Reinforcement Learning with Verifiable Rewards (RLVR) is a central paradigm for turning large language models (LLMs) into reliable problem solvers, especially in logic-heavy domains. Despite its empirical success, it remains unclear whether RLVR elicits novel capabilities or merely sharpens the distribution over existing knowledge. We study this by formalizing over-sharpening, a phenomenon where the policy collapses onto limited modes, suppressing valid alternatives. At a high level, we discover finite-batch updates intrinsically bias learning toward sampled modes, triggering a collapse that propagates globally via semantic coupling. To mitigate this, we propose inverse-success advantage calibration to prioritize difficult queries and distribution-level calibration to diversify sampling via a memory network. Empirical evaluations validate that our strategies can effectively improve generalization.

The proliferation of XR devices has made egocentric hand pose estimation a vital task, yet this perspective is inherently challenged by frequent finger occlusions. To address this, we propose a novel approach that leverages the rich information in dorsal hand skin deformation, unlocked by recent advances in dense visual featurizers. We introduce a dual-stream delta encoder that learns pose by contrasting features from a dynamic hand with a baseline relaxed position. Our evaluation demonstrates that, using only cropped dorsal images, our method reduces the Mean Per Joint Angle Error (MPJAE) by 18% in self-occluded scenarios (fingers >= 50% occluded) compared to state-of-the-art techniques that depend on the whole hand's geometry and large model backbones. Consequently, our method not only enhances the reliability of downstream tasks like index finger pinch and tap estimation in occluded scenarios but also unlocks new interaction paradigms, such as detecting isometric force for a surface "click" without visible movement while minimizing model size.

We propose a novel way to evaluate sycophancy of LLMs in a direct and neutral way, mitigating various forms of uncontrolled bias, noise, or manipulative language, deliberately injected to prompts in prior works. A key novelty in our approach is the use of LLM-as-a-judge, evaluation of sycophancy as a zero-sum game in a bet setting. Under this framework, sycophancy serves one individual (the user) while explicitly incurring cost on another. Comparing four leading models - Gemini 2.5 Pro, ChatGpt 4o, Mistral-Large-Instruct-2411, and Claude Sonnet 3.7 - we find that while all models exhibit sycophantic tendencies in the common setting, in which sycophancy is self-serving to the user and incurs no cost on others, Claude and Mistral exhibit "moral remorse" and over-compensate for their sycophancy in case it explicitly harms a third party. Additionally, we observed that all models are biased toward the answer proposed last. Crucially, we find that these two phenomena are not independent; sycophancy and recency bias interact to produce `constructive interference' effect, where the tendency to agree with the user is exacerbated when the user's opinion is presented last.

Resolving team conflicts requires not only task-specific competence, but also social intelligence to find common ground and build consensus. As AI agents increasingly collaborate on complex work, they must develop coordination capabilities to function as effective teammates. Yet we hypothesize that current agents lack these capabilities. To test this, we introduce CooperBench, a benchmark of over 600 collaborative coding tasks across 12 libraries in 4 programming languages. Each task assigns two agents different features that can be implemented independently but may conflict without proper coordination. Tasks are grounded in real open-source repositories with expert-written tests. Evaluating state-of-the-art coding agents, we observe the curse of coordination: agents achieve on average 30% lower success rates when working together compared to performing both tasks individually. This contrasts sharply with human teams, where adding teammates typically improves productivity. Our analysis reveals three key issues: (1) communication channels become jammed with vague, ill-timed, and inaccurate messages; (2) even with effective communication, agents deviate from their commitments; and (3) agents often hold incorrect expectations about others' plans and communication. Through large-scale simulation, we also observe rare but interesting emergent coordination behavior including role division, resource division, and negotiation. Our research presents a novel benchmark for collaborative coding and calls for a shift from pursuing individual agent capability to developing social intelligence.

