This work focuses on Bayesian optimization (BO) under reward model uncertainty. We propose the first BO algorithm that achieves no-regret guarantee in a general reward setting, requiring only Lipschitz continuity of the objective function and accommodating a broad class of measurement noise. The core of our approach is a novel surrogate model, termed as infinite Gaussian process ($\infty$-GP). It is a Bayesian nonparametric model that places a prior on the space of reward distributions, enabling it to represent a substantially broader class of reward models than classical Gaussian process (GP). The $\infty$-GP is used in combination with Thompson Sampling (TS) to enable effective exploration and exploitation. Correspondingly, we develop a new TS regret analysis framework for general rewards, which relates the regret to the total variation distance between the surrogate model and the true reward distribution. Furthermore, with a truncated Gibbs sampling procedure, our method is computationally scalable, incurring minimal additional memory and computational complexities compared to classical GP. Empirical results demonstrate state-of-the-art performance, particularly in settings with non-stationary, heavy-tailed, or other ill-conditioned rewards.

The energy and latency costs of deep neural network inference are increasingly driven by deployment rather than training, motivating hardware-specialized alternatives to arithmetic-heavy models. Field-Programmable Gate Arrays (FPGAs) provide an attractive substrate for such specialization, yet existing FPGA-based neural approaches are fragmented and difficult to compare. We present BitLogic, a fully gradient-based, end-to-end trainable framework for FPGA-native neural networks built around Lookup Table (LUT) computation. BitLogic replaces multiply-accumulate operations with differentiable LUT nodes that map directly to FPGA primitives, enabling native binary computation, sparse connectivity, and efficient hardware realization. The framework offers a modular functional API supporting diverse architectures, along with learned encoders, hardware-aware heads, and multiple boundary-consistent LUT relaxations. An automated Register Transfer Level (RTL) export pipeline translates trained PyTorch models into synthesizable HDL, ensuring equivalence between software and hardware inference. Experiments across standard vision benchmarks and heterogeneous hardware platforms demonstrate competitive accuracy and substantial gains in FPGA efficiency, including 72.3% test accuracy on CIFAR-10 achieved with fewer than 0.3M logic gates, while attaining sub-20 ns single-sample inference using only LUT resources.

Self-attention dominates the computational and memory cost of long-context LLM inference across both prefill and decode phases. To address this challenge, we introduce Sketch&Walk Attention, a training-free sparse attention method that determines sparsity with lightweight sketches and deterministic walk. Sketch&Walk applies Hadamard sketching to get inexpensive approximations of attention scores, then aggregates these estimates across layers via a walk mechanism that captures attention influence beyond direct interactions between tokens. The accumulated walk scores are used to select top-k attention blocks, enabling dynamic sparsity with a single training-free algorithm that applies uniformly to both the prefill and decode phases, together with custom sparse attention kernels. Across a wide range of models and tasks, Sketch&Walk maintains near-lossless accuracy at 20% attention density and can slightly outperform dense attention in some settings, while achieving up to 6x inference speedup.

Generative model-based policies have shown strong performance in imitation-based robotic manipulation by learning action distributions from demonstrations. However, in long-horizon tasks, visually similar observations often recur across execution stages while requiring distinct actions, which leads to ambiguous predictions when policies are conditioned only on instantaneous observations, termed multi-modal action ambiguity (MA2). To address this challenge, we propose the Trace-Focused Diffusion Policy (TF-DP), a simple yet effective diffusion-based framework that explicitly conditions action generation on the robot's execution history. TF-DP represents historical motion as an explicit execution trace and projects it into the visual observation space, providing stage-aware context when current observations alone are insufficient. In addition, the induced trace-focused field emphasizes task-relevant regions associated with historical motion, improving robustness to background visual disturbances. We evaluate TF-DP on real-world robotic manipulation tasks exhibiting pronounced multi-modal action ambiguity and visually cluttered conditions. Experimental results show that TF-DP improves temporal consistency and robustness, outperforming the vanilla diffusion policy by 80.56 percent on tasks with multi-modal action ambiguity and by 86.11 percent under visual disturbances, while maintaining inference efficiency with only a 6.4 percent runtime increase. These results demonstrate that execution-trace conditioning offers a scalable and principled approach for robust long-horizon robotic manipulation within a single policy.

