Patch-based methods are widely used in 3D medical image segmentation to address memory constraints in processing high-resolution volumetric data. However, these approaches often neglect the patch's location within the global volume, which can limit segmentation performance when anatomical context is important. In this paper, we investigate the role of location context in patch-based 3D segmentation and propose a novel attention mechanism, LocBAM, that explicitly processes spatial information. Experiments on BTCV, AMOS22, and KiTS23 demonstrate that incorporating location context stabilizes training and improves segmentation performance, particularly under low patch-to-volume coverage where global context is missing. Furthermore, LocBAM consistently outperforms classical coordinate encoding via CoordConv. Code is publicly available at https://github.com/compai-lab/2026-ISBI-hooft

Multi-modal object tracking has attracted considerable attention by integrating multiple complementary inputs (e.g., thermal, depth, and event data) to achieve outstanding performance. Although current general-purpose multi-modal trackers primarily unify various modal tracking tasks (i.e., RGB-Thermal infrared, RGB-Depth or RGB-Event tracking) through prompt learning, they still overlook the effective capture of spatio-temporal cues. In this work, we introduce a novel multi-modal tracking framework based on a mamba-style state space model, termed UBATrack. Our UBATrack comprises two simple yet effective modules: a Spatio-temporal Mamba Adapter (STMA) and a Dynamic Multi-modal Feature Mixer. The former leverages Mamba's long-sequence modeling capability to jointly model cross-modal dependencies and spatio-temporal visual cues in an adapter-tuning manner. The latter further enhances multi-modal representation capacity across multiple feature dimensions to improve tracking robustness. In this way, UBATrack eliminates the need for costly full-parameter fine-tuning, thereby improving the training efficiency of multi-modal tracking algorithms. Experiments show that UBATrack outperforms state-of-the-art methods on RGB-T, RGB-D, and RGB-E tracking benchmarks, achieving outstanding results on the LasHeR, RGBT234, RGBT210, DepthTrack, VOT-RGBD22, and VisEvent datasets.

Designing good reflection questions is pedagogically important but time-consuming and unevenly supported across teachers. This paper introduces a reflection-in-reflection framework for automated generation of reflection questions with large language models (LLMs). Our approach coordinates two role-specialized agents, a Student-Teacher and a Teacher-Educator, that engage in a Socratic multi-turn dialogue to iteratively refine a single question given a teacher-specified topic, key concepts, student level, and optional instructional materials. The Student-Teacher proposes candidate questions with brief rationales, while the Teacher-Educator evaluates them along clarity, depth, relevance, engagement, and conceptual interconnections, responding only with targeted coaching questions or a fixed signal to stop the dialogue. We evaluate the framework in an authentic lower-secondary ICT setting on the topic, using GPT-4o-mini as the backbone model and a stronger GPT- 4-class LLM as an external evaluator in pairwise comparisons of clarity, relevance, depth, and overall quality. First, we study how interaction design and context (dynamic vs.fixed iteration counts; presence or absence of student level and materials) affect question quality. Dynamic stopping combined with contextual information consistently outperforms fixed 5- or 10-step refinement, with very long dialogues prone to drift or over-complication. Second, we show that our two-agent protocol produces questions that are judged substantially more relevant and deeper, and better overall, than a one-shot baseline using the same backbone model.

Current remote sensing change detection (CD) methods mainly rely on specialized models, which limits the scalability toward modality-adaptive Earth observation. For homogeneous CD, precise boundary delineation relies on fine-grained spatial cues and local pixel interactions, whereas heterogeneous CD instead requires broader contextual information to suppress speckle noise and geometric distortions. Moreover, difference operator (e.g., subtraction) works well for aligned homogeneous images but introduces artifacts in cross-modal or geometrically misaligned scenarios. Across different modality settings, specialized models based on static backbones or fixed difference operations often prove insufficient. To address this challenge, we propose UniRoute, a unified framework for modality-adaptive learning by reformulating feature extraction and fusion as conditional routing problems. We introduce an Adaptive Receptive Field Routing MoE (AR2-MoE) module to disentangle local spatial details from global semantic context, and a Modality-Aware Difference Routing MoE (MDR-MoE) module to adaptively select the most suitable fusion primitive at each pixel. In addition, we propose a Consistency-Aware Self-Distillation (CASD) strategy that stabilizes unified training under data-scarce heterogeneous settings by enforcing multi-level consistency. Extensive experiments on five public datasets demonstrate that UniRoute achieves strong overall performance, with a favorable accuracy-efficiency trade-off under a unified deployment setting.

