让下一篇重要论文，不再错过。

We study two variants of the mirror descent-ascent (MDA) algorithm for solving min-max problems on the space of measures: simultaneous and alternating. We work under assumptions of convexity-concavity and relative smoothness of the payoff function with respect to a suitable Bregman divergence, defined on the space of measures via flat derivatives. We establish non-asymptotic convergence rates to mixed Nash equilibria, measured in the Nikaidô-Isoda error, proving an $\mathcal{O}(N^{-1/2})$ rate for simultaneous MDA and an improved $\mathcal{O}(N^{-2/3})$ rate for alternating MDA. The main technical contribution is an infinite-dimensional dual space analysis that relates Bregman divergences on measures to dual Bregman divergences on spaces of bounded continuous functions, allowing us to control asymmetric commutator terms created by alternating updates. The results substantially generalize prior analyses restricted to bilinear objectives and also apply to nonlinear convex-concave problems on measure spaces, thereby providing a unified theoretical foundation for MDA in mean-field min-max optimization.

Linear attention has recently gained significant attention for long-context inference due to its constant decoding cost with respect to context length. However, existing serving systems typically serve linear attention by recurrently computing and updating a large linear attention state in every decoding step. Since the state is much larger than the per-token key and value, recurrent decoding incurs substantial memory access and becomes inefficient for serving linear attention. In this paper, we propose KVBuffer, an IO-aware serving mechanism for linear attention. By buffering recent keys and values, KVBuffer enables serving systems to compute linear attention outputs in more flexible and memory-efficient ways. For decoding, KVBuffer enables chunkwise computation, which reduces average memory access and decoding latency by deferring state updates and applying them in batch. For speculative decoding, KVBuffer verifies draft tokens in parallel and avoids storing temporary states. For short contexts, KVBuffer computes attention outputs directly from buffered keys and values, without creating or updating the linear attention state. We implement KVBuffer in SGLang for Qwen3.5 and Nemotron-3 series models. Our evaluations show that KVBuffer can speedup batch decoding by up to 1.47x and increase the maximum number of serving requests by 5x for speculative decoding when verifying four draft tokens.

Weather fronts are transition regions between air masses with contrasting thermodynamic properties and are associated with mid-latitude precipitation and severe weather. We investigate machine learning (ML) methods for detecting weather fronts and quantify the contribution of reanalysis data to performance. We compare two configurations of the same ML model for pixel-wise frontal classification over the North Atlantic and Europe: (1) using only geostationary infrared satellite observations, and (2) additionally using 28 ERA5 reanalysis fields. The satellite-only model achieves a neighbourhood critical success index at a 150 km tolerance (CSI@150) of $0.402$, which lies within the range of agreement between analyses from different meteorological services. The satellite$+$ERA5 model achieves $0.575$, an improvement of $0.173$. Our analysis shows that low-level winds are the least predictable from the satellite representation ($R^2 = 0.23$, compared with $0.88$ for temperature), while removing winds or moisture causes the largest decreases in detection performance.

Robot reinforcement learning increasingly runs environments outside the trainer: in another process, on another machine, in another language. Many faults at this boundary raise no error. We make the boundary a wire protocol and ask what notices a fault in it. On Pendulum, HalfCheetah, Hopper and a 7-joint arm (Pusher), we injected up to 23 faults into what crosses the wire and 2 into the trainer's log, from one value to every value. Unless every value was wrong, a learning-curve threshold caught at most one run in ten, and a comparison with ten fault-free seeds at most 21%. A clean evaluation of all 1,060 final policies agreed with the curves: of the 728 boundary-fault runs both curve checks passed, 723 scored no more than three standard deviations below fault-free seeds, though nearly all ended on other weights and, for reward faults, logged other returns. Checks of the run itself, not of its outcome, saw every fault. A step ledger kept on both sides and compared step by step caught and located every fault that crossed, on Pendulum, Hopper and Pusher on one machine and on Pendulum across two, with no fault-free run and, on Pendulum, a median 3% more training time. An episode record kept at the environment caught the faults in the trainer's log. We list such faults found in practice, in Gymnasium 1.3 and our own code and, reported by others, in five multi-agent libraries, and recommend a step ledger and episode counts kept at the environment.

Anomaly detection (AD) is often framed as a representation-learning problem, where strong performance is expected to require learned features or AD-specific optimization. However, this view can overlook the intrinsic separability of anomalies from normal references, even in the absence of learned representations. We therefore investigate how far untrained neural networks can go in anomaly detection. To this end, we propose PRISM, a framework combining fixed random nonlinear features with normal-reference Mahalanobis scoring for tabular detection and dense visual localization. Across 47 tabular datasets in ADBench, PRISM-T achieves the highest mean AUROC (86.4%) and the second-highest mean AUPRC (65.9%) among the compared methods. On MVTec AD and VisA, covering 27 visual categories, PRISM-V achieves mean image-level and pixel-level AUROCs of 90.4% and 96.4%, respectively. Notably, both models achieve these results without neural weight optimization or pretrained representations. A theoretical analysis further characterizes how nonlinear transformations affect the mean and covariance contributions to the Mahalanobis score gap, while empirical analyses examine the resulting score separability. These findings establish PRISM as both a competitive anomaly detector and a prior-to-learning baseline for assessing the added value of learned representations in anomaly detection.