Continual learning aims to enable neural networks to acquire new knowledge on sequential tasks. However, the key challenge in such settings is to learn new tasks without catastrophically forgetting previously learned tasks. We propose the Fisher-Orthogonal Projected Natural Gradient Descent (FOPNG) optimizer, which enforces Fisher-orthogonal constraints on parameter updates to preserve old task performance while learning new tasks. Unlike existing methods that operate in Euclidean parameter space, FOPNG projects gradients onto the Fisher-orthogonal complement of previous task gradients. This approach unifies natural gradient descent with orthogonal gradient methods within an information-geometric framework. We provide theoretical analysis deriving the projected update, describe efficient and practical implementations using the diagonal Fisher, and demonstrate strong results on standard continual learning benchmarks such as Permuted-MNIST, Split-MNIST, Rotated-MNIST, Split-CIFAR10, and Split-CIFAR100. Our code is available at https://github.com/ishirgarg/FOPNG.

Retrieval-Augmented Generation (RAG) has emerged as a powerful technique for enhancing the quality of responses in Question-Answering (QA) tasks. However, existing approaches often struggle with retrieving contextually relevant information, leading to incomplete or suboptimal answers. In this paper, we introduce Structured-Semantic RAG (SSRAG), a hybrid architecture that enhances QA quality by integrating query augmentation, agentic routing, and a structured retrieval mechanism combining vector and graph based techniques with context unification. By refining retrieval processes and improving contextual grounding, our approach improves both answer accuracy and informativeness. We conduct extensive evaluations on three popular QA datasets, TruthfulQA, SQuAD and WikiQA, across five Large Language Models (LLMs), demonstrating that our proposed approach consistently improves response quality over standard RAG implementations.

Large-scale livestock operations pose significant risks to human health and the environment, while also being vulnerable to threats such as infectious diseases and extreme weather events. As the number of such operations continues to grow, accurate and scalable mapping has become increasingly important. In this work, we present an infrastructure-first, explainable pipeline for identifying and characterizing Concentrated Animal Feeding Operations (CAFOs) from aerial and satellite imagery. Our method (i) detects candidate infrastructure (e.g., barns, feedlots, manure lagoons, silos) with a domain-tuned YOLOv8 detector, then derives SAM2 masks from these boxes and filters component-specific criteria; (ii) extracts structured descriptors (e.g., counts, areas, orientations, and spatial relations) and fuses them with deep visual features using a lightweight spatial cross-attention classifier; and (iii) outputs both CAFO type predictions and mask-level attributions that link decisions to visible infrastructure. Through comprehensive evaluation, we show that our approach achieves state-of-the-art performance, with Swin-B+PRISM-CAFO surpassing the best performing baseline by up to 15\%. Beyond strong predictive performance across diverse U.S. regions, we run systematic gradient--activation analyses that quantify the impact of domain priors and show how specific infrastructure (e.g., barns, lagoons) shapes classification decisions. We release code, infrastructure masks, and descriptors to support transparent, scalable monitoring of livestock infrastructure, enabling risk modeling, change detection, and targeted regulatory action. Github: https://github.com/Nibir088/PRISM-CAFO.

We present a family of open-source Music Foundation Models designed to advance large-scale music understanding and generation across diverse tasks and modalities. Our framework consists of four major components: (1) HeartCLAP, an audio-text alignment model; (2) HeartTranscriptor, a robust lyric recognition model optimized for real-world music scenarios; and (3) HeartCodec, a low-frame-rate (12.5 Hz) yet high-fidelity music codec tokenizer that captures long-range musical structure while preserving fine-grained acoustic details and enabling efficient autoregressive modeling; (4) HeartMuLa, an LLM-based song generation model capable of synthesizing high-fidelity music under rich, user-controllable conditions (e.g., textual style descriptions, lyrics, and reference audio). In addition, it provides two specialized modes: (i) fine-grained musical attribute control, which allows users to specify the style of different song sections (e.g., intro, verse, chorus) using natural language prompts; and (ii) short, engaging music generation, which is suitable as background music for short videos. Lastly, HeartMuLa improves significantly when scaled to 7B parameters. For the first time, we show that a Suno-level, commercial-grade system can be reproduced using academic-scale data and GPU resources. We expect these foundation models to serve as strong baselines for future research and to facilitate practical applications in multimodal content production.

Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning and prediction across different domains. Yet, their ability to infer temporal regularities from structured behavioral data remains underexplored. This paper presents a systematic study investigating whether LLMs can predict time intervals between recurring user actions, such as repeated purchases, and how different levels of contextual information shape their predictive behavior. Using a simple but representative repurchase scenario, we benchmark state-of-the-art LLMs in zero-shot settings against both statistical and machine-learning models. Two key findings emerge. First, while LLMs surpass lightweight statistical baselines, they consistently underperform dedicated machine-learning models, showing their limited ability to capture quantitative temporal structure. Second, although moderate context can improve LLM accuracy, adding further user-level detail degrades performance. These results challenge the assumption that "more context leads to better reasoning". Our study highlights fundamental limitations of today's LLMs in structured temporal inference and offers guidance for designing future context-aware hybrid models that integrate statistical precision with linguistic flexibility.

Language places subtle constraints on how we make inductive inferences. Developmental evidence by Gelman et al. (2002) has shown children (4 years and older) to differentiate among generic statements ("Bears are daxable"), universally quantified NPs ("all bears are daxable") and indefinite plural NPs ("some bears are daxable") in extending novel properties to a specific member (all > generics > some), suggesting that they represent these types of propositions differently. We test if these subtle differences arise in general purpose statistical learners like Vision Language Models, by replicating the original experiment. On tasking them through a series of precondition tests (robust identification of categories in images and sensitivities to all and some), followed by the original experiment, we find behavioral alignment between models and humans. Post-hoc analyses on their representations revealed that these differences are organized based on inductive constraints and not surface-form differences.

Large Language Models are rapidly emerging as web-native interfaces to social platforms. On the social web, users frequently have ambiguous and dynamic goals, making complex intent understanding-rather than single-turn execution-the cornerstone of effective human-LLM collaboration. Existing approaches attempt to clarify user intents through sequential or parallel questioning, yet they fall short of addressing the core challenge: modeling the logical dependencies among clarification questions. Inspired by the Cognitive Load Theory, we propose Prism, a novel framework for complex intent understanding that enables logically coherent and efficient intent clarification. Prism comprises four tailored modules: a complex intent decomposition module, which decomposes user intents into smaller, well-structured elements and identifies logical dependencies among them; a logical clarification generation module, which organizes clarification questions based on these dependencies to ensure coherent, low-friction interactions; an intent-aware reward module, which evaluates the quality of clarification trajectories via an intent-aware reward function and leverages Monte Carlo Sample to simulate user-LLM interactions for large-scale,high-quality training data generation; and a self-evolved intent tuning module, which iteratively refines the LLM's logical clarification capability through data-driven feedback and optimization. Prism consistently outperforms existing approaches across clarification interactions, intent execution, and cognitive load benchmarks. It achieves stateof-the-art logical consistency, reduces logical conflicts to 11.5%, increases user satisfaction by 14.4%, and decreases task completion time by 34.8%. All data and code are released.

As language-model agents evolve from passive chatbots into proactive assistants that handle personal data, evaluating their adherence to social norms becomes increasingly critical, often through the lens of Contextual Integrity (CI). However, existing CI benchmarks are largely text-centric and primarily emphasize negative refusal scenarios, overlooking multimodal privacy risks and the fundamental trade-off between privacy and utility. In this paper, we introduce MPCI-Bench, the first Multimodal Pairwise Contextual Integrity benchmark for evaluating privacy behavior in agentic settings. MPCI-Bench consists of paired positive and negative instances derived from the same visual source and instantiated across three tiers: normative Seed judgments, context-rich Story reasoning, and executable agent action Traces. Data quality is ensured through a Tri-Principle Iterative Refinement pipeline. Evaluations of state-of-the-art multimodal models reveal systematic failures to balance privacy and utility and a pronounced modality leakage gap, where sensitive visual information is leaked more frequently than textual information. We will open-source MPCI-Bench to facilitate future research on agentic CI.

