Noticias de IA - ítems

arXiv:2604.13226v3 Announce Type: replace-cross Abstract: Large Language Models (LLMs) rely heavily on Key-Value (KV) caching to minimize inference latency. However, standard KV caches are context-dependent: reusing a cached document in a new context requires recomputing KV states to account for shifts in attention distribution. Existing solutions such as CacheBlend, EPIC, and SAM-KV mitigate this issue by selectively recomputing a subset of tokens; however, they still incur non-negligible computational overhead (FLOPs) and increased Time-to-First-Token (TTFT) latency. In this paper, we propose KV Packet, a recomputation-free cache reuse framework that treats cached documents as immutable "packets" wrapped in light-weight trainable soft-token adapters, which are trained via self-supervised distillation to bridge context discontinuities. Experiments on Llama-3.1 and Qwen-3 demonstrate that the proposed KV Packet method achieves near-zero FLOPs and lower TTFT than recomputation-based baselines, while retaining F1 scores comparable to those of the full recomputation baseline. Code and experimental results are available at https://github.com/ChuangtaoChen-TUM/KVPacket.

arXiv:2604.13082v3 Announce Type: replace-cross Abstract: Large language models today still make mistakes on multi-digit arithmetic when they do not have tools, and part of the reason is how numbers are split into tokens. Arithmetic is also a setting where the correct answer to every input and the algorithm that produces it are already known, so training transformers on it lets us study what they learn against a ground truth. These models often fit their training set long before they generalize to held-out inputs, and this delay, known as grokking, decreases as the training set grows. But, how does the representation of numbers change the amount of data a transformer needs to generalize? We study one-step Collatz prediction, in which a transformer reads the digits of an integer in a chosen base and outputs the digits of the next term in its Collatz sequence. We find that the long delay appears only when the training set has few different integers, and that the numeral base determines how many integers are sufficient. In bases where each output digit depends only on a few neighbouring input digits, because the carry between positions depends only on the current digit, a transformer needs several times fewer training integers. We predicted this ordering before training, and input length does not explain it. The same carry structure predicts which bases are easy for a second Collatz map, for integer division, for a decoder-only transformer, and across three model sizes. Finally, a linear probe recovers the low-order residues of the input as accurately from a randomly initialized encoder as from a trained one. Thus, high probe accuracy on such features does not reliably show that a model has learned to compute them. More broadly, we show that the choice of numeral representation affects sample efficiency in ways we can predict from the structure of the task before training.

arXiv:2604.11674v3 Announce Type: replace-cross Abstract: Many everyday robot manipulation skills are affordance-dependent, with success determined by whether the robot contacts the functional object region required by the subsequent action. Current simulation data generators obtain contacts from generic grasp estimators or per-object manual contact annotations, but generic estimators rank stable grasps without task semantics and often select contacts that are misaligned with the downstream action, while manual contact annotations must be rewritten for each new object and task. To solve these challenges, we introduce AffordSim, a scalable data generator and benchmark that integrates open-vocabulary 3D affordance prediction into simulation-based trajectory generation. Given a natural-language task description, AffordSim synthesizes a task-relevant scene, emits affordance queries, grounds them on object surfaces, samples region-conditioned grasps, and selects executable candidates with motion planning. It further randomizes object pose, texture, lighting, image noise, and cross-viewpoint backgrounds for sim-to-real transfer. We instantiate AffordSim as a 50-task benchmark across diverse manipulation skills, five robot embodiments, and 500+ rigid and articulated objects. AffordSim achieves 93% of the trajectory collection success rate of manual contact annotations on affordance-critical tasks and 89% on hard composite tasks. Vision-language-action policies trained on AffordSim data transfer zero-shot to a real Franka FR3, reaching 24% average success.

