), and (iii) graded iconicity ratings. We assess 13 state-of-the-art VLMs in zero- and few-shot settings on Sign Language of the Netherlands and compare them to human baselines. On phonological form prediction, VLMs recover some handshape and location detail but remain below human performance; on transparency, they are far from human baselines; and only top models correlate moderately with human iconicity ratings. Interestingly, models with stronger phonological form prediction correlate better with human iconicity judgment, indicating shared sensitivity to visually grounded structure. Our findings validate these diagnostic tasks and motivate human-centric signals and embodied learning methods for modelling iconicity and improving visual grounding in multimodal models.

To drive progress in science and engineering, large language models (LLMs) must be able to process large amounts of numerical data and solve long calculations efficiently. This is currently only possible through the use of external tools or extensive reasoning chains, either weakening the numerical representations of LLMs or limiting the length of problems they can solve. We show that frontier LLMs require excessive amounts of reasoning tokens to solve even basic calculations, which is exacerbated by their tokenization strategies that split single numbers into multiple tokens. This motivates the need for efficient and effective single-token number encodings. We introduce a set of desiderata for such encodings and show that existing approaches fail to fulfill them. To address these shortcomings, we propose BitTokens, a novel encoding strategy that represents any number as a single token using its IEEE 754 binary floating-point representation. Through extensive experiments we show that our BitTokens allow even small language models to learn algorithms that solve basic arithmetic operations nearly perfectly. This newly gained efficiency could expand the length and complexity of problems language models can solve.

Foundation models are reshaping medical imaging, yet their application in echocardiography remains limited, hindered by a heavy reliance on private datasets that prevent reproducible comparison. Echocardiography poses unique challenges, including noisy acquisitions, high frame redundancy, and limited diverse public datasets. To address this, we introduce CardioBench, a comprehensive benchmark for echocardiography foundation models. Specifically, CardioBench unifies eight publicly available datasets into a standardized suite spanning four regression and five classification tasks, covering functional, structural, diagnostic, and view recognition endpoints. Leveraging this framework, we evaluate several leading foundation models, including cardiac-specific, biomedical, and general-purpose encoders, under consistent zero-shot, probing, and alignment protocols. Our analysis reveals that while general-purpose encoders transfer well and often close the gap with probing, they struggle significantly with fine-grained distinctions like view classification and subtle pathology recognition. Results indicate that models capturing temporal cardiac dynamics perform best on functional tasks, while retrieval-based approaches generalize more consistently across datasets. By releasing preprocessing, splits, and public evaluation pipelines, CardioBench establishes a reproducible reference point to guide the architectural design of future echocardiography and possibly other medical imaging foundation models.

Designing dense rewards is crucial for reinforcement learning (RL), yet in robotics it often demands extensive manual effort and lacks scalability. One promising solution is to view task progress as a dense reward signal, as it quantifies the degree to which actions advance the system toward task completion over time. We present TimeRewarder, a simple yet effective reward learning method that derives progress estimation signals from passive videos, including robot demonstrations and human videos, by modeling temporal distances between frame pairs. We then demonstrate how TimeRewarder can supply step-wise proxy rewards to guide reinforcement learning. In our comprehensive experiments on ten challenging Meta-World tasks, we show that TimeRewarder dramatically improves RL for sparse-reward tasks, achieving nearly perfect success in 9/10 tasks with only 200,000 environment interactions per task. This approach outperformed previous methods and even the manually designed environment dense reward on both the final success rate and sample efficiency. Moreover, we show that TimeRewarder pretraining can exploit real-world human videos, highlighting its potential as a scalable approach to rich reward signals from diverse video sources.

This article initiates the study of a basic question about model pruning. Given a vector $s$ of importance scores assigned to model components, how many of the scored components could be discarded without sacrificing performance? We propose Effective Model Pruning (EMP), which derives the desired sparsity directly from the score distribution using the notion of effective sample size from particle filtering, also known as the inverse Simpson index. Rather than prescribe a pruning criterion, EMP supplies a universal adaptive threshold derived from the distribution of the score $s$ over the model components: EMP maps $s$ to a number $N_{eff}=N_{eff}(s)$, called the effective sample size. The $N-N_{eff}$ lowest scoring components are discarded. A tight lower bound on the effective mass $s_{eff}$ (the sum of retained normalized scores) in terms of $N_{eff}$ is derived. This process yields models with a provable upper bound on the loss change relative to the original dense model. Numerical experiments are performed demonstrating this phenomenon across a variety of network architectures including MLPs, CNNs, Transformers, LLMs, and KAN. It is also shown that EMP addresses a rich set of pruning criteria such as weight magnitude, attention score, KAN importance score, and even feature-level signals such as image pixels.

Reinforcement learning (RL) struggles to scale to large, combinatorial action spaces common in many real-world problems. This paper introduces a novel framework for training discrete diffusion models as highly effective policies in these complex settings. Our key innovation is an efficient online training process that ensures stable and effective policy improvement. By leveraging policy mirror descent (PMD) to define an ideal, regularized target policy distribution, we frame the policy update as a distributional matching problem, training the expressive diffusion model to replicate this stable target. This decoupled approach stabilizes learning and significantly enhances training performance. Our method achieves state-of-the-art results and superior sample efficiency across a diverse set of challenging combinatorial benchmarks, including DNA sequence generation, RL with macro-actions, and multi-agent systems. Experiments demonstrate that our diffusion policies attain superior performance compared to other baselines.

