Hyperbolic space is quickly gaining traction as a promising geometry for hierarchical and robust representation learning. A core open challenge is the development of a mathematical formulation of hyperbolic neural networks that is both efficient and captures the key properties of hyperbolic space. The Lorentz model of hyperbolic space has been shown to enable both fast forward and backward propagation. However, we prove that, with the current formulation of Lorentz linear layers, the hyperbolic norms of the outputs scale logarithmically with the number of gradient descent steps, nullifying the key advantage of hyperbolic geometry. We propose a new Lorentz linear layer grounded in the well-known ``distance-to-hyperplane" formulation. We prove that our formulation results in the usual linear scaling of output hyperbolic norms with respect to the number of gradient descent steps. Our new formulation, together with further algorithmic efficiencies through Lorentzian activation functions and a new caching strategy results in neural networks fully abiding by hyperbolic geometry while simultaneously bridging the computation gap to Euclidean neural networks. Code available at: https://github.com/robertdvdk/hyperbolic-fully-connected.

Representation learning is central to many downstream tasks such as search, clustering, classification, and reranking. State-of-the-art sequence encoders typically collapse a variable-length token sequence to a single vector using a pooling operator, most commonly a special [CLS] token or mean pooling over token embeddings. In this paper, we identify systematic weaknesses of these pooling strategies: [CLS] tends to concentrate information toward the initial positions of the sequence and can under-represent distributed evidence, while mean pooling can dilute salient local signals, sometimes leading to worse short-context performance. To address these issues, we introduce Landmark (LMK) pooling, which partitions a sequence into chunks, inserts landmark tokens between chunks, and forms the final representation by mean-pooling the landmark token embeddings. This simple mechanism improves long-context extrapolation without sacrificing local salient features, at the cost of introducing a small number of special tokens. We empirically demonstrate that LMK pooling matches existing methods on short-context retrieval tasks and yields substantial improvements on long-context tasks, making it a practical and scalable alternative to existing pooling methods.

To promote cooperation in Multi-Agent Reinforcement Learning, the reward signals of all agents can be aggregated together, forming global rewards that are commonly known as the fully cooperative setting. However, global rewards are usually noisy because they contain the contributions of all agents, which have to be resolved in the credit assignment process. On the other hand, using local reward benefits from faster learning due to the separation of agents' contributions, but can be suboptimal as agents myopically optimize their own reward while disregarding the global optimality. In this work, we propose a method that combines the merits of both approaches. By using a graph of interaction between agents, our method discerns the individual agent contribution in a more fine-grained manner than a global reward, while alleviating the cooperation problem with agents' local reward. We also introduce a practical approach for approximating such a graph. Our experiments demonstrate the flexibility of the approach, enabling improvements over the traditional local and global reward settings.

Spatial covariance matrices of EEG signals are Symmetric Positive Definite (SPD) and lie on a Riemannian manifold, yet the theoretical connection between embedding geometry and optimization dynamics remains unexplored. We provide a formal analysis linking embedding choice to gradient conditioning and numerical stability for SPD manifolds, establishing three theoretical results: (1) BWSPD's $\sqrtκ$ gradient conditioning (vs $κ$ for Log-Euclidean) via Daleckii-Kre\uın matrices provides better gradient conditioning on high-dimensional inputs ($d \geq 22$), with this advantage reducing on low-dimensional inputs ($d \leq 8$) where eigendecomposition overhead dominates; (2) Embedding-Space Batch Normalization (BN-Embed) approximates Riemannian normalization up to $O(\varepsilon^2)$ error, yielding $+26\%$ accuracy on 56-channel ERP data but negligible effect on 8-channel SSVEP data, matching the channel-count-dependent prediction; (3) bi-Lipschitz bounds prove BWSPD tokens preserve manifold distances with distortion governed solely by the condition ratio $κ$. We validate these predictions via a unified Transformer framework comparing BWSPD, Log-Euclidean, and Euclidean embeddings within identical architecture across 1,500+ runs on three EEG paradigms (motor imagery, ERP, SSVEP; 36 subjects). Our Log-Euclidean Transformer achieves state-of-the-art performance on all datasets, substantially outperforming classical Riemannian classifiers and recent SPD baselines, while BWSPD offers competitive accuracy with similar training time.