The scarcity of training data presents a fundamental challenge in applying deep learning to archaeological artifact classification, particularly for the rare types of Chinese porcelain. This study investigates whether synthetic images generated through Stable Diffusion with Low-Rank Adaptation (LoRA) can effectively augment limited real datasets for multi-task CNN-based porcelain classification. Using MobileNetV3 with transfer learning, we conducted controlled experiments comparing models trained on pure real data against those trained on mixed real-synthetic datasets (95:5 and 90:10 ratios) across four classification tasks: dynasty, glaze, kiln and type identification. Results demonstrate task-specific benefits: type classification showed the most substantial improvement (5.5\% F1-macro increase with 90:10 ratio), while dynasty and kiln tasks exhibited modest gains (3-4\%), suggesting that synthetic augmentation effectiveness depends on the alignment between generated features and task-relevant visual signatures. Our work contributes practical guidelines for deploying generative AI in archaeological research, demonstrating both the potential and limitations of synthetic data when archaeological authenticity must be balanced with data diversity.

While Large Language Models demonstrate remarkable proficiency in high-level semantic planning, they remain limited in handling fine-grained, low-level web component manipulations. To address this limitation, extensive research has focused on enhancing model grounding capabilities through techniques such as Reinforcement Learning. However, rather than compelling agents to adapt to human-centric interfaces, we propose constructing interaction interfaces specifically optimized for agents. This paper introduces Component Interface for Agent (CI4A), a semantic encapsulation mechanism that abstracts the complex interaction logic of UI components into a set of unified tool primitives accessible to agents. We implemented CI4A within Ant Design, an industrial-grade front-end framework, covering 23 categories of commonly used UI components. Furthermore, we developed a hybrid agent featuring an action space that dynamically updates according to the page state, enabling flexible invocation of available CI4A tools. Leveraging the CI4A-integrated Ant Design, we refactored and upgraded the WebArena benchmark to evaluate existing SoTA methods. Experimental results demonstrate that the CI4A-based agent significantly outperforms existing approaches, achieving a new SoTA task success rate of 86.3%, alongside substantial improvements in execution efficiency.

Diffusion models have seen widespread adoption for text-driven human motion generation and related tasks due to their impressive generative capabilities and flexibility. However, current motion diffusion models face two major limitations: a representational gap caused by pre-trained text encoders that lack motion-specific information, and error propagation during the iterative denoising process. This paper introduces Reconstruction-Anchored Diffusion Model (RAM) to address these challenges. First, RAM leverages a motion latent space as intermediate supervision for text-to-motion generation. To this end, RAM co-trains a motion reconstruction branch with two key objective functions: self-regularization to enhance the discrimination of the motion space and motion-centric latent alignment to enable accurate mapping from text to the motion latent space. Second, we propose Reconstructive Error Guidance (REG), a testing-stage guidance mechanism that exploits the diffusion model's inherent self-correction ability to mitigate error propagation. At each denoising step, REG uses the motion reconstruction branch to reconstruct the previous estimate, reproducing the prior error patterns. By amplifying the residual between the current prediction and the reconstructed estimate, REG highlights the improvements in the current prediction. Extensive experiments demonstrate that RAM achieves significant improvements and state-of-the-art performance. Our code will be released.

