Despite state-of-the-art vision-language models (VLMs) have demonstrated strong reasoning capabilities, their performance in multilingual mathematical reasoning remains underexplored, particularly when compared to human performance. To bridge this gap, we introduce M3Kang, the first massively multilingual, multimodal mathematical reasoning dataset for VLMs. It is derived from the Kangaroo Math Competition, the world's largest mathematics contest, which annually engages over six million participants under the age of 18 across more than 90 countries. M3Kang includes 1,747 unique multiple-choice problems organized by grade-level difficulty, with translations into 108 culturally diverse languages, some of them including diagrams essential for solving them. Using this dataset, we conduct extensive benchmarking on both closed- and open-source SOTA models. We observe that, despite recent advances, models still struggle with basic math and diagram-based reasoning, with performance scaling with language presence and model size, but not with grade level. We also find that multilingual techniques can be effectively extended to the multimodal setting, resulting in significant improvements over baseline approaches. Our analysis also incorporates performance data from over 68,000 students, enabling direct comparison with human performance. We are open-sourcing M3Kang, including the English-only subset M2Kang, along with the framework and codebase used to construct the dataset.

Code-mixing is increasingly prevalent in interactions between humans and large language models, yet existing work often reduces it to a translation or convertibility problem, making it difficult to assess whether a model's switching behavior is context-appropriate and aligned with human conventions. We introduce ChiEngMixBench, the first benchmark designed to evaluate code-mixing ability in authentic community contexts, built upon a general construction pipeline that enables scalable dataset development across domains and bilingual pairs. ChiEngMixBench formulates code-mixing as a cognitive alignment problem, characterized by two complementary signals: Spontaneity and Naturalness. Empirical evaluation shows that our metrics can systematically distinguish code-mixing performance across models. Beyond benchmarking, we further uncover an implicitly emergent Terminology Layering Strategy, a phenomenon consistent with the Matrix Language Frame (MLF) theory, indicating structured cognitive alignment between multilingual large language models and human communication.

Fine-tuning a task-specific multilingual large language model (LLM) involves training the model on a multilingual dataset with examples in all the required languages. Updating one or more supported languages with additional data or adding support for a new language involves retraining the model, which can be computationally inefficient and creates a severe maintenance bottleneck. Recent research on merging multilingual multitask models has shown promise in terms of improved quality, but its computational and maintenance efficiency remains unstudied. In this work, we provide the first focused analysis of this merging strategy from an efficiency perspective, evaluating it across three independent tasks. We demonstrate significant efficiency gains while maintaining parity in terms of quality: this merging approach reduces the initial training time by up to 50\%. We also demonstrate that updating an individual language and re-merging as part of model maintenance reduces training costs by more than 60\%, compared to re-training the full multilingual model. We show this on both public and proprietary industry datasets confirming that the approach works well for industrial use cases in addition to academic settings already studied in previous work.

Human motion reconstruction from monocular videos is a fundamental challenge in computer vision, with broad applications in AR/VR, robotics, and digital content creation, but remains challenging under frequent occlusions in real-world settings. Existing regression-based methods are efficient but fragile to missing observations, while optimization- and diffusion-based approaches improve robustness at the cost of slow inference speed and heavy preprocessing steps. To address these limitations, we leverage recent advances in generative masked modeling and present MoRo: Masked Modeling for human motion Recovery under Occlusions. MoRo is an occlusion-robust, end-to-end generative framework that formulates motion reconstruction as a video-conditioned task, and efficiently recover human motion in a consistent global coordinate system from RGB videos. By masked modeling, MoRo naturally handles occlusions while enabling efficient, end-to-end inference. To overcome the scarcity of paired video-motion data, we design a cross-modality learning scheme that learns multi-modal priors from a set of heterogeneous datasets: (i) a trajectory-aware motion prior trained on MoCap datasets, (ii) an image-conditioned pose prior trained on image-pose datasets, capturing diverse per-frame poses, and (iii) a video-conditioned masked transformer that fuses motion and pose priors, finetuned on video-motion datasets to integrate visual cues with motion dynamics for robust inference. Extensive experiments on EgoBody and RICH demonstrate that MoRo substantially outperforms state-of-the-art methods in accuracy and motion realism under occlusions, while performing on-par in non-occluded scenarios. MoRo achieves real-time inference at 70 FPS on a single H200 GPU.

