Zhoutong Ye, Chengwen Zhang, Zixuan Wang, Xufeng Jian, Zhe He, Qingyang Wan, Xutong Wang, Jiaqi Liu 等 12 位作者
We present IMUDINO, the first foundation model (FM) for inertial sensing to achieve competitive performance across diverse sensing tasks including human activity recognition (HAR), health, motion capture, and odometry. Prior inertial FMs often focus on one subfield of inertial sensing and sometimes underperform state-of-the-art supervised methods (e.g., in HAR), making their scope considerably narrower than FMs for vision and natural language. Moreover, they often require fixed input length and sampling rate, and generalize poorly to unseen sensor locations. To build a general-purpose foundation model for inertial sensing, we adapt the DINO self-supervised learning (SSL) paradigm to inertial data through a set of semantics-preserving motion transformations that simulate variations in motion pattern and sensor setup. For model architecture, we use a plain transformer encoder with frequency-aware positional encoding that natively supports varying input length and sampling rate. This simple setup proved remarkably effective for learning global and dense features generalizable across users, sensor setups, tasks, and, in the case of robot odometry, embodiments. We hope IMUDINO can serve as a backbone for inertial sensing and kickstart the development of general-purpose inertial FMs.
We validate our previous hypothesis that a language model holds relational patterns apart from the subject domains that instantiate them (Nordfors, 2026). These archetypal contexts are patterns of relations that people recognize across many fields, such as the tipping point or winner takes all, held once and applied across domains. We show this inside the model, Qwen3.8-27B. One set of neurons carries the archetype, say the tipping point; another carries the topic, say politics, ecology or cell biology; a third carries the language, such as English, Spanish or Chinese. Silence the tipping point's neurons, and the model no longer identifies the pattern but still knows what the text is about. Silence the topic's, and it loses the topic but still names the pattern. In the neuron maps the layers organize themselves into a sandwich: the outer layers carry languages, the layers inside them topics, the layers at the heart archetypes. The sandwich emerges spontaneously during pretraining and appears in every model we mapped. Archetypes are the architecture for analogy, and we strongly suggest that this structure of archetypes and topics is the mechanism of a language model's intelligence: we see it self-organize as the model is trained. The neurons, the texts and the code are released with this paper as its reproduction package, so that anyone can repeat the test inside the model.
High-Dimensional Robust Mean Estimation from Doubly Corrupted and Heterogeneous Batches
Maryam Aliakbarpour, Vladimir Braverman, Yuhan Liu, Junze Yin
AI 速览：针对用户级任意污染与样本级小比例污染并存的双重污染模型，作者提出SoS算法在高维协方差受限分布下达到min-max最优的ℓ2误差，并扩展到用户局部均值有偏移的异构情形。
Modern data systems often require aggregating information from multiple sources: some might be entirely untrustworthy, while the remaining presumably reliable sources may still suffer from a small amount of corruption. We study robust mean estimation under this practical scenario, which is formulated through the double corruption model: samples are divided among $N$ users, each with $n$ samples. $\varepsilon$ fraction of the users could be arbitrarily corrupted, while each of the remaining ``good'' users still have $α$-fraction ($α>0$) of corrupted samples. The goal is to estimate the unknown mean $μ$ with small $\ell_2$ error. This is a challenging scenario because we must handle two types of adversarial corruption: user-level corruption and sample-level corruption. When the unknown distribution has covariance matrix bounded by the identity matrix, we design a sum-of-squares (SoS) algorithm that achieves an $\ell_2$ error of $\sqrtα$ with probability at least $2/3$ as long as the total number of samples $nN=\widetildeΩ(d/α)$. The estimation error and sample complexity are min-max optimal under the mild condition. For comparison, treating all $nN$ samples as $(\varepsilon+α)$-corrupted and applying standard robust mean estimation only achieves an error of $\sqrt{\varepsilon+α}$. Thus, our new algorithm removes the adverse effect of untrustworthy users and achieves the optimal error as if only sample-level corruption is present. Also, our SoS algorithm extends to the heterogeneous setting where each user's local mean could be additionally shifted by $\sqrtβ$ in $\ell_2$ distance. We show that an error of $O(\sqrt{α+β})$ can be achieved with probability at least $2/3$ when $nN=\widetildeΩ(d/(α+β))$.
Many high-stakes regression decisions ask three things of a model at once: a point estimate to act on, an interval that bounds the risk of acting, and the option to abstain when neither can be trusted. Deep learning provides these in isolation: prediction-interval methods return an interval, and at best a value estimate inside it, for every input, while selective-prediction methods learn when to abstain, but only for point predictions and with no statement of how far the truth may lie from the estimate. We introduce SelPI, which trains interval, value and abstention heads jointly under a single loss with an explicit acceptance-rate budget. A natural objection is that interval width already tells us when to abstain. We show that it does not: across 18 tabular benchmarks, width predicts large point errors well (AUROC 0.72) but is anti-correlated with interval misses (AUROC 0.38), so abstaining on wide intervals discards precisely the inputs the interval got right. SelPI therefore selects with two gates, one trained on the interval loss and one on the value loss, coupled by a Fr'echet penalty that ties their separate budgets to the joint acceptance rate they deliver. With the base predictor held fixed, the learned gate beats width thresholding on a proper interval score on all 18 tabular datasets, closing a median 87% of the gap to an oracle selector, and it also beats selection by deep-ensemble disagreement on every dataset. The advantage grows with the heteroscedasticity of the task, and we report the settings in which it disappears.