Summarizing Indian legal court judgments is a complex task not only due to the intricate language and unstructured nature of the legal texts, but also since a large section of the Indian population does not understand the complex English in which legal text is written, thus requiring summaries in Indian languages. In this study, we aim to improve the summarization of Indian legal text to generate summaries in both English and Hindi (the most widely spoken Indian language), by injecting domain knowledge into diverse summarization models. We propose a framework to enhance extractive neural summarization models by incorporating domain-specific pre-trained encoders tailored for legal texts. Further, we explore the injection of legal domain knowledge into generative models (including Large Language Models) through continual pre-training on large legal corpora in English and Hindi. Our proposed approaches achieve statistically significant improvements in both English-to-English and English-to-Hindi Indian legal document summarization, as measured by standard evaluation metrics, factual consistency metrics, and legal domain-specific metrics. Furthermore, these improvements are validated through domain experts, demonstrating the effectiveness of our approaches.

Large Language Models (LLMs) need to be in accordance with human values-being helpful, harmless, and honest (HHH)-is important for safe deployment. Existing works use Supervised Fine-Tuning (SFT) and Mixture-of-Experts (MoE) to align LLMs. However, these works face challenges in multi-objective settings, such as SFT leading to interference between conflicting objectives, while MoEs suffer from miscalibrated routing. We term this failure mode Axis Collapse, marked by (1) disjoint feature spaces causing catastrophic forgetting, and (2) unreliable inference from misrouted experts. To resolve this, we propose AlignX, a two-stage framework. Stage 1 uses prompt-injected fine-tuning to extract axis-specific task features, mitigating catastrophic forgetting. Stage 2 deploys a MoCaE module that calibrates expert routing using fractal and natural geometry, improving inference reliability. AlignX achieves significant gains on Alpaca (Helpfulness), BeaverTails (Harmlessness), and TruthfulQA (Honesty), with +171.5% win rate, +110.1% in truthfulness-informativeness, and 4.3% fewer safety violations. It also reduces latency and memory usage by over 35% compared to prior MoEs. Results across four LLMs validate its generalizability.

The dynamics of gradient-based training in neural networks often exhibit nontrivial structures; hence, understanding them remains a central challenge in theoretical machine learning. In particular, a concept of feature unlearning, in which a neural network progressively loses previously learned features over long training, has gained attention. In this study, we consider the infinite-width limit of a two-layer neural network updated with a large-batch stochastic gradient, then derive differential equations with different time scales, revealing the mechanism and conditions for feature unlearning to occur. Specifically, we utilize the fast-slow dynamics: while an alignment of first-layer weights develops rapidly, the second-layer weights develop slowly. The direction of a flow on a critical manifold, determined by the slow dynamics, decides whether feature unlearning occurs. We give numerical validation of the result, and derive theoretical grounding and scaling laws of the feature unlearning. Our results yield the following insights: (i) the strength of the primary nonlinear term in data induces the feature unlearning, and (ii) an initial scale of the second-layer weights mitigates the feature unlearning. Technically, our analysis utilizes Tensor Programs and the singular perturbation theory.