Recent advances in text-to-music generation (TTM) have yielded high-quality results, but often at the cost of extensive compute and the use of large proprietary internal data. To improve the affordability and openness of TTM training, an open-source generative model backbone that is more training- and data-efficient is needed. In this paper, we constrain the number of trainable parameters in the generative model to match that of the MusicGen-small benchmark (with about 300M parameters), and replace its Transformer backbone with the emerging class of state-space models (SSMs). Specifically, we explore different SSM variants for sequence modeling, and compare a single-stage SSM-based design with a decomposable two-stage SSM/diffusion hybrid design. All proposed models are trained from scratch on a purely public dataset comprising 457 hours of CC-licensed music, ensuring full openness. Our experimental findings are three-fold. First, we show that SSMs exhibit superior training efficiency compared to the Transformer counterpart. Second, despite using only 9% of the FLOPs and 2% of the training data size compared to the MusicGen-small benchmark, our model achieves competitive performance in both objective metrics and subjective listening tests based on MusicCaps captions. Finally, our scaling-down experiment demonstrates that SSMs can maintain competitive performance relative to the Transformer baseline even at the same training budget (measured in iterations), when the model size is reduced to four times smaller. To facilitate the democratization of TTM research, the processed captions, model checkpoints, and source code are available on GitHub via the project page: https://lonian6.github.io/ssmttm/.

Achieving bound consistency for the no-overlap constraint is known to be NP-complete. Therefore, several polynomial-time tightening techniques, such as edge finding, not-first-not-last reasoning, and energetic reasoning, have been introduced for this constraint. In this work, we derive the first bound-consistent algorithm for the no-overlap constraint. By building on the no-overlap MDD defined by Ciré and van Hoeve, we extract bounds of the time window of the jobs, allowing us to tighten start and end times in time polynomial in the number of nodes of the MDD. Similarly, to bound the size and time-complexity, we limit the width of the MDD to a threshold, creating a relaxed MDD that can also be used to relax the bound-consistent filtering. Through experiments on a sequencing problem with time windows and a just-in-time objective ($1 \mid r_j, d_j, \bar{d}_j \mid \sum E_j + \sum T_j$), we observe that the proposed filtering, even with a threshold on the width, achieves a stronger reduction in the number of nodes visited in the search tree compared to the previously proposed precedence-detection algorithm of Ciré and van Hoeve. The new filtering also appears to be complementary to classical propagation methods for the no-overlap constraint, allowing a substantial reduction in both the number of nodes and the solving time on several instances.

Recognizing and navigating client resistance is critical for effective mental health counseling, yet detecting such behaviors is particularly challenging in text-based interactions. Existing NLP approaches oversimplify resistance categories, ignore the sequential dynamics of therapeutic interventions, and offer limited interpretability. To address these limitations, we propose PsyFIRE, a theoretically grounded framework capturing 13 fine-grained resistance behaviors alongside collaborative interactions. Based on PsyFIRE, we construct the ClientResistance corpus with 23,930 annotated utterances from real-world Chinese text-based counseling, each supported by context-specific rationales. Leveraging this dataset, we develop RECAP, a two-stage framework that detects resistance and fine-grained resistance types with explanations. RECAP achieves 91.25% F1 for distinguishing collaboration and resistance and 66.58% macro-F1 for fine-grained resistance categories classification, outperforming leading prompt-based LLM baselines by over 20 points. Applied to a separate counseling dataset and a pilot study with 62 counselors, RECAP reveals the prevalence of resistance, its negative impact on therapeutic relationships and demonstrates its potential to improve counselors' understanding and intervention strategies.

Movie dubbing is the task of synthesizing speech from scripts conditioned on video scenes, requiring accurate lip sync, faithful timbre transfer, and proper modeling of character identity and emotion. However, existing methods face two major limitations: (1) high-quality multimodal dubbing datasets are limited in scale, suffer from high word error rates, contain sparse annotations, rely on costly manual labeling, and are restricted to monologue scenes, all of which hinder effective model training; (2) existing dubbing models rely solely on the lip region to learn audio-visual alignment, which limits their applicability to complex live-action cinematic scenes, and exhibit suboptimal performance in lip sync, speech quality, and emotional expressiveness. To address these issues, we propose FunCineForge, which comprises an end-to-end production pipeline for large-scale dubbing datasets and an MLLM-based dubbing model designed for diverse cinematic scenes. Using the pipeline, we construct the first Chinese television dubbing dataset with rich annotations, and demonstrate the high quality of these data. Experiments across monologue, narration, dialogue, and multi-speaker scenes show that our dubbing model consistently outperforms SOTA methods in audio quality, lip sync, timbre transfer, and instruction following. Code and demos are available at https://anonymous.4open.science/w/FunCineForge.