We study the problem of collision-free humanoid traversal in cluttered indoor scenes, such as hurdling over objects scattered on the floor, crouching under low-hanging obstacles, or squeezing through narrow passages. To achieve this goal, the humanoid needs to map its perception of surrounding obstacles with diverse spatial layouts and geometries to the corresponding traversal skills. However, the lack of an effective representation that captures humanoid-obstacle relationships during collision avoidance makes directly learning such mappings difficult. We therefore propose Humanoid Potential Field (HumanoidPF), which encodes these relationships as collision-free motion directions, significantly facilitating RL-based traversal skill learning. We also find that HumanoidPF exhibits a surprisingly negligible sim-to-real gap as a perceptual representation. To further enable generalizable traversal skills through diverse and challenging cluttered indoor scenes, we further propose a hybrid scene generation method, incorporating crops of realistic 3D indoor scenes and procedurally synthesized obstacles. We successfully transfer our policy to the real world and develop a teleoperation system where users could command the humanoid to traverse in cluttered indoor scenes with just a single click. Extensive experiments are conducted in both simulation and the real world to validate the effectiveness of our method. Demos and code can be found in our website: https://axian12138.github.io/CAT/.

Diffusion-based language models (DLLMs) offer non-sequential, block-wise generation and richer data reuse compared to autoregressive (AR) models, but existing code DLLMs still lag behind strong AR baselines under comparable budgets. We revisit this setting in a controlled study and introduce Stable-DiffCoder, a block diffusion code model that reuses the Seed-Coder architecture, data, and training pipeline. To enable efficient knowledge learning and stable training, we incorporate a block diffusion continual pretraining (CPT) stage enhanced by a tailored warmup and block-wise clipped noise schedule. Under the same data and architecture, Stable-DiffCoder overall outperforms its AR counterpart on a broad suite of code benchmarks. Moreover, relying only on the CPT and supervised fine-tuning stages, Stable-DiffCoder achieves stronger performance than a wide range of \~8B ARs and DLLMs, demonstrating that diffusion-based training can improve code modeling quality beyond AR training alone. Moreover, diffusion-based any-order modeling improves structured code modeling for editing and reasoning, and through data augmentation, benefits low-resource coding languages.

The development of native computer-use agents (CUA) represents a significant leap in multimodal AI. However, their potential is currently bottlenecked by the constraints of static data scaling. Existing paradigms relying primarily on passive imitation of static datasets struggle to capture the intricate causal dynamics inherent in long-horizon computer tasks. In this work, we introduce EvoCUA, a native computer use agentic model. Unlike static imitation, EvoCUA integrates data generation and policy optimization into a self-sustaining evolutionary cycle. To mitigate data scarcity, we develop a verifiable synthesis engine that autonomously generates diverse tasks coupled with executable validators. To enable large-scale experience acquisition, we design a scalable infrastructure orchestrating tens of thousands of asynchronous sandbox rollouts. Building on these massive trajectories, we propose an iterative evolving learning strategy to efficiently internalize this experience. This mechanism dynamically regulates policy updates by identifying capability boundaries -- reinforcing successful routines while transforming failure trajectories into rich supervision through error analysis and self-correction. Empirical evaluations on the OSWorld benchmark demonstrate that EvoCUA achieves a success rate of 56.7%, establishing a new open-source state-of-the-art. Notably, EvoCUA significantly outperforms the previous best open-source model, OpenCUA-72B (45.0%), and surpasses leading closed-weights models such as UI-TARS-2 (53.1%). Crucially, our results underscore the generalizability of this approach: the evolving paradigm driven by learning from experience yields consistent performance gains across foundation models of varying scales, establishing a robust and scalable path for advancing native agent capabilities.