AI-text detectors achieve high accuracy on in-domain benchmarks, but often struggle to generalize across different generation conditions such as unseen prompts, model families, or domains. While prior work has reported these generalization gaps, there are limited insights about the underlying causes. In this work, we present a systematic study aimed at explaining generalization behavior through linguistic analysis. We construct a comprehensive benchmark that spans 6 prompting strategies, 7 large language models (LLMs), and 4 domain datasets, resulting in a diverse set of human- and AI-generated texts. Using this dataset, we fine-tune classification-based detectors on various generation settings and evaluate their cross-prompt, cross-model, and cross-dataset generalization. To explain the performance variance, we compute correlations between generalization accuracies and feature shifts of 80 linguistic features between training and test conditions. Our analysis reveals that generalization performance for specific detectors and evaluation conditions is significantly associated with linguistic features such as tense usage and pronoun frequency.

The ability to construct mental models of the world is a central aspect of understanding. Similarly, visual understanding can be viewed as the ability to construct a representative model of the system depicted in an image. This work explores the capacity of Vision Language Models (VLMs) to recognize and simulate the systems and mechanisms depicted in images using the Im2Sim methodology. The VLM is given a natural image of a real-world system (e.g., cities, clouds, vegetation) and is tasked with describing the system and writing code that simulates and generates it. This generative code is then executed to produce a synthetic image, which is compared against the original. This approach is tested on various complex emergent systems, ranging from physical systems (waves, lights, clouds) to vegetation, cities, materials, and geological formations. Through analysis of the models and images generated by the VLMs, we examine their understanding of the systems in images. The results show that leading VLMs (GPT, Gemini) have the ability to understand and model complex, multi-component systems across multiple layers of abstraction and a wide range of domains. At the same time, the VLMs exhibit limited ability to replicate fine details and low-level arrangements of patterns in the image. These findings reveal an interesting asymmetry: VLMs combine high-level, deep visual understanding of images with limited perception of fine details.

Recent 3D-aware head generative models based on 3D Gaussian Splatting achieve real-time, photorealistic and view-consistent head synthesis. However, a fundamental limitation persists: the deep entanglement of illumination and intrinsic appearance prevents controllable relighting. Existing disentanglement methods rely on strong assumptions to enable weakly supervised learning, which restricts their capacity for complex illumination. To address this challenge, we introduce HeadLighter, a novel supervised framework that learns a physically plausible decomposition of appearance and illumination in head generative models. Specifically, we design a dual-branch architecture that separately models lighting-invariant head attributes and physically grounded rendering components. A progressive disentanglement training is employed to gradually inject head appearance priors into the generative architecture, supervised by multi-view images captured under controlled light conditions with a light stage setup. We further introduce a distillation strategy to generate high-quality normals for realistic rendering. Experiments demonstrate that our method preserves high-quality generation and real-time rendering, while simultaneously supporting explicit lighting and viewpoint editing. We will publicly release our code and dataset.

Novel metaphor comprehension involves complex semantic processes and linguistic creativity, making it an interesting task for studying language models (LMs). This study investigates whether surprisal, a probabilistic measure of predictability in LMs, correlates with annotations of metaphor novelty in different datasets. We analyse the surprisal of metaphoric words in corpus-based and synthetic metaphor datasets using 16 causal LM variants. We propose a cloze-style surprisal method that conditions on full-sentence context. Results show that LM surprisal yields significant moderate correlations with scores/labels of metaphor novelty. We further identify divergent scaling patterns: on corpus-based data, correlation strength decreases with model size (inverse scaling effect), whereas on synthetic data it increases (quality-power hypothesis). We conclude that while surprisal can partially account for annotations of metaphor novelty, it remains limited as a metric of linguistic creativity. Code and data are publicly available: https://github.com/OmarMomen14/surprisal-metaphor-novelty