This study investigates the feature representations produced by publicly available open source medical vision-language models (VLMs). While medical VLMs are expected to capture diagnostically relevant features, their learned representations remain underexplored, and standard evaluations like classification accuracy do not fully reveal if they acquire truly discriminative, lesion-specific features. Understanding these representations is crucial for revealing medical image structures and improving downstream tasks in medical image analysis. This study aims to investigate the feature distributions learned by medical VLMs and evaluate the impact of medical specialization. We analyze the feature distribution of multiple image modalities extracted by some representative medical VLMs across lesion classification datasets on multiple modalities. These distributions were compared them with non-medical VLMs to assess the domain-specific medical training. Our experiments showed that medical VLMs can extract discriminative features that are effective for medical classification tasks. Moreover, it was found that non-medical VLMs with recent improvement with contextual enrichment such as LLM2CLIP produce more refined feature representations. Our results imply that enhancing text encoder is more crucial than training intensively on medical images when developing medical VLMs. Notably, non-medical models are particularly vulnerable to biases introduced by overlaied text strings on images. These findings underscore the need for careful consideration on model selection according to downstream tasks besides potential risks in inference due to background biases such as textual information in images.

Video summarization is a crucial technique for social understanding, enabling efficient browsing of massive multimedia content and extraction of key information from social platforms. Most existing unsupervised summarization methods rely on Generative Adversarial Networks (GANs) to enhance keyframe selection and generate coherent, video summaries through adversarial training. However, such approaches primarily exploit unimodal features, overlooking the guiding role of semantic information in keyframe selection, and often suffer from unstable training. To address these limitations, we propose a novel Semantic-Guided Unsupervised Video Summarization method. Specifically, we design a novel frame-level semantic alignment attention mechanism and integrate it into a keyframe selector, which guides the Transformer-based generator within the adversarial framework to better reconstruct videos. In addition, we adopt an incremental training strategy to progressively update the model components, effectively mitigating the instability of GAN training. Experimental results demonstrate that our approach achieves superior performance on multiple benchmark datasets.

Identifying unique polyps in colon capsule endoscopy (CCE) images is a critical yet challenging task for medical personnel due to the large volume of images, the cognitive load it creates for clinicians, and the ambiguity in labeling specific frames. This paper formulates this problem as a multi-instance learning (MIL) task, where a query polyp image is compared with a target bag of images to determine uniqueness. We employ a multi-instance verification (MIV) framework that incorporates attention mechanisms, such as variance-excited multi-head attention (VEMA) and distance-based attention (DBA), to enhance the model's ability to extract meaningful representations. Additionally, we investigate the impact of self-supervised learning using SimCLR to generate robust embeddings. Experimental results on a dataset of 1912 polyps from 754 patients demonstrate that attention mechanisms significantly improve performance, with DBA L1 achieving the highest test accuracy of 86.26\% and a test AUC of 0.928 using a ConvNeXt backbone with SimCLR pretraining. This study underscores the potential of MIL and self-supervised learning in advancing automated analysis of Colon Capsule Endoscopy images, with implications for broader medical imaging applications.

In recent years, significant progress has been made on algorithms for learning optimal decision trees, primarily in the context of binary features. Extending these methods to continuous features remains substantially more challenging due to the large number of potential splits for each feature. Recently, an elegant exact algorithm was proposed for learning optimal decision trees with continuous features; however, the rapidly increasing computational time limits its practical applicability to shallow depths (typically 3 or 4). It relies on a depth-first search optimization strategy that fully optimizes the left subtree of each split before exploring the corresponding right subtree. While effective in finding optimal solutions given sufficient time, this strategy can lead to poor anytime behavior: when interrupted early, the best-found tree is often highly unbalanced and suboptimal. In such cases, purely greedy methods such as C4.5 may, paradoxically, yield better solutions. To address this limitation, we propose an anytime, yet complete approach leveraging limited discrepancy search, distributing the computational effort more evenly across the entire tree structure, and thus ensuring that a high-quality decision tree is available at any interruption point. Experimental results show that our approach outperforms the existing one in terms of anytime performance.