Large Language Model (LLM) safety is inherently pluralistic, reflecting variations in moral norms, cultural expectations, and demographic contexts. Yet, existing alignment datasets such as ANTHROPIC-HH and DICES rely on demographically narrow annotator pools, overlooking variation in safety perception across communities. Demo-SafetyBench addresses this gap by modeling demographic pluralism directly at the prompt level, decoupling value framing from responses. In Stage I, prompts from DICES are reclassified into 14 safety domains (adapted from BEAVERTAILS) using Mistral 7B-Instruct-v0.3, retaining demographic metadata and expanding low-resource domains via Llama-3.1-8B-Instruct with SimHash-based deduplication, yielding 43,050 samples. In Stage II, pluralistic sensitivity is evaluated using LLMs-as-Raters-Gemma-7B, GPT-4o, and LLaMA-2-7B-under zero-shot inference. Balanced thresholds (delta = 0.5, tau = 10) achieve high reliability (ICC = 0.87) and low demographic sensitivity (DS = 0.12), confirming that pluralistic safety evaluation can be both scalable and demographically robust.

Post-training pruning is an effective approach for reducing the size and inference cost of large language models (LLMs), but existing methods often face a trade-off between pruning quality and computational efficiency. Heuristic pruning methods are efficient but sensitive to activation outliers, while reconstruction-based approaches improve fidelity at the cost of heavy computation. In this work, we propose a lightweight post-training pruning framework based on first-order statistical properties of model weights and activations. During pruning, channel-wise statistics are used to calibrate magnitude-based importance scores, reducing bias from activation-dominated channels. After pruning, we apply an analytic energy compensation to correct distributional distortions caused by weight removal. Both steps operate without retraining, gradients, or second-order information. Experiments across multiple LLM families, sparsity patterns, and evaluation tasks show that the proposed approach improves pruning performance while maintaining computational cost comparable to heuristic methods. The results suggest that simple statistical corrections can be effective for post-training pruning of LLMs.

We give new differentially private algorithms for the classic problems of learning decision lists and large-margin halfspaces in the PAC and online models. In the PAC model, we give a computationally efficient algorithm for learning decision lists with minimal sample overhead over the best non-private algorithms. In the online model, we give a private analog of the influential Winnow algorithm for learning halfspaces with mistake bound polylogarithmic in the dimension and inverse polynomial in the margin. As an application, we describe how to privately learn decision lists in the online model, qualitatively matching state-of-the art non-private guarantees.

Graph neural networks (GNNs) naturally align with sparse operators and unstructured discretizations, making them a promising paradigm for physics-informed machine learning in computational mechanics. Motivated by discrete physics losses and Hierarchical Deep Learning Neural Network (HiDeNN) constructions, we embed finite-element (FEM) computations at nodes and Gauss points directly into message-passing layers and propose a numerically consistent FEM-Informed Hypergraph Neural Networks (FHGNN). Similar to conventional physics-informed neural networks (PINNs), training is purely physics-driven and requires no labeled data: the input is a node element hypergraph whose edges encode mesh connectivity. Guided by empirical results and condition-number analysis, we adopt an efficient variational loss. Validated on 3D benchmarks, including cyclic loading with isotropic/kinematic hardening, the proposed method delivers substantially improved accuracy and efficiency over recent, competitive PINN variants. By leveraging GPU-parallel tensor operations and the discrete representation, it scales effectively to large elastoplastic problems and can be competitive with, or faster than, multi-core FEM implementations at comparable accuracy. This work establishes a foundation for scalable, physics-embedded learning in nonlinear solid mechanics.

Quadruped robots are increasingly deployed in unstructured environments. Safe locomotion in these settings requires long-horizon goal progress, passability over uneven terrain and static constraints, and collision avoidance against high-speed dynamic obstacles. A single system cannot fully satisfy all three objectives simultaneously: planning-based decisions can be too slow, while purely reactive decisions can sacrifice goal progress and passability. To resolve this conflict, we propose UEREBot (Unstructured-Environment Reflexive Evasion Robot), a hierarchical framework that separates slow planning from instantaneous reflexive evasion and coordinates them during execution. UEREBot formulates the task as a constrained optimal control problem blueprint. It adopts a spatial--temporal planner that provides reference guidance toward the goal and threat signals. It then uses a threat-aware handoff to fuse navigation and reflex actions into a nominal command, and a control barrier function shield as a final execution safeguard. We evaluate UEREBot in Isaac Lab simulation and deploy it on a Unitree Go2 quadruped equipped with onboard perception. Across diverse environments with complex static structure and high-speed dynamic obstacles, UEREBot achieves higher avoidance success and more stable locomotion while maintaining goal progress than representative baselines, demonstrating improved safety--progress trade-offs.