arXiv:2604.08552v5 Announce Type: replace-cross Abstract: Descriptive scientific metadata in public repositories are often incomplete and inconsistent with community standards and ontologies, limiting data FAIRness. Large language models (LLM) offer a promising approach to automatically standardizing such metadata when the LLM is provided with knowledge of the relevant standards in machine-actionable form. Entry of metadata standards through prompt engineering offers only fixed snapshots of the relevant standards, however, and it relies on an LLM's pretrained knowledge to interpret and satisfy constraints related to the standard. We evaluate whether giving an LLM access to metadata specifications and authoritative terminology at runtime improves automated metadata standardization. We present ARMS, a tool-using LLM agent that retrieves complete CEDAR metadata templates and that queries authoritative biomedical terminology services dynamically at execution time. We compared ARMS with a prompt-engineering approach on 839 legacy metadata records describing experiment datasets with expert-standardized records serving as the reference standard. Across the corpus, ARMS outperformed the prompt-engineering approach in converting legacy metadata to standards-compliant form, increasing precision from 0.56 to 0.93 and recall from 0.51 to 0.85, with improvements across all field categories and assay types. The largest gains occurred for ontology-constrained fields where precision increased from 0.36 to 0.92. Our findings show that access to appropriate metadata standards can substantially improve LLM-based standardization of legacy metadata. ARMS improves metadata standardization by giving LLMs runtime access to the authoritative resources that define valid metadata. Representation of metadata standards in machine-actionable form enhances the task of rectifying legacy metadata using LLM technology, especially when those standards can be queried dynamically.

arXiv:2604.08362v3 Announce Type: replace-cross Abstract: The emergence of Large Language Models (LLMs) has illuminated the potential for a general-purpose user simulator. However, existing benchmarks remain constrained to isolated scenarios, narrow action spaces, or synthetic data, failing to capture the holistic nature of authentic human behavior. To bridge this gap, we introduce OmniBehavior, the first predictive user simulation benchmark constructed entirely from real-world data, integrating long-horizon, cross-scenario, and heterogeneous behavioral patterns into a unified framework. Based on this benchmark, we first provide empirical evidence that previous datasets with isolated scenarios suffer from tunnel vision, whereas real-world decision-making relies on long-term, cross-scenario causal chains. Extensive evaluations of state-of-the-art LLMs reveal that current models struggle to accurately predict these complex behaviors, with performance plateauing even as context windows expand. Crucially, a systematic comparison between simulated and authentic behaviors uncovers a fundamental structural bias: LLMs tend to converge towards a positive average person, exhibiting hyper-activity, persona homogenization, and a utopian bias. This results in the loss of individual differences and long-tail behaviors, highlighting critical directions for future high-fidelity simulation research.

arXiv:2604.00004v3 Announce Type: replace-cross Abstract: The extension of context windows in Large Language Models is typically facilitated by scaling positional encodings followed by Continued Pre-Training (CPT). While effective, this paradigm is notoriously data-hungry and computationally expensive, requiring massive long-text corpora to recalibrate the model to the shifted positional distribution. We propose LinearARD, a restoration distillation method that restores Rotary Position Embedding (RoPE)-scaled students through attention-structure consistency with a frozen native-RoPE teacher. Rather than next-token prediction or opaque hidden-state matching, LinearARD aligns row-wise distributions of dense $Q/Q$, $K/K$, and $V/V$ self-relation matrices from the final attention layer to directly supervise attention dynamics. To remove the quadratic memory bottleneck of $n \times n$ relation maps, we introduce a linear-activation-memory kernel that stores only per-token log-sum-exp statistics and recomputes logits in the backward pass to obtain exact Kullback-Leibler divergence gradients. Across LLaMA2-7B, LLaMA3-8B, and Mistral-7B-v0.1 extended to 32K context, LinearARD recovers 93.1%/94.2%/94.3% of native short-context performance and achieves strong long-context robustness on RULER using only 4.25M tokens---amounting to just 1.6% (a $\sim$60$\times$ reduction) of the 256M-token budget required by state-of-the-art baselines. Under this severely constrained budget, CPT and LongReD remain near zero on RULER, demonstrating that relation-level restoration provides a substantially better efficiency-quality tradeoff. Our code is available at https://github.com/gracefulning/LinearARD.