Yifan Sun, Panqi Chen, Weifeng Shi, Xiao Fu, Lei Cheng
Reconstructing three-dimensional (3D) complex ocean acoustic fields from sparse measurements is challenging due to highly oscillatory wave structures and restricted sensing patterns/routes. Learning-based approaches fall into two categories. First, purely data-driven models can learn complex field patterns, but do not explicitly exploit wave physics that may enhance interpretability and reduce training costs. Second, physics-informed methods impose a specified physical equation as a reconstruction constraint, but often require environment-dependent parameters that are difficult to acquire. As a result, practical 3D acoustic field reconstruction that is both physically grounded and robust to uncertain environmental information remains largely unexplored. This work takes an alternative route: instead of solving a fully specified wave equation, it extracts a structured representation from wave physics and learns its components from data. From the Helmholtz equation and normal mode propagation, we derive a structured spatial-angular tensor representation, termed the S/c representation. Under a low-rank specialization, it takes the form of a rank-(L,L,1) block-term tensor decomposition, which connects structured sampling patterns/routes used in ocean sensing to sufficient conditions for field recovery. Building on this tensor representation, we propose TIDAL, a neural reconstructor that uses the S/c representation as an inductive bias for structured decoding and aligns information aggregation with the sampling geometry. The neural learner removes the need for known model parameters required by conventional physics-based recovery methods. Experiments across three structured sampling geometries show that TIDAL achieves the best reconstruction performance and fastest inference among baselines, with benefits from the physics-derived representation under limited data and model capacity.
Aviation authorities hold decades of accident reports whose narratives record how investigators reconstructed each accident. Computational work on these archives has mostly sorted reports into fixed taxonomies. In this work, we propose a hybrid framework in which a text-based corpus identifies the leading factors of a failure mode. Physics-based models then quantify how likely that failure is at a stated operating point. Ordered causal chains extracted from 56{,}202 American accident reports with a local large language model aggregate into a causation knowledge graph. On an evaluation sample of 2{,}000 reports, the chains recover 90.8\% of investigator-flagged causes. Each physical factor whose drivers the record carries receives an occurrence model calibrated to the accident population. A leaky noisy-OR fitted to the corpus combines the factors into a failure mode probability that supports intervention. For engine power loss, the carburetor icing model separates 1{,}205 icing accidents from other accidents using temperature and dewpoint alone, reaching 0.61 in area under the receiver operating characteristic curve against a density ratio ceiling of 0.64. Routine weather hours give no valid control, because flying concentrates in fair weather. On later years the physics adds 0.03 to that area for occurrence and nothing for severity, so its value lies in interpretability. On four named accidents, abduction gives a probability of necessity of 0.29 to 0.50 for the heat omission, under monotonicity. The Fast Causal Inference algorithm does not license the largest backdoor adjustment on the learned networks.
Can a simple labeled-pixel contrastive branch improve semi-supervised segmentation with a strong pretrained encoder? We study this question with PixCon, which adds a class-wise memory bank to a DINOv2-Base implementation of UniMatch V2. The bank stores ground-truth-labeled features only where the student prediction agrees with the label. A multi-positive contrastive loss uses these features to train the shared encoder-decoder; the head and bank are discarded at inference. This simplifies earlier labeled-pixel banks by removing confidence thresholds and learned ranking. On Pascal VOC with 183 labels, PixCon improves all three paired seeds by 0.20, 0.64 and 1.83 mIoU points, raising the mean from 87.01 to 87.90. With 92 labels the mean gain is 0.40 points; single-seed differences on Cityscapes and ADE20K range from -0.08 to +0.14 points. A predicted-label contrastive control reaches nearly the same Pascal accuracy, so these results support the added branch rather than an accuracy benefit from its correctness filter. An exact gradient decomposition distinguishes mislabeled-positive counts from their similarity-weighted influence. Separate teacher diagnostics find retained-pixel error rates of 1.8% on Pascal and 10.6% on ADE20K. These diagnostics do not measure the training bank or predict final accuracy. PixCon adds 15.3-19.7% iteration time in the measured runs and no inference computation. We report paired seeds, training cost and the limits of the evidence.
Off-policy deep reinforcement learning trains on random samples from a replay buffer of past experience. In physical robot settings, that buffer is dominated by routine: thousands of near-identical transitions from normal operation, while the rare events that define the edges of the task are rarely sampled. We propose scoring every stored transition by its statistical leverage, which measures how unusual it is relative to the buffer's own data, computed from the stored numbers alone, with no labels, rewards, or value network. We draw each mini-batch from a mixture of this score and uniform sampling. First, the natural way to build the score along the directions where the data varies most turns out to be blind when actions are stored as one-hot indicators, because indicator columns always win the variance ranking; we instead select the feature columns with the heaviest tails (largest kurtosis) where small far-away populations are revealed. Second, the importance weights that prioritized replay uses to keep gradients unbiased cancel the method exactly: with full correction, training matches uniform sampling, so we drop the correction and deliberately optimize the rebalanced objective. Third, when the rare situations of a log are relational, visible only in how a stretch of behavior relates to its goal, not in any single stored coordinate-per-transition scores are blind; applying the same tail rule to short trajectory windows described relative to their goal restores the signal, and on a three-joint arm that must push an object around an obstacle it raises the rare-task success from 0.49 to 0.69 against a labeled-oracle ceiling of 0.74.