We present RoboPhD, a system where AI agents autonomously conduct research to improve Text-to-SQL performance. RoboPhD implements a closed-loop evolution cycle with two coordinated components: a SQL Generation agent composed of a database analysis script and SQL generation instructions, and an Evolution agent that designs new versions based on performance feedback. Central to the framework is an ELO-based selection mechanism enabling survival-of-the-fittest dynamics while handling non-transitivity in performance. Starting from a naive 70-line baseline, RoboPhD evolves agents through iterative cross-pollination, discovering effective techniques without any external guidance on the Text-to-SQL domain. Our best agent, evolved to 1500 lines over 18 iterations, autonomously discovered strategies such as size-adaptive database analysis that adjusts depth based on schema complexity and SQL generation patterns for column selection, evidence interpretation, and aggregation. Evolution provides the largest gains on cheaper models: while we improve by 2.3 points over a strong Claude Opus 4.5 naive baseline, we show an improvement of 8.9 points over the weaker Claude Haiku model. This enables 'skip a tier' deployment: evolved Haiku exceeds naive Sonnet accuracy, and evolved Sonnet exceeds naive Opus, both at lower cost. The full system achieves 73.67% accuracy on the BIRD test set, demonstrating that AI can autonomously build a strong agentic system with only a trivial human-provided starting point.

Automated Essay Scoring (AES) has gained increasing attention in recent years, yet research on Arabic AES remains limited due to the lack of publicly available datasets. To address this, we introduce LAILA, the largest publicly available Arabic AES dataset to date, comprising 7,859 essays annotated with holistic and trait-specific scores on seven dimensions: relevance, organization, vocabulary, style, development, mechanics, and grammar. We detail the dataset design, collection, and annotations, and provide benchmark results using state-of-the-art Arabic and English models in prompt-specific and cross-prompt settings. LAILA fills a critical need in Arabic AES research, supporting the development of robust scoring systems.

Large Language Models can develop reasoning capabilities through online fine-tuning with rule-based rewards. However, recent studies reveal a critical constraint: reinforcement learning succeeds only when the base model already assigns non-negligible probability to correct answers -- a property we term 'latent solvability'. This work investigates the emergence of chemical reasoning capabilities and what these prerequisites mean for chemistry. We identify two necessary conditions for RL-based chemical reasoning: 1) Symbolic competence, and 2) Latent chemical knowledge. We propose mid-stage scientific training (MiST): a set of mid-stage training techniques to satisfy these, including data-mixing with SMILES/CIF-aware pre-processing, continued pre-training on 2.9B tokens, and supervised fine-tuning on 1B tokens. These steps raise the latent-solvability score on 3B and 7B models by up to 1.8x, and enable RL to lift top-1 accuracy from 10.9 to 63.9% on organic reaction naming, and from 40.6 to 67.4% on inorganic material generation. Similar results are observed for other challenging chemical tasks, while producing interpretable reasoning traces. Our results define clear prerequisites for chemical reasoning training and highlight the broader role of mid-stage training in unlocking reasoning capabilities.

This letter extends the exactly sparse Gaussian variational inference (ESGVI) algorithm for state estimation in two complementary directions. First, ESGVI is generalized to operate on matrix Lie groups, enabling the estimation of states with orientation components while respecting the underlying group structure. Second, factors are introduced to accommodate heavy-tailed and skewed noise distributions, as commonly encountered in ultra-wideband (UWB) localization due to non-line-of-sight (NLOS) and multipath effects. Both extensions are shown to integrate naturally within the ESGVI framework while preserving its sparse and derivative-free structure. The proposed approach is validated in a UWB localization experiment with NLOS-rich measurements, demonstrating improved accuracy and comparable consistency. Finally, a Python implementation within a factor-graph-based estimation framework is made open-source (https://github.com/decargroup/gvi_ws) to support broader research use.