Recent advances in text-to-image (T2I) diffusion models have enabled increasingly realistic synthesis of vehicle damage, raising concerns about their reliability in automated insurance workflows. The ability to generate crash-like imagery challenges the boundary between authentic and synthetic data, introducing new risks of misuse in fraud or claim manipulation. To address these issues, we propose HERS (Hidden-Pattern Expert Learning for Risk-Specific Damage Adaptation), a framework designed to improve fidelity, controllability, and domain alignment of diffusion-generated damage images. HERS fine-tunes a base diffusion model via domain-specific expert adaptation without requiring manual annotation. Using self-supervised image-text pairs automatically generated by a large language model and T2I pipeline, HERS models each damage category, such as dents, scratches, broken lights, or cracked paint, as a separate expert. These experts are later integrated into a unified multi-damage model that balances specialization with generalization. We evaluate HERS across four diffusion backbones and observe consistent improvements: plus 5.5 percent in text faithfulness and plus 2.3 percent in human preference ratings compared to baselines. Beyond image fidelity, we discuss implications for fraud detection, auditability, and safe deployment of generative models in high-stakes domains. Our findings highlight both the opportunities and risks of domain-specific diffusion, underscoring the importance of trustworthy generation in safety-critical applications such as auto insurance.

Arabic is a linguistically and culturally rich language with a vast vocabulary that spans scientific, religious, and literary domains. Yet, large-scale lexical datasets linking Arabic words to precise definitions remain limited. We present MURAD (Multi-domain Unified Reverse Arabic Dictionary), an open lexical dataset with 96,243 word-definition pairs. The data come from trusted reference works and educational sources. Extraction used a hybrid pipeline integrating direct text parsing, optical character recognition, and automated reconstruction. This ensures accuracy and clarity. Each record aligns a target word with its standardized Arabic definition and metadata that identifies the source domain. The dataset covers terms from linguistics, Islamic studies, mathematics, physics, psychology, and engineering. It supports computational linguistics and lexicographic research. Applications include reverse dictionary modeling, semantic retrieval, and educational tools. By releasing this resource, we aim to advance Arabic natural language processing and promote reproducible research on Arabic lexical semantics.

Large language models have enabled automated algorithm design (AAD) by generating optimization algorithms directly from natural-language prompts. While evolutionary frameworks such as LLaMEA demonstrate strong exploratory capabilities across the algorithm design space, their search dynamics are entirely driven by fitness feedback, leaving substantial information about the generated code unused. We propose a mechanism for guiding AAD using feedback constructed from graph-theoretic and complexity features extracted from the abstract syntax trees of the generated algorithms, based on a surrogate model learned over an archive of evaluated solutions. Using explainable AI techniques, we identify features that substantially affect performance and translate them into natural-language mutation instructions that steer subsequent LLM-based code generation without restricting expressivity. We propose LLaMEA-SAGE, which integrates this feature-driven guidance into LLaMEA, and evaluate it across several benchmarks. We show that the proposed structured guidance achieves the same performance faster than vanilla LLaMEA in a small controlled experiment. In a larger-scale experiment using the MA-BBOB suite from the GECCO-MA-BBOB competition, our guided approach achieves superior performance compared to state-of-the-art AAD methods. These results demonstrate that signals derived from code can effectively bias LLM-driven algorithm evolution, bridging the gap between code structure and human-understandable performance feedback in automated algorithm design.

