Papers

Learning long-range dependencies with recurrent neural networks (RNNs) is difficult because of the vanishing gradient problem. Gated recurrent neural networks like GRUs and LSTMs address this problem and are commonly explained by their ability to create additive copy paths through time. This explanation is correct, but incomplete: useful RNN gate values must themselves be learned, which is typically done with another RNN. We study this gate-learning problem through carefully controlled sequence classification tasks. Our analysis shows that gated architectures struggle when all training samples require long-range memory and thus equally suffer from the vanishing gradient problem. We intervene by adding short-dependency samples which let gated models learn solutions that capture long-dependency samples, but also generalize across different class labels. When properly learned, gated RNNs match or even exceed the test accuracy of modern architectures like Mamba and xLSTM. Overall, our results highlight gate learning as a central part of learning long-range memory and show how training data can support this process by exposing the same task over shorter dependencies.

Tensor parallelism ends every layer with an all-reduce that sits between two dependent matrix multiplications, so its cost cannot be hidden and grows as compute outpaces interconnect bandwidth. Sending 4-bit instead of 16-bit values would cut this cost by nearly 4x, but the all-reduce must add the values it carries, and rounding the sum injects an error at every call. Model quality depends on how these errors accumulate over layers, yet existing methods treat each call on its own. We present Echo, a 4-bit all-reduce for training and inference that controls the accumulated error. Echo moves data in 4 bits and sums in high precision, so each value is quantized at most twice per call for any number of GPUs. The summation reveals each injected error, and Echo feeds it back into later calls, so old errors cancel instead of accumulating. We prove that the accumulated error stays bounded independently of depth. On A100 GPUs, Echo is 2.09x faster than the 16-bit NCCL all-reduce on 1 GB messages. It speeds up the time to first token by up to 1.16x and the training step by up to 1.09x on an idle interconnect, and by up to 1.44x and 1.38x as background traffic grows. It keeps the average accuracy over seven benchmarks within 0.4 points of 16-bit inference on Qwen3-8B. In 10B-token pretraining, its final loss is less than 0.9% above that of 16-bit training. Echo needs no FP8 or FP4 hardware.

Image-based crack analysis is essential for automated infrastructure inspection to reduce the cost, subjectivity, and safety risks of manual surveys. Despite advances in visual crack detection, methods that translate predictions into physically meaningful and engineering-interpretable crack condition information remain limited. To bridge this gap, this study proposes Inspect Any Crack (IAC), an innovative end-to-end crack recognition and condition assessment framework that automatically derives physical geometric measurements and continuous severity scores from concrete-crack imagery. At its core, IAC introduces SACv2-Tiny, an efficient crack segmentation model adapted from a vision foundation model. Trained on the challenging OmniCrack30K dataset, SACv2-Tiny achieved a $62.49\%$ F1-score and $45.45\%$ IoU on the test set and a mean zero-shot F1-score of $69.89\%$ across three unseen datasets, demonstrating robust segmentation that enables downstream analysis with limited post-processing. IAC further processes the predicted masks to quantify physical crack width and affected area using ground-sampling-distance calibration. Validation at 41 caliper-referenced locations yielded a mean relative width error of $7.33\%$. Beyond geometric quantification, IAC introduces a novel continuous crack-condition metric, the Crack Severity Index (CSI), bounded on $[0,100]$, which combines maximum crack width and affected-area ratio into a unified, standards-informed severity measure. Through these developments, IAC advances crack inspection beyond pixel-level recognition toward interpretable and automated infrastructure condition assessment.

We extend the study of Hamilton-Jacobi equations on networks from a single equation to a weakly coupled system of two equations on a star graph, and we develop a numerical scheme for it. Instead of prescribing the boundary values and imposing a junction condition, we compute the solution on each branch from its outer vertex toward the junction and determine the boundary values for which these branchwise solutions agree at the junction. On each branch, we replace the right-hand sides by their cell averages and solve the resulting problems exactly on every cell. This gives an iteration whose limits we compute in closed form. From these limits, the values of the discrete solution at the junction are affine functions of the boundary values on the branch, with a $2 \times 2$ matrix whose rows sum to one and whose determinant factors into explicit contributions of the cells. We use these formulas to characterize the boundary values that give solutions continuous at the junction and monotone on every branch: when the matrices are nonsingular, the admissible boundary values of one branch form a convex polygon, bounded by conditions on the difference of the two components and by positivity conditions, and the boundary values on the other branches follow from them. We also treat the singular case, and we illustrate the results with numerical experiments.