arXiv:2603.28963v3 Announce Type: replace-cross Abstract: Simulation with realistic traffic agents is essential for validating autonomous driving systems. Existing data-driven simulators learn agent behavior from higher-level abstractions such as 3D bounding boxes and polylines, inferred by upstream perception pipelines. These lossy abstractions discard sensory context that directly shapes agent behavior, limiting the distributional realism that simulation aims to reproduce. To address this limitation, we propose AutoWorld, a traffic simulation framework that grounds agent behavior in raw sensor observations through a self-supervised world model trained on LiDAR occupancy data. Given world model samples, AutoWorld constructs a coarse-to-fine predictive scene context as input to a multi-agent motion generation model. Furthermore, we designed a motion-aware latent supervision objective that enriches AutoWorld's latent representation of scene dynamics. To better exploit this latent space during inference, AutoWorld employs a cascaded Determinantal Point Process framework to guide diversity-aware sampling across both the world model and motion model. Experiments on nuPlan and Waymo demonstrate that AutoWorld improves generation realism, with larger gains in partially observed scenarios where trajectory abstractions are most limited, reducing collision rates by 36%. AutoWorld also remains physically plausible under planner-controlled ego behavior and supports controllable, diverse safety-critical scenario generation.

arXiv:2603.25924v2 Announce Type: replace-cross Abstract: Multimodal AI systems are evaluated by downstream task accuracy, but high accuracy does not mean the underlying data is coherent. A model can score well on Visual Question Answering (VQA) while its inputs contradict each other. We introduce the Multimodal Coherence Score (MCS), a metric that evaluates fusion quality independent of any downstream model. MCS decomposes coherence into four dimensions, identity, spatial, semantic, and decision, with weights learned via Nelder-Mead optimization. We evaluate on 1,000 Visual Genome images using DETR, CLIP, and ViLT, and validate on 150 COCO images with no retraining. Across three fusion architectures, MCS discriminates quality with higher sensitivity than task accuracy alone (Spearman rho = 0.093 vs. 0.071). Perturbation experiments confirm each dimension responds independently to its failure mode with zero cross-talk. MCS is lightweight, requires no human annotation, and tells you not just that something broke, but what broke.

arXiv:2603.18773v2 Announce Type: replace-cross Abstract: LLM post-training pipelines that combine supervised fine-tuning and reinforcement learning are difficult to configure under realistic compute budgets: the configuration space is high-dimensional and heterogeneous, stages are strongly coupled, and each end-to-end evaluation is expensive. We propose AutoPipe, a budget-aware two-phase framework for configuration selection in LLM post-training. Offline, AutoPipe learns a dataset-conditioned learning-to-rank surrogate from historical runs, capturing within-dataset preferences and providing transferable guidance toward promising regions of the configuration space. Online, for a new dataset, AutoPipe uses the offline guidance to steer Bayesian optimization and models dataset-specific deviations with a Gaussian-process residual surrogate. To reduce evaluation cost, each trial is early-stopped and scored by a learned predictor that maps early training signals to a low-cost proxy for final post-training performance. Experiments on biomedical reasoning tasks show that AutoPipe achieves performance comparable to full-fidelity HPO baselines at a fraction of their online cost, while outperforming offline-only approaches and early-stopped HPO variants.

arXiv:2603.18558v3 Announce Type: replace-cross Abstract: Long-form video question answering requires reasoning over extended temporal contexts, making frame selection a critical bottleneck for multi-modal large language models (MLLMs) bound by finite context windows. Within the controlled frame-budget regime that governs practical deployment, prior selectors score frames against a single global query embedding; as a result, compositional multimodal questions that involve temporal ordering or cross-modal cues such as "what happens on screen right after the narrator mentions the reaction?" are flattened into a representation that loses sub-event ordering and modality bindings. We introduce HiMu, a training-free framework for compositional multimodal frame selection. A single text-only LLM call decomposes the query into a hierarchical logic tree whose leaves are atomic predicates, each routed to a lightweight expert spanning vision (CLIP, open-vocabulary detection, OCR) and audio (speech recognition and non-speech sound matching). Expert signals are normalized, smoothed to align across modalities, and composed bottom-up through fuzzy-logic operators that enforce temporal sequencing and adjacency, yielding a continuous per-frame satisfaction curve. Under the standard 16-frame budget on Video-MME, LongVideoBench, and HERBench-Lite, HiMu achieves state-of-the-art accuracy among frame selection methods and improves over uniform sampling across seven diverse MLLMs as a drop-in module, matching the accuracy of uniform sampling at 4x its frame budget, without retraining and without multiple iterative MLLM calls during selection.