Interpretability is significant in computational pathology, leading to the development of multimodal information integration from histopathological image and corresponding text data.However, existing multimodal methods have limited interpretability due to the lack of high-quality dataset that support explicit reasoning and inference and simple reasoning process.To address the above problems, we introduce a novel multimodal pathology large language model with strong reasoning capabilities.To improve the generation of accurate and contextually relevant textual descriptions, we design a semantic reward strategy integrated with group relative policy optimization.We construct a high-quality pathology visual question answering (VQA) dataset, specifically designed to support complex reasoning tasks.Comprehensive experiments conducted on this dataset demonstrate that our method outperforms state-of-the-art methods, even when trained with only 20% of the data.Our method also achieves comparable performance on downstream zero-shot image classification task compared with CLIP.

Achieving pronunciation proficiency in a second language (L2) remains a challenge, despite the development of Computer-Assisted Pronunciation Training (CAPT) systems. Traditional CAPT systems often provide unintuitive feedback that lacks actionable guidance, limiting its effectiveness. Recent advancements in audio-language models (ALMs) offer the potential to enhance these systems by providing more user-friendly feedback. In this work, we investigate ALMs for chat-based pronunciation training by introducing L2-Arctic-plus, an English dataset with detailed error explanations and actionable suggestions for improvement. We benchmark cascaded ASR+LLMs and existing ALMs on this dataset, specifically in detecting mispronunciation and generating actionable feedback. To improve the performance, we further propose to instruction-tune ALMs on L2-Arctic-plus. Experimental results demonstrate that our instruction-tuned models significantly outperform existing baselines on mispronunciation detection and suggestion generation in terms of both objective and human evaluation, highlighting the value of the proposed dataset.

Reliable drone detection is challenging due to limited annotated real-world data, large appearance variability, and the presence of visually similar distractors such as birds. To address these challenges, this paper introduces SimD3, a large-scale high-fidelity synthetic dataset designed for robust drone detection in complex aerial environments. Unlike existing synthetic drone datasets, SimD3 explicitly models drones with heterogeneous payloads, incorporates multiple bird species as realistic distractors, and leverages diverse Unreal Engine 5 environments with controlled weather, lighting, and flight trajectories captured using a 360 six-camera rig. Using SimD3, we conduct an extensive experimental evaluation within the YOLOv5 detection framework, including an attention-enhanced variant termed Yolov5m+C3b, where standard bottleneck-based C3 blocks are replaced with C3b modules. Models are evaluated on synthetic data, combined synthetic and real data, and multiple unseen real-world benchmarks to assess robustness and generalization. Experimental results show that SimD3 provides effective supervision for small-object drone detection and that Yolov5m+C3b consistently outperforms the baseline across in-domain and cross-dataset evaluations. These findings highlight the utility of SimD3 for training and benchmarking robust drone detection models under diverse and challenging conditions.

Artificial Intelligence-Generated Content (AIGC) has made significant strides, with high-resolution text-to-image (T2I) generation becoming increasingly critical for improving users' Quality of Experience (QoE). Although resource-constrained edge computing adequately supports fast low-resolution T2I generations, achieving high-resolution output still faces the challenge of ensuring image fidelity at the cost of latency. To address this, we first investigate the performance of super-resolution (SR) methods for image enhancement, confirming a fundamental trade-off that lightweight learning-based SR struggles to recover fine details, while diffusion-based SR achieves higher fidelity at a substantial computational cost. Motivated by these observations, we propose an end-edge collaborative generation-enhancement framework. Upon receiving a T2I generation task, the system first generates a low-resolution image based on adaptively selected denoising steps and super-resolution scales at the edge side, which is then partitioned into patches and processed by a region-aware hybrid SR policy. This policy applies a diffusion-based SR model to foreground patches for detail recovery and a lightweight learning-based SR model to background patches for efficient upscaling, ultimately stitching the enhanced ones into the high-resolution image. Experiments show that our system reduces service latency by 33% compared with baselines while maintaining competitive image quality.

