Noticias de IA - ítems

arXiv:2610.35113v1 Announce Type: cross Abstract: Learning rate decay (LRD) is critical for training high-performance models, yet most designs remain empirically driven. Although the Universal Budget Aware (UBA) LRD offers a theoretical framework, its reliance on SGD dynamics limits its applicability to modern Adam-type optimizers and leaves its key `optimization difficulty parameter' opaque. To bridge this gap, we propose the Universal Training Adaptive (UTA) LRD. Unlike UBA's fixed offline trajectory anchored solely to iterations, UTA treats LRD as a closed-loop shape-and-scale problem. By continuous-time formulation, UTA is training-state adaptive with segment-wise optimal shape. This adaptivity links external iterations with internal training states, enabling a near-optimal path in both progression shape and learning rate (LR) scale, a property unattainable by other LRDs, including UBA. Empirically, we demonstrate consistent improvements across language models (LLaMA 60M to 1B) and vision models (ResNet50, ViT-B) on ImageNet with plug-and-play ease and negligible overhead. Extensive ablation studies further confirm UTA's robustness across dimensions, including peak LR, final decay ratio, its parameter $\varphi_0$, component efficacy and generalization to fine-tuning scenarios, underscoring its practical implementability.

arXiv:2610.35112v1 Announce Type: cross Abstract: Grokking describes understanding something deeply and intuitively. In neural networks, it is the phenomenon where a network generalizes long after memorizing its training data. This delay has been linked to a fall in the effective rank of the network's weight matrices, but whether this rank compression drives grokking or follows it is still unresolved. We study a mean-field MLP and a one-layer transformer trained on modular arithmetic and observe that: (i) Compression follows the onset, not memorization. Every transformer memorizes its training data within 250 steps, while the grokking onset (the first step at which test accuracy reaches 95%) varies from 6,500 to 21,000 steps. The steepest fall in effective rank comes about 900 steps after the onset, however late the onset occurs, and networks that memorize without grokking compress far less. (ii) Subspace crystallization. We call the span of the fewest top singular directions of a weight matrix that preserve 99% of test accuracy its sufficient subspace. Its size can shrink as the margin between the correct and the best wrong output grows, so we also count the directions needed to keep the outputs within 10%. In the transformer, this count falls during the transition from a quarter or more of the available directions to about a tenth, which we call subspace crystallization. (iii) Imposing low rank does not speed up grokking. Against a factorized full-rank control, capping the rank of the transformer's MLP input weights delays the onset, and capping the rank of its embedding stops 15 of 20 runs from grokking. Compression is therefore a signature of grokking rather than its cause.

arXiv:2610.35082v1 Announce Type: cross Abstract: Agent Skills allow LLM agents to combine reusable capabilities, but checking each Skill in isolation cannot rule out policy violations when they are composed. We term such failures Conflict of Skills (CoS): artifacts or assessment signals produced for one purpose may be reused downstream without preserving the restrictions or authorization needed for their new use. Using controlled executions of publicly available Skills and matched single-Skill controls, we identify four recurring mechanisms behind CoS. Building on these findings, we introduce the Conflict of Skills Attack (CoSA), which adds locally admitted Skills to an existing pool to supply missing business operations (capability supplementation) or reshape artifact exchange, decision reuse, and target/action bindings (relationship reshaping). CoSA leaves existing Skills and runtime permissions unchanged. To evaluate this threat under autonomous Skill selection, we introduce CoSA-Bench, built from real-world Skills and controlled services, with 288 tasks across recruiting, finance, IT, and business, covering data leakage, unauthorized state modification, and unsafe execution. Across five LLMs and three agent harnesses, additions capped at 5% of each original pool raise the policy-violation rate from 35.02% to 58.45%, an increase of 23.43 percentage points. Trace analysis links compliant-to-violating transitions to cross-Skill dependencies involving the additions. Finally, we develop a lightweight harness-level guard that checks information flow and authorization before consequential actions. On OpenClaw recruiting tasks across five models, the guard reduces violations from 65.28% to 6.67%, while policy-compliant task completion on matched benign controls changes from 66.67% to 65.00%.

