Transformer architectural modifications frequently measure a statistic on internal representations, observe that it is large, remove it architecturally, and report an improvement. Interpreting that sequence requires evaluating whether the statistic exceeds its noise floor, whether it survives a geometric null, and whether the training gain requires the specific direction removed. We formulate these requirements as three reusable checks and apply them to self-attention removal, which deletes the component of an attention head's output aligned with its own value vector. Across nine multi-head pretrained models (5,408 heads, 124M to 6.9B parameters), an earlier-position causal null accounts for 48% to 63% of the head-value cosine alignment, with no monotone scale trend. Under grouped-query attention, this alignment is specific to each head's own key-value group and reverses across groups (+0.24 to +0.27 within-group, -0.11 to -0.19 across-group). Split-half reliability demonstrates that the per-head intervention effect is resolvable in GPT-2 ($r_\Delta = 0.80$) but attenuated in Pythia-160M (0.42) and Pythia-410M (0.53), whereas structural self-cosine reliability is near-perfect in all three ($r_{\text{stat}} > 0.99$); which statistic predicts its own intervention is model-dependent. In a pre-registered 3-arm by 8-seed paired factorial at $4.0 \times 10^8$ tokens per run (399,900,672 tokens), a fixed arbitrary rank-one direction does not differ from baseline (+0.0011 nats, $p = 0.197$), while self-directed removal lowers validation loss (-0.0029 nats, Holm-corrected $p = 0.0023$) and outperforms the fixed arbitrary control directly (-0.0040 nats, $p = 0.0018$). At 50.9M scale the training benefit requires the own-value direction rather than a fixed arbitrary one, but a large cosine alignment alone does not establish redundancy in frozen models.
Data-science and ML-engineering agents are evaluated on benchmarks such as MLE-bench, ALE-Bench, SWE-fficiency and AlgoTune, whose tasks yield a quantity (a Kaggle metric, contest score or speedup). When one stops separating the best models, the usual remedy is a harder version. Re-scoring released runs of eight benchmarks, including SWE-bench Verified and Terminal-Bench for breadth, we find much of this saturation lies in the measurement, not the tasks. Most agent benchmarks reduce a run to one pass/fail bit, which fails in two ways. It is too coarse: neighbouring top-ten agents on SWE-bench Verified differ by less than its tests' errors, and correcting them reorders neighbours no more than deleting a random sixth of each agent's passes. It is capped at the reference: ALE-Bench's top ten, separated by contest score, tie when graded against the median human, although nearly half the runs beat that human. On fixed tasks, a higher bar only moves the problem; graded credit fixes it. We derive a clipped log-ratio score with dated anchors and cost as a separate axis. In a Terminal-Bench re-run, pass/fail-tied models differ over tenfold on the quantity the tests discard.
Traffic cameras continuously record urban road-user behavior, providing proactive safety evidence beyond crash reports. Yet safety-critical events - conflicts, near misses, unsafe turns, sudden braking, and crosswalk interactions - are rare and brief, buried in hours of routine footage. Existing traffic-video pipelines extract detections, trajectories, and risk scores but process footage uniformly, without adapting to an analyst's query. Vision-language methods improve scene understanding but often lack traffic-specific verification of participant-level conflicts. We introduce AgenticConflictLens, an agentic, query-conditioned framework for efficient, auditable conflict extraction from long fixed-camera traffic videos. Given a natural-language safety query, a Traffic-Aware Controller combines structured traffic memory with specialized conflict-analysis tools to choose where to search, which road users to examine, and which evidence to verify or reject. A coarse-to-fine search retrieves candidate intervals and applies detailed verification selectively. The framework produces conflict records with refined time intervals, participant pairs, semantic regions, trajectories, motion evidence, uncertainty, rejection reasons, and supporting clips. On 96 annotated hours from Bellevue and NYC DOT, AgenticConflictLens achieves the highest conflict-existence F1 and temporal IoU on both datasets: 0.84 F1/0.77 tIoU on Bellevue and 0.80 F1/0.73 tIoU on NYC DOT, while inspecting only 8.7% and 11.9% of video, respectively. These results support scalable, evidence-grounded urban traffic-safety analysis.