Large language model (LLM) agents increasingly rely on external memory stores to retain information across conversations, typically via dense vector embeddings searched with approximate nearest-neighbor methods. The fruit fly's olfactory circuit computes a locality-sensitive hash of sensory input using a strikingly different mechanism - sparse random projection followed by winner-take-all competition - and prior work (Dasgupta et al., 2017) showed this scheme, dubbed FlyHash, preserves neighbor structure better per bit than conventional random-hyperplane hashing (SimHash) on several benchmarks. We apply FlyHash, to our knowledge for the first time, to LLM agent memory retrieval, and ask three questions a practitioner actually needs answered: does the quality advantage hold under a correctly matched-bit protocol (it does, robustly, across 10 seeds and four datasets spanning two modalities, with one specific, previously undocumented comparison pitfall identified and corrected along the way); does that advantage translate into real memory and latency savings, including against the dense-vector indexing (HNSW) practitioners default to today (no: at equal bits FlyHash's codes are larger and slower to compare than SimHash's, and HNSW dominates both hash methods at ordinary scale; a sparse-set comparison we introduce recovers most but not all of a naive latency gap, leaving hashing's practical case narrower and more specific than the quality result alone implies); and does the advantage survive genuine, non-circular agent-memory relevance labels rather than the nearest-neighbor-by-embedding proxy this literature otherwise relies on (it does). We report all three answers precisely, including the two that narrow rather than flatter the headline result, and release our implementation, a working agent-memory demo, and a full account of the methodological pitfalls we found and corrected.
Sepehr Mousavi, Levi Lingsch, Siddhartha Mishra, Laura De Lorenzis
Neural operators fit their representation of functions to the task they are trained on, so every new task rebuilds a reduced space from labeled data. Following the self-supervised pre-training that made representations shareable in natural language processing and computer vision, we decouple representation learning from operator learning. A functional autoencoder is pre-trained on analytic functions synthesized on the fly, with no solver and no labels, to represent functions on arbitrary point clouds, and the latent space it induces is frozen and reused across tasks. The representation is exactly equivariant to translation, dilation, and amplitude, and reconstructs solutions of partial differential equations (PDEs) and natural images on geometries it never saw. Being task-agnostic, the latent space can be interpreted in its own right. The whole space is used and decodes to valid fields, it preserves the ordering of distances between physical fields, its geodesics interpolate physical evolutions, and edit directions transfer across fields. With frozen representation maps, a dataset is encoded once and a neural operator is trained entirely on cached embeddings, which saves memory and training time. On 28 tasks in fluid dynamics, solid mechanics, acoustics, and the Helmholtz and Poisson equations, this paradigm can exceed the accuracy of state-of-the-art end-to-end architectures, with up to an order of magnitude higher throughput and memory savings that fit models nearly two orders of magnitude larger on consumer-grade GPUs. At inference, latent rollouts run up to an order of magnitude faster than regular rollouts, and chunked inference processes functions of over half a billion points on a single consumer-grade GPU.
Justin Tahmassebpur, Asadullah Bhuiyan, Hyejin Kim, Omri Lesser
Learning from imperfect data is a central theme in machine learning, connecting practical questions of robustness to fundamental questions of learnability. Here we examine attribute noise: learning from corrupted inputs while keeping the labels intact, a setting that has received considerably less analytical attention than its label-noise counterpart. We consider two types of corruption models: additive noise and replacement noise. Through experiments with multi-layer perceptrons (MLPs) on corrupted classification datasets, we find that neural networks remain robust, maintaining well-above-chance accuracy even when inputs are >90% corrupted -- far beyond human recognition. To understand this robustness, we analyze infinite-width networks in the heavy-corruption regime using a mean-field-inspired approach and derive a leading-order decision rule for the classification outcome: the network implements a prototype rule, the nearest-class-mean, assigning each test point to the class whose training-set average it most closely resembles. This leading-order decision rule is universal across a broad range of MLP architectures, holding for any depth, as well as a wide class of activation functions and noise distributions. The same centroid mechanism closely matches finite-width network behavior in our experiments and provides an interpretable and analytically tractable account of why learning can succeed even when individual training examples carry almost no signal.