Question Answering (QA) over regulatory documents is inherently challenging due to the need for multihop reasoning across legally interdependent texts, a requirement that is particularly pronounced in the healthcare domain where regulations are hierarchically structured and frequently revised through amendments and cross-references. Despite recent progress in retrieval-augmented and graph-based QA methods, systematic evaluation in this setting remains limited, especially for low-resource languages such as Vietnamese, due to the lack of benchmark datasets that explicitly support multihop reasoning over healthcare regulations. In this work, we introduce the Vietnamese Healthcare Regulations-Multihop Reasoning Dataset (ViHERMES), a benchmark designed for multihop QA over Vietnamese healthcare regulatory documents. ViHERMES consists of high-quality question-answer pairs that require reasoning across multiple regulations and capture diverse dependency patterns, including amendment tracing, cross-document comparison, and procedural synthesis. To construct the dataset, we propose a controlled multihop QA generation pipeline based on semantic clustering and graph-inspired data mining, followed by large language model-based generation with structured evidence and reasoning annotations. We further present a graph-aware retrieval framework that models formal legal relations at the level of legal units and supports principled context expansion for legally valid and coherent answers. Experimental results demonstrate that ViHERMES provides a challenging benchmark for evaluating multihop regulatory QA systems and that the proposed graph-aware approach consistently outperforms strong retrieval-based baselines. The ViHERMES dataset and system implementation are publicly available at https://github.com/ura-hcmut/ViHERMES.

Deep research agents have emerged as powerful tools for automating complex intellectual tasks through multi-step reasoning and web-based information seeking. While recent efforts have successfully enhanced these agents by scaling depth through increasing the number of sequential thinking and tool calls, the potential of scaling width via parallel tool calling remains largely unexplored. In this work, we propose the Wide and Deep research agent, a framework designed to investigate the behavior and performance of agents when scaling not only depth but also width via parallel tool calling. Unlike existing approaches that rely on complex multi-agent orchestration to parallelize workloads, our method leverages intrinsic parallel tool calling to facilitate effective coordination within a single reasoning step. We demonstrate that scaling width significantly improves performance on deep research benchmarks while reducing the number of turns required to obtain correct answers. Furthermore, we analyze the factors driving these improvements through case studies and explore various tool call schedulers to optimize parallel tool calling strategy. Our findings suggest that optimizing the trade-off between width and depth is a critical pathway toward high-efficiency deep research agents. Notably, without context management or other tricks, we obtain 62.2% accuracy with GPT-5-Medium on BrowseComp, surpassing the original 54.9% reported by GPT-5-High.

This paper focuses on the scalable robot learning for manipulation in the dexterous robot arm-hand systems, where the remote human-robot interactions via augmented reality (AR) are established to collect the expert demonstration data for improving efficiency. In such a system, we present a unified framework to address the general manipulation task problem. Specifically, the proposed method consists of two phases: i) In the first phase for pretraining, the policy is created in a behavior cloning (BC) manner, through leveraging the learning data from our AR-based remote human-robot interaction system; ii) In the second phase, a contrastive learning empowered reinforcement learning (RL) method is developed to obtain more efficient and robust policy than the BC, and thus a projection head is designed to accelerate the learning progress. An event-driven augmented reward is adopted for enhancing the safety. To validate the proposed method, both the physics simulations via PyBullet and real-world experiments are carried out. The results demonstrate that compared to the classic proximal policy optimization and soft actor-critic policies, our method not only significantly speeds up the inference, but also achieves much better performance in terms of the success rate for fulfilling the manipulation tasks. By conducting the ablation study, it is confirmed that the proposed RL with contrastive learning overcomes policy collapse. Supplementary demonstrations are available at https://cyberyyc.github.io/.

