Papers

Despite its elusive character, coherent elastic neutrino-nucleus scattering (CEvNS) detection is evolving as a precision measurement from different neutrino measurements. In this note we show how CEvNS experiments using the Spallation Neutron Source (SNS) are robust measurements that enter into a precision era. Despite the interplay of nuclear and elementary particle physics, we demonstrate that CEvNS current measurements are accurately described by using only one extra parameter, the number of neutrons, $N$. By making this test, we illustrate that current CEvNS are precise enough to reconstruct the number of neutrons in the nuclear target. At the same time, determining the number of nucleons in the target material can be used as a benchmark for the precision and accuracy of this kind of detectors.

We present the first performance-portable implementation of a loop integral library for high-energy particle physics. The new code is based on the QCDLoop package, which provides an essential component in the software stack for perturbative precision calculations at collider experiments. Our implementation extends QCDLoop to modern computing architectures by means of the Kokkos portability framework, which allows to target most existing CPUs and GPUs.

The paper studies $l^p$-wedge sums of metric spaces and provides a two-sided estimate for the pointed Gromov-Hausdorff distance between them. Using an upper bound for the Gromov-Hausdorff distance, the existence of an isometric embedding of $l^\infty$ into the Gromov-Hausdorff space of all non-empty metric spaces is proved.

Autoregressive video diffusion models have become a leading approach to real-time interactive video generation and video world modeling. However, existing methods typically allocate computation uniformly across space when generating the next video chunk, updating all patches with the same denoising schedule and number of function evaluations. This ignores the spatial heterogeneity of video content: static, simple, or recurring regions are often easier to denoise, while other regions require further refinement or additional context. A uniform budget thus leads to redundant computation on easy patches, whereas reducing the fixed step count can leave difficult patches insufficiently refined. We introduce \methodname{}, a general framework that dynamically allocates denoising steps to individual patches: easy patches stop early to improve efficiency, while difficult patches receive more refinement steps. This design accelerates generation with negligible quality degradation. Our method is compatible with most existing few-step autoregressive diffusion models, requiring only a lightweight routing module with modest training. Experiments demonstrate that \methodname{} reduces the total number of denoising steps applied across video patches by about 50% while maintaining visual quality comparable to the four-step baseline.

Prior outdoor LiDAR registration works are commonly limited by the pair-wise input paradigm, which neglects temporal correlations inherent within streaming sequences. In this paper, we propose a novel multi-frame outdoor point cloud registration network for streaming LiDAR scans enhanced with pose-related memory buffers. The key observation is that long-term temporal LiDAR sequences can provide rich global contextual information to complete sparse measurements, filter outliers, and address low-overlap problems, thereby boosting registration performance. Specifically, two hybrid memory buffers are designed, including an implicit memory feature buffer and an explicit memory pose buffer, to store and dynamically update pose-related temporal features. Moreover, a novel dynamic history weighting module is developed to adaptively fuse current and history pose-related features. Extensive experiments on three outdoor datasets, including KITTI, nuScenes, and Apollo-Southbay, demonstrate state-of-the-art performance of MemReg, surpassing all previous pair-wise methods and also multi-frame SLAM systems. Our method also generalizes surprisingly well to multiview indoor registration scenarios with rather competitive performance on 3DMatch, 3DLoMatch, and ScanNet. Code will be released at https://github.com/IRMVLab/MemReg.

We study the sensitivity of future high-luminosity $e^+e^-$ colliders to the electromagnetic and weak dipole moments of the $\tau$ lepton within the Standard Model Effective Field Theory. We analyse several complementary processes across the FCC-ee operating stages and show that the exceptional luminosity and clean environment of future $e^+e^-$ colliders allow for substantial improvements over current limits from LEP and the LHC. The $e^+e^- \to \tau^+\tau^-$ process provides excellent sensitivity to weak dipole operators through resonant $Z$ production and $Z$-pole energy scans, while $\gamma\gamma \to \tau^+\tau^-$ yields the strongest constraint on the anomalous magnetic moment owing to the dominance of linear interference effects in the near-threshold region. Higher-energy stages improve the sensitivity to the $\tau$ electric dipole moment. The radiative processes $e^+e^- \to \tau^+\tau^-\gamma$ and $H\to\tau^+\tau^-\gamma$ provide complementary, though subleading, probes of $\tau$ dipole operators. Combining all channels, we estimate the statistical reach at the 95% CL to be $|\Delta a_\tau|<2.8\times10^{-5}$, $|d_\tau|<2.5\times10^{-18}\,e\,{\rm cm}$, $|\Delta a_\tau^Z|<1.0\times10^{-5}$, and $|d_\tau^Z|<4.4\times10^{-19}\,e\,{\rm cm}$, improving upon the current bounds by several orders of magnitude. These projections quantify the potential statistical reach of the FCC-ee program, motivating dedicated studies incorporating detector effects, systematic uncertainties, and higher-order theoretical predictions.

