搜索结果：last

Neural networks are typically optimized with variants of stochastic gradient descent. Under a squared loss, however, the optimal solution to the linear last layer weights is known in closed-form. We propose to leverage this during optimization, treating the last layer as a function of the backbone parameters, and optimizing solely for these parameters. We show this is equivalent to alternating between gradient descent steps on the backbone and closed-form updates on the last layer. We adapt the method for the setting of stochastic gradient descent, by trading off the loss on the current batch against the accumulated information from previous batches. We provide theoretical analyses showing convergence of the method to an optimal solution in the neural tangent kernel regime, as well as quantifying the gains compared to standard SGD in a one-step analysis. Finally, we demonstrate the effectiveness of our approach compared with SGD and Adam on a squared loss in several regression tasks, including neural operators and causal inference.

Recent AI systems have achieved strong results on a wide range of benchmarks, yet these gains have not translated into economically meaningful deployment across many professional domains. We argue that this gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows. This paper introduces Agents' Last Exam (ALE), a benchmark designed to evaluate AI agents on long horizon, economically valuable, real world tasks with verifiable outcomes. Developed in collaboration with 250+ industry experts, ALE covers non-physical industries defined with reference to O*NET / SOC 2018 (the U.S. federal occupational taxonomy). It is organized around a task taxonomy with 55 sub fields grouped into 13 industry clusters covering 1K+ tasks. Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is below 1%. ALE is designed as a living benchmark: its task pool grows continuously as new workflows and industries are onboarded. More broadly, ALE is intended not merely as another leaderboard, but as an instrument for closing the g

Meridian circles played a fundamental role in astronomy since their invention in 1704. Then, at the end of the XX century, this function had been taken over by space astrometry, when the Hipparcos mission demonstrated the advantages of space astrometry by achieving the milliarcsecond level of accuracy. This historical sketch describes the development of the last meridian circles at Pulkovo Observatory (St. Petersburg) in the last quarter of the XX century.

We introduce a deterministic variational formulation for training Bayesian last layer neural networks. This yields a sampling-free, single-pass model and loss that effectively improves uncertainty estimation. Our variational Bayesian last layer (VBLL) can be trained and evaluated with only quadratic complexity in last layer width, and is thus (nearly) computationally free to add to standard architectures. We experimentally investigate VBLLs, and show that they improve predictive accuracy, calibration, and out of distribution detection over baselines across both regression and classification. Finally, we investigate combining VBLL layers with variational Bayesian feature learning, yielding a lower variance collapsed variational inference method for Bayesian neural networks.

It is believed that, under very general conditions, bi-infinite geodesics (or bigeodesics) do not exist for planar first and last passage percolation (LPP) models. However, if one endows the model with a natural dynamics, thereby gradually perturbing the geometry, then it is plausible that there could exist a non-trivial set $\mathscr{T}$ of exceptional times at which such bigeodesics exist. For dynamical exponential LPP, we show that $\mathscr{T}$ is "very close" to being non-trivial; namely, we obtain an $Ω( 1/\log n)$ lower bound on the probability that there exists a random time $t\in [0,1]$ at which a non-trivial geodesic of length $n$ passes through the origin at its midpoint; note that if the above probability were $Ω(1)$, then it would imply the non-triviality of $\mathscr{T}$. We conjecture that, even if $\mathscr{T} eq \emptyset$, it a.s. has Hausdorff dimension exactly zero.

We take the perspective of an advanced high school student trying to understand the proof of Fermat's Last Theorem for the first time. We collect definitions and statements needed to summarise how Fermat's Last Theorem was first proved in 1995 as well as to see how the argument has simplified since then. We include a current timeline, outlines of proofs, background material and recent developments, as well as to organise the content in a way that is beginner friendly, offering a preview of what students may expect to learn more deeply in the future.