The rapid evolution of diffusion models has democratized face swapping but also raises concerns about privacy and identity security. Existing proactive defenses, often adapted from image editing attacks, prove ineffective in this context. We attribute this failure to an oversight of the structural resilience and the unique static conditional guidance mechanism inherent in face swapping systems. To address this, we propose VoidFace, a systemic defense method that views face swapping as a coupled identity pathway. By injecting perturbations at critical bottlenecks, VoidFace induces cascading disruption throughout the pipeline. Specifically, we first introduce localization disruption and identity erasure to degrade physical regression and semantic embeddings, thereby impairing the accurate modeling of the source face. We then intervene in the generative domain by decoupling attention mechanisms to sever identity injection, and corrupting intermediate diffusion features to prevent the reconstruction of source identity. To ensure visual imperceptibility, we perform adversarial search in the latent manifold, guided by a perceptual adaptive strategy to balance attack potency with image quality. Extensive experiments show that VoidFace outperforms existing defenses across various diffusion-based swapping models, while producing adversarial faces with superior visual quality.

Despite the widespread success of Graph Neural Networks (GNNs), understanding the reasons behind their specific predictions remains challenging. Existing explainability methods face a trade-off that gradient-based approaches are computationally efficient but often ignore structural interactions, while game-theoretic techniques capture interactions at the cost of high computational overhead and potential deviation from the model's true reasoning path. To address this gap, we propose FSX (Message Flow Sensitivity Enhanced Structural Explainer), a novel hybrid framework that synergistically combines the internal message flows of the model with a cooperative game approach applied to the external graph data. FSX first identifies critical message flows via a novel flow-sensitivity analysis: during a single forward pass, it simulates localized node perturbations and measures the resulting changes in message flow intensities. These sensitivity-ranked flows are then projected onto the input graph to define compact, semantically meaningful subgraphs. Within each subgraph, a flow-aware cooperative game is conducted, where node contributions are evaluated fairly through a Shapley-like value that incorporates both node-feature importance and their roles in sustaining or destabilizing the identified critical flows. Extensive evaluation across multiple datasets and GNN architectures demonstrates that FSX achieves superior explanation fidelity with significantly reduced runtime, while providing unprecedented insights into the structural logic underlying model predictions--specifically, how important sub-structures exert influence by governing the stability of key internal computational pathways.

Document extraction is a core component of digital workflows, yet existing vision-language models (VLMs) predominantly favor high-resource languages. Thai presents additional challenges due to script complexity from non-latin letters, the absence of explicit word boundaries, and the prevalence of highly unstructured real-world documents, limiting the effectiveness of current open-source models. This paper presents Typhoon OCR, an open VLM for document extraction tailored for Thai and English. The model is fine-tuned from vision-language backbones using a Thai-focused training dataset. The dataset is developed using a multi-stage data construction pipeline that combines traditional OCR, VLM-based restructuring, and curated synthetic data. Typhoon OCR is a unified framework capable of text transcription, layout reconstruction, and document-level structural consistency. The latest iteration of our model, Typhoon OCR V1.5, is a compact and inference-efficient model designed to reduce reliance on metadata and simplify deployment. Comprehensive evaluations across diverse Thai document categories, including financial reports, government forms, books, infographics, and handwritten documents, show that Typhoon OCR achieves performance comparable to or exceeding larger frontier proprietary models, despite substantially lower computational cost. The results demonstrate that open vision-language OCR models can achieve accurate text extraction and layout reconstruction for Thai documents, reaching performance comparable to proprietary systems while remaining lightweight and deployable.