Diffusion-based trajectory planners have demonstrated strong capability for modeling the multimodal nature of human driving behavior, but their reliance on iterative stochastic sampling poses critical challenges for real-time, safety-critical deployment. In this work, we present RAPiD, a deterministic policy extraction framework that distills a pretrained diffusion-based planner into an efficient policy while eliminating diffusion sampling. Using score-regularized policy optimization, we leverage the score function of a pre-trained diffusion planner as a behavior prior to regularize policy learning. To promote safety and passenger comfort, the policy is optimized using a critic trained to imitate a predictive driver controller, providing dense, safety-focused supervision beyond conventional imitation learning. Evaluations demonstrate that RAPiD achieves competitive performance on closed-loop nuPlan scenarios with an 8x speedup over diffusion baselines, while achieving state-of-the-art generalization among learning-based planners on the interPlan benchmark. The official website of this work is: https://github.com/ruturajreddy/RAPiD.

Multi-turn conversation has emerged as a predominant interaction paradigm for Large Language Models (LLMs). Users often employ follow-up questions to refine their intent, expecting LLMs to adapt dynamically. However, recent research reveals that LLMs suffer a substantial performance drop in multi-turn settings compared to single-turn interactions with fully specified instructions, a phenomenon termed ``Lost in Conversation'' (LiC). While this prior work attributes LiC to model unreliability, we argue that the root cause lies in an intent alignment gap rather than intrinsic capability deficits. In this paper, we first demonstrate that LiC is not a failure of model capability but rather a breakdown in interaction between users and LLMs. We theoretically show that scaling model size or improving training alone cannot resolve this gap, as it arises from structural ambiguity in conversational context rather than representational limitations. To address this, we propose to decouple intent understanding from task execution through a Mediator-Assistant architecture. By utilizing an experience-driven Mediator to explicate user inputs into explicit, well-structured instructions based on historical interaction patterns, our approach effectively bridges the gap between vague user intent and model interpretation. Experimental results demonstrate that this method significantly mitigates performance degradation in multi-turn conversations across diverse LLMs.

Grasping under limited sensing remains a fundamental challenge for real-world robotic manipulation, as vision and high-resolution tactile sensors often introduce cost, fragility, and integration complexity. This work demonstrates that reliable multifingered grasping can be achieved under extremely minimal sensing by relying solely on uniaxial fingertip force feedback and joint proprioception, without vision or multi-axis/tactile sensing. To enable such blind grasping, we employ an efficient teacher-student training pipeline in which a reinforcement-learned teacher exploits privileged simulation-only observations to generate demonstrations for distilling a transformer-based student policy operating under partial observation. The student policy is trained to act using only sensing modalities available at real-world deployment. We validate the proposed approach on real hardware across 18 objects, including both in-distribution and out-of-distribution cases, achieving a 98.3~$\%$ overall grasp success rate. These results demonstrate strong robustness and generalization beyond the simulation training distribution, while significantly reducing sensing requirements for real-world grasping systems.

Flat regions of the neural network loss landscape have long been hypothesized to correlate with better generalization properties. A closely related but distinct problem is training models that are robust to internal perturbations to their weights, which may be an important need for future low-power hardware platforms. In this paper, we explore the usage of two methods, sharpness-aware minimization (SAM) and random-weight perturbation (RWP), to find minima robust to a variety of random corruptions to weights. We consider the problem from two angles: generalization (how do we reduce the noise-robust generalization gap) and optimization (how do we maximize performance from optimizers when subject to strong perturbations). First, we establish, both theoretically and empirically, that an over-regularized RWP training objective is optimal for noise-robust generalization. For small-magnitude noise, we find that SAM's adversarial objective further improves performance over any RWP configuration, but performs poorly for large-magnitude noise. We link the cause of this to a vanishing-gradient effect, caused by unevenness in the loss landscape, affecting both SAM and RWP. Lastly, we demonstrate that dynamically adjusting the perturbation strength to match the evolution of the loss landscape improves optimizing for these perturbed objectives.