Researchers and engineers improve recommendation systems by performing A/B experiments with new model architectures, optimization methods, and data features. However, running full-scale experiments (e.g., full-model training with online data) for every feature change is prohibitively computationally intensive, while the result from a single small-scale experiment is often misleading. To address this challenge, we show that modern recommendation models exhibit feature-invariant scaling: the change of a model's performance follows a predictable scaling-law even when different feature changes are applied to the baseline model. We introduce an efficient method that exploits scaling-laws and uses statistical inference to predict (with confidence intervals) the full-scale effect of an A/B experiment in recommendation systems. This enables practitioners to efficiently filter out neutral or regressive feature changes with only small-scale experimentation. We validate our method on three distinct recommendation model families---Wukong, HSTU, and LLM-based recommendation models---across more than 40 architectural, optimization, and data changes. Our method accurately distinguishes beneficial feature changes from neutral and regressive ones while using only a fraction of the compute required for full-scale evaluation. It also identifies several changes that improve HSTU's state-of-the-art performance on MovieLens-20M. Our results demonstrate that scaling laws can be practically exploited to make compute-efficient decisions for improving industrial-scale recommendation systems.
Catalysis underpins chemical manufacturing, yet knowledge for catalyst design remains dispersed across experimental observations, computational data and catalysis subdisciplines. To overcome this challenge, we present Composable Catalytic Intelligence (CatAI), a physics-grounded modular AI framework that learns 1.8 million structure-property relations and organizes them into 30 physicochemical encoders. This reusable intermediate layer enables task-specific models to guide catalyst discovery for targeted reactions and objectives. Across benchmark tasks, CatAI achieves a 7.8% lower task-averaged prediction error than the next-best models. In experimental validation, CatAI identified a new confined-passable pocket catalyst applicable in industrial homogeneous polymer synthesis with 9 times the activity of the commercial benchmark. In parallel, it guided autonomous closed-loop optimization of a highly active Cu-based geminal-atom photocatalyst for C-O coupling reactions surpassing the traditional Ni-based systems. This approach connects property-level data to catalytic objectives, supporting physicochemical knowledge reuse across mechanistically distinct systems through computational screening and iterative model-experiment feedback.
Classical neural scaling laws model excess risk as a separable power law in model capacity $N$ and dataset size $D$. At finite budgets, however, empirical learning curves and compute-optimal allocations depart from this log-linear frontier. This paper analyzes how spectral and domain heterogeneity govern these finite-scale departures. Along the capacity axis, Mellin--Barnes analysis of covariance resolvents yields a three-term expansion where competing spectral and regularization modes produce finite-scale curvature and logarithmic transitions. Because fitting multiple power laws on narrow pilot grids is ill-conditioned, we eliminate the loss floor via first-order Aitken differencing to obtain a well-conditioned closed-form initializer. Along the data axis, KKT duality reduces $K$-domain mixing to scalar water-filling, yielding finite-budget bottleneck bounds, an online sampling curriculum, and a simplex Hessian metric for cross-domain transfer. Along the iso-compute frontier $C=6ND$, we prove that the optimal allocation exponent drifts in proportion to capacity and data exponent variances, while Schur-complement reduction separates intrinsic drift from frozen-hyperparameter bias. Across 45 operator spectra, Pythia (70M--12B), Chinchilla (245 runs), and RegMix (17 domains), our estimators achieve $5.89\%$ holdout MAPE, cut high-compute allocation regret by up to $76.4\%$, and reduce cross-scale mixture RMSE by $59.4\%$ with zero selection regret.
Large language models increasingly evaluate, route, and moderate news, so a model's judgment of a headline has become a gatekeeping decision. We ask whether that judgment follows the fact being reported or the way the fact is framed. We combine an ecological audit of 63,926 Chinese news headlines with a controlled experiment. The experiment covers 30 real events, 30 synthetic events, and 29 natural headline pairs, tested on four Chinese models. Changing only the headline frame shifted both perceived misleadingness and event severity, and the direction and ordering of the shifts were the same for DeepSeek, MiMo, Dots, and LongCat. Emotionally intensified headlines raised severity ratings by 0.65 to 0.88 points, whereas positively reframed headlines lowered them by 0.34 to 0.52 points. Supplying the article body reduced this severity bias by 54 to 80 percent, and all four reductions were significant at p < .01. A small human reference study showed the same direction of effect. Headline-only LLM evaluation therefore carries a measurable robustness gap, and framing sensitivity belongs among the reliability properties that deployment and governance check.