arXiv:2610.35070v1 Announce Type: cross Abstract: Connected and automated vehicles could allow reassigning lane directions and speed limits on every road frequently, but the benefit hinges on how traffic reroutes, and one-way roads can isolate junctions. We formulate network-wide dynamic lane allocation and directional speed limit control with one-way roads, lane clearing before each reversal, and a navigable graph kept strongly connected at all times. An action-resolution algorithm grounded in graph orientation theory admits a proposal closing a direction only if the network stays strongly connected and otherwise relaxes or reverts it, keeping all junctions mutually reachable within the model, even during clearing. All road controllers share one transformer whose attention, Global Endpoint Relation-Aware Spatio-Temporal Self-Attention (GERASTSA), spans all road-time tokens, biased by the endpoint relations between roads. It is pretrained on action-conditioned traffic forecasting and fine-tuned with per-road proximal policy optimization under a mixed reward that also credits neighboring roads and the network. On a 6 x 6 grid with commuter rush hours, the learned policy makes about two thirds of roads one-way in the peak direction and cuts daily delay by 90% against the initial configuration held fixed and by 85% against a tuned two-way load balancer. Restricted to two-way roads at test time, it gains only about as much as this balancer, whereas its learned speed limits add delay. Pretraining roughly halves the updates needed to surpass the fixed configuration. Of the ablations, only removing the relational bias or the spatial embedding clearly matters, multiplying delay about fivefold.

arXiv:2610.35045v1 Announce Type: cross Abstract: Dense retrievers are commonly trained with hard negatives selected from lexical or dense retrieval results. Retrieval proximity, however, does not indicate whether a candidate conflicts with explicit domain structure. We introduce CA-HNM, a constraint-aware hard-negative mining framework that combines retrieval-based candidate generation with deterministic incompatibility rules derived from a structured domain model. The framework removes known positives, constructs a local typed context around the query target, records relational, scope, target, and level mismatches, and uses the resulting candidates to fine-tune a standard dual encoder. Structured information is required only during offline negative construction. We evaluate CA-HNM on MOOCCubeX and Course-Skill Atlas with two encoders, two training losses, and ten seeds. Relative to a retrieval-rank-matched control, CA-HNM-Pure has higher mean NDCG@10 in six of eight dataset--encoder--loss configurations. Four hierarchical 95% bootstrap intervals are entirely above zero, mostly under triplet loss. In a separate experiment, CA-HNM-Mixed shows slightly higher mean NDCG@10 than a matched-mixed control on both collections. Structural variants show that changes in available structure can also change candidate eligibility and negative-pool composition. The results indicate that the contribution of constraint-aware selection depends on the training objective and the composition of the negative pool.

arXiv:2610.35032v1 Announce Type: cross Abstract: Self-supervised learning has emerged as a powerful paradigm for Path Representation Learning (PRL) in road networks. PRL aims to learn informative representations of paths traversed by moving objects, such as vehicles, where each path consists of a sequence of connected road segments. These representations are essential for various downstream tasks, including travel-time estimation and path ranking. However, existing methods often suffer from ineffective view generation and inefficient training procedures due to their reliance on heuristic augmentations and computationally intensive contrastive learning. To address these issues, we propose AutoPRL, an effective and efficient self-supervised PRL framework that integrates three key components. The Automatic Data Augmentation (AutoDA) module employs variational autoencoders to generate semantically meaningful and diverse feature-level path views by modeling each topology-aware node embedding with a Gaussian latent distribution, thereby mitigating information loss and enhancing representation quality. Second, the Cross-View Decoupled Contrastive (CVDC) learning mechanism improves training efficiency by disentangling the optimization of positive and negative pairs across node- and path-level views, enabling more stable contrastive learning. Third, an Adaptive Co-Training (ACT) module dynamically adjusts uncertainty-based loss weights to balance multiple objectives, ensuring efficient and scalable training. Extensive experiments on three real-world datasets demonstrate that AutoPRL achieves strong and consistent performance in both effectiveness and efficiency across multiple downstream tasks.

arXiv:2610.35016v1 Announce Type: cross Abstract: Characterizing nitrogen-vacancy (NV) ensembles requires connecting their collective spectra to individual centers that experience different local environments. We develop a symbolic-regression pipeline that carries single-center spin structure into a calibrated ensemble model by generating individual spectral responses and combining them with explicit weights. A constrained search selects Hermitian operator terms from single-center optically detected magnetic resonance spectra, fits their coefficients, and retains the selected expressions as source models for the ensemble. Ensemble calibration then accounts for resonance shifts, local disorder, orientation, and optical response while keeping the individual spectral contributions accessible. On public single-center and ensemble data, the zero-field ensemble held-frequency root-mean-square error decreases from 0.008835 to 0.003872 in normalized photoluminescence, a 56.2% improvement over matched uncorrected composition. In controlled QuTiP ensembles of 128 sites spanning four NV orientations, models calibrated at 0 and 3 mT predict an independently held-out 5.17 mT spectrum with errors of 0.000207-0.000228 across three random seeds, comparable to the 0.000200 noise standard deviation. Resolving the simulated response into its individual-center contributions shows how local spectral diversity combines into the ensemble line shape. By retaining the source Hamiltonians while calibrating sample-dependent spectral differences, the pipeline connects local spin physics to collective optical readout and makes the contributions to an ensemble line shape available for physical interpretation.