Indoor mobile robot navigation requires fast responsiveness and robust semantic understanding, yet existing methods struggle to provide both. Classical geometric approaches such as SLAM offer reliable localization but depend on detailed maps and cannot interpret human-targeted cues (e.g., signs, room numbers) essential for indoor reasoning. Vision-Language-Action (VLA) models introduce semantic grounding but remain strictly reactive, basing decisions only on visible frames and failing to anticipate unseen intersections or reason about distant textual cues. Vision-Language Models (VLMs) provide richer contextual inference but suffer from high computational latency, making them unsuitable for real-time operation on embedded platforms. In this work, we present IROS, a real-time navigation framework that combines VLM-level contextual reasoning with the efficiency of lightweight perceptual modules on low-cost, on-device hardware. Inspired by Dual Process Theory, IROS separates fast reflexive decisions (System One) from slow deliberative reasoning (System Two), invoking the VLM only when necessary. Furthermore, by augmenting compact VLMs with spatial and textual cues, IROS delivers robust, human-like navigation with minimal latency. Across five real-world buildings, IROS improves decision accuracy and reduces latency by 66% compared to continuous VLM-based navigation.

Controlling the behavior of large language models (LLMs) at inference time is essential for aligning outputs with human abilities and safety requirements. \emph{Activation steering} provides a lightweight alternative to prompt engineering and fine-tuning by directly modifying internal activations to guide generation. This research advances the literature in three significant directions. First, while previous work demonstrated the technical feasibility of steering emotional tone using automated classifiers, this paper presents the first human evaluation of activation steering concerning the emotional tone of LLM outputs, collecting over 7,000 crowd-sourced ratings from 190 participants via Prolific ($n=190$). These ratings assess both perceived emotional intensity and overall text quality. Second, we find strong alignment between human and model-based quality ratings (mean $r=0.776$, range $0.157$--$0.985$), indicating automatic scoring can proxy perceived quality. Moderate steering strengths ($λ\approx 0.15$) reliably amplify target emotions while preserving comprehensibility, with the strongest effects for disgust ($η_p^2 = 0.616$) and fear ($η_p^2 = 0.540$), and minimal effects for surprise ($η_p^2 = 0.042$). Finally, upgrading from Alpaca to LlaMA-3 yielded more consistent steering with significant effects across emotions and strengths (all $p < 0.001$). Inter-rater reliability was high (ICC $= 0.71$--$0.87$), underscoring the robustness of the findings. These findings support activation-based control as a scalable method for steering LLM behavior across affective dimensions.

Occluded traffic agents pose a significant challenge for autonomous vehicles, as hidden pedestrians or vehicles can appear unexpectedly, yet this problem remains understudied. Existing learning-based methods, while capable of inferring the presence of hidden agents, often produce redundant occupancy predictions where a single agent is identified multiple times. This issue complicates downstream planning and increases computational load. To address this, we introduce MatchInformer, a novel transformer-based approach that builds on the state-of-the-art SceneInformer architecture. Our method improves upon prior work by integrating Hungarian Matching, a state-of-the-art object matching algorithm from object detection, into the training process to enforce a one-to-one correspondence between predictions and ground truth, thereby reducing redundancy. We further refine trajectory forecasts by decoupling an agent's heading from its motion, a strategy that improves the accuracy and interpretability of predicted paths. To better handle class imbalances, we propose using the Matthews Correlation Coefficient (MCC) to evaluate occupancy predictions. By considering all entries in the confusion matrix, MCC provides a robust measure even in sparse or imbalanced scenarios. Experiments on the Waymo Open Motion Dataset demonstrate that our approach improves reasoning about occluded regions and produces more accurate trajectory forecasts than prior methods.

Recent advancements in Generative Artificial Intelligence (GenAI) have significantly enhanced the capabilities of both image generation and editing. However, current approaches often treat these tasks separately, leading to inefficiencies and challenges in maintaining spatial consistency and semantic coherence between generated content and edits. Moreover, a major obstacle is the lack of structured control over object relationships and spatial arrangements. Scene graph-based methods, which represent objects and their interrelationships in a structured format, offer a solution by providing greater control over composition and interactions in both image generation and editing. To address this, we introduce SimGraph, a unified framework that integrates scene graph-based image generation and editing, enabling precise control over object interactions, layouts, and spatial coherence. In particular, our framework integrates token-based generation and diffusion-based editing within a single scene graph-driven model, ensuring high-quality and consistent results. Through extensive experiments, we empirically demonstrate that our approach outperforms existing state-of-the-art methods.