Hybrid reasoning models such as the Qwen3 series switch between a THINK mode and a NO-THINK mode through a few chat-template tokens. A popular recipe trains such models with reinforcement learning in NO-THINK mode and reports improved reasoning performance. We provide evidence that much of this gain can be explained by thinking leakage: the gain comes largely from partially reactivating the reasoning abilities of the original model's THINK mode instead of learning enhanced reasoning behavior. We train Qwen3-1.7B, 4B, and 8B with GRPO in NO-THINK mode. As accuracy improves, responses become longer and more reflective, are increasingly judged by the original model to resemble its own thinking, and the model's internal activations shift toward its THINK mode. We then investigate at the level of weights, activations, and attention heads: scaling the RL update further elicits THINK-mode behavior and approaches THINK-mode accuracy but does not exceed it; in Qwen3-8B, replacing a low-dimensional THINK-related activation subspace of the RL-trained model with its original-model value removes most of the RL gain, and the reverse exchange transfers a considerable fraction of the gain to the original model; and the RL-induced shift passes through the attention heads that read the empty think block. We further identify a possible mechanism in the GRPO training signal: responses favored by the advantage-weighted update are more THINK-like, and a gradient step computed from this signal moves the model toward its THINK state. Thus, gains in NO-THINK mode can largely reflect reactivation of reasoning abilities already present in the original model.

Humanoid face robots can now move their lips in sync with speech and show detailed facial expressions. Give such a robot the face of a real person and drive it with generated speech, and you get a face forgery that happens in front of the camera: the camera records a real object, and no pixel is edited. We call this a RoboFake. In this paper we ask what current deepfake detectors do when they see a robot face. We test six public detectors trained on FaceForensics++ on video of a physical lip-syncing face robot, on a synthesized video of the same robot, and on rendered robot faces. All thresholds are fixed on FaceForensics++, and every result is compared with real human faces that went through the same processing as the robot videos. At a threshold that flags 5% of real faces, three detectors flag almost every physical robot frame, while two flag fewer than 7%. A matched human control and masking show that the response comes from the face, not from the processing or the background. However, it does not track synthesis: with only the face visible, three of the six detectors give the physical robot a higher fake score than its synthesized video. The response is also fragile, and light JPEG compression is enough to push most robot frames below the threshold for some detectors. Our results suggest that current detectors flag today's robots because their faces do not look human, not because they recognize a forgery. We release our evaluation protocol and code, and state three testable hypotheses about realistic, interactive RoboFakes.

The scalability and performance of edge federated learning are fundamentally constrained by synchronization bottlenecks. Clients operate under heterogeneous compute, time-varying communication conditions, and, in some settings, privacy constraints, yet synchronization is often still handled with fixed schedules. However, synchronization decisions are often made outside the client, even though the client directly observes local progress, resource state, and communication conditions. We present a client-side reinforcement learning synchronization framework in which each client decides during training whether to upload, download, synchronize bidirectionally, or defer communication. The framework places a lightweight scheduler alongside the local model and formulates synchronization as a resource-aware control problem driven by local learning and system state. We study this framework on two representative edge workloads, spatiotemporal traffic forecasting and image classification. Across both settings, adaptive client-side synchronization improves the accuracy-cost trade-off, reducing system cost in traffic forecasting while maintaining competitive cost and improving final accuracy in image classification. We also consider privacy-constrained settings, where synchronization must account for not only latency and energy but also privacy expenditure. Overall, our results show that synchronization in edge federated learning is better treated as an adaptive client-side control problem than only as a server-side coordination mechanism.

Recent advances in multi-modal large language models have shifted toward the "thinking with images" paradigm to improve reasoning faithfulness, yet existing methods rely heavily on explicit coordinate- based grounding. We identify two fundamental limitations in this approach, i.e., a semantic gap where sparse regressions miss high-dimensional latent structures, and a resolution gap due to the granularity mismatch between patch-level tokens and pixel-level coordinates. To address these, we propose TiCO (Thinking with images without COordinates). It reformulates visual reasoning as a semantic evidence discovery process within the latent space, and evidence is injected as token-level regions without pixel coordinates. Specifically, TiCO utilizes a dedicated probe token to estimate similarity distributions and performs importance sampling over spatially connected components to generate diverse, semantically dominant evidence rollouts. Extensive experiments across multiple high-resolution and general visual reasoning benchmarks demonstrate that our TiCO consistently outperforms state-of-the-art coordinate-based methods with significantly improved inference efficiency.