$Φ$asmaNet: A Staged Spectral Token-Refinement Architecture for Fast Gravitational-Wave Signal-to-Noise Ratio Prediction
Ioannis Tsioulis, Anastasios Tefas, Nikolaos Stergioulas
AI 速览：作者提出紧凑神经网络 ΦasmaNet，以分阶段谱精炼为显式架构先验，快速预测进动高阶模式双黑洞信号的相位平均最优信噪比。
Quantifying selection effects in gravitational-wave population studies requires millions of costly signal-to-noise ratio (SNR) calculations. We present $Φ$asmaNet, a compact neural architecture for structured tabular regression, with staged spectral refinement as an explicit architectural prior. Learned input tokens are refined through ordered attention-based $Φ$-Stages, each augmented by a $Φ$-Refractor subspace-mixing module, then revisited through gated top-down $Φ$-Feedback. Applied to the 11-dimensional parameters of precessing, higher-mode binary black hole signals (IMRPhenomXPHM), it predicts phase-averaged optimal SNR for a single detector at fixed Advanced LIGO design sensitivity. With approximately 732k parameters, it outperforms MLP, 1D-ResNet and Transformer-encoder baselines under matched training across $10^5$--$10^6$ samples. At $10^6$ samples, it achieves $100-\mathrm{MAPE}=98.73\%$ on the held-out test set, $99.29\%$ after rank-preserving monotonic (PCHIP) calibration, and $99.39\%$ ($R^2=0.9998$) in the detection-relevant regime $ρ>8$, with stable accuracy across independent dataset realizations. On a single GPU, it sustains approximately $1.2\times10^5$ SNR evaluations per second and processes a $10^6$-injection catalogue end-to-end in approximately 8.3 s, about two orders of magnitude faster than conventional waveform-and-inner-product calculations on a 16-core CPU. Exact distance rescaling through $ρ\propto1/d_L$ supports repeated catalogue-wide detectability evaluations. The fully documented pipeline is intended as a reproducible benchmark, while results on public tabular datasets indicate competitiveness beyond gravitational-wave applications. Although demonstrated for a fixed single-detector configuration, the construction supports extensions to detector networks and epoch-specific sensitivities. (Abridged)
Power law graph attention: exact generalization of scaled dot-product attention, empirical collapse at inference
Burc Gokden
AI 速览：PLDR-LLM 的幂律图注意力（PLGA）在理论上精确包含缩放点积注意力（SDPA），但作者报告其头评分算子在推理时趋于输入不变，导致推理坍缩为每头使用常数算子的广义 SDPA。
The Large Language Model from Power Law Decoder Representations (PLDR-LLM) and its attention, Power Law Graph Attention (PLGA), replace the fixed bilinear form of scaled dot-product attention (SDPA) with a learned, input-generated bilinear operator $G_{\mathrm{LM}}$, built from a positive tensor $A_{\mathrm{LM}}$ by elementwise power laws. The architecture is fully specified, verified against pinned reference releases; claims are labeled theorem, conditional theorem, measurement, or conjecture. Unconditionally: PLGA contains SDPA exactly at $G_{\mathrm{LM}}=I$; $A_{\mathrm{LM}}$ and $A_{P}$ are strictly entrywise positive, with Perron-Frobenius structure on $A_{\mathrm{LM}}$; the DAG regularizer has the NOTEARS walk-counting form and positivity obstructs exact acyclicity; and, under nonresonance (satisfied by standard rotary frequencies), a commutant criterion identifies which operators preserve relative-position dependence. An inference-collapse theorem: exact input invariance of the head score operators $G_{\mathrm{LM}}$ collapses inference to generalized SDPA with a constant operator per head. Measured invariance: relative fluctuations of $10^{-6}$ and below; perturbation bounds quantify but do not certify cached inference; the assembled proxy misses the decoding margin. A conditional three-stage mechanism (rotary twirl, concentration, row-map contraction) is measured on a released checkpoint. Blockwise training and scoring under the global Gram are stated with explicit target exposure; on tested samples, block and sequential scoring select identical answers and agree on the published TruthfulQA probability-mass metric within $5\times10^{-5}$ per item. Self-organized criticality enters as a phenomenological framework with an intrinsic order parameter; open claims become falsifiable conjectures. Selected proof cores are machine-checked in Lean 4.
Two methods are presented that perturb a given learning rate schedule in stochastic gradient descent with cross entropy loss. Both methods use an extra forward pass to estimate various derivatives of the minibatch loss. The first method selects a value $η$ by linearising the softmax values and minimizing the squared error $\tfrac 1 2 ||y - q - ηq'||_F^2$, where $q$ is the matrix representing the softmax values of the minibatch and $y$ is the matrix with one-hot encodings of the labels. The second method considers a quadratic Taylor approximation of the minibatch loss under the outer approximation of the Hessian. Thus, it minimizes $L(0) + ηL'(0) + \tfrac 1 2 η^2 L''(0)$. Here, the derivatives $L'$ and $L''$ are computed from the derivatives of the logits (using the outer approximation of the Hessian of the loss). We observe that both methods give very similar values of $η$. Finally, these estimated $η$ are used to slightly increase or decrease the learning rate of an existing schedule. The method is evaluated on Cifar10 and Cifar100 datasets for training Resnets, where they decrease the loss faster than the schedule that is perturbed and increase monotonicity. For Cifar10 it also leads to a small improvement of test accuracy, but in Cifar100 test accuracy remains the same.