In this paper, we present the first attempt to estimate the necessity of debulking coronary artery calcifications from computed tomography (CT) images. We formulate this task as a Multiple-instance Learning (MIL) problem. The difficulty of this task lies in that physicians adjust their focus and decision criteria for device usage according to tabular data representing each patient's condition. To address this issue, we propose a hypernetwork-based adaptive aggregation transformer (HyperAdAgFormer), which adaptively modifies the feature aggregation strategy for each patient based on tabular data through a hypernetwork. The experiments using the clinical dataset demonstrated the effectiveness of HyperAdAgFormer. The code is publicly available at https://github.com/Shiku-Kaito/HyperAdAgFormer.

On-policy reinforcement learning (RL) methods widely used for language model post-training, like Group Relative Policy Optimization (GRPO), often suffer from limited exploration and early saturation due to low sampling diversity. While off-policy data can help, current approaches that mix entire trajectories cause significant policy mismatch and instability. In this work, we propose the $\textbf{S}$ingle-sample Mix-p$\textbf{O}$licy $\textbf{U}$nified $\textbf{P}$aradigm (SOUP), a framework that unifies off- and on-policy learning within individual samples at the token level. It confines off-policy influence to the prefix of a generated sequence sampled from historical policies, while the continuation is generated on-policy. Through token-level importance ratios, SOUP effectively leverages off-policy information while preserving training stability. Extensive experiments demonstrate that SOUP consistently outperforms standard on-policy training and existing off-policy extensions. Our further analysis clarifies how our fine-grained, single-sample mix-policy training can improve both exploration and final performance in LLM RL.

For contact-intensive tasks, the ability to generate policies that produce comprehensive tactile-aware motions is essential. However, existing data collection and skill learning systems for dexterous manipulation often suffer from low-dimensional tactile information. To address this limitation, we propose DexTac, a visuo-tactile manipulation learning framework based on kinesthetic teaching. DexTac captures multi-dimensional tactile data-including contact force distributions and spatial contact regions-directly from human demonstrations. By integrating these rich tactile modalities into a policy network, the resulting contact-aware agent enables a dexterous hand to autonomously select and maintain optimal contact regions during complex interactions. We evaluate our framework on a challenging unimanual injection task. Experimental results demonstrate that DexTac achieves a 91.67% success rate. Notably, in high-precision scenarios involving small-scale syringes, our approach outperforms force-only baselines by 31.67%. These results underscore that learning multi-dimensional tactile priors from human demonstrations is critical for achieving robust, human-like dexterous manipulation in contact-rich environments.

LLM-based multi-agent simulations are increasingly adopted across application domains, but remain difficult to scale due to GPU memory pressure. Each agent maintains private GPU-resident states, including models, prefix caches, and adapters, which quickly exhaust device memory as the agent count grows. We identify two key properties of these workloads: sparse agent activation and an estimable agent invocation order. Based on an analysis of representative workload classes, we introduce invocation distance, a unified abstraction that estimates the relative order in which agents will issue future LLM requests. Leveraging this abstraction, we present ScaleSim, a memory-efficient LLM serving system for large-scale multi-agent simulations. ScaleSim enables proactive prefetching and priority-based eviction, supports diverse agent-specific memory through a modular interface, and achieves up to 1.74x speedup over SGLang on simulation benchmarks.

We study fixed-confidence best-arm identification (BAI) where a cheap but potentially biased proxy (e.g., LLM judge) is available for every sample, while an expensive ground-truth label can only be acquired selectively when using a human for auditing. Unlike classical multi-fidelity BAI, the proxy is biased (arm- and context-dependent) and ground truth is selectively observed. Consequently, standard multi-fidelity methods can mis-select the best arm, and uniform auditing, though accurate, wastes scarce resources and is inefficient. We prove that without bias correction and propensity adjustment, mis-selection probability may not vanish (even with unlimited proxy data). We then develop an estimator for the mean of each arm that combines proxy scores with inverse-propensity-weighted residuals and form anytime-valid confidence sequences for that estimator. Based on the estimator and confidence sequence, we propose an algorithm that adaptively selects and audits arms. The algorithm concentrates audits on unreliable contexts and close arms and we prove that a plug-in Neyman rule achieves near-oracle audit efficiency. Numerical experiments confirm the theoretical guarantees and demonstrate the superior empirical performance of the proposed algorithm.