Hyperscale artificial intelligence (AI) data centers are emerging as a structurally new class of electricity demand. Unlike conventional loads, they materialize as spatially concentrated, transmission-level point loads with rapid deployment timelines and converter-dominated dynamics. Their integration is governed by two defining properties: forecast uncertainty in load realization and multi-scale temporal variation spanning sub-seconds to planning horizons. We develop a unified framework that links reliability and affordability to these dimensions through a four-quadrant structure: (reliability, affordability) $\times$ (uncertainty, variation), which integrates probabilistic resource adequacy, generation and transmission expansion, operational security, and market design. Measured traces from a public generative-AI workload dataset show strongly tonal GPU power spectra rather than smooth broadband behavior: harmonic combs at the training-step cadence, and a distinct sub-second periodicity in inference. Across these traces the dominant variance band shifts with workload type (seconds for training and offline inference; tens of seconds and longer for online inference). We report this as a characterization of the measured workloads rather than a universal signature, and take it to indicate that such loads warrant a spectral description rather than a peak-demand one. These signatures directly affect reserve requirements, ramping products, voltage stability, ride-through performance, and harmonic interactions in converter-rich grids. Taken together, the analytical framework and empirical results establish a system-level basis for assessing the reliability and cost implications of large AI data center integration across planning and operational horizons.

Under the conditional framework breaking the longstanding O(n log n) integer multiplication barrier (originating from OpenAI Preprint #109), the asymptotic time complexity exponent T(n) = O(n (log n)^{1-\kappa}) is governed by bounding the second moment E[X^2] < 1 of recursive branching processes over trilinear tensor carrier graphs. In this work, we present a constructive combinatorial optimization over the alternating carrier bipartite matching graphs of the Schönhage trilinear tensor decomposition at dimensions h=23 and h=25. Through an alternating augmenting path search, we uncover a non-trivial cluster-unlocking cascade that eliminates 185 carrier roles (95 roles in h=23 and 90 roles in h=25), reducing the global physical wire width by 377,890 wires from the reviewed baseline W = 177,284,805 down to W = 176,906,915. Evaluated over the exact 10^{-14} rational grid, this wire reduction certifies an improved second-moment coefficient AB = 5,164,059 / 125,000,000,000 and establishes the sharpened conditional exponent \kappa = 826,215,307 / 20,000,000,000,000 \approx 4.131076535 \times 10^{-5}, strictly surpassing the preceding record on this architecture (\kappa \approx 4.123864 \times 10^{-5}). All 47 polynomial constraints and 7 rational security margins are audited in exact Q-arithmetic with zero floating-point error, and the fundamental wire reduction identity is formally verified in the Lean 4 interactive theorem prover without unproven axioms.

Whole-slide images (WSIs) are gigapixel-scale digital scans of histopathology tissue and serve as the primary basis for cancer diagnosis. Existing histopathology generative models operate at the patch scale, producing fixed-size tiles of at most 1024$\times$1024 pixels which encompass only a small fraction of the WSI. In contrast, text reports for clinical diagnoses and slide-level diagnostic labels – the supervision that exists at scale – operate at the whole-slide level. We present BigCell, the first approach to generate coherent whole-slide histopathology images at native gigapixel resolution. BigCell trains a flow-matching transformer at 2048$\times$2048 pixels on over 120,000 publicly available WSIs spanning multiple organ types. Conditioning on embeddings from pathology foundation models and using an inference-time sliding-window pipeline, BigCell generates coherent images of up to 32,768$\times$32,768 pixels, outperforming prior tile generators on both tile-level and slide-level fidelity metrics. To make gigapixel synthesis practical, we introduce an adaptive scheduling strategy that prioritizes important slide regions, speeding up image generation by over 3\ttimes with minimal quality degradation. Furthermore, we propose the first text-to-image synthesis at 8k resolution by training a separate model that produces dense conditioning grids directly from pathology text reports. We demonstrate that augmenting with synthetic samples improves slide-level few-shot classification performance by up to $18 \%$ over baselines trained on real data.