A language model that learns a new fact by fine-tuning can repeat it but cannot reason with it: taught that a person is a citizen of a new country, it still cannot say, without chain-of-thought, who the head of state of that person's country is. This two-hop curse is widely treated as a reasoning deficit, and the strongest knowledge editors answer it with synthesized reasoning data. We show that the reasoning is intact and the writing is wrong. Fine-tuning writes a new fact as an answer to be said, whereas the second hop reads the bridge ("that person's country") as an entity, from the middle layers on: patching the model's own representation of the entity into the bridge at those layers largely restores the second hop, yet after fine-tuning the bridge there is no closer to the new entity than before the edit. We therefore propose read-equivalent writing (Legible): one loss term makes the bridge phrase, at the layer where the second hop finishes reading it, match the model's own reading of the new entity, and keeps the same relation's other subjects unchanged. Phrases and targets come from the edited model itself, and no multi-hop training data is used. Across Qwen2.5, Llama-3 and Qwen3, Legible (I) composes learned facts more often than the model composes its own knowledge, on MQuAKE's two-hop-curse questions; (II) reaches 1.8-2.6x the bridge accuracy of CaKE, whose reasoning data composes no better than plain fine-tuning, and has the best average over composition, accuracy and locality against four families of writers; and (III) lifts the curse on RippleEdits, on the benchmark of Balesni et al., who named the curse, and at 14B.
Can a search over innate drives find intrinsic rewards that help a reinforcement-learning agent learn, transfer to a task it never saw, and resist an exploit that was absent during the search? Each candidate reward combines four generic events (a new view, a pickup, a drop and a toggle), each with a weight, a relief and a refill rate. We ran three pre-registered stages; the search protocol was locked in a public, timestamped commit before any search result existed. CMA-ES and an equal-budget random search used 1,560 PPO training runs, and the final evaluation and tamper test used 340. On the held-out task, KeyCorridorS3R2, the CMA-ES-selected reward (E) beat task-only PPO, episodic novelty and the random-search-selected reward (R) in all 3 registered comparisons, all of them meaningful by the registered size rule. Among these generic rewards E transferred best; the earlier paper's satiation reward, which on this task is given hand-wired key and locked-door events, did better (20 of 20 seeds solved against 4 of 20; descriptive). On UnlockPickup, where that reward uses generic events, E's mean area exceeded it by 0.276 (descriptive). On the training tasks there was no registered difference between E and R: both E-over-R tests missed. A free reset action, absent during the search, collapsed E's learning on DoorKey-8x8 (mean area 0.829 without it, 0.136 with it; registered) to below task-only PPO (descriptive). The searched drives helped on every task tested, but a reward given task-specific events did better on the held-out task, and selection gave E no protection against an exploit absent from the search.
Minsu Kim, Ye-Sung Kim, Hyeseong Jeon, Wooseok Hyung, Joshua Lee, Chang-Hwan Im
Electroencephalography (EEG) provides a non-invasive measure of ongoing neural activity, but building general-purpose EEG models remains challenging due to the heterogeneity of subjects, devices, and electrode montages. Existing EEG foundation models predominantly rely on reconstruction-based objectives defined on the observed signal, which contains both neural and non-neural components. We introduce the Spherical Prior EEG Representation Architecture (SPERA), an EEG foundation model that adopts the joint-embedding predictive architecture (JEPA) to predict in latent space. SPERA introduces a Legendre-polynomial spatial prior, incorporated into attention to encode varying scalp electrode geometries. Two further components adapt the model to EEG: factorized temporal and spatial attention interleaved with periodic full-attention blocks, and a relational spectral regularizer aligning latent similarity structure with spectral views. Pretrained on approximately 80,000 hours of EEG from 29,048 subjects across 106 datasets, SPERA achieves the highest average balanced accuracy across nine downstream tasks spanning clinical, cognitive, and BCI applications. SPERA further exhibits strong parameter efficiency under linear probing and robustness across varying recording conditions, suggesting its potential as a general-purpose backbone for diverse EEG analyses.
Low curvature trace, an effective rank condition, is a common explanation for why randomized zeroth-order methods scale to high-dimensional tasks such as language model fine-tuning. Can deterministic methods achieve the same savings? We show that they cannot. We study finding $\varepsilon$-stationary points of $L$-smooth nonconvex functions on $\mathbb{R}^d$ from exact function values, assuming that below the starting value the Hessian is dominated by a fixed matrix with trace at most $LE$; this is a fixed-matrix form of the effective rank condition used to explain zeroth-order fine-tuning. When $ET\le d$, the Gaussian two-point method that such fine-tuning uses needs only $O(ET)$ calls if it rejects steps that raise $f$, where $T=LΔ/\varepsilon^2$ is the iteration count of gradient descent from initial gap $Δ$, and we prove that every randomized method needs $Ω(ET)$ calls, so this cost is optimal up to constants. In contrast, every deterministic method needs $\widetildeΩ(d\min\{T,\sqrt{ET}\})$ calls, and adaptive subspace search attains $O(d\min\{T,\sqrt{ET}\})$: a deterministic method pays about $d$ calls per discovered direction, and the trace budget allows about $\sqrt{ET}$ directions. Determinism thus multiplies the cost by a factor of order $d/(E+\sqrt{ET})$ up to logarithmic factors, which grows linearly with $d$ for fixed $E$ and $T$; once $E$ exceeds $T$ by a logarithmic factor, deterministic methods need $Ω(dT)$ calls, as without any trace budget. Our lower bounds hold for $C^\infty$ functions with the matrix bound on all of $\mathbb{R}^d$, even when each call also returns gradient projections priced by rank.