We study semi-supervised stochastic optimization when labeled data is scarce but predictions from pre-trained models are available. PPI and SVRG both reduce variance through control variates -- PPI uses predictions, SVRG uses reference gradients. We show they are mathematically equivalent and develop PPI-SVRG, which combines both. Our convergence bound decomposes into the standard SVRG rate plus an error floor from prediction uncertainty. The rate depends only on loss geometry; predictions affect only the neighborhood size. When predictions are perfect, we recover SVRG exactly. When predictions degrade, convergence remains stable but reaches a larger neighborhood. Experiments confirm the theory: PPI-SVRG reduces MSE by 43--52\% under label scarcity on mean estimation benchmarks and improves test accuracy by 2.7--2.9 percentage points on MNIST with only 10\% labeled data.

Block-coordinate descent (BCD) is the method of choice to solve numerous large scale optimization problems, however their theoretical study for non-convex optimization, has received less attention. In this paper, we present a new block-coordinate descent (BCD) framework to tackle non-convex composite optimization problems, ensuring decrease of the objective function and convergence to a solution. This framework is general enough to include variable metric proximal gradient updates, proximal Newton updates, and alternated minimization updates. This generality allows to encompass three versions of the most used solvers in the sparse precision matrix estimation problem, deemed Graphical Lasso: graphical ISTA, Primal GLasso, and QUIC. We demonstrate the value of this new framework on non-convex sparse precision matrix estimation problems, providing convergence guarantees and up to a $100$-fold reduction in the number of iterations required to reach state-of-the-art estimation quality.

Text clustering is today the most popular paradigm for topic modelling, both in academia and industry. Despite clustering topic models' apparent success, we identify a number of issues in Top2Vec and BERTopic, which remain largely unsolved. Firstly, these approaches are unreliable at discovering natural clusters in corpora, due to extreme sensitivity to sample size and hyperparameters, the default values of which result in suboptimal behaviour. Secondly, when estimating term importance, BERTopic ignores the semantic distance of keywords to topic vectors, while Top2Vec ignores word counts in the corpus. This results in, on the one hand, less coherent topics due to the presence of stop words and junk words, and lack of variety and trust on the other. In this paper, I introduce a new approach, \textbf{Topeax}, which discovers the number of clusters from peaks in density estimates, and combines lexical and semantic indices of term importance to gain high-quality topic keywords. Topeax is demonstrated to be better at both cluster recovery and cluster description than Top2Vec and BERTopic, while also exhibiting less erratic behaviour in response to changing sample size and hyperparameters.

4D radar has emerged as a critical sensor for autonomous driving, primarily due to its enhanced capabilities in elevation measurement and higher resolution compared to traditional 3D radar. Effective integration of 4D radar with cameras requires accurate extrinsic calibration, and the development of radar-based perception algorithms demands large-scale annotated datasets. However, existing calibration methods often employ separate targets optimized for either visual or radar modalities, complicating correspondence establishment. Furthermore, manually labeling sparse radar data is labor-intensive and unreliable. To address these challenges, we propose 4D-CAAL, a unified framework for 4D radar-camera calibration and auto-labeling. Our approach introduces a novel dual-purpose calibration target design, integrating a checkerboard pattern on the front surface for camera detection and a corner reflector at the center of the back surface for radar detection. We develop a robust correspondence matching algorithm that aligns the checkerboard center with the strongest radar reflection point, enabling accurate extrinsic calibration. Subsequently, we present an auto-labeling pipeline that leverages the calibrated sensor relationship to transfer annotations from camera-based segmentations to radar point clouds through geometric projection and multi-feature optimization. Extensive experiments demonstrate that our method achieves high calibration accuracy while significantly reducing manual annotation effort, thereby accelerating the development of robust multi-modal perception systems for autonomous driving.