Augmentation-based self-supervised learning can recover the content shared across views without separating its underlying factors into individual coordinates. We ask whether augmentations can be designed to make those factors identifiable. Our approach allows content factors to fluctuate across augmented views while preserving their mean for each source instance. Each instance receives a fixed, recorded set of augmentation strengths that controls how much its factors fluctuate. When factors respond differently to changes in these strengths, their patterns of variation provide information that distinguishes them. Under suitable decoder and support assumptions, we prove that these differences identify individual factors in a conditional Gaussian model, up to permutation, scaling, and translation. We then develop a learner that uses the recorded strengths to estimate these patterns jointly across instances, enabling learning from few views per instance. Synthetic experiments demonstrate near-perfect factor recovery through nonlinear mixing. Controlled dSprites experiments show improved coordinate alignment using augmentations that target known factors, even without telling the learner which augmentation corresponds to which representation coordinate. These results suggest a route to component identification in self-supervised learning: augmentation design can supply identifying information beyond shared content alone.

Regularization-based continual learning protects each parameter in proportion to an estimate of its importance, yet these estimates come from finite samples and heuristics and are used as if they were exact. We report four studies that start from a biologically inspired mechanism and end with a statistical explanation of it. Motivated by dendritic nanotubes, we let transient lateral links diffuse importance tags between hidden units. Links between co-active units gave no significant gain, but random links with lifetimes of 25 or 200 steps raised final accuracy on 20-task Permuted MNIST by 1.5 to 2.6 points. We show that random diffusion implements linear shrinkage of importance toward the layer mean in expectation. A simulation of the link process alone predicts shrinkage weights of 0.23 and 0.52 for these lifetimes, and graph-free shrinkage at 0.25 and 0.5 reproduces both the accuracy and the dispersion of the consolidated importance, while every activity-dependent rule did worse than random links. A five-seed screen across EWC, SI, and MAS found nominal gains for SI and MAS on Permuted MNIST and losses on task-incremental Split CIFAR-100, where forgetting was small and shrinkage cost plasticity. Varying the number of importance samples then separates two mechanisms on Permuted MNIST. For EWC, the gain fell by about ten points per tenfold increase in samples, from 19.6 points with 10 samples to -0.4 with 1000, a variance effect. For MAS, a gain of about one point persisted at 1000 samples, consistent with a bias in the importance measure rather than noise. Shrinkage had no detectable effect for MAS on a domain-incremental CIFAR-100 benchmark. Finally, we show that a standard heteroscedastic empirical Bayes rule shrinks the largest Fisher-type estimates most, which may explain its poor results, and point to a log-scale alternative.

Deploying large language model (LLM) inference at scale requires jointly selecting hardware layout, parallelism strategy, scheduling policy, admission control, and request routing to meet latency and throughput targets. These choices interact: a routing decision that improves cache locality can concentrate load, and an admission threshold that triggers only at saturation allows queues to grow deep enough to degrade critical-tier traffic. Evaluating candidates on GPU clusters is too expensive for iterative exploration, yet existing predictive tools either ignore batching and queueing dynamics or model only fragments of the pipeline. In this paper, we present BLIS, a deterministic, CPU-only discrete-event simulator that models the full inference serving pipeline (admission, routing, scheduling, prefill/decode execution, KV-cache management, and preemption). BLIS pairs analytical basis functions with learned scaling coefficients and overhead terms, fit once globally, to predict step latency without per-model or per-GPU profiling. Across 36 experiments spanning six models and three GPU types, BLIS achieves 6.7% median end-to-end error while running around 200 times faster than real execution. In a configuration search over 420 serving configurations for Llama-2-70B, BLIS identifies a 4-GPU deployment that meets a 300ms mean TTFT service-level objective (SLO), whereas competing tools either require double the GPUs or recommend configurations that fail under load. Two policies evolved entirely in simulation were validated on real hardware and merged into llm-d, a widely adopted open-source platform for distributed LLM serving: an admission policy that reduces critical-tier TTFT P90 by up to 82%, and a flow-control policy that reduces E2E P50 latency by 32-48% on overloaded workloads.