arXiv:2610.34987v1 Announce Type: cross Abstract: GLP-1 medications are increasingly used for weight management, and many people share their experiences online. These narratives show why people take the drugs and what they experience, but they remain unstructured and disconnected from formal pharmacovigilance. We therefore present AskGLP1, a system that transforms online patient narratives into a knowledge graph (KG) and answers questions about medication experiences. Through an LLM-assisted pipeline grounded in the MedDRA terminology, AskGLP1 screens nearly 8.5 million Reddit posts for medication experiences and distills 163,135 of them, with 2,506 WebMD reviews, into a KG of 385,790 drug-side-effect and drug-purpose relations. We find that GLP-1 experiences are source- and drug-specific. Online narratives share the most common gastrointestinal symptoms with FAERS, but they foreground everyday burdens such as fatigue and food aversion, whereas FAERS reports more often include injection-site reactions and selected serious diagnoses such as pancreatitis. We further find that posts report less nausea and vomiting for tirzepatide than for semaglutide, more fatigue for Wegovy than Ozempic and for Zepbound than Mounjaro. Evaluations show that the agent answers with verifiable evidence and declines requests for medical advice. Our findings reveal gaps between what FAERS captures and what people report online, and they call for patient-centered approaches to pharmacovigilance. Code, prompts, and evaluation questions are available at https://github.com/ZhaoqianXue/AskGLP1.

arXiv:2610.34976v1 Announce Type: cross Abstract: Camera-based bird's-eye view (BEV) perception methods such as Fast-BEV have demonstrated promising results with fast inference speed and competitive accuracy. However, their robustness to real-world corruptions remains insufficiently explored, posing risks for autonomous driving deployment where environmental variations are inevitable. In this paper, we propose Sparse Recovery Bird's-Eye View Model Soup (SR-BEVS), a sparse recovery learning approach that simultaneously improves clean accuracy and corruption robustness without requiring additional data or architectural modifications. Our method identifies pruning candidates based on layer-wise weight magnitudes, applies random pruning to create diverse sparse models, and retrains each model with a randomly sampled learning rate. By averaging these diverse recovered models into a single dense model, we achieve enhanced robustness against various corruptions while introducing zero inference overhead. We further provide a mode connectivity analysis demonstrating that our pruning-based diversification preserves the connected low-loss basin necessary for stable weight averaging. Experiments on the nuScenes dataset demonstrate up to 33.8% improvement in corruption robustness and a 1.8% gain in clean accuracy over the Fast-BEV baseline, with consistent improvements across sensor noise, weather distortions, and illumination changes. Our simple yet effective approach may be applicable to other BEV models.

arXiv:2610.34971v1 Announce Type: cross Abstract: Pretrained transformers spend a large share of their inference compute inside self-attention, yet the per-head query-key interaction is structurally low-rank by construction. We ask a diagnostic question: across layers and architectures, which heads actually use the full rank budget on real data, and which do not? We propose SVD-Attention, a post-hoc method that decomposes the per-head interaction matrix $M_h$ and scores every singular mode by both its spectral magnitude and its activation variance across tokens on a small calibration set. The score keeps the modes the head can amplify and that real data actually activates, and discards the rest; the resulting factorization is exact for the retained modes and needs no finetuning. Across seven pretrained models and five task types, the diagnostic finds a task-dependent pattern: redundancy sits in the early to middle layers for sentiment and image classification and for generation, and in the late layers for dense prediction and extractive question answering. It is also predictive: truncating the layers it flags costs at most $1.75$ F1, while truncating the same number of layers elsewhere in the same model costs up to $50$ F1. SVD-Attention is an analysis tool rather than a compression method: acting on it reduces inference FLOPs by $0.8$-$1.5\%$ on its own and by $8$-$11\%$ when composed with weight pruning, with $0$-$4\%$ change in task quality.