Mikhail Rudomanov, Ivan Fomin, Alexander V. Gasnikov, Fedor S. Stonyakin
We study the minimization of the average of $n$ convex $(L_0,L_1)$-smooth functions: the Hessian norm of each function is bounded by an affine function of its gradient norm. We assume known upper bounds $R$ on the distance from the starting point to a minimizer and $δ$ on the difference between the minimum of the average and the average of the infima of the components. We derive upper and lower bounds on the oracle complexity of reaching accuracy $\varepsilon$ in function value in expectation. Up to logarithmic factors, both bounds are of order $n(1+(L_1R)^{2/3}) +\sqrt{\frac{n(L_0+L_1^2δ)R^2}{\varepsilon}+n(L_1R)^2}$. The upper bound is attained by the proposed method YETI. The method combines ideas of accelerated proximal methods and variance reduction methods with a heuristic for choosing additional points. The number of oracle calls made by the heuristic is bounded, which preserves the upper bound on the total number of queries needed to reach the target accuracy. In the experiments, YETI is compared with first-order methods on synthetic problems and Poisson regression problems. We also examine the stopping criterion and the effect of the heuristic separately. The lower bound is obtained by an explicit construction of families of instances that are hard for any first-order method.
Reinforcement learning environments are now a primary lever for improving large language model (LLM) capabilities in post-training, yet most agentic benchmarks remain static: the world moves only when the agent acts, the reward is a terminal verdict, and the pass bar is set arbitrarily. We introduce StoreBench, a live-commerce environment in which an agent runs a mid-size online apparel store on a production-grade commerce backend, testing long-horizon planning and economic judgment under uncertainty. Customers order around the clock, suppliers reprice and fail, and market shocks arrive with partial or no warning. The agent acts through the same 29 merchant tools a human operator would use, under a windowed operation budget that makes simulated time a function of actions taken, so model latency cannot influence simulated time. Pass thresholds are calibrated against scripted anchor policies, the reward is hardened against a catalogue of reward hacks, and every episode replays identically given a sequence of actions. We evaluate seven frontier LLMs on 11 scenarios of 30 to 45 days and a full simulated year, over three world seeds at matched reasoning effort. No model matches the scripted smart-triage policy on average: the best, DeepSeek-V4-Pro, passes 49% of task-seed cells against the heuristic's 97%. Human experts working through the same tools and budgets outscore every model (mean composite 0.708 vs. 0.700). Over a full simulated year under the Claude Code harness, most models show dramatic performance improvement. In a GRPO post-training run, Qwen3.5-27B trained on only five disjoint tasks raises its mean composite on the held-out evaluation tasks from 0.136 to 0.373. We release five example training-split tasks, ten sample trajectories, and the scoring and verification tooling; the full environment and evaluation suite are withheld to keep the benchmark uncontaminated.
Boris Radovič, Istemi Ekin Akkus, Marco Canini, Ruichuan Chen, Alice Dethise, Ivica Rimac, Veljko Pejović
Training models over data distributed across edge devices presents a trade-off: centralizing data simplifies training but raises privacy and regulatory concerns, while full local training may exceed the devices' compute and memory capacity. Split learning addresses this challenge by partitioning the model between clients and the server, allowing part of the computation to be offloaded. However, existing approaches typically rely on static split configurations or simplified assumptions that fail to capture the heterogeneity and complex dynamics of real deployments. We present DynSL, a framework that dynamically allocates model layers between clients and the server by continuously monitoring system metrics and periodically solving an optimization problem that accounts for client and server computation and communication. When beneficial, the system reconfigures the split, enabling the offloading strategy to adapt to changing device capabilities and system conditions. Extensive experiments on a testbed of 24 heterogeneous edge devices show that DynSL reduces training time by up to $\sim\!$30% over the strongest adaptive baseline and by more than 50% over a suboptimal uniform split, with no loss -- and sometimes a gain -- in accuracy.
How fast AI inference prices have fallen depends on how an index holds quality constant. We build quality-adjusted price indices for text and code inference sold through APIs, joining 21,024 posted prices to a capability scale fitted to 4,605 benchmark scores. From January 2025 to July 2026, matched-model methods record an annual fall of 0.07 in log price per token. Our pre-specified hedonic index falls at 0.51 a year in logs (95% interval 0.11 to 0.89), entirely through the rising capability of the models on sale. Across three treatments of measurement error and two anchor sets it runs from 0.34 to 0.89. Excluding six anchors flagged for possible contamination leaves model rankings intact (rank correlation 0.998) yet lowers the rate by 0.40 to 0.51 under two treatments. Agreement on rankings therefore does not license claims about levels or increments. Per completed task, the data do not establish a quality-adjusted decline.