We introduce Sprecher Networks (SNs), a family of trainable architectures derived from David Sprecher's 1965 constructive form of the Kolmogorov-Arnold representation. Each SN block implements a "sum of shifted univariate functions" using only two shared learnable splines per block, a monotone inner spline $ϕ$ and a general outer spline $Φ$, together with a learnable shift parameter $η$ and a mixing vector $λ$ shared across all output dimensions. Stacking these blocks yields deep, compositional models; for vector-valued outputs we append an additional non-summed output block. We also propose an optional lateral mixing operator enabling intra-block communication between output channels with only $O(d_{\mathrm{out}})$ additional parameters. Owing to the vector (not matrix) mixing weights and spline sharing, SNs scale linearly in width, approximately $O(\sum_{\ell}(d_{\ell-1}+d_{\ell}+G))$ parameters for $G$ spline knots, versus $O(\sum_{\ell} d_{\ell-1}d_{\ell})$ for dense MLPs and $O(G\sum_{\ell} d_{\ell-1}d_{\ell})$ for edge-spline KANs. This linear width-scaling is particularly attractive for extremely wide, shallow models, where low depth can translate into low inference latency. Finally, we describe a sequential forward implementation that avoids materializing the $d_{\mathrm{in}}\times d_{\mathrm{out}}$ shifted-input tensor, reducing peak forward-intermediate memory from quadratic to linear in layer width, relevant for memory-constrained settings such as on-device/edge inference; we demonstrate deployability via fixed-point real-time digit classification on resource-constrained embedded device with only 4 MB RAM. We provide empirical demonstrations on supervised regression, Fashion-MNIST classification (including stable training at 25 hidden layers with residual connections and normalization), and a Poisson PINN, with controlled comparisons to MLP and KAN baselines.

This paper examines the exploration-exploitation trade-off in reinforcement learning with verifiable rewards (RLVR), a framework for improving the reasoning of Large Language Models (LLMs). Recent studies suggest that RLVR can elicit strong mathematical reasoning in LLMs through two seemingly paradoxical mechanisms: spurious rewards, which suppress exploitation by rewarding outcomes unrelated to the ground truth, and entropy minimization, which suppresses exploration by pushing the model toward more confident and deterministic outputs, highlighting a puzzling dynamic: both discouraging exploitation and discouraging exploration improve reasoning performance, yet the underlying principles that reconcile these effects remain poorly understood. We focus on two fundamental questions: (i) how policy entropy relates to performance, and (ii) whether spurious rewards yield gains, potentially through the interplay of clipping bias and model contamination. Our results show that clipping bias under spurious rewards reduces policy entropy, leading to more confident and deterministic outputs, while entropy minimization alone is insufficient for improvement. We further propose a reward-misalignment model explaining why spurious rewards can enhance performance beyond contaminated settings. Our findings clarify the mechanisms behind spurious-reward benefits and provide principles for more effective RLVR training.

Data is a critical asset for training large language models (LLMs), alongside compute resources and skilled workers. While some training data is publicly available, substantial investment is required to generate proprietary datasets, such as human preference annotations or to curate new ones from existing sources. As larger datasets generally yield better model performance, two natural questions arise. First, how can data owners make informed decisions about curation strategies and data sources investment? Second, how can multiple data owners collaboratively pool their resources to train superior models while fairly distributing the benefits? This problem, data valuation, which is not specific to large language models, has been addressed by the machine learning community through the lens of cooperative game theory, with the Shapley value being the prevalent solution concept. However, computing Shapley values is notoriously expensive for data valuation, typically requiring numerous model retrainings, which can become prohibitive for large machine learning models. In this work, we demonstrate that this computational challenge is dramatically simplified for LLMs trained with Direct Preference Optimization (DPO). We show how the specific mathematical structure of DPO enables scalable Shapley value computation. We believe this observation unlocks many applications at the intersection of data valuation and large language models.