While Text-to-SQL remains the dominant approach for database interaction, real-world analytics increasingly require the flexibility of general-purpose programming languages such as Python or Pandas to manage file-based data and complex analytical workflows. Despite this growing need, the reliability of Text-to-Python in core data retrieval remains underexplored relative to the mature SQL ecosystem. To address this gap, we introduce BIRD-Python, a benchmark designed for cross-paradigm evaluation. We systematically refined the original dataset to reduce annotation noise and align execution semantics, thereby establishing a consistent and standardized baseline for comparison. Our analysis reveals a fundamental paradigmatic divergence: whereas SQL leverages implicit DBMS behaviors through its declarative structure, Python requires explicit procedural logic, making it highly sensitive to underspecified user intent. To mitigate this challenge, we propose the Logic Completion Framework (LCF), which resolves ambiguity by incorporating latent domain knowledge into the generation process. Experimental results show that (1) performance differences primarily stem from missing domain context rather than inherent limitations in code generation, and (2) when these gaps are addressed, Text-to-Python achieves performance parity with Text-to-SQL. These findings establish Python as a viable foundation for analytical agents-provided that systems effectively ground ambiguous natural language inputs in executable logical specifications. Resources are available at https://anonymous.4open.science/r/Bird-Python-43B7/.

Adding memory to pretrained language models typically requires architectural changes or weight modification. We present Prometheus Mind, which retrofits memory to a frozen Qwen3-4B using 11 modular adapters (530MB, 7% overhead) -- fully reversible by removing the adapters. Building this system required solving four problems: (1) Extraction -- we develop Contrastive Direction Discovery (CDD), which finds semantic directions via minimal pairs without labeled data. (2) Training -- end-to-end optimization collapses; stage-wise training of each adapter on simple proxy tasks succeeds. (3) Injection -- learned encoders fail to generalize; we find that lm_head-weight rows already provide the mapping we need, requiring no training. (4) Hidden state collapse -- transformers make ``wife'' and ``brother'' 0.98+ similar; we train projections to recover distinction (0.98 $\rightarrow$ 0.09). On PrometheusExtract-132 (132 cases), the system achieves 94.4% retrieval on clean inputs (n=54, 95% CI: [84.9%, 98.1%]), degrading to 19.4% on informal inputs with ellipsis, filler words, or implicit subjects (n=36). The primary bottleneck is relation classification (47.3% accuracy), responsible for most extraction errors.

We introduce AfriEconQA, a specialized benchmark dataset for African economic analysis grounded in a comprehensive corpus of 236 World Bank reports. The task of AfriEconQA is to answer complex economic queries that require high-precision numerical reasoning and temporal disambiguation from specialized institutional documents. The dataset consists of 8,937 curated QA instances, rigorously filtered from a pool of 10018 synthetic questions to ensure high-quality evidence-answer alignment. Each instance is composed of: (1) a question requiring reasoning over economic indicators, (2) the corresponding evidence retrieved from the corpus, (3) a verified ground-truth answer, and (4) source metadata (e.g., URL and publication date) to ensure temporal provenance. AfriEconQA is the first benchmark focused specifically on African economic analysis, providing a unique challenge for Information Retrieval (IR) systems, as the data is largely absent from the pretraining corpora of current Large Language Models (LLMs). We operationalize this dataset through an 11-experiment matrix, benchmarking a zero-shot baseline (GPT-5 Mini) against RAG configurations using GPT-4o and Qwen 32B across five distinct embedding and ranking strategies. Our results demonstrate a severe parametric knowledge gap, where zero-shot models fail to answer over 90 percent of queries, and even state-of-the-art RAG pipelines struggle to achieve high precision. This confirms AfriEconQA as a robust and challenging benchmark for the next generation of domain-specific IR and RAG systems. The AfriEconQA dataset and code will be made publicly available upon publication.

Explainable AI (XAI) is frequently positioned as a technical problem of revealing the inner workings of an AI model. This position is affected by unexamined onto-epistemological assumptions: meaning is treated as immanent to the model, the explainer is positioned outside the system, and a causal structure is presumed recoverable through computational techniques. In this paper, we draw on Barad's agential realism to develop an alternative onto-epistemology of XAI. We propose that interpretations are material-discursive performances that emerge from situated entanglements of the AI model with humans, context, and the interpretative apparatus. To develop this position, we read a comprehensive set of XAI methods through agential realism and reveal the assumptions and limitations that underpin several of these methods. We then articulate the framework's ethical dimension and propose design directions for XAI interfaces that support emergent interpretation, using a speculative text-to-music interface as a case study.