Text-to-motion generation typically relies on dataset-specific skeletal representations, with models trained separately for benchmarks with different joint layouts. We investigate whether image diffusion architectures can model 3D human motion through a common image-compatible representation that accommodates heterogeneous skeletal layouts while preserving motion-specific kinematic structure. To this end, we organize root-centered joint trajectories into topology-informed joint–time images and model global root dynamics through a complementary latent stream. A motion autoencoder conditions local-motion decoding on the root latent, while a text-conditioned Scalable Interpolant Transformer (SiT) jointly generates both latent streams, with supervision on decoded 3D joint positions further constraining generation in motion space. In addition to dataset-specific models, we train a single motion autoencoder on HumanML3D and KIT-ML, followed by a single text-conditioned SiT shared across both datasets, preserving their respective skeletal definitions without skeleton retargeting to a common joint layout. Experiments on HumanML3D and KIT-ML show that our dataset-specific models achieve state-of-the-art performance, while shared training yields further gains on both benchmarks, demonstrating the benefits of shared modeling across heterogeneous skeletal layouts.

Slider-based control aims to give users continuous, fine-grained control over the intensity of individual edits in instruction-based image editing. However, existing approaches fail to deliver reliable continuous controllability: traversing a slider can produce abrupt visual jumps and non-monotonic changes, while jointly controlled sliders can interfere with one another. To systematically study these limitations, we introduce SlideBench, a comprehensive benchmark for continuous image editing spanning diverse editing categories and equipped with a unified suite of metrics quantifying monotonic progression, perceptual continuity, and inter-slider disentanglement. Building on these observations, we propose SliderEdit++, which explicitly regularizes the intermediate editing trajectory rather than supervising only its endpoints. Our method combines velocity- and semantic-space trajectory regularization with spectral control to encourage smooth, predictable slider traversal, together with a disentanglement objective that reduces interference between independently controlled edits. On SlideBench, SliderEdit++ achieves state-of-the-art performance, improving the overall score by 18.5% relative to the strongest baseline. At the same time, using a few-step editing backbone makes our method approximately 3x to 20x faster than existing slider-based approaches, enabling substantially more responsive interactive editing.

Looped transformers reuse the same layer multiple times, allowing inference compute to be adjusted without changing the model. Recent adaptive-depth methods train the model across multiple loop depths from the beginning of training. We introduce fixed-first training, which varies model depth only during a short final phase of training. At 200M parameters, training a model of fixed-depth 2 for the first 67% of the run reduces modeled training compute by 45% while keeping mean negative log-likelihood (NLL) across inference depths within 0.5% of training across depths throughout. At 1B parameters, training initially at depth 4 reduces modeled compute by 31% to 37% and wall-clock time by 32% to 39%, while keeping final NLL within 0.8% and downstream differences within the observed seed variation. These results show that training over the full range of depths can be reserved for a short finishing phase with only small changes in performance.

We consider a multi-hop wireless network where energy-harvesting nodes rely solely on ambient radio-frequency interference as their power source. The goal is to deliver a packet from source to destination with minimum latency while guaranteeing a target end-to-end outage probability under Rayleigh fading. This gives rise to a joint routing and power allocation problem in which interference plays a dual role: it provides harvestable energy, but also impairs link reliability. We first transform the intricate, multiplicative end-to-end outage constraint into an additive form. Then, we propose two optimal algorithms. The first, HOPE, minimizes the bottleneck harvesting latency, i.e., the longest energy-accumulation time along the path, by reducing the problem to a least-cost path computation on a graph with interference-aware link weights. The second, THOPE, extends the objective to total end-to-end latency, adding the cumulative transmission delay, and likewise admits a closed-form least-cost path solution. Both algorithms are solved in polynomial time via Dijkstra's algorithm and yield explicit expressions for the optimal transmit powers. Extensive simulations reveal that HOPE and THOPE reduce the average total latency by up to an order of magnitude compared with shortest-path and minimum-energy benchmarks, while exhibiting sharply concentrated delay distributions that are largely insensitive to interferer density.

Diffusion models provide expressive trajectory priors for multi-robot motion planning (MRMP), but scaling them to large teams requires coordinating individual motions without expensive joint computation. We introduce Akule, a sparse interaction diffusion framework that augments a frozen single-robot prior with three learned modules: a reusable pairwise correction model $R_{\phi}$, a contextual selector $U_{\psi}$ that determines which corrections to evaluate, and a composer $G_{\omega}$ that combines their effects through context-dependent signed weights. This decomposition separates which interactions merit computation from how their effects should be combined into coordinated denoising updates. We train $R_{\phi}$ using matched pairwise trajectory optimizations, then learn $G_{\omega}$ and $U_{\psi}$ from physically optimized joint trajectories without requiring a prescribed ground-truth interaction graph. The resulting joint proposals are passed to a downstream search-based planner for validation and repair. Across diverse MRMP benchmarks, Akule achieves competitive planning success, complete-planner speedups of up to $25\times$, zero-shot population transfer, and successful planning with up to $256$ robots. Ablations further demonstrate that learned selection and composition reduce pre-repair conflicts and downstream repair effort relative to heuristic selection and uniform summation.