Unified multimodal models (UMMs) have opened up new possibilities for generating rich interleaved text-image content. However, they generally lack native support for coordinating the two modalities, and translating coarse user requests into coherent multimodal compositions also requires explicit planning. We propose UniWeaver, a structured, model-agnostic harness framework that integrates multimodal planning, generation, and visual feedback to produce coherent interleaved outputs. To train UMMs for this task, we construct UniWeaver-10K, a dataset of 10,000 tasks across 20 task families, paired with execution trajectories guided by UniWeaver. We also introduce UniWeaver-Bench, a task-oriented benchmark systematically assessing text coverage, writing quality, visual coverage, and image-text alignment. Experiments show that UniWeaver achieves state-of-the-art performance and improves interleaved text-image generation across multiple UMM backbones.
Photorealistic and metrically consistent world representations are a primary requirement for LiDAR--inertial--visual SLAM systems that serve robots over continuous operation. 3D Gaussian splatting makes such representations attainable by supplying an explicit, renderable, and queryable scene model. We present GEMOR-GS (Global Extrinsic-calibration-refined Mapping and Online Relocalization for 3D Gaussian-Splatting SLAM) to make the Gaussian map geometrically accurate, globally refined, and re-enterable, providing a better appearance for robot perception and decision-making. To improve the 3DGS map, the LiDAR--camera extrinsic is first estimated offline by direct-method alignment over dedicated calibration segments and reused by all following online and offline processes. Under the calibrated geometry, keyframe poses and Gaussian parameters are refined alternately under a gated differentiable-rendering and LiDAR--projection objective. The final global reconstruction is then rebuilt from the refined keyframe schema with all online constraints lifted. For the robot SLAM lifecycle, the method recovers the pose from the 3DGS map after tracking loss through an XFeat bypass map, RANSAC-PnP coarse localization, and Gaussian RGB-D warp refinement. Experiments on one self-collected and five public sequences quantify each element: the offline reconstruction wins all paired keyframes against the online map by 14.3--19.9 dB, improves PSNR by approximately 9 dB over Gaussian-LIC2, and achieves 86--100% held-out relocalization success tied to extrinsic quality. The results show that GEMOR-GS improves the geometric accuracy, appearance fidelity, and re-enterability of the 3DGS map, serving as a dependable robot world model for SLAM and downstream tasks.
Communities in the United States and around the world are contesting data centers, the physical infrastructure of AI. These disputes pose environmental justice issues: residents must weigh unevenly distributed burdens using evidence scattered across metrics, corporate claims, policy documents, and public testimony. A formative study with 10 users showed that dashboards and single-agent AI can settle questions before people examine that evidence. Here we present EJAI, a source-grounded system that links an interactive map, retrieved evidence, missing-data statements, and two stakeholder personas that allow people to inspect evidence and form their own policy recommendations. In a small functional benchmark of 21 queries, retrieved evidence supported 98.4% of factual claims and the system acknowledged missing evidence in 32 of 33 cases. Persona alignment carried a grounding cost: the Corporate Executive omitted some evidence that conflicted with its assigned priorities. In a within-subjects study, 15 participants acting as residents prepared for a simulated town hall. Some checked cited AI claims while treating map metrics as official, some made conditional recommendations with safeguards such as water-use disclosure, and some found the personas simplified or biased. Our results suggest that civic AI can support preparation for public reasoning when it makes evidence, gaps, and disagreement open to inspection, and that corpus imbalance and two-sided personas remain open justice concerns.
Recursive self-improvement (RSI) seeks to improve an agent using its own task experience. Its practical value depends on whether performance gains justify the full cost of self-improvement and task execution. We evaluate this trade-off on continual task streams, scoring each task before its feedback is used for subsequent updates. We define the \emph{RSI Tax} as the token cost of the self-improvement process and include it alongside task-execution costs in end-to-end evaluation. Across six methods and two language models, a bounded-note baseline matches or exceeds the observed accuracy of Meta-Harness, EvoSkill, and RRSI, with mean total-token savings of 65.6-71.0\%. We decompose the RSI Tax into verification, exploration, and writing costs, and track subsequent use of the associated state or artifact. For these three RSI methods, verification accounts for 74.8--93.6% of the aggregate RSI Tax. A substantial share of the RSI Tax is also associated with states or artifacts that are not used later in the evaluated streams. These results expose a practical limitation of costly RSI procedures: improvements in the deployed agent need not justify the expense of the improvement process. They motivate RSI objectives that jointly optimize the RSI Tax and expected task-execution cost while ensuring continued growth in agent capabilities across the task stream.