Reliable human--robot collaboration in emergency scenarios requires autonomous systems that can detect humans, infer navigation goals, and operate safely in dynamic environments. This paper presents HumanDiffusion, a lightweight image-conditioned diffusion planner that generates human-aware navigation trajectories directly from RGB imagery. The system combines YOLO-11 based human detection with diffusion-driven trajectory generation, enabling a quadrotor to approach a target person and deliver medical assistance without relying on prior maps or computationally intensive planning pipelines. Trajectories are predicted in pixel space, ensuring smooth motion and a consistent safety margin around humans. We evaluate HumanDiffusion in simulation and real-world indoor mock-disaster scenarios. On a 300-sample test set, the model achieves a mean squared error of 0.02 in pixel-space trajectory reconstruction. Real-world experiments demonstrate an overall mission success rate of 80% across accident-response and search-and-locate tasks with partial occlusions. These results indicate that human-conditioned diffusion planning offers a practical and robust solution for human-aware UAV navigation in time-critical assistance settings.

Scribble-supervised methods have emerged to mitigate the prohibitive annotation burden in medical image segmentation. However, the inherent sparsity of these annotations introduces significant ambiguity, which results in noisy pseudo-label propagation and hinders the learning of robust anatomical boundaries. To address this challenge, we propose SDT-Net, a novel dual-teacher, single-student framework designed to maximize supervision quality from these weak signals. Our method features a Dynamic Teacher Switching (DTS) module to adaptively select the most reliable teacher. This selected teacher then guides the student via two synergistic mechanisms: high-confidence pseudo-labels, refined by a Pick Reliable Pixels (PRP) mechanism, and multi-level feature alignment, enforced by a Hierarchical Consistency (HiCo) module. Extensive experiments on the ACDC and MSCMRseg datasets demonstrate that SDT-Net achieves state-of-the-art performance, producing more accurate and anatomically plausible segmentation.

Achieving human-level performance in Vision-and-Language Navigation (VLN) requires an embodied agent to jointly understand multimodal instructions and visual-spatial context while reasoning over long action sequences. Recent works, such as NavCoT and NavGPT-2, demonstrate the potential of Chain-of-Thought (CoT) reasoning for improving interpretability and long-horizon planning. Moreover, multimodal extensions like OctoNav-R1 and CoT-VLA further validate CoT as a promising pathway toward human-like navigation reasoning. However, existing approaches face critical drawbacks: purely textual CoTs lack spatial grounding and easily overfit to sparse annotated reasoning steps, while multimodal CoTs incur severe token inflation by generating imagined visual observations, making real-time navigation impractical. In this work, we propose FantasyVLN, a unified implicit reasoning framework that preserves the benefits of CoT reasoning without explicit token overhead. Specifically, imagined visual tokens are encoded into a compact latent space using a pretrained Visual AutoRegressor (VAR) during CoT reasoning training, and the model jointly learns from textual, visual, and multimodal CoT modes under a unified multi-CoT strategy. At inference, our model performs direct instruction-to-action mapping while still enjoying reasoning-aware representations. Extensive experiments on LH-VLN show that our approach achieves reasoning-aware yet real-time navigation, improving success rates and efficiency while reducing inference latency by an order of magnitude compared to explicit CoT methods.

We introduce GazeD, a new 3D gaze estimation method that jointly provides 3D gaze and human pose from a single RGB image. Leveraging the ability of diffusion models to deal with uncertainty, it generates multiple plausible 3D gaze and pose hypotheses based on the 2D context information extracted from the input image. Specifically, we condition the denoising process on the 2D pose, the surroundings of the subject, and the context of the scene. With GazeD we also introduce a novel way of representing the 3D gaze by positioning it as an additional body joint at a fixed distance from the eyes. The rationale is that the gaze is usually closely related to the pose, and thus it can benefit from being jointly denoised during the diffusion process. Evaluations across three benchmark datasets demonstrate that GazeD achieves state-of-the-art performance in 3D gaze estimation, even surpassing methods that rely on temporal information. Project details will be available at https://aimagelab.ing.unimore.it/go/gazed.