Variance-based methods recover a Gaussian linear structural equation model (SEM) with equal or suitably constrained error variances, either by recursively identifying the source nodes with smallest conditional variance (top-down method) or sink nodes with highest conditional variance (bottom-up method). We show that the use of conditional variance is incidental: replacing it with the conditional risk, i.e., the minimal expected loss of linear prediction under the negative log-likelihood of the noise distribution, extends both methods to non-Gaussian noise. Existing methods are the least-squares special case. This yields two main results. First, exploiting the noise distribution improves structure recovery: the matched-loss procedures recover the SEM more accurately than their least-squares counterparts and the LiNGAM methods, which are designed for non-Gaussianity. Second, the two methods, interchangeable in the Gaussian setting, separate under non-Gaussian noise: the top-down method is correct for every noise density with finite differential entropy, whereas the bottom-up method requires a regularity condition and can fail without it, for instance under mean-zero exponential or Cauchy noise. Simulations confirm both results.

An LLM-directed excavator can execute mechanically feasible motion whose destination or authority no longer matches the operator's approval. We propose execution-bound authorization for the approval–dispatch–execution interval. A one-use lease binds the action, generated motion, mission contract and admission tolerance; atomic dispatch rechecks current authority and measured joint drift. Periodic supervision detects revocation and requests a quintic stop with a carry-pitch endpoint, selected through sampled joint, stroke, flow and reference-acceleration checks. The framework connects exact motion identity to dispatch integrity and load-aware interruption. Evaluation combines a 672-case interface suite, four loaded digging cases in synchronous simulation, unloaded MuJoCo/Pinocchio transfers and two command-level LLM planners. The gateway admitted all 120 benign interface cases and produced zero unauthorized dispatches in the attack suite. Loaded dispatch-window tests produced zero unauthorized deposition. During simulated dump-phase revocation, maximum closed-site deposition decreased from 1.23 to 0.33\,m$^3$; loaded-slew stops used less than 75% of nominal swing torque. These results demonstrate motion-bound authorization and quantify the retention and stopping behavior of supervised excavation skills.

We perform the first comprehensive experimental characterization of power consumed by real HBM2 DRAM chips. Our extensive characterization yields 1) the empirical basis for developing new, accurate HBM power models or validating existing power models and 2) key insights into structural and data-pattern-dependent variation of power in HBM2 chips. Out of our 13 new findings and 6 key takeaway messages, we highlight: 1) a read operation consumes significantly higher current than any other operation and 2) the toggled data bus signals in a data transfer burst significantly affect current draw where toggling all HBM2-internal bank group bus signals causes 29.4% higher power consumption than keeping them constant. Leveraging our empirical insights into HBM2 power consumption, we engineer an access pattern that yields 23.6% higher power consumption on average across 36 tested HBM2 chips than the highest-power-consuming JEDEC-specified current measurement loop. Our results have important implications for HBM-based system design and validation. Relying solely on JEDEC- or vendor-specified test loops to bound HBM power could cause system failures and/or thermal throttling in real-world high-power AI/ML workloads. To improve 3D-stacked DRAM power modeling in open source architectural simulation infrastructures, we develop a new HBM power model. Our model predicts the power of 36 real HBM2 chips across diverse workloads (including a real LLM decode workload) with 7.6 – 8.1% mean absolute percentage error, substantially more accurately than state-of-the-art HBM power models. The power consumption predicted by our model for a power-intensive microbenchmark closely matches (with 2.4% to 5.3% error across HBM3E's specified supply voltage range) the average HBM3E memory power consumed by three tested NVIDIA H200 GPUs. We publicly release our power model at https://github.com/CMU-SAFARI/HBM-Power.