Recently, the rapid advancement of multimodal domains has driven a data-centric paradigm shift in graph ML, transitioning from text-attributed to multimodal-attributed graphs. This advancement significantly enhances data representation and expands the scope of graph downstream tasks, such as modality-oriented tasks, thereby improving the practical utility of graph ML. Despite its promise, limitations exist in the current neural paradigms: (1) Neglect Context in Modality Alignment: Most existing methods adopt topology-constrained or modality-specific operators as tokenizers. These aligners inevitably neglect graph context and inhibit modality interaction, resulting in suboptimal alignment. (2) Lack of Adaptation in Modality Fusion: Most existing methods are simple adaptations for 2-modality graphs and fail to adequately exploit aligned tokens equipped with topology priors during fusion, leading to poor generalizability and performance degradation. To address the above issues, we propose LION (c\underline{LI}ff\underline{O}rd \underline{N}eural paradigm) based on the Clifford algebra and decoupled graph neural paradigm (i.e., propagation-then-aggregation) to implement alignment-then-fusion in multimodal-attributed graphs. Specifically, we first construct a modality-aware geometric manifold grounded in Clifford algebra. This geometric-induced high-order graph propagation efficiently achieves modality interaction, facilitating modality alignment. Then, based on the geometric grade properties of aligned tokens, we propose adaptive holographic aggregation. This module integrates the energy and scale of geometric grades with learnable parameters to improve modality fusion. Extensive experiments on 9 datasets demonstrate that LION significantly outperforms SOTA baselines across 3 graph and 3 modality downstream tasks.

Metric learning is central to retrieval, yet its effects on embedding geometry and optimization dynamics are not well understood. We introduce a diagnostic framework, VARIANCE (intra-/inter-class variance) and GREEDINESS (active ratio and gradient norms), to compare seven representative losses, i.e., Contrastive, Triplet, N-pair, InfoNCE, ArcFace, SCL, and CCL, across five image-retrieval datasets. Our analysis reveals that Triplet and SCL preserve higher within-class variance and clearer inter-class margins, leading to stronger top-1 retrieval in fine-grained settings. In contrast, Contrastive and InfoNCE compact embeddings are achieved quickly through many small updates, accelerating convergence but potentially oversimplifying class structures. N-pair achieves a large mean separation but with uneven spacing. These insights reveal a form of efficiency-granularity trade-off and provide practical guidance: prefer Triplet/SCL when diversity preservation and hard-sample discrimination are critical, and Contrastive/InfoNCE when faster embedding compaction is desired.

Illicit financial activities such as money laundering often manifest through recurrent topological patterns in transaction networks. Detecting these patterns automatically remains challenging due to the scarcity of labeled real-world data and strict privacy constraints. To address this, we investigate whether Graph Autoencoders (GAEs) can effectively learn and distinguish topological patterns that mimic money laundering operations when trained on synthetic data. The analysis consists of two phases: (i) data generation, where synthetic samples are created for seven well-known illicit activity patterns using parametrized generators that preserve structural consistency while introducing realistic variability; and (ii) model training and validation, where separate GAEs are trained on each pattern without explicit labels, relying solely on reconstruction error as an indicator of learned structure. We compare three GAE implementations based on three distinct convolutional layers: Graph Convolutional (GAE-GCN), GraphSAGE (GAE-SAGE), and Graph Attention Network (GAE-GAT). Experimental results show that GAE-GCN achieves the most consistent reconstruction performance across patterns, while GAE-SAGE and GAE-GAT exhibit competitive results only in few specific patterns. These findings suggest that graph-based representation learning on synthetic data provides a viable path toward developing AI-driven tools for detecting illicit behaviors, overcoming the limitations of financial datasets.