Dengue, a mosquito-borne disease, continues to pose a persistent public health challenge in urban areas, particularly in tropical regions such as Singapore. Effective and affordable control requires anticipating where transmission risks are likely to emerge so that interventions can be deployed proactively rather than reactively. This study introduces a novel framework that uncovers and exploits latent transmission links between urban regions, mined directly from publicly available dengue case data. Instead of treating cases as isolated reports, we model how hotspot formation in one area is influenced by epidemic dynamics in neighboring regions. While mosquito movement is highly localized, long-distance transmission is often driven by human mobility, and in our case study, the learned network aligns closely with commuting flows, providing an interpretable explanation for citywide spread. These hidden links are optimized through gradient descent and used not only to forecast hotspot status but also to verify the consistency of spreading patterns, by examining the stability of the inferred network across consecutive weeks. Case studies on Singapore during 2013-2018 and 2020 show that four weeks of hotspot history are sufficient to achieve an average F-score of 0.79. Importantly, the learned transmission links align with commuting flows, highlighting the interpretable interplay between hidden epidemic spread and human mobility. By shifting from simply reporting dengue cases to mining and validating hidden spreading dynamics, this work transforms open web-based case data into a predictive and explanatory resource. The proposed framework advances epidemic modeling while providing a scalable, low-cost tool for public health planning, early intervention, and urban resilience.

Large language models (LLMs) possess extensive world knowledge, yet methods for effectively eliciting this knowledge remain underexplored. Nationality and region prediction tasks require understanding of not only linguistic features but also cultural and historical background, making LLM world knowledge particularly valuable. However, conventional LLM prompting methods rely on direct reasoning approaches, which have limitations in applying abstract linguistic rules. We propose LLM Associative Memory Agents (LAMA), a novel framework that leverages LLM world knowledge as associative memory. Rather than directly inferring nationality from names, LAMA recalls famous individuals with the same name and aggregates their nationalities through indirect reasoning. A dual-agent architecture comprising a Person Agent and a Media Agent, specialized in different knowledge domains, recalls famous individuals in parallel, generating Top-1 predictions through voting and Top-K predictions through conditional completion. On a 99-country nationality prediction task, LAMA achieved 0.817 accuracy, substantially outperforming conventional LLM prompting methods and neural models. Our experiments reveal that LLMs exhibit higher reliability in recalling concrete examples than in abstract reasoning, that recall-based approaches are robust to low-frequency nationalities independent of data frequency distributions, and that the dual-agent architecture functions complementarily to produce synergistic effects. These results demonstrate the effectiveness of a new multi-agent system that retrieves and aggregates LLM knowledge rather than prompting reasoning.

GPU code optimization is a key performance bottleneck for HPC workloads as well as large-model training and inference. Although compiler optimizations and hand-written kernels can partially alleviate this issue, achieving near-hardware-limit performance still relies heavily on manual code refactoring and parameter tuning. Recent progress in LLM-agent-based kernel generation and optimization has been reported, yet many approaches primarily focus on direct code rewriting, where parameter choices are often implicit and hard to control, or require human intervention, leading to unstable performance gains. This paper introduces a template-based rewriting layer on top of an agent-driven iterative loop: kernels are semantically refactored into explicitly parameterizable templates, and template parameters are then optimized via search-based autotuning, yielding more stable and higher-quality speedups. Experiments on a set of real-world kernels demonstrate speedups exceeding 3x in the best case. We extract representative CUDA kernels from SGLang as evaluation targets; the proposed agentic tuner iteratively performs templating, testing, analysis, and planning, and leverages profiling feedback to execute constrained parameter search under hardware resource limits. Compared to agent-only direct rewriting, the template-plus-search design significantly reduces the randomness of iterative optimization, making the process more interpretable and enabling a more systematic approach toward high-performance configurations. The proposed method can be further extended to OpenCL, HIP, and other backends to deliver automated performance optimization for real production workloads.