Kai Shao, Ling Jian, Liming Liu, Serguey V. Bogomolov, Xinmin Wang
Multiproduct pipelines transport refined products to delivery stations in sequential batches, and their detailed scheduling must meet market demand while respecting nonlinear hydraulic constraints. Retaining the full hydraulic nonlinearity yields a mixed-integer nonlinear program that is computationally intractable, while the linearization commonly adopted for solver-based optimization degrades hydraulic accuracy without resolving the heavy computational burden. This paper develops an action-decoupled proximal policy optimization (PPO) framework that efficiently produces detailed schedules while preserving hydraulic fidelity. The scheduling process is cast as a semi-Markov decision process (SMDP) with flexible action durations. To handle the discrete-continuous hybrid action space, a cascaded dual-actor architecture decouples the discrete operating duration from the continuous flow decisions, eliminating the gradient interference that destabilizes conventional coupled-action policies and thereby yielding high-quality schedules far more efficiently. Validated on a real-world multiproduct pipeline in China and benchmarked against a single-actor baseline and a linearized MILP solved by Gurobi, the proposed framework achieves delivery accuracy close to the exact optimum while substantially improving flow rate stability and substantially reducing solution time.
Mahmoud S. Mayaleh, Liban Ali Achkir, Nijat Mehdiyev, Ahmet Ozdil
Lip reading plays a crucial role in applications such as speech recognition, assistive technologies for the hearing-impaired, and biometric authentication. However, performance often degrades when the speaker varies and when certain speech sounds are visually similar. To address these challenges, we propose LipLingo, a robust lip-reading system based on the LipNet architecture, enhanced with a standardized preprocessing pipeline for consistent mouth-region representation, a rotational validation strategy for improved generalization to unseen speakers, and a flexible similarity evaluation for more reliable sentence-level assessment. The model architecture combines 3D convolutional layers for spatio-temporal feature extraction and bidirectional recurrent layers for sequence modeling, optimized using the Adam optimizer and Connectionist Temporal Classification (CTC) loss. On the GRID dataset, LipLingo achieved 95.0\% character-level accuracy (CER 4.99\%) and 87.2\% sentence-level accuracy. At the corpus level, the system obtained a (WER 13.1\%), outperforming the baseline LipNet and demonstrating competitive performance relative to recent 3D-ResNet + Bi-GRU + MixUp approaches. These results provide insights into strategies that enhance lip-reading robustness and highlight directions for future work toward real-world deployment.
Skin cancer classifiers need large sets of annotated images, which are scarce, especially for rare lesion types. AI-generated images can fill the gap, but a generator produces images of widely varying quality, so selecting which images to retain is important. That choice usually rests on a single quality score, most often the Fréchet Inception Distance (FID), which has not been tested against what a classifier needs and cannot tell whether the generator memorizes its training photographs. We test that practice with a tournament of 16 quality metrics. Each metric alone selects a training subset from a pool of 275,472 images generated by a diffusion model (EDM2), and a classifier trained on each subset scores the metric per class. The tournament shows that no single metric serves every class: FID computed on Inception features gives the best classifier for one lesion type and no measurable advantage over random sampling for another, and the generation setting it prefers is also the one closest to the training images. We turn these per-class measurements into a safety-aware evaluation algorithm that weights each metric by its measured value in each class and removes images too close to a real training image. On the held-out ISIC-2018 test set, the algorithm improves Macro-$F_2$, a recall-weighted measure of classification performance, over the FID baseline for a Swin-B classifier by +0.0136 to +0.0164 across four safety thresholds (Bonferroni-corrected $p < 0.02$) and over random sampling by about +0.03, though not for ResNet-50. Tightening the threshold improves a memorization-safety score with no measurable change in utility, but every curated set remains less safe than random sampling. The results show a modest but statistically significant gain on the primary classifier, and offer a practical, safety-aware procedure for evaluating and using synthetic data.
Many overparameterized models that have achieved remarkable success are singular models. That is, the set of optimal parameters consists of multiple points, or the Hessian of the generalization error at such parameters is degenerate, reflecting redundancy in the model parametrization. Therefore, understanding the effective dimension and intrinsic model complexity, beyond the apparent dimension of the parameter space, is essential for studying the generalization ability of such overparameterized models. In this work, we focus on the real log canonical threshold (RLCT) and its localized counterpart, the local RLCT, which are fundamental quantities for characterizing the generalization of singular models. We clarify the theoretical relationship between the local RLCT and existing two key metrics related to generalization under some localization assumption: (I) the rank of the Hessian matrix of the generalization error and (II) the covering number of a family of loss functions. This result provides novel insights into effective dimension and intrinsic complexity of overparameterized models.