We consider Brownian last passage percolation evolving dynamically via a discrete resampling procedure. Using $Γ_{(0,0)}^{(n,n),r}$ to denote a geodesic from $(0,0)$ to $(n,n)$ at time $r$, we prove that the expected total number of coarse-grained changes (or "switches") accumulated by $Γ_{(0,0)}^{(n,n),r}$ away from its endpoints during a time interval $[s,t]$ is at most $n^{5/3+o(1)}(t-s)$; we expect the exponent $5/3$ to be tight. Using the above estimate, we establish that the set $\mathscr{T}$ of exceptional times at which a non-trivial bi-infinite geodesic exists a.s. has Hausdorff dimension at most $1/2$. Further, for any fixed direction $θ$, we show that the set $\mathscr{T}^θ\subseteq \mathscr{T}$ of times at which a non-trivial bi-infinite geodesic directed along $θ$ exists a.s. has Hausdorff dimension equal to $0$.

The last success problem is an optimal stopping problem that aims to maximize the probability of stopping on the last success in a sequence of independent $n$ Bernoulli trials. In the classical setting where complete information about the distributions is available, Bruss~\cite{B00} provided an optimal stopping policy that ensures a winning probability of $1/e$. However, assuming complete knowledge of the distributions is unrealistic in many practical applications. This paper investigates a variant of the last success problem where samples from each distribution are available instead of complete knowledge of them. When a single sample from each distribution is allowed, we provide a deterministic policy that guarantees a winning probability of $1/4$. This is best possible by the upper bound provided by Nuti and Vondrák~\cite{NV23}. Furthermore, for any positive constant $ε$, we show that a constant number of samples from each distribution is sufficient to guarantee a winning probability of $1/e-ε$.

We examine the Foreign Exchange (FX) spot price spreads with and without Last Look on the transaction. We assume that brokers are risk-neutral and they quote spreads so that losses to latency arbitrageurs (LAs) are recovered from other traders in the FX market. These losses are reduced if the broker can reject, ex-post, loss-making trades by enforcing the Last Look option which is a feature of some trading venues in FX markets. For a given rejection threshold the risk-neutral broker quotes a spread to the market so that her expected profits are zero. When there is only one venue, we find that the Last Look option reduces quoted spreads. If there are two venues we show that the market reaches an equilibrium where traders have no incentive to migrate. The equilibrium can be reached with both venues coexisting, or with only one venue surviving. Moreover, when one venue enforces Last Look and the other one does not, counterintuitively, it may be the case that the Last Look venue quotes larger spreads.

In 1949 Fano published his last paper on $3$-folds with canonical sectional curves. There he constructed and described a $3$-fold of the type $X^{22}_3$ in ${\mathbb P}^{13}$ with canonical curve section, which we like to call Fano's last Fano. We report on Fano's construction, providing various (in our opinion missing) proofs, in modern language and trying to use results and techniques available at that time. Then we construct Fano's with modern tools, in particular via the Hilbert scheme of zero cycles on a rational surface; as a consequence we easily point out the corresponding example in the Mori-Mukai classification.

This paper documents the places in mainland Europe at which the sun sets latest, by Coordinated Universal Time (UTC), on any given day (distortion due to differences in local standard times is ignored). In contradiction to the naïve assumption that the sun always sets latest at the westernmost point, the point of last sunset changes cyclically over the course of a year due to the changing orientation of the axis of the Earth with respect to the sun. Specifically, between the winter and summer solstices the last sunset shifts successively from Cabo de Sao Vicente (Portugal) to Cabo da Roca (Portugal) to Cabo Tourinan (Spain) to a site near Aglapsvik (Norway) to a location in the Norwegian municipality of Masoy south of Havoysund; and it shifts back again between the summer and winter solstices. There are two days in the year (April 24th and August 18th) on which the last sunset of mainland Europe (shared in those days effectively by Cabo Tourinan and the Aglapsvik area) coincides with the last sunset of mainland Africa, at a site in Western Sahara near Cap Blanc. A similar analysis of the first Spanish sunrise shows that from April 22nd to August 20th it occurs on the Costa Brava at