Nonlinear state space models are typically evaluated sequentially, limiting their ability to exploit modern parallel hardware. Recent work has shown that state space models can be evaluated in parallel by reformulating evaluation as a root-finding problem, which can be solved with a parallel form of Newton's method. However, exact Newton methods require expensive multiplication of dense Jacobian matrices, and quasi-Newton methods based on diagonal approximations, while more scalable, are often slow to converge because they neglect interactions across state dimensions. We introduce a novel quasi-Newton approach based on Broyden's method, which captures coupling terms with diagonal-plus-low-rank approximations to blocks of the Jacobian. We construct the approximations from trajectory secants, avoiding automatic differentiation while retaining the favorable scaling of diagonal approximations and remaining compatible with parallel scan through rank compression. We prove finite-step recovery of the exact trajectory under the compressed scan, and show that convergence is tracked by the compression error and by the block approximation error along the current trajectory error, rather than in spectral norm. Empirically, we show that our method is advantageous when interactions are approximately low rank or when Jacobian evaluation is expensive: on low-rank RNNs, parallel MCMC algorithms, and denoising diffusion models, it converges in fewer iterations than other quasi-Newton methods and is the fastest parallel method when Jacobian evaluation is expensive or state dimensions are large.
Data circulation in vocational education has long been constrained by unclear ownership, a lack of trust, and insufficient incentives. The Internet of Data provides a new underlying technology for enabling trusted cross-domain data circulation. Building on an Internet of Data based trusted data space for vocational education, this paper develops a multi-stakeholder operational model encompassing a value network, a product portfolio, revenue allocation, and governance rules. Emphasizing data sovereignty protection and incentive compatibility, the model provides a reference for the sustainable operation of regional vocational education data ecosystems.
Accurate inertial odometry is critical for autonomous systems but suffers from unbounded drift caused by the integration of sensor noise and biases. In approaches such as TLIO, neural pseudo-measurements are trained independently of their downstream filter. This paper introduces Differentiable Filtering for Inertial Odometry (DFIO), a learning framework that embeds the update stages of an error-state Kalman filter directly into a computational graph. Following standalone pre-training, DFIO jointly optimizes the displacement and covariance output heads using posterior-position supervision through differentiable measurement updates. The proposed system is validated on public pedestrian motion datasets featuring both calibrated and uncalibrated inertial measurements. Experimental results on a pedestrian dataset demonstrate that DFIO reduces mean absolute and one-second relative translation errors by 13.2% and 39.6% relative to the TLIO baseline, and by 7.7% and 17.4% relative to its pre-trained counterpart. Furthermore, evaluations on uncalibrated smartphone inertial data indicate that the proposed method effectively mitigates translational drift and maintains trajectory alignment under highly noisy sensor conditions. Source code will be made publicly available.
Diabetic retinopathy (DR) lesions vary widely in shape, and microaneurysms (MAs) span only a few pixels. We hypothesize that modeling the spatial organization of lesion regions, with global context and shape-adaptive scanning, improves the segmentation of such small and irregular lesions. To test this, we propose D$^2$Mamba, which couples non-causal state space duality with deformable Mamba scanning. We evaluate it with five-fold cross-validation on the DDR dataset, comparing against representative segmentation baselines under an identical training protocol, and conduct ablations on each component. D$^2$Mamba achieved the best overall performance, with 33.87% mIoU, 2.99 percentage points above the strongest baseline. The largest gain appeared on MAs, the smallest and most easily missed lesions, where Dice improved by 10.06 percentage points. These results confirm our hypothesis that spatially adaptive state space modeling benefits tiny-lesion segmentation. Code and pretrained weights are available at https://github.com/PUREPANEL1/D2Mamba.
We propose a novel conditional entropy function as a basis for quantifying predictive uncertainty in Evidential Deep Learning (EDL) for glaucoma detection from retinal fundus images. Using the EyePACS-AIROGS-light-V2 and RIM-ONE-DL datasets for validation, we found that our proposed EDL model regularized by conditional entropy function produced uncertainty estimates with substantially lower calibration errors than Deep Ensembles, MC Dropout, post-hoc calibration and the conventional EDL framework based on Dempster uncertainty. To further probe our models' interpretability, we analyzed Grad-CAM heatmaps using a focus index representing spatial concentration. Compared with the other uncertainty quantification methods, EDL models exhibited greater separation between the index distributions of correctly and incorrectly classified fundus images. In particular, false negatives were associated with localized Grad-CAM activations that were displaced from the optic disc, consistent with its established clinical importance in glaucoma assessment. Overall, our findings suggest that our proposed EDL framework for glaucoma detection provides not only enhanced uncertainty calibration but also admits a more physiologically consistent interpretability.