Automatic Music Transcription (AMT) -- the task of converting music audio into note representations -- has seen rapid progress, driven largely by deep learning systems. Due to the limited availability of richly annotated music datasets, much of the progress in AMT has been concentrated on classical piano music, and even a few very specific datasets. Whether these systems can generalize effectively to other musical contexts remains an open question. Complementing recent studies on distribution shifts in sound (e.g., recording conditions), in this work we investigate the musical dimension -- specifically, variations in genre, dynamics, and polyphony levels. To this end, we introduce the MDS corpus, comprising three distinct subsets -- (1) Genre, (2) Random, and (3) MAEtest -- to emulate different axes of distribution shift. We evaluate the performance of several state-of-the-art AMT systems on the MDS corpus using both traditional information-retrieval and musically-informed performance metrics. Our extensive evaluation isolates and exposes varying degrees of performance degradation under specific distribution shifts. In particular, we measure a note-level F1 performance drop of 20 percentage points due to sound, and 14 due to genre. Generally, we find that dynamics estimation proves more vulnerable to musical variation than onset prediction. Musically informed evaluation metrics, particularly those capturing harmonic structure, help identify potential contributing factors. Furthermore, experiments with randomly generated, non-musical sequences reveal clear limitations in system performance under extreme musical distribution shifts. Altogether, these findings offer new evidence of the persistent impact of the Corpus Bias problem in deep AMT systems.

Envy shapes competitiveness and cooperation in human groups, yet its role in large language model interactions remains largely unexplored. As LLMs increasingly operate in multi-agent settings, it is important to examine whether they exhibit envy-like preferences under social comparison. We evaluate LLM behavior across two scenarios: (1) a point-allocation game testing sensitivity to relative versus absolute payoff, and (2) comparative evaluations across general and contextual settings. To ground our analysis in psychological theory, we adapt four established psychometric questionnaires spanning general, domain-specific, workplace, and sibling-based envy. Our results reveal heterogeneous envy-like patterns across models and contexts, with some models sacrificing personal gain to reduce a peer's advantage, while others prioritize individual maximization. These findings highlight competitive dispositions as a design and safety consideration for multi-agent LLM systems.

In automotive sensor fusion systems, smart sensors and Vehicle-to-Everything (V2X) modules are commonly utilized. Sensor data from these systems are typically available only as processed object lists rather than raw sensor data from traditional sensors. Instead of processing other raw data separately and then fusing them at the object level, we propose an end-to-end cross-level fusion concept with Transformer, which integrates highly abstract object list information with raw camera images for 3D object detection. Object lists are fed into a Transformer as denoising queries and propagated together with learnable queries through the latter feature aggregation process. Additionally, a deformable Gaussian mask, derived from the positional and size dimensional priors from the object lists, is explicitly integrated into the Transformer decoder. This directs attention toward the target area of interest and accelerates model training convergence. Furthermore, as there is no public dataset containing object lists as a standalone modality, we propose an approach to generate pseudo object lists from ground-truth bounding boxes by simulating state noise and false positives and negatives. As the first work to conduct cross-level fusion, our approach shows substantial performance improvements over the vision-based baseline on the nuScenes dataset. It demonstrates its generalization capability over diverse noise levels of simulated object lists and real detectors.

Contact-rich tasks pose significant challenges for robotic systems due to inherent uncertainty, complex dynamics, and the high risk of damage during interaction. Recent advances in learning-based control have shown great potential in enabling robots to acquire and generalize complex manipulation skills in such environments, but ensuring safety, both during exploration and execution, remains a critical bottleneck for reliable real-world deployment. This survey provides a comprehensive overview of safe learning-based methods for robot contact-rich tasks. We categorize existing approaches into two main domains: safe exploration and safe execution. We review key techniques, including constrained reinforcement learning, risk-sensitive optimization, uncertainty-aware modeling, control barrier functions, and model predictive safety shields, and highlight how these methods incorporate prior knowledge, task structure, and online adaptation to balance safety and efficiency. A particular emphasis of this survey is on how these safe learning principles extend to and interact with emerging robotic foundation models, especially vision-language models (VLMs) and vision-language-action models (VLAs), which unify perception, language, and control for contact-rich manipulation. We discuss both the new safety opportunities enabled by VLM/VLA-based methods, such as language-level specification of constraints and multimodal grounding of safety signals, and the amplified risks and evaluation challenges they introduce. Finally, we outline current limitations and promising future directions toward deploying reliable, safety-aligned, and foundation-model-enabled robots in complex contact-rich environments. More details and materials are available at our \href{ https://github.com/jack-sherman01/Awesome-Learning4Safe-Contact-rich-tasks}{Project GitHub Repository}.