When a user tells an AI system that someone "should not" take an action, the system ought to treat this as a prohibition. Yet many large language models do the opposite: they interpret negated instructions as affirmations. We audited 16 models across 14 ethical scenarios and found that open-source models endorse prohibited actions 77% of the time under simple negation and 100% under compound negation -- a 317% increase over affirmative framing. Commercial models fare better but still show swings of 19-128%. Agreement between models drops from 74% on affirmative prompts to 62% on negated ones, and financial scenarios prove twice as fragile as medical ones. These patterns hold under deterministic decoding, ruling out sampling noise. We present case studies showing how these failures play out in practice, propose the Negation Sensitivity Index (NSI) as a governance metric, and outline a tiered certification framework with domain-specific thresholds. The findings point to a gap between what current alignment techniques achieve and what safe deployment requires: models that cannot reliably distinguish "do X" from "do not X" should not be making autonomous decisions in high-stakes contexts.

In this paper, we address a fundamental gap between pre-training and fine-tuning of deep neural networks: while pre-training has shifted from unimodal to multimodal learning with enhanced visual understanding, fine-tuning predominantly remains unimodal, limiting the benefits of rich pre-trained representations. To bridge this gap, we propose a novel approach that transforms unimodal datasets into multimodal ones using Multimodal Large Language Models (MLLMs) to generate synthetic image captions for fine-tuning models with a multimodal objective. Our method employs carefully designed prompts incorporating class labels and domain context to produce high-quality captions tailored for classification tasks. Furthermore, we introduce a supervised contrastive loss function that explicitly encourages clustering of same-class representations during fine-tuning, along with a new inference technique that leverages class-averaged text embeddings from multiple synthetic captions per image. Extensive experiments across 13 image classification benchmarks demonstrate that our approach outperforms baseline methods, with particularly significant improvements in few-shot learning scenarios. Our work establishes a new paradigm for dataset enhancement that effectively bridges the gap between multimodal pre-training and fine-tuning. Our code is available at https://github.com/s-enmt/MMFT.

Large language models allocate uniform computation across all tokens, ignoring that some sequences are trivially predictable while others require deep reasoning. We introduce ConceptMoE, which dynamically merges semantically similar tokens into concept representations, performing implicit token-level compute allocation. A learnable chunk module identifies optimal boundaries by measuring inter-token similarity, compressing sequences by a target ratio $R$ before they enter the compute-intensive concept model. Crucially, the MoE architecture enables controlled evaluation: we reallocate saved computation to match baseline activated FLOPs (excluding attention map computation) and total parameters, isolating genuine architectural benefits. Under these conditions, ConceptMoE consistently outperforms standard MoE across language and vision-language tasks, achieving +0.9 points on language pretraining, +2.3 points on long context understanding, and +0.6 points on multimodal benchmarks. When converting pretrained MoE during continual training with layer looping, gains reach +5.5 points, demonstrating practical applicability. Beyond performance, ConceptMoE reduces attention computation by up to $R^2\times$ and KV cache by $R\times$. At $R=2$, empirical measurements show prefill speedups reaching 175\% and decoding speedups up to 117\% on long sequences. The minimal architectural modifications enable straightforward integration into existing MoE, demonstrating that adaptive concept-level processing fundamentally improves both effectiveness and efficiency of large language models.

Large Reasoning Models (LRMs) achieve explicit chain-of-thought expansion by imitating deep thinking behaviors of humans, demonstrating excellent performance in complex task scenarios. However, the deep-thinking mode often leads to unnecessarily lengthy reasoning and resource inefficiency when handling simple tasks. This overthinking phenomenon may arise from the generation preference triggered by the reward function during post-training. Existing research attempts to mitigate overthinking from the perspective of prompt design or model training, but generally underestimates the importance of task difficulty awareness, which makes it difficult for LRMs to effectively allocate reasoning resources. In this paper, we propose Difficulty-aware Policy Optimization (DiPO), a reinforcement learning-based LRM training framework. DiPO encourages LRM to spontaneously model task complexity, and integrates them into reinforcement learning framework to adjust the generation preferences introduced by post-training. A difficulty modeling method based on model self-reasoning is proposed, which significantly reduces the dependence on manual annotation and formalize task complexity. We further develop a difficulty-signal-enhanced reward function that incorporates a penalty for lengthy reasoning while considering reasoning performance and output format. Experimental results indicate that DiPO enables the model to spontaneously adjust inference overhead, significantly reducing redundant tokens without losing performance due to thought compression.