Self-supervised training is monitored without labels, through the loss and statistics of the learned features. We report a case where we know the encoders lost part of what they had learned, so we can check whether these signals noticed. We trained LeJEPA encoders on synthetic images with known generating factors. In nine of ten seeds, the readability of those factors fell abruptly and stayed down, nearly always along the same direction in factor space. Meanwhile the loss kept falling, and most of the eleven label-free signals we tracked kept improving and rated the declining runs better than runs that did not decline. Standard change-point detection rarely flagged the decline near its onset, and selecting checkpoints by these signals mostly chose ones after it.

Large language models are classically trained in stages: pretraining on raw text followed by post-training for instruction following and reasoning. However, this separation creates a fundamental limitation: many desirable behaviors such as safety, factuality, overall generation quality, and reasoning ability are only added at a late stage, even though the patterns learned earlier strongly shape a model's capabilities. To tackle this issue, we introduce a new way to pretrain and mid-train models that incorporates these behaviors earlier. We utilize an existing strong, post-trained model to both rewrite data and to judge policy model rollouts, thus introducing thinking and reinforcement earlier in training. We further improve reasoning at all stages by introducing the concept of self-improving post-thinking: improving the model's reasoning and self-judging ability by reasoning after generation. In our experiments, we show these approaches can give strong gains across multiple axes.

State-space models (SSMs) recover low-dimensional latent dynamics from high-dimensional observations, but latents are generally not identifiable from observational data alone. Interventional state-space models (iSSMs) use perturbations to break observational symmetries and identify parameters up to permutation, scaling, and shift. However, prior work restricts identifiability to linear latent dynamics and piecewise-linear emissions, excluding many behaviors characteristic of neural circuits. We extend iSSMs to a substantially broader model class: we prove that (i) affine identifiability holds under general injective emission functions and Gaussian latent transitions, and (ii) with interventions, identifiability up to permutation, scaling, and shift is preserved under general nonlinear latent dynamics. We validate our theory on simulated FitzHugh-Nagumo dynamics, and then demonstrate two applications: adjudicating between competing models of decision-making; and predicting responses to unseen micro-stimulation in macaque and human neural recordings, with proof-of-concept generalization to a held-out stimulation site in the latter. Finally, simulations yield practical guidance for designing stimulation protocols that improve identifiability.

Many quantum software development kits provide a suite of circuit optimisation passes. These passes have been highly optimised and tested in isolation. However, the order in which they are applied is left to the user, or else defined in general-purpose default pass sequences. While general-purpose sequences miss opportunities for optimisation which are particular to individual circuits, designing pass sequences bespoke to particular circuits requires exceptional knowledge about quantum circuit design and optimisation. Here we propose and demonstrate training a reinforcement learning agent to compose optimisation-pass sequences. In particular the agent's action space consists of passes for two-qubit gate count reduction used in default PyTKET pass sequences. For the circuits in our diverse test set, the (mean, median) fraction of two-qubit gates removed by the agent is $(57.7\%, \ 56.7 \%)$, compared to $(41.8 \%, \ 50.0 \%)$ for the next best default pass sequence. In doing so our model demonstrates the ability to forgo reward to improve the final two-qubit gate count. Our approach clearly separates the problem of designing circuit optimisation passes from the problem of using them optimally in practice. We expect this separation of concerns to facilitate innovations in both areas.

Generative creative interfaces routinely report the user's position within a space of possibilities. Such displays imply that the quantities shown were measured from the artefact; in a growing class of systems they are instead predicted by a language model describing its own output, and the two are indistinguishable on screen. Most creative domains lack an agreed measurement against which such a self-report can be checked; audio admits one. In a measurement study of a working LLM-driven timbre interface, what a user can cause proves separable from what the interface reports. The first is exactly characterisable: the reachable set is a zonotope in parameter space whose dimension increases by one per sound placed, so the space the model may act within is demonstrably the user's own. The second is inaccurate. The displayed position differs from measurement by approximately three times the width of that region, and the error has two independent sources: model-predicted coordinates, whose accuracy varies twofold between the two models evaluated, and the interpolation step, which arises from combining synthesiser parameters and is unaffected by model quality. End to end, 27 of 36 refinement instructions produced no perceptible change, and under the weaker model the facility performed no action while reporting convergence. Effective agency (what a user can cause) and epistemic agency (whether they can know it) are therefore separable and independently engineered: structurally enforced constraints survive model failure, whereas guarantees resting on the model's own report do not. The failure mode that remains is silence presented as success. We close with a short protocol for building and auditing such displays.

Deep neural networks trained with different architectures, objectives, and datasets have been reported to converge on similar visual representations. However, what remains unknown is which visual properties models actually converge on and which factors may underlie this convergence. To address this, we decompose the object similarity structure of 162 diverse vision models into a small set of non-negative dimensions. To determine universal versus model-specific dimensions, we then estimate how often each dimension reappears across models. In contrast to model-specific dimensions, universal dimensions are more interpretable and more strongly driven by conceptual image properties, indicating the relevance of interpretability and semantic content as implicit factors driving universality across models. Differences in architecture, objective function, training data, model size, and model performance do not explain the emergence of universal dimensions. However, models with more universal dimensions also better predict macaque IT activity and human similarity judgments, suggesting that universality reflects representations relevant to biological vision. These findings have important implications for understanding the emergent representations underlying deep neural network models and their alignment with biological vision.