The acquisition cost for large, annotated motion datasets remains a critical bottleneck for skeletal-based Human Activity Recognition (HAR). Although Text-to-Motion (T2M) generative models offer a compelling, scalable source of synthetic data, their training objectives, which emphasize general artistic motion, and dataset structures fundamentally differ from HAR's requirements for kinematically precise, class-discriminative actions. This disparity creates a significant domain gap, making generalist T2M models ill-equipped for generating motions suitable for HAR classifiers. To address this challenge, we propose KineMIC (Kinetic Mining In Context), a transfer learning framework for few-shot action synthesis. KineMIC adapts a T2M diffusion model to an HAR domain by hypothesizing that semantic correspondences in the text encoding space can provide soft supervision for kinematic distillation. We operationalize this via a kinetic mining strategy that leverages CLIP text embeddings to establish correspondences between sparse HAR labels and T2M source data. This process guides fine-tuning, transforming the generalist T2M backbone into a specialized few-shot Action-to-Motion generator. We validate KineMIC using HumanML3D as the source T2M dataset and a subset of NTU RGB+D 120 as the target HAR domain, randomly selecting just 10 samples per action class. Our approach generates significantly more coherent motions, providing a robust data augmentation source that delivers a +23.1% accuracy points improvement. Animated illustrations and supplementary materials are available at https://lucazzola.github.io/publications/kinemic.

Contact-rich manipulation is difficult for robots to execute and requires accurate perception of the environment. In some scenarios, vision is occluded. The robot can then no longer obtain real-time scene state information through visual feedback. This is called ``blind manipulation". In this manuscript, a novel physically-driven contact cognition method, called ``Contact SLAM", is proposed. It estimates the state of the environment and achieves manipulation using only tactile sensing and prior knowledge of the scene. To maximize exploration efficiency, this manuscript also designs an active exploration policy. The policy gradually reduces uncertainties in the manipulation scene. The experimental results demonstrated the effectiveness and accuracy of the proposed method in several contact-rich tasks, including the difficult and delicate socket assembly task and block-pushing task.

The continuous scaling of deep neural networks has fundamentally transformed machine learning, with larger models demonstrating improved performance across diverse tasks. This growth in model size has dramatically increased the computational resources required for the training process. Consequently, distributed approaches, such as Federated Learning and Split Learning, have become essential paradigms for scalable deployment. However, existing Split Learning approaches assume client homogeneity and uniform split points across all participants. This critically limits their applicability to real-world IoT systems where devices exhibit heterogeneity in computational resources. To address this limitation, this paper proposes Hetero-SplitEE, a novel method that enables heterogeneous IoT devices to train a shared deep neural network in parallel collaboratively. By integrating heterogeneous early exits into hierarchical training, our approach allows each client to select distinct split points (cut layers) tailored to its computational capacity. In addition, we propose two cooperative training strategies, the Sequential strategy and the Averaging strategy, to facilitate this collaboration among clients with different split points. The Sequential strategy trains clients sequentially with a shared server model to reduce computational overhead. The Averaging strategy enables parallel client training with periodic cross-layer aggregation. Extensive experiments on CIFAR-10, CIFAR-100, and STL-10 datasets using ResNet-18 demonstrate that our method maintains competitive accuracy while efficiently supporting diverse computational constraints, enabling practical deployment of collaborative deep learning in heterogeneous IoT ecosystems.