Sparse autoencoders (SAEs) offer a natural path toward comparable explanations across different representation spaces. However, current SAEs are trained per modality, producing dictionaries whose features are not directly understandable and whose explanations do not transfer across domains. In this study, we introduce LUCID (Learning Unified vision-language sparse Codes for Interpretable concept Discovery), a unified vision-language sparse autoencoder that learns a shared latent dictionary for image patch and text token representations, while reserving private capacity for modality-specific details. We achieve feature alignment by coupling the shared codes with a learned optimal transport matching objective without the need of labeling. LUCID yields interpretable shared features that support patch-level grounding, establish cross-modal neuron correspondence, and enhance robustness against the concept clustering problem in similarity-based evaluation. Leveraging the alignment properties, we develop an automated dictionary interpretation pipeline based on term clustering without manual observations. Our analysis reveals that LUCID's shared features capture diverse semantic categories beyond objects, including actions, attributes, and abstract concepts, demonstrating a comprehensive approach to interpretable multimodal representations.

Current analysis of additive manufactured niobium-based copper alloys relies on hand annotation due to varying contrast, noise, and image artifacts present in micrographs, slowing iteration speed in alloy development. We present a filtering and segmentation algorithm for detecting precipitates in FIB cross-section micrographs, optimized using linear genetic programming (LGP), which accounts for the various artifacts. To this end, the optimization environment uses a domain-specific language for image processing to iterate on solutions. Programs in this language are a list of image-filtering blocks with tunable parameters that sequentially process an input image, allowing for reliable generation and mutation by a genetic algorithm. Our environment produces optimized human-interpretable MATLAB code representing an image filtering pipeline. Under ideal conditions--a population size of 60 and a maximum program length of 5 blocks--our system was able to find a near-human accuracy solution with an average evaluation error of 1.8% when comparing segmentations pixel-by-pixel to a human baseline using an XOR error evaluation. Our automation work enabled faster iteration cycles and furthered exploration of the material composition and processing space: our optimized pipeline algorithm processes a 3.6 megapixel image in about 2 seconds on average. This ultimately enables convergence on strong, low-activation, precipitation hardened copper alloys for additive manufactured fusion reactor parts.

The ICAP framework defines four cognitive engagement levels: Passive, Active, Constructive, and Interactive, where increased cognitive engagement can yield improved learning. However, personalizing learning activities that elicit the optimal level of cognitive engagement remains a key challenge in intelligent tutoring systems (ITS). In this work, we develop and evaluate a system that adaptively scaffolds cognitive engagement by dynamically selecting worked examples in two different ICAP modes: (active) Guided examples and (constructive) Buggy examples. We compare Bayesian Knowledge Tracing (BKT) and Deep Reinforcement Learning (DRL) as adaptive methods against a non-adaptive baseline method for selecting example type in a logic ITS. Our experiment with 113 students demonstrates that both adaptive policies significantly improved student performance on test problems. BKT yielded the largest improvement in posttest scores for low prior knowledge students, helping them catch up with their high prior knowledge peers, whereas DRL yielded significantly higher posttest scores among high prior knowledge students. This paper contributes new insights into the complex interactions of cognitive engagement and adaptivity and their results on learning outcomes.

Diabetic Retinopathy (DR) is an eye disease which arises due to diabetes mellitus. It might cause vision loss and blindness. To prevent irreversible vision loss, early detection through systematic screening is crucial. Although researchers have developed numerous automated deep learning-based algorithms for DR screening, their clinical applicability remains limited, particularly in lesion segmentation. Our method provides pixel-level annotations for lesions, which practically supports Ophthalmologist to screen DR from fundus images. In this work, we segmented four types of DR-related lesions: microaneurysms, soft exudates, hard exudates, and hemorrhages on 757 images from DDR dataset. To enhance lesion segmentation, an attention mechanism was integrated with DeepLab-V3+. Compared to the baseline model, the Attention-DeepLab model increases mean average precision (mAP) from 0.3010 to 0.3326 and the mean Intersection over Union (IoU) from 0.1791 to 0.1928. The model also increased microaneurysm detection from 0.0205 to 0.0763, a clinically significant improvement. The detection of microaneurysms is the earliest visible symptom of DR.