Accurate multi-needle localization in intraoperative CT images is crucial for optimizing seed placement in pelvic seed implant brachytherapy. However, this task is challenging due to poor image contrast and needle adhesion. This paper presents a novel approach that reframes needle localization as a tip-handle detection and matching problem to overcome these difficulties. An anchor-free network, based on HRNet, is proposed to extract multi-scale features and accurately detect needle tips and handles by predicting their centers and orientations using decoupled branches for heatmap regression and polar angle prediction. To associate detected tips and handles into individual needles, a greedy matching and merging (GMM) method designed to solve the unbalanced assignment problem with constraints (UAP-C) is presented. The GMM method iteratively selects the most probable tip-handle pairs and merges them based on a distance metric to reconstruct 3D needle paths. Evaluated on a dataset of 100 patients, the proposed method demonstrates superior performance, achieving higher precision and F1 score compared to a segmentation-based method utilizing the nnUNet model,thereby offering a more robust and accurate solution for needle localization in complex clinical scenarios.

Vision Transformers (ViTs) achieve state-of-the-art performance in semantic segmentation but are hindered by high computational and memory costs. To address this, we propose STEP (SuperToken and Early-Pruning), a hybrid token-reduction framework that combines dynamic patch merging and token pruning to enhance efficiency without significantly compromising accuracy. At the core of STEP is dCTS, a lightweight CNN-based policy network that enables flexible merging into superpatches. Encoder blocks integrate also early-exits to remove high-confident supertokens, lowering computational load. We evaluate our method on high-resolution semantic segmentation benchmarks, including images up to 1024 x 1024, and show that when dCTS is applied alone, the token count can be reduced by a factor of 2.5 compared to the standard 16 x 16 pixel patching scheme. This yields a 2.6x reduction in computational cost and a 3.4x increase in throughput when using ViT-Large as the backbone. Applying the full STEP framework further improves efficiency, reaching up to a 4x reduction in computational complexity and a 1.7x gain in inference speed, with a maximum accuracy drop of no more than 2.0%. With the proposed STEP configurations, up to 40% of tokens can be confidently predicted and halted before reaching the final encoder layer.

Generative recommendation plays a crucial role in personalized systems, predicting users' future interactions from their historical behavior sequences. A critical yet underexplored factor in training these models is data augmentation, the process of constructing training data from user interaction histories. By shaping the training distribution, data augmentation directly and often substantially affects model generalization and performance. Nevertheless, in much of the existing work, this process is simplified, applied inconsistently, or treated as a minor design choice, without a systematic and principled understanding of its effects. Motivated by our empirical finding that different augmentation strategies can yield large performance disparities, we conduct an in-depth analysis of how they reshape training distributions and influence alignment with future targets and generalization to unseen inputs. To systematize this design space, we propose GenPAS, a generalized and principled framework that models augmentation as a stochastic sampling process over input-target pairs with three bias-controlled steps: sequence sampling, target sampling, and input sampling. This formulation unifies widely used strategies as special cases and enables flexible control of the resulting training distribution. Our extensive experiments on benchmark and industrial datasets demonstrate that GenPAS yields superior accuracy, data efficiency, and parameter efficiency compared to existing strategies, providing practical guidance for principled training data construction in generative recommendation. Our code is available at https://github.com/snap-research/GenPAS.

3D Gaussian Splatting (3DGS) is rapidly gaining popularity for its photorealistic rendering quality and real-time performance, but it generates massive amounts of data. Hence compressing 3DGS data is necessary for the cost effectiveness of 3DGS models. Recently, several anchor-based neural compression methods have been proposed, achieving good 3DGS compression performance. However, they all rely on uniform scalar quantization (USQ) due to its simplicity. A tantalizing question is whether more sophisticated quantizers can improve the current 3DGS compression methods with very little extra overhead and minimal change to the system. The answer is yes by replacing USQ with lattice vector quantization (LVQ). To better capture scene-specific characteristics, we optimize the lattice basis for each scene, improving LVQ's adaptability and R-D efficiency. This scene-adaptive LVQ (SALVQ) strikes a balance between the R-D efficiency of vector quantization and the low complexity of USQ. SALVQ can be seamlessly integrated into existing 3DGS compression architectures, enhancing their R-D performance with minimal modifications and computational overhead. Moreover, by scaling the lattice basis vectors, SALVQ can dynamically adjust lattice density, enabling a single model to accommodate multiple bit rate targets. This flexibility eliminates the need to train separate models for different compression levels, significantly reducing training time and memory consumption.

Multi-Target Multi-Camera Tracking (MTMC) is an essential computer vision task for automating large-scale surveillance. With camera calibration and depth information, the targets in the scene can be projected into 3D space, offering unparalleled levels of automatic perception of a 3D environment. However, tracking in the 3D space requires replacing all 2D tracking components from the ground up, which may be infeasible for existing MTMC systems. In this paper, we present an approach for extending any online 2D multi-camera tracking system into 3D space by utilizing depth information to reconstruct a target in point-cloud space, and recovering its 3D box through clustering and yaw refinement following tracking. We also introduced an enhanced online data association mechanism that leverages the target's local ID consistency to assign global IDs across frames. The proposed framework is evaluated on the 2025 AI City Challenge's 3D MTMC dataset, achieving 3rd place on the leaderboard.