Social-physical human-robot interaction (spHRI) is difficult to study: building and programming robots that integrate multiple interaction modalities is costly and slow, while VR-based prototypes often lack physical contact, breaking users' visuo-tactile expectations. We present XR$^3$, a co-located dual-VR-headset platform for HRI research in which an attendee and a hidden operator share the same physical space while experiencing different virtual embodiments. The attendee sees an expressive virtual robot that interacts face-to-face in a shared virtual environment. In real time, the robot's upper-body motion, head and gaze behavior, and facial expressions are mapped from the operator's tracked limbs and face signals. Because the operator is co-present and calibrated in the same coordinate frame, the operator can also touch the attendee, enabling perceived robot touch synchronized with the robot's visible hands. Finger and hand motion is mapped to the robot avatar using inverse kinematics to support precise contact. Beyond motion retargeting, XR$^3$ supports social retargeting of multiple nonverbal cues that can be experimentally varied while keeping physical interaction constant. We detail the system design and calibration, and demonstrate the platform in a touch-based Wizard-of-Oz study, lowering the barrier to prototyping and evaluating embodied, contact-based robot behaviors.

Reinforcement learning has become a central paradigm for improving LLM reasoning. However, existing methods use a single policy to produce both inference responses and training optimization trajectories. The objective conflict between generating stable inference responses and diverse training trajectories leads to insufficient exploration, which harms reasoning capability. In this paper, to address the problem, we propose R$^2$PO (Residual Rollout Policy Optimization), which introduces a lightweight Residual Rollout-Head atop the policy to decouple training trajectories from inference responses, enabling controlled trajectory diversification during training while keeping inference generation stable. Experiments across multiple benchmarks show that our method consistently outperforms baselines, achieving average accuracy gains of 3.4% on MATH-500 and 1.3% on APPS, while also reducing formatting errors and mitigating length bias for stable optimization. Our code is publicly available at https://github.com/RRPO-ARR/Code.

Gait recognition is emerging as a promising technology and an innovative field within computer vision, with a wide range of applications in remote human identification. However, existing methods typically rely on complex architectures to directly extract features from images and apply pooling operations to obtain sequence-level representations. Such designs often lead to overfitting on static noise (e.g., clothing), while failing to effectively capture dynamic motion regions, such as the arms and legs. This bottleneck is particularly challenging in the presence of intra-class variation, where gait features of the same individual under different environmental conditions are significantly distant in the feature space. To address the above challenges, we present a Languageguided and Motion-aware gait recognition framework, named LMGait. To the best of our knowledge, LMGait is the first method to introduce natural language descriptions as explicit semantic priors into the gait recognition task. In particular, we utilize designed gait-related language cues to capture key motion features in gait sequences. To improve cross-modal alignment, we propose the Motion Awareness Module (MAM), which refines the language features by adaptively adjusting various levels of semantic information to ensure better alignment with the visual representations. Furthermore, we introduce the Motion Temporal Capture Module (MTCM) to enhance the discriminative capability of gait features and improve the model's motion tracking ability. We conducted extensive experiments across multiple datasets, and the results demonstrate the significant advantages of our proposed network. Specifically, our model achieved accuracies of 88.5%, 97.1%, and 97.5% on the CCPG, SUSTech1K, and CASIAB datasets, respectively, achieving state-of-the-art performance. Homepage: https://dingwu1021.github.io/LMGait/

Extractive summarization (ES) aims to generate a concise summary by selecting a subset of sentences from a document while maximizing relevance and minimizing redundancy. Although modern ES systems achieve high accuracy using powerful neural models, their deployment typically relies on CPU or GPU infrastructures that are energy-intensive and poorly suited for real-time inference in resource-constrained environments. In this work, we explore the feasibility of implementing McDonald-style extractive summarization on a low-power CMOS coupled oscillator-based Ising machine (COBI) that supports integer-valued, all-to-all spin couplings. We first propose a hardware-aware Ising formulation that reduces the scale imbalance between local fields and coupling terms, thereby improving robustness to coefficient quantization: this method can be applied to any problem formulation that requires k of n variables to be chosen. We then develop a complete ES pipeline including (i) stochastic rounding and iterative refinement to compensate for precision loss, and (ii) a decomposition strategy that partitions a large ES problem into smaller Ising subproblems that can be efficiently solved on COBI and later combined. Experimental results on the CNN/DailyMail dataset show that our pipeline can produce high-quality summaries using only integer-coupled Ising hardware with limited precision. COBI achieves 3-4.5x runtime speedups compared to a brute-force method, which is comparable to software Tabu search, and two to three orders of magnitude reductions in energy, while maintaining competitive summary quality. These results highlight the potential of deploying CMOS Ising solvers for real-time, low-energy text summarization on edge devices.