Vertical Federated Learning (VFL) enables collaborative analysis across parties holding complementary feature views of the same samples, yet existing approaches are largely restricted to distributed variants of $k$-means, requiring centralized coordination or the exchange of feature-dependent numerical statistics, and exhibiting limited robustness under heterogeneous views or adversarial behavior. We introduce VertCoHiRF, a fully decentralized framework for vertical federated clustering based on structural consensus across heterogeneous views, allowing each agent to apply a base clustering method adapted to its local feature space in a peer-to-peer manner. Rather than exchanging feature-dependent statistics or relying on noise injection for privacy, agents cluster their local views independently and reconcile their proposals through identifier-level consensus. Consensus is achieved via decentralized ordinal ranking to select representative medoids, progressively inducing a shared hierarchical clustering across agents. Communication is limited to sample identifiers, cluster labels, and ordinal rankings, providing privacy by design while supporting overlapping feature partitions and heterogeneous local clustering methods, and yielding an interpretable shared Cluster Fusion Hierarchy (CFH) that captures cross-view agreement at multiple resolutions.We analyze communication complexity and robustness, and experiments demonstrate competitive clustering performance in vertical federated settings.

Oversmoothing is a fundamental limitation of deep graph convolutional networks (GCNs), causing node representations to collapse as depth increases. While many prior approaches mitigate this effect through architectural modifications or residual mechanisms, the underlying spectral cause of oversmoothing is often left implicit. We propose Laplacian-LoRA, a simple and interpretable low-rank spectral adaptation of standard GCNs. Rather than redesigning message passing, Laplacian-LoRA introduces a learnable, spectrally anchored correction to the fixed Laplacian propagation operator, selectively weakening contraction while preserving stability and the low-pass inductive bias. Across multiple benchmark datasets and depths, Laplacian-LoRA consistently delays the onset of oversmoothing, extending the effective depth of GCNs by up to a factor of two. Embedding variance diagnostics confirm that these gains arise from delayed representational collapse, while learned spectral analysis demonstrates that the correction is smooth, bounded, and well behaved. Our results show that oversmoothing is a depth-dependent spectral phenomenon that can be systematically delayed through modest, low-rank adaptation of the graph propagation operator.

World models enable agents to plan by imagining future states, but existing approaches operate from a single viewpoint, typically egocentric, even when other perspectives would make planning easier; navigation, for instance, benefits from a bird's-eye view. We introduce Cross-View World Models (XVWM), trained with a cross-view prediction objective: given a sequence of frames from one viewpoint, predict the future state from the same or a different viewpoint after an action is taken. Enforcing cross-view consistency acts as geometric regularization: because the input and output views may share little or no visual overlap, to predict across viewpoints, the model must learn view-invariant representations of the environment's 3D structure. We train on synchronized multi-view gameplay data from Aimlabs, an aim-training platform providing precisely aligned multi-camera recordings with high-frequency action labels. The resulting model gives agents parallel imagination streams across viewpoints, enabling planning in whichever frame of reference best suits the task while executing from the egocentric view. Our results show that multi-view consistency provides a strong learning signal for spatially grounded representations. Finally, predicting the consequences of one's actions from another viewpoint may offer a foundation for perspective-taking in multi-agent settings.