Nepali speakers regularly mix English words into conversation, but multilingual speech recognisers often struggle with language switches and the writing script used for English words. We study how to adapt SraVaani, a 449M-parameter hybrid TDT-CTC model, using a 172.8-hour corpus covering six Nepali speech domains. Its pretrained tokenizer does not support digits and several common symbols. Rather than replace it or discard these recordings, we generate Nepali and English readings for numbers and use zero-shot ASR output to select a plausible form. Full fine-tuning reduces test word error rate (WER) from 48.6% to 17.1%; a 12-epoch run reaches 16.3%, compared with 16.4% for fine-tuned Whisper large-v3-turbo. Freezing the encoder increases WER to 25.5%, while excluding numeric clips raises WER on numeric test speech from 14.3% to 21.3%. We also evaluate on FLEURS using both its original references and mixed-script alternatives. The comparison shows that transcription conventions explain part, but not all, of the out-of-domain performance loss.
Curvilinear structures such as retinal vessels and surface cracks need both pixel-level accuracy and structural continuity. Pixel-wise supervision does not model the connections between pixels, and topology-based losses still produce ordinary masks. Connectivity-based methods predict these connections directly, but they use binary connectivity targets over a fixed local range. Such labels give the same value to thin segments and to thick regions, and they cannot describe connections at larger offsets. To address these limitations, we propose Topology-preserving Multi-offset Affinity Learning (TMAL). TMAL uses soft affinities that emphasize thin regions near the boundary, and it predicts directional affinities at several offsets to reach further. It then combines these affinities with weights that come from the decoder features, so local continuity and wider context are used together. An auxiliary segmentation branch adds region-level supervision and keeps the overlap accuracy high. On the retinal vessel dataset DRIVE and the crack inspection dataset DeepCrack, TMAL reaches the highest Dice, IoU, and clDice scores, and it greatly reduces the connected-component error $\Delta\beta_0$. Code is available at https://github.com/suho-han/TMAL.
Remote sensing imagery acquired by Unmanned Aerial Vehicles (UAVs) supports wide range of applications; however, deploying machine learning (ML) models on resource-constrained aerial platforms introduces challenges beyond predictive accuracy, particularly when deployment data differ from the distributions observed during training. Such shifts are common in remote sensing due to geographic, sensor, seasonal, and other visual variations. A practical solution is to equip deployed models with mechanisms that detect emerging distribution shifts and support subsequent responsive actions. This paper presents an empirical case study of three open-source detection tools: (i) Alibi-Detect; (ii) NannyML; and (iii) PyTorch-OOD. Using cross-dataset remote-sensing shifts, we evaluate both detection effectiveness and computational latency of these tools. Our results reveal differences in detection behavior and processing cost across methods and implementations, highlighting trade-offs between detection capability and response-time. These findings motivate further work toward lightweight, and latency-aware monitoring mechanisms for resource-constrained systems.
Small language models (SLMs) running on mobile devices can keep user data private and work without connectivity. But they perform well only when fine-tuned for the target task and sized for the target device, a process that today requires repeated expert tuning. We present {\it Mobile AI Factory}, an agentic system that builds a fine-tuned, quantized SLM for a given task and phone. Given a task specification and a target phone, an orchestrator LLM chooses whether to tune hyperparameters, mine real data, or generate synthetic examples, while fixed rules roll back changes that hurt accuracy. The factory searches from small to large models across base models and quantizations, adapts its accuracy goal as it searches, and uses real measurements of the target phone across inference engines and backends. We evaluate Mobile AI Factory on ten tasks and four Android phones. On an unseen testset, its on-device models outperform two five-shot frontier models on five tasks, by up to 0.34, and match them on three more. Its search comes within 0.03 of exhaustively training all 27 model variants on an entity extraction task, while using 19$\times$ less GPU time. Phone measurements also matter: the best configuration changes across phones. Without phone measurements, Claude Opus 5.5, given the full hardware and model specifications, misses the app's requirements in 25\% of easy and moderate cases and in 60\% of hard ones. Code is available at: https://github.com/evanly-gh/SLM_Factory.