LLM data engineering requires iterative experimentation to determine how data selection, rewards, and curricula affect model learning, making it a natural setting for AI-for-AI (AI4AI) research. Can agents use experimental feedback to progressively improve strategies, and do the resulting strategies improve target-model performance on held-out data? We study how agents use experimental feedback to develop data-engineering strategies under fixed training and evaluation protocols. Our Agentic Data Engineering (ADE) framework lets agents pursue, revise, or combine research directions using cumulative findings. Strategies are selected by in-loop scores and evaluated on held-out benchmarks. Across four tasks-math and code data selection, math reward design, and math curriculum learning-selected strategies achieve absolute gains over generic baselines of up to 13.34% on AIME25 Pass@32 and 5.85% averaged across four held-out benchmarks under fixed training conditions. In-loop scores improve over successive experiments, while held-out trajectories show that later in-loop gains do not always improve generalization. Comparisons of research organization and feedback access favor combining evidence analysis with cumulative findings. Research trajectories show how earlier findings inform strategy revisions and combinations across research directions.

Physically connected humanoids need consistent dynamics for internal actuator loops, inter-robot forces, and changing ground supports. We extend RoboCompiler with member-wise dynamics assembly, compiling reduced dynamics from rigid bodies, moving actuator parts, and reflected rotor inertias. Compiled body Jacobians and analytic constraint curvature yield an explicit Coriolis factor satisfying the power identity for floating-base quasi-velocities. A two-level reduction retains reusable robot models while composing their supports and hand connections; local inertial updates have rank at most six, and prescribed-acceleration constraint wrenches are affine in actuator efforts. These interfaces support actuator-aware tracking, coupled force allocation, payload updates, and support-mode switching for five hand-linked humanoids. Across 264 feasible states in eight humanoid configurations, normalised acceleration discrepancies remain within $1.6\!\times\!10^{-10}$ against a hard-constraint reference and $1.4\!\times\!10^{-10}$ against a compiler-independent reference; MuJoCo simulations achieve 0.73~mm root-mean-square (RMS) pelvis error in the ring dance and complete side steps through 20 support-mode changes.

Information of the tire-road friction coefficient is critical for intelligent vehicle systems, as friction establishes the limits at which the vehicle can be manoeuvred without loss of control. While there is extensive literature on online friction estimation, no previous studies have explored post-processing and mapping friction coefficients along a driven trajectory based on recorded sensor data. Mapped friction information could be shared by connected vehicles, enabling proactive measures in low-friction areas. This paper presents an offline post-processing method, FrictionMap, for mapping friction coefficients from recorded driving data. The method operates on commonly available on-board vehicle sensor information, and requires only a basic vehicle parameter set and a slip threshold as prior knowledge. FrictionMap runs an extended Kalman filter combined with a Rauch-Tung-Striebel smoother to estimate vehicle dynamics. Data points where the friction limit is observable are identified by filtering points based on tire slip. A Gaussian process is fitted to these points to acquire a continuous estimate of the friction coefficient along the trajectory. FrictionMap is validated via a simulation study, and shown to outperform commonly used online friction estimation techniques. The code implementation and the simulation benchmark are released as open-source after paper acceptance to enable further research.

Automatic speech recognition produces words but drops the non-verbal content of a recording: the breaths, coughs, sighs and laughter of the speaker, and the horns, music or barking around them. We describe Speech Event Detector, a 5.8-million-parameter convolutional-Transformer model that outputs probabilities every 40 ms for 54 event classes (20 speaker vocalizations and 34 background sounds), together with post-processing that converts these probabilities into timestamped events and a procedure for merging the events into a Whisper transcript. The model is trained on material drawn from seven public datasets whose labels come in incompatible forms: single detected regions, free-text tags with timestamps, temporally strong ontology labels, symbols embedded in transcripts, isolated recordings and environmental recordings. We unify them using rule-based label mapping, recover timestamps for transcript-embedded labels through CTC forced alignment, locating 6,360 of 8,620 sounds, and combine about 168 hours of real annotated training recordings with synthetic mixtures and augmentation using real background noise. Per-class thresholds for two operating profiles are tuned on real held-out recordings, and performance is estimated using two-fold cross-validation. On whole held-out recordings from four datasets, considering the 48 classes with at least ten held-out events, the sensitive profile detects an average of 49 percent of a class's events, covering 67 percent of all 9,367 events, at a macro precision of 0.30. The balanced profile achieves a macro recall of 0.37 at a precision of 0.36, with macro event F1 scores of 0.32 and 0.33, respectively. The detector runs about 39 times faster than real time on eight CPU threads.