Activation steering has emerged as a promising approach for efficiently adapting large language models (LLMs) to downstream behaviors. However, most existing steering methods rely on a single static direction per task or concept, making them inflexible under task variation and inadequate for complex tasks that require multiple coordinated capabilities. To address this limitation, we propose STEER2ADAPT, a lightweight framework that adapts LLMs by composing steering vectors rather than learning new ones from scratch. In many domains (e.g., reasoning or safety), tasks share a small set of underlying concept dimensions. STEER2ADAPT captures these dimensions as a reusable, low-dimensional semantic prior subspace, and adapts to new tasks by dynamically discovering a linear combination of basis vectors from only a handful of examples. Experiments across 9 tasks and 3 models in both reasoning and safety domains demonstrate the effectiveness of STEER2ADAPT, achieving an average improvement of 8.2%. Extensive analyses further show that STEER2ADAPT is a data-efficient, stable, and transparent inference-time adaptation method for LLMs.

Executing complex terminal tasks remains a significant challenge for open-weight LLMs, constrained by two fundamental limitations. First, high-fidelity, executable training environments are scarce: environments synthesized from real-world repositories are not diverse and scalable, while trajectories synthesized by LLMs suffer from hallucinations. Second, standard instruction tuning uses expert trajectories that rarely exhibit simple mistakes common to smaller models. This creates a distributional mismatch, leaving student models ill-equipped to recover from their own runtime failures. To bridge these gaps, we introduce TermiGen, an end-to-end pipeline for synthesizing verifiable environments and resilient expert trajectories. Termi-Gen first generates functionally valid tasks and Docker containers via an iterative multi-agent refinement loop. Subsequently, we employ a Generator-Critic protocol that actively injects errors during trajectory collection, synthesizing data rich in error-correction cycles. Fine-tuned on this TermiGen-generated dataset, our TermiGen-Qwen2.5-Coder-32B achieves a 31.3% pass rate on TerminalBench. This establishes a new open-weights state-of-the-art, outperforming existing baselines and notably surpassing capable proprietary models such as o4-mini. Dataset is avaiable at https://github.com/ucsb-mlsec/terminal-bench-env.

Immersive applications such as virtual and augmented reality impose stringent requirements on frame rate, latency, and synchronization between physical and virtual environments. To meet these requirements, an edge server must render panoramic content, predict user head motion, and transmit a portion of the scene that is large enough to cover the user viewport while remaining within wireless bandwidth constraints. Each portion produces two feedback signals: prediction feedback, indicating whether the selected portion covers the actual viewport, and transmission feedback, indicating whether the corresponding packets are successfully delivered. Prior work models this problem as a multi-armed bandit with two-level bandit feedback, but fails to exploit the fact that prediction feedback can be retrospectively computed for all candidate portions once the user head pose is observed. As a result, prediction feedback constitutes full-information feedback rather than bandit feedback. Motivated by this observation, we introduce a two-level hybrid feedback model that combines full-information and bandit feedback, and formulate the portion selection problem as an online learning task under this setting. We derive an instance-dependent regret lower bound for the hybrid feedback model and propose AdaPort, a hybrid learning algorithm that leverages both feedback types to improve learning efficiency. We further establish an instance-dependent regret upper bound that matches the lower bound asymptotically, and demonstrate through real-world trace driven simulations that AdaPort consistently outperforms state-of-the-art baseline methods.

Evaluating the real-world capabilities of AI systems requires grounding benchmark performance in human-interpretable measures of task difficulty. Existing approaches that rely on direct human task completion time annotations are costly, noisy, and difficult to scale across benchmarks. In this work, we propose BRIDGE, a unified psychometric framework that learns the latent difficulty scale from model responses and anchors it to human task completion time. Using a two-parameter logistic Item Response Theory model, we jointly estimate latent task difficulty and model capability from model performance data across multiple benchmarks. We demonstrate that latent task difficulty varies linearly with the logarithm of human completion time, allowing human task completion time to be inferred for new benchmarks from model performance alone. Leveraging this alignment, we forecast frontier model capabilities in terms of human task length and independently reproduce METR's exponential scaling results, with the 50% solvable task horizon doubling approximately every 6 months.