Weakly Supervised Semantic Segmentation (WSSS), which relies only on image-level labels, has attracted significant attention for its cost-effectiveness and scalability. Existing methods mainly enhance inter-class distinctions and employ data augmentation to mitigate semantic ambiguity and reduce spurious activations. However, they often neglect the complex contextual dependencies among image patches, resulting in incomplete local representations and limited segmentation accuracy. To address these issues, we propose the Context Patch Fusion with Class Token Enhancement (CPF-CTE) framework, which exploits contextual relations among patches to enrich feature representations and improve segmentation. At its core, the Contextual-Fusion Bidirectional Long Short-Term Memory (CF-BiLSTM) module captures spatial dependencies between patches and enables bidirectional information flow, yielding a more comprehensive understanding of spatial correlations. This strengthens feature learning and segmentation robustness. Moreover, we introduce learnable class tokens that dynamically encode and refine class-specific semantics, enhancing discriminative capability. By effectively integrating spatial and semantic cues, CPF-CTE produces richer and more accurate representations of image content. Extensive experiments on PASCAL VOC 2012 and MS COCO 2014 validate that CPF-CTE consistently surpasses prior WSSS methods.

We present PCL-Reasoner-V1.5, a 32-billion-parameter large language model (LLM) for mathematical reasoning. The model is built upon Qwen2.5-32B and refined via supervised fine-tuning (SFT) followed by reinforcement learning (RL). A central innovation is our proposed offline RL method, which provides superior training stability and efficiency over standard online RL methods such as GRPO. Our model achieves state-of-the-art performance among models post-trained on Qwen2.5-32B, attaining average accuracies of 90.9% on AIME 2024 and 85.6% on AIME 2025. Our work demonstrates offline RL as a stable and efficient paradigm for advancing reasoning in LLMs. All experiments were conducted on Huawei Ascend 910C NPUs.

Optimizing the advertiser's cumulative value of winning impressions under budget constraints poses a complex challenge in online advertising, under the paradigm of AI-Generated Bidding (AIGB). Advertisers often have personalized objectives but limited historical interaction data, resulting in few-shot scenarios where traditional reinforcement learning (RL) methods struggle to perform effectively. Large Language Models (LLMs) offer a promising alternative for AIGB by leveraging their in-context learning capabilities to generalize from limited data. However, they lack the numerical precision required for fine-grained optimization. To address this limitation, we introduce GRPO-Adaptive, an efficient LLM post-training strategy that enhances both reasoning and numerical precision by dynamically updating the reference policy during training. Built upon this foundation, we further propose DARA, a novel dual-phase framework that decomposes the decision-making process into two stages: a few-shot reasoner that generates initial plans via in-context prompting, and a fine-grained optimizer that refines these plans using feedback-driven reasoning. This separation allows DARA to combine LLMs' in-context learning strengths with precise adaptability required by AIGB tasks. Extensive experiments on both real-world and synthetic data environments demonstrate that our approach consistently outperforms existing baselines in terms of cumulative advertiser value under budget constraints.

Optimally sequencing experimental assays in drug discovery is a high-stakes planning problem under severe uncertainty and resource constraints. A primary obstacle for standard reinforcement learning (RL) is the absence of an explicit environment simulator or transition data $(s, a, s')$; planning must rely solely on a static database of historical outcomes. We introduce the Implicit Bayesian Markov Decision Process (IBMDP), a model-based RL framework designed for such simulator-free settings. IBMDP constructs a case-guided implicit model of transition dynamics by forming a nonparametric belief distribution using similar historical outcomes. This mechanism enables Bayesian belief updating as evidence accumulates and employs ensemble MCTS planning to generate stable policies that balance information gain toward desired outcomes with resource efficiency. We validate IBMDP through comprehensive experiments. On a real-world central nervous system (CNS) drug discovery task, IBMDP reduced resource consumption by up to 92\% compared to established heuristics while maintaining decision confidence. To rigorously assess decision quality, we also benchmarked IBMDP in a synthetic environment with a computable optimal policy. Our framework achieves significantly higher alignment with this optimal policy than a deterministic value iteration alternative that uses the same similarity-based model, demonstrating the superiority of our ensemble planner. IBMDP offers a practical solution for sequential experimental design in data-rich but simulator-poor domains.