Reinforcement learning (RL) is a machine learning paradigm in which an agent learns a decision-making policy by interacting with an environment, guided by a scalar reward rather than explicit supervision. Supervised and unsupervised learning have become indispensable tools in high-energy physics (HEP), but they learn from fixed datasets, whereas many HEP tasks are closed-loop decisions: the data an algorithm observes depend on its choices, the feedback is sparse or delayed, and the objective is often measured on a real machine or in a stochastic simulation, not by a differentiable model. This survey reviews the application of RL to such tasks: accelerator control and feedback, data quality monitoring, jet grooming, clustering, and representation, charged particle tracking, beyond-the-Standard-Model (BSM) and string model building, detector design, trigger control, analysis, and formal quantum field theory computation. It provides a self-contained introduction to the underlying RL methods, from model-free and model-based to multi-agent, offline, inverse, and bandit approaches, reviews each domain against its traditional methods, and summarizes the head-to-head comparisons. The applications range from policies deployed on real accelerators and evaluated on recorded collision data to exploratory studies in formal theory. RL is most advantageous for repeated tasks with a fast simulator, decisions inside fast control loops, operation under faults and corrupted measurements, physics-based rewards in place of labels, and sequential discrete constructions with a reusable policy, while Bayesian optimization, genetic algorithms, and differentiable methods remain competitive elsewhere. We conclude with the open challenges that recur across domains, including reward design, sim-to-real transfer, safety, the lack of shared benchmarks, and the statistical validity of policy-driven decisions.
Studies of neural network generalization and robustness often summarize smoothness with a single number, such as a Lipschitz constant or a Hessian curvature statistic. However, Lipschitz constants summarize global worst-case sensitivity and Hessian statistics summarize local curvature; neither captures how model responses vary with the size, direction, and location of perturbations encountered in real-world deployment. To measure smoothness under these perturbations, we introduce a finite-difference roughness spectrum that profiles model responses across these conditions. Our roughness spectrum generalizes scalar smoothness measures, bringing Lipschitz sensitivity and directional Hessian curvature into a common framework across perturbation scales. We connect this spectrum to classical Besov theory, which describes how a function's variation changes with scale. Our theoretical analysis also proves that a larger output shift can cause less task damage than a smaller one, and that local curvature alone can miss the blocks most vulnerable to finite perturbations. In image classification, Roughness Source Control (RSC) suppresses margin roughness along noise directions and improves noise robustness in a comparison of 80 checkpoints across four model--dataset combinations. In post-training quantization, we identify blocks that need higher precision by measuring task-loss increases from perturbations matched to the size and direction of each block's quantization error. Our mixed-precision allocation achieves lower perplexity than the Hessian-based HAWQ-V2 baseline under the same GPTQ weights and bit budgets in all 14 settings across seven held-out large language models.
Sayantan Choudhury, Xiaoran Cheng, Martin Takáč, Sen Na, Mladen Kolar
Most first-order optimizers treat matrix-valued parameters as vectors, ignoring the intrinsic geometry of hidden-layer weights in neural networks. Muon addresses this mismatch by updating along the polar factor of a momentum matrix, but its theoretical understanding has lagged behind practice. In particular, practical implementations incorporate Nesterov momentum, compute the polar factor only approximately, and operate with stochastic gradients that may be heavy-tailed. We close this gap by developing a convergence theory for Muon with Nesterov momentum and inexact polar decomposition in non-convex matrix optimization under heavy-tailed noise. Our analysis builds on a unified framework for inexact polar decomposition that captures practical iterative approximations such as Newton-Schulz and quantifies how their errors propagate through the optimization dynamics. Under uniform bounds on polar errors, we establish an optimal iteration and sample complexity of $O \left(\varepsilon^{\frac{-(3α-2)}{(α-1)}} \right)$ for finding an $\varepsilon$-stationary point, where $α\in(1,2]$~denotes the heavy-tail index. For the inexact-polar setting with $σ_1=0$, we also provide guarantees that do not require prior knowledge of $α$. We analyze a randomized low-rank polar decomposition that is substantially more efficient than full-space methods while remaining compatible with our theory. Numerical experiments further demonstrate the effectiveness of the proposed inexact and randomized~variants.
The rapid growth of electric vehicles (EVs) in urban environments demands intelligent, decentralised charging management strategies that operate under strict real-time constraints. This paper proposes an Edge-Intelligent architecture in which each EV selects a charging station (CS) by querying a shared, centrally trained Deep Q-Network (DQN) policy over a lightweight MQTT publish-subscribe backbone; independent per-EV edge inference is the deployment target, not the configuration evaluated here. The reported simulation evaluates that shared DQN policy applied sequentially to each vehicle's observations within a single SUMO run, using MQTT for KPI telemetry while observations are obtained through TraCI. The system is evaluated on two real-world road networks from OpenStreetMap (Rabat Agdal-Hay Riad and Tangier, Morocco); the primary network carries four geolocated public charging stations with heterogeneous power levels (50 kW and 150 kW). Three traffic demand profiles (morning, midday, evening peak) are simulated, and four decision policies are compared: an Adaptive Heuristic, a Tabular Q-Learning agent, the proposed DQN agent, and a Double DQN variant. A multi-seed statistical protocol (7 seeds, Wilcoxon signed-rank test, Cohen's d) evaluates the results. The DQN agent achieves a 6.8% relative improvement in charge rate and a 5.4% increase in delivered energy over the Adaptive Heuristic under Evening Peak demand ($p = 0.016$, Cohen's $d = 0.99$ and $1.51$), while showing no advantage in waiting time; on the second network the heuristic leads on waiting time, so the ranking is not stable across geographies. A scalability analysis shows linear state-space growth with station count and sub-millisecond inference; all runs include intermittent station outages under a Markov failure process.