Hawaii researchers are giving old fishing nets and recycled plastic a second life by mixing them into asphalt roads。 Early tests found these roads didn't release more plastic particles than standard pavement, with tire wear overwhelming any plastic signal from the recycled material。 If future studies confirm the roads are durable, the technology co

The cosmic microwave background (CMB) fluctuations effectively measure the basic properties of the universe during the recombination epoch. CMB measurements fix the distance to the surface of last scatter, the sound horizon of the baryon-photon fluid and the fraction of the energy density in relativistic species. We show that the microwave background observations can also very effectively constrain the thickness of the last scattering surface, which is directly related to the ratio of the small-scale E-mode polarization signal to the small-scale temperature signal. The current cosmological data enables a 0.1\% measurement of the thickness of the surface of last scatter: $19 \pm 0.065$ Mpc. This constraint is relatively model-independent, so it can provide a new metric for systematic errors and an independent test of the $Λ{\rm CDM}$ model. On the other hand, it is sensitive to models which affect the reionization history of the universe such as models with annihilating dark matter and varying fundamental constants (e.g., the fine-structure constant, $α_{\rm EM}$, and electron rest mass, $m_{\rm e}$) and as such can be used as a viable tool to constrain them.

We study last passage percolation in a half-quadrant, which we analyze within the framework of Pfaffian Schur processes. For the model with exponential weights, we prove that the fluctuations of the last passage time to a point on the diagonal are either GSE Tracy-Widom distributed, GOE Tracy-Widom distributed, or Gaussian, depending on the size of weights along the diagonal. Away from the diagonal, the fluctuations of passage times follow the GUE Tracy-Widom distribution. We also obtain a two-dimensional crossover between the GUE, GOE and GSE distribution by studying the multipoint distribution of last passage times close to the diagonal when the size of the diagonal weights is simultaneously scaled close to the critical point. We expect that this crossover arises universally in KPZ growth models in half-space. Along the way, we introduce a method to deal with diverging correlation kernels of point processes where points collide in the scaling limit.

The goal of this paper is to study the theory of last multipliers in the framework of complex manifolds with a fixed holomorphic volume form. The motivation of our study is based on the equivalence between a holomorphic ODE system and an associated real ODE system and we are interested how we can relate holomorphic last multipliers with real last multipliers. Also, we consider some applications of our study for holomorphic gradient vector fields on holomorphic Riemannain manifolds as well as for holomorphic Hamiltonian vector fields and holomorphic Poisson bivector fields on holomorphic Poisson manifolds.

Last three years have seen new developments in the theory of last passage percolation, which has variety applications to random permutations, random growth and random vicious walks. It turns out that a few class of models have determinant formulas for the probability distribution, which can be analyzed asymptotically. One of the tools for the asymptotic analysis has been the Riemann-Hilbert method. In this paper, we survey the use of Riemann-Hilbert method in the last passage percolation problems.

In this paper we study the sequences defined by the last and the last non-zero digits of $n^n$ in base $b$. For the sequence given by the last digits of $n^n$ in base $b$, we prove its periodicity using different techniques than those used by W. Sierpinski and R. Hampel. In the case of the sequence given by the last non-zero digits of $n^n$ in base $b$ (which had been studied only for $b=10$) we show the non-periodicity of the sequence when $b$ is an odd prime power and when it is even and square-free. We also show that if $b=2^{2^s}$ the sequence is periodic and conjecture that this is the only such case.

In this paper will be proved the existence of a formula to reduce a tetration of base $2^{k}$ and $5^{k}$ $\mod 10^{n}$. Indeed, last digits of a tetration are the same starting from a certain hyper-exponent; In order to compute the last digits of those expressions we reduce them $\mod 10^{n}$. Lots of different formulas will be derived, for different cases of $k$ (where $k$ is the exponent of the base of the tetration). This kind of operation is fascinating, because the tetration grows very fast. But using these formulas we can actually have informations about the last digits of those expressions. It's possible to use these results on a software in order to reduce tetrations $\mod 10^{n}$ faster.