AI slop has flooded social media feeds with overprocessed and derivative content–Jesus fusing with a shrimp, a humanoid orange chubby cat shoplifting, and a "hero" dog rescuing a human child after a plane crash. This might be no laughing matter, as waves of media coverage have taken a distinctly negative bent, emphasizing its low quality, ethical concerns, and threats to human creativity. Despite extensive attention from politicians and technologists, the perspectives of everyday users–the very audience most exposed to this phenomenon–remain largely unexplored. This descriptive work explored how Reddit users experience, conceptualize and contest AI slop. Inductive thematic analysis of posts and comments across six slop-related subreddits identified four themes covering constitutional defects, algorithmic overreach, user resistance, and entertainment outcomes–ranging from anxieties about eroding agency, moral norms, creativity, and existential questions to defensive distinctions between quality AI content and slop. We derive a user-driven definition: AI slop is audience-contested AI-generated content characterized by constitutional defects and enabled by platform logics. We operationalize slop through a facet-theoretical framework to inform future research into this fuzzy category.

Safety-critical Cellular Internet of Things (CIoT) systems increasingly rely on microservice architectures orchestrated across heterogeneous edge and cloud platforms. Although this paradigm improves modularity and scalability, the fault-tolerance mechanisms of contemporary orchestration platforms remain predominantly reactive, i.e., failed components are restarted or rescheduled only after an outage has already occurred. Such recovery can produce multi-second End to End Service Outage Time (E2E SOT), which can grow further under container restart backoff. This paper proposes an Availability Management Framework (AMF), a generic monitor-predict-recover loop that complements reactive recovery with machine learning-based anomaly detection and make-before-break switchover to a standby replica. Candidate ML models are evaluated offline under three protocols, including a change of operating regime, in order to select a model for live deployment. AMF is instantiated and evaluated on a three-tier tele-operated driving (ToD) testbed under a staged resource-exhaustion scenario. Relative to Kubernetes-only recovery, AMF completes switchover during the anomaly window, reduces E2E SOT, and avoids the tens-of-seconds outages associated with repeated in-place restarts under backoff. On this non-real-time testbed, residual E2E SOT lies inside healthy-path jitter and is therefore limited by path timing rather than by lateness of the recovery decision. The paper whows that coupling a validated ML based anomaly detector with proactive switchover can remove restart- and backoff-dominated outage from a critical path.

Digital technologies are reshaping how families negotiate care, authority, and responsibility, yet these dynamics remain understudied in religiously situated family life. Through interviews with 38 adult members of Bangladeshi Muslim families, we examine how family members guide one another across religious, digital, and AI-mediated contexts. We find that authority did not simply follow age or family hierarchy: older members often remained sources of religious and familial guidance, while younger members increasingly provided expertise around digital risks, information credibility, and AI. Families also actively shaped digital environments and negotiated the role AI could occupy in everyday relationships. We develop the concepts of household epistemic governance, counter-curation, and domain-dependent authority to explain these shifting relations of knowledge and responsibility. Our findings suggest opportunities for family technologies that support negotiated boundaries, credible information practices, and culturally situated forms of care.

Pixel-space world models have recently emerged as a promising approach for GUI agents, predicting future screens prior to execution and subsequently re-perceiving them for action selection. However, this two-stage paradigm faces two key challenges in GUI environments: (L1) the high computational cost of pixel reconstruction and (L2) the need to reconstruct large and diverse appearance-level details beyond what is required for action selection. In practice, this screen-generation objective can induce an identity shortcut problem, where future-state prediction largely preserves the current state rather than capturing action-induced changes. In this work, we reformulate GUI world modeling by shifting the prediction target from full-screen reconstruction to action-induced future information in latent space. Concretely, we leverage a large multimodal model as the latent world model, whose shared representations support two dedicated heads for both action selection and predictive transition modeling. The action head directly translates these latent representations into action selection, while the transition head captures action-conditioned future information. To mitigate the identity shortcut, we further introduce an action-aware contrastive objective that encourages the transition representations to capture discriminative action-induced changes. Extensive experiments across offline and online GUI settings show that our method achieves up to 31.9% and 40.6% relative gains in success rate over pixel-space world models, respectively, while substantially reducing inference latency by up to 98.9%. Our code is available at https://github.com/sjyu001/latent2world.