As the commercial surgical guide design software usually does not support the export of implant position for pre-implantation data, existing methods have to scan the post-implantation data and map the implant to pre-implantation space to get the label of implant position for training. Such a process is time-consuming and heavily relies on the accuracy of registration algorithm. Moreover, not all hospitals have paired CBCT data, limitting the construction of multi-center dataset. Inspired by the way dentists determine the implant position based on the neighboring tooth texture, we found that even if the implant area is masked, it will not affect the determination of the implant position. Therefore, we propose to mask the implants in the post-implantation data so that any CBCT containing the implants can be used as training data. This paradigm enables us to discard the registration process and makes it possible to construct a large-scale multi-center implant dataset. On this basis, we proposes ImplantFairy, a comprehensive, publicly accessible dental implant dataset with voxel-level 3D annotations of 1622 CBCT data. Furthermore, according to the area variation characteristics of the tooth's spatial structure and the slope information of the implant, we designed a slope-aware implant position prediction network. Specifically, a neighboring distance perception (NDP) module is designed to adaptively extract tooth area variation features, and an implant slope prediction branch assists the network in learning more robust features through additional implant supervision information. Extensive experiments conducted on ImplantFairy and two public dataset demonstrate that the proposed RegFreeNet achieves the state-of-the-art performance.

Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated promising gains in enhancing the reasoning capabilities of large language models. However, its dependence on domain-specific verifiers significantly restricts its applicability to open and general domains. Recent efforts such as RLPR have extended RLVR to general domains, enabling training on broader datasets and achieving improvements over RLVR. However, a notable limitation of these methods is their tendency to overfit to reference answers, which constrains the model's ability to generate diverse outputs. This limitation is particularly pronounced in open-ended tasks such as writing, where multiple plausible answers exist. To address this, we propose DARL, a simple yet effective reinforcement learning framework that encourages the generation of diverse answers within a controlled deviation range from the reference while preserving alignment with it. Our framework is fully compatible with existing general reinforcement learning methods and can be seamlessly integrated without additional verifiers. Extensive experiments on thirteen benchmarks demonstrate consistent improvements in reasoning performance. Notably, DARL surpasses RLPR, achieving average gains of 1.3 points on six reasoning benchmarks and 9.5 points on seven general benchmarks, highlighting its effectiveness in improving both reasoning accuracy and output diversity.

Tool-Integrated Reasoning (TIR) has significantly enhanced the capabilities of Large Language Models (LLMs), yet current agents tend to exhibit cognitive offloading, redundantly invoking external tools even for simple tasks. In this paper, we suggest that true agentic intelligence requires not just tool invocation, but the adaptive wisdom to discern when to use them. We propose AdaTIR, a framework that shifts the paradigm from static tool invocation to difficulty-aware reasoning internalization. By introducing a difficulty-aware efficiency reward, AdaTIR dynamically adjusts tool budgets based on task complexity--internalizing reasoning for simple tasks while selectively invoking tools for complex tasks. Furthermore, we identify a sign reversal problem where tool penalties outweigh correctness rewards, mistakenly penalizing correct rollouts with negative advantages. To resolve this, we propose Clipped Advantage Shaping (CAS), which ensures that correctness remains the primary objective while using efficiency as a secondary constraint. Empirical results demonstrate that AdaTIR reduces tool calls by up to 97.6% on simple tasks and 28.2% on complex challenges while maintaining or enhancing accuracy. Notably, AdaTIR successfully internalizes reasoning, outperforming baselines by 4.8% on AIME 2024 even when tool access is strictly disabled.

Training Large Reasoning Model (LRM) is usually unstable and unpredictable, especially on hard problems or weak foundation models. We found that the current post-training scaling strategy can still improve on these cases. We propose CoScale-RL, a novel scaling strategy with better data and computational efficiency. We first scale up solutions to make problems solvable. The core idea is to collect multiple solutions for each problem, rather than simply enlarging the dataset. Then, we scale up rollout computation to stabilize Reinforcement Learning. We further leverage a model merge technique called Re-distillation to sustain or even improve computational efficiency when scaling up. Our method significantly improves data and computational efficiency, with an average 3.76$\times$ accuracy improvement on four benchmarks. CoScale-RL is able to improve an LRM's ability boundary without an extensive SFT dataset. Our method provides a new scaling direction to further improve LRM's reasoning ability.