Interpretable graph learning has recently emerged as a popular research topic in machine learning. The goal is to identify the important nodes and edges of an input graph that are crucial for performing a specific graph reasoning task. A number of studies have been conducted in this area, and various benchmark datasets have been proposed to facilitate evaluation. Among them, one of the most challenging is the Spurious-Motif benchmark, introduced at ICLR 2022. The datasets in this synthetic benchmark are deliberately designed to include spurious correlations, making it particularly difficult for models to distinguish truly relevant structures from misleading patterns. As a result, existing methods exhibit significantly worse performance on this benchmark compared to others. In this paper, we focus on improving interpretability on the challenging Spurious-Motif datasets. We demonstrate that the self-reflection technique, commonly used in large language models to tackle complex tasks, can also be effectively adapted to enhance interpretability in datasets with strong spurious correlations. Specifically, we propose a self-reflection framework that can be integrated with existing interpretable graph learning methods. When such a method produces importance scores for each node and edge, our framework feeds these predictions back into the original method to perform a second round of evaluation. This iterative process mirrors how large language models employ self-reflective prompting to reassess their previous outputs. We further analyze the reasons behind this improvement from the perspective of graph representation learning, which motivates us to propose a fine-tuning training method based on this feedback mechanism.

Vision-language models (VLMs) have advanced rapidly, but their ability to capture spatial relationships remains a blindspot. Current VLMs are typically built with contrastive language-image pretraining (CLIP) style image encoders. The training recipe often flattens images into 1D patch sequences, discarding the 2D structure necessary for spatial reasoning. We argue that this lack of spatial awareness is a missing dimension in VLM design and a bottleneck for applications requiring spatial grounding, such as robotics and embodied AI. To address this, we investigate (i) image encoders trained with alternative objectives and (ii) 2D positional encodings. Our experiments show that these architectural choices can lead to improved spatial reasoning on several benchmarks.

Large Language Models (LLMs), despite their remarkable capabilities across NLP tasks, struggle with phonologically-grounded phenomena like rhyme detection and generation. This is even more evident in lower-resource languages such as Modern Greek. In this paper, we present a hybrid system that combines LLMs with deterministic phonological algorithms to achieve accurate rhyme identification/analysis and generation. Our approach implements a comprehensive taxonomy of Greek rhyme types, including Pure, Rich, Imperfect, Mosaic, and Identical Pre-rhyme Vowel (IDV) patterns, and employs an agentic generation pipeline with phonological verification. We evaluate multiple prompting strategies (zero-shot, few-shot, Chain-of-Thought, and RAG-augmented) across several LLMs including Claude 3.7 and 4.5, GPT-4o, Gemini 2.0 and open-weight models like Llama 3.1 8B and 70B and Mistral Large. Results reveal a significant "Reasoning Gap": while native-like models (Claude 3.7) perform intuitively (40\% accuracy in identification), reasoning-heavy models (Claude 4.5) achieve state-of-the-art performance (54\%) only when prompted with Chain-of-Thought. Most critically, pure LLM generation fails catastrophically (under 4\% valid poems), while our hybrid verification loop restores performance to 73.1\%. We release our system and a corpus of 40,000+ rhymes, derived from the Anemoskala and Interwar Poetry corpora, to support future research.

Recent research suggests that looped Transformers have superior reasoning capabilities compared to standard deep architectures. Current approaches to training single-head looped architectures on benchmark tasks frequently fail or yield suboptimal performance due to a highly non-convex and irregular loss landscape. In these settings, optimization often stagnates in poor local minima and saddle points of the loss landscape, preventing the model from discovering the global minimum point. The internal mechanisms of these single-head looped transformer models remain poorly understood, and training them from scratch remains a significant challenge. In this paper, we propose a novel training framework that leverages Tsallis entropy and Hamiltonian dynamics to transform the geometry of the loss landscape. By treating the parameter updates as a physical flow, we successfully trained a single-head looped Transformer with model dimension $d = 8$ to solve induction head task with input sequence length of 1000 tokens. This success reveals the internal mechanism behind the superior reasoning capability.