The generalization capabilities of robotic manipulation policies are heavily influenced by the choice of visual representations. Existing approaches typically rely on representations extracted from pre-trained encoders, using two dominant types of features: global features, which summarize an entire image via a single pooled vector, and dense features, which preserve a patch-wise embedding from the final encoder layer. While widely used, both feature types mix task-relevant and irrelevant information, leading to poor generalization under distribution shifts, such as changes in lighting, textures, or the presence of distractors. In this work, we explore an intermediate structured alternative: Slot-Based Object-Centric Representations (SBOCR), which group dense features into a finite set of object-like entities. This representation permits to naturally reduce the noise provided to the robotic manipulation policy while keeping enough information to efficiently perform the task. We benchmark a range of global and dense representations against intermediate slot-based representations, across a suite of simulated and real-world manipulation tasks ranging from simple to complex. We evaluate their generalization under diverse visual conditions, including changes in lighting, texture, and the presence of distractors. Our findings reveal that SBOCR-based policies outperform dense and global representation-based policies in generalization settings, even without task-specific pretraining. These insights suggest that SBOCR is a promising direction for designing visual systems that generalize effectively in dynamic, real-world robotic environments.

Training a unified language model that adapts between intuitive System 1 and deliberative System 2 remains challenging due to interference between their cognitive modes. Recent studies have thus pursued making System 2 models more efficient. However, these approaches focused on output control, limiting what models produce. We argue that this paradigm is misaligned: output length is merely a symptom of the model's cognitive configuration, not the root cause. In this work, we shift the focus to capability control, which modulates \textit{how models think} rather than \textit{what they produce}. To realize this, we leverage existing Instruct and Thinking checkpoints through dynamic parameter interpolation, without additional training. Our pilot study establishes that linear interpolation yields a convex, monotonic Pareto frontier, underpinned by representation continuity and structural connectivity. Building on this, we propose \textbf{DAMI} (\textbf{D}yn\textbf{A}mic \textbf{M}odel \textbf{I}nterpolation), a framework that estimates a query-specific Reasoning Intensity $λ(q)$ to configure cognitive depth. For training-based estimation, we develop a preference learning method encoding accuracy and efficiency criteria. For zero-shot deployment, we introduce a confidence-based method leveraging inter-model cognitive discrepancy. Experiments on five mathematical reasoning benchmarks demonstrate that DAMI achieves higher accuracy than the Thinking model while remaining efficient, effectively combining the efficiency of System 1 with the reasoning depth of System 2.

A classical approach to the multibody systems (MBS) modeling is to use absolute coordinates, i.e., a set of (possibly redundant) coordinates that describe the absolute position and orientation of the individual bodies with respect to an inertial frame (IFR). A well-known problem for the time integration of the equations of motion (EOM) is the lack of a singularity-free parameterization of spatial motions, which is usually tackled by using unit quaternions. Lie group integration methods were proposed as an alternative approach to the singularity-free time integration. At the same time, Lie group formulations of EOM naturally respect the geometry of spatial motions during integration. Lie group integration methods, operating directly on the configuration space Lie group, are incompatible with standard formulations of the EOM, and cannot be implemented in existing MBS simulation codes without a major restructuring. The contribution of this paper is twofold: (1) A framework for interfacing Lie group integrators to standard EOM formulations is presented. It allows describing MBS in terms of various absolute coordinates and at the same using Lie group integration schemes. (2) A method for consistently incorporating the geometry of rigid body motions into the evaluation of EOM in absolute coordinates integrated with standard vector space integration schemes. The direct product group and the semidirect product group SO(3)xR3 and the semidirect product group SE(3) are used for representing rigid body motions. The key element is the local-global transitions (LGT) transition map, which facilitates the update of (global) absolute coordinates in terms of the (local) coordinates on the Lie group. This LGT map is specific to the absolute coordinates, the local coordinates on the Lie group, and the Lie group used to represent rigid body configurations.