Multi-constraint planning involves identifying, evaluating, and refining candidate plans while satisfying multiple, potentially conflicting constraints. Existing large language model (LLM) approaches face fundamental limitations in this domain. Pure reasoning paradigms, which rely on long natural language chains, are prone to inconsistency, error accumulation, and prohibitive cost as constraints compound. Conversely, LLMs combined with coding- or solver-based strategies lack flexibility: they often generate problem-specific code from scratch or depend on fixed solvers, failing to capture generalizable logic across diverse problems. To address these challenges, we introduce the Scalable COde Planning Engine (SCOPE), a framework that disentangles query-specific reasoning from generic code execution. By separating reasoning from execution, SCOPE produces solver functions that are consistent, deterministic, and reusable across queries while requiring only minimal changes to input parameters. SCOPE achieves state-of-the-art performance while lowering cost and latency. For example, with GPT-4o, it reaches 93.1% success on TravelPlanner, a 61.6% gain over the best baseline (CoT) while cutting inference cost by 1.4x and time by ~4.67x. Code is available at https://github.com/DerrickGXD/SCOPE.

The ubiquity of machine learning (ML) and the demand for ever-larger models bring an increase in energy consumption and environmental impact. However, little is known about the energy scaling laws in ML, and existing research focuses on training cost -- ignoring the larger cost of inference. Furthermore, tools for measuring the energy consumption of ML do not provide actionable feedback. To address these gaps, we developed Energy Consumption Optimiser (ECOpt): a hyperparameter tuner that optimises for energy efficiency and model performance. ECOpt quantifies the trade-off between these metrics as an interpretable Pareto frontier. This enables ML practitioners to make informed decisions about energy cost and environmental impact, while maximising the benefit of their models and complying with new regulations. Using ECOpt, we show that parameter and floating-point operation counts can be unreliable proxies for energy consumption, and observe that the energy efficiency of Transformer models for text generation is relatively consistent across hardware. These findings motivate measuring and publishing the energy metrics of ML models. We further show that ECOpt can have a net positive environmental impact and use it to uncover seven models for CIFAR-10 that improve upon the state of the art, when considering accuracy and energy efficiency together.

Attention mechanisms have become a core component of deep learning models, with Channel Attention and Spatial Attention being the two most representative architectures. Current research on their fusion strategies primarily bifurcates into sequential and parallel paradigms, yet the selection process remains largely empirical, lacking systematic analysis and unified principles. We systematically compare channel-spatial attention combinations under a unified framework, building an evaluation suite of 18 topologies across four classes: sequential, parallel, multi-scale, and residual. Across two vision and nine medical datasets, we uncover a "data scale-method-performance" coupling law: (1) in few-shot tasks, the "Channel-Multi-scale Spatial" cascaded structure achieves optimal performance; (2) in medium-scale tasks, parallel learnable fusion architectures demonstrate superior results; (3) in large-scale tasks, parallel structures with dynamic gating yield the best performance. Additionally, experiments indicate that the "Spatial-Channel" order is more stable and effective for fine-grained classification, while residual connections mitigate vanishing gradient problems across varying data scales. We thus propose scenario-based guidelines for building future attention modules. Code is open-sourced at https://github.com/DWlzm.

In the analysis of complex physical systems, the objective often extends beyond merely computing a numerical solution to capturing the precise crossover between different regimes and extracting parameters containing meaningful information. However, standard numerical solvers and conventional deep learning approaches, such as Physics-Informed Neural Networks (PINNs), typically operate as black boxes that output solution fields without disentangling the solution into its interpretable constituent parts. In this work, we propose GlueNN, a physics-informed learning framework that decomposes the global solution into interpretable, patchwise analytic components. Rather than approximating the solution directly, GlueNN promotes the integration constants of local asymptotic expansions to learnable, scale-dependent coefficient functions. By constraining these coefficients with the differential equation, the network effectively performs regime transition, smoothly interpolating between asymptotic limits without requiring ad hoc boundary matching. We demonstrate that this coefficient-centric approach reproduces accurate global solutions in various examples and thus directly extracts physical information that is not explicitly available through standard numerical integration.