We give a reduction from a smooth-domain theorem excluding interior critical points of first positive Neumann eigenfunctions to a non-strict hot-spots conclusion on arbitrary bounded simply connected planar domains with compact Neumann resolvent and simple first positive eigenvalue. The main analytic step is spectral convergence under exhaustion by conformal images of concentric disks. A reflection extension, conformal invariance of Dirichlet energy, and compactness of the Sobolev embedding give a direct proof. A truncation argument justifies square integrability of the transferred extensions without any boundary regularity. The smooth-domain theorem proved by OpenAI is an external input; the approximation argument is proved here in full. The paper was generated by ChatGPT with prompts from the listed author.

Software-based microarchitectural attacks have mainly exploited timing differences to infer information from DRAM. However, defenses and optimizations can prevent timing-based approaches from resolving precise microarchitectural behaviors. We show that unprivileged millisecond- and milliampere-resolution power sensors on modern Arm SoCs can recover DRAM mapping functions and reveal the data dependence of memory accesses without timers. We present HOMARD, a methodology that reads DRAM activity from this telemetry, and two attack primitives. HOMARD-RE is a hierarchy-aware linear-algebraic method that reverse-engineers DRAM addressing functions from power alone. HOMARD-DATA recovers the Hamming weight of a fetched cache line through joint-power rail templating and secret byte values through correlation power analysis on the same rails. Across an AMD ZCU102, a Raspberry Pi 5, and a multi-chip-based LPDDR4 NVIDIA Jetson TX2, HOMARD-RE succeeds where row-buffer timing methods fail. On the Pi 5, HOMARD-DATA classifies 9 Hamming-weight levels at 81.5% accuracy and recovers secret bytes in minutes by correlating rail current with data-dependent bit toggling between neighbors in a row. Both primitives work with an architectural load or a mispredicted Spectre-RSB path on which no value is architecturally read. These results establish unprivileged power telemetry as an independent side channel for profiled microarchitectural attacks, particularly when timing falls silent.

Resonant dispersive wave (RDW) emission in gas-filled anti-resonant hollow-core fibers (AR-HCF) offers a wavelength-tunable route to ultraviolet (UV) and vacuum ultraviolet (VUV) frequency comb generation. RDW emission is expected to be coherent fundamentally, but various noise sources are expected to limit the coherence, including amplitude-to-phase fluctuations due to input pulse intensity noise and plasma formation. It is critical, therefore, to establish the limits of the coherence of UV RDW pulses to determine their applicability for precision spectroscopy. Here, we demonstrate the coherence of RDW emission at 260~nm generated with an amplified Yb-doped fiber system at a 2-MHz repetition rate with a self-heterodyne beat measurement between the RDW and the fourth harmonic of the source. We isolate the RDW contribution to the frequency noise by coherently subtracting the carrier-envelope offset noise contribution measured simultaneously with an f-2f interferometer. We observe a beatnote linewidth below 10~kHz over timescales longer than 100~ms, and we find a conservative upper bound of 1~rad on the integrated RMS phase deviations over a frequency range spanning 1 to 30~kHz. The dominant frequency noise contributions are consistent with environmental noise of the interferometers and AR-HCF coupling fluctuations. These results provide strong evidence that RDW emission is coherent under well-controlled experimental conditions, confirming its potential for UV and VUV frequency comb generation.

For a positive integer $d$, let $\Delta_d^L$ denote the supremum of the densities of lattice sphere packings in $\mathbb R^d$. We prove that, for every sufficiently large $d$, $\Delta_d^L \gg d^2\log\log d\,2^{-d}$. Abuya-Gargava-Zhao obtained the same bound along a sparse sequence of dimensions by combining Venkatesh's cyclotomic symmetries with Klartag's stochastic ellipsoid method. We introduce a multi-cyclotomic gluing construction that allows this argument to work in every sufficiently large dimension.

Physics foundation models are an increasingly popular approach for building versatile emulators for spatiotemporal systems. They require tokenisation layers that lift physical field data into a latent representation suitable for dynamic processing. A typical choice for this layer is a convolutional encoder, which lifts field information in a locality-aware way and can be trained very efficiently. However, this choice cannot be applied to a large class of scientific computing problems where data are represented on unstructured meshes, for example simulations on domains with curved boundaries. Such data typically require graph neural network architectures, which are more expensive to train and, in the form currently used, encode information differently from their CNN counterparts, thus limiting the portability of physics emulators between these data regimes. In this work we introduce Mesh-Native Convolution (MNC), a locality-aware tokenisation layer that encodes unstructured mesh-based data in a way that is analogous to CNNs. This approach leads to models that can be trained once on mesh data and then deployed to arbitrary unstructured meshes, even if these are irregular, with minimal retraining. We show that this tokenisation approach can be used to retrofit grid-pretrained foundation models efficiently on unstructured meshes and provide theoretical results to confirm the consistency of this approach with classical CNN encoders. We demonstrate the favourable performance of this retrofitting approach, which in many cases reaches acceptable accuracy even in zero-shot deployment, and outperforms existing baselines after only a few epochs of fine-tuning, thus providing an efficient bridge between the two data regimes.

Geopolitical and strategic intelligence reports constitute a rich source of expert knowledge and have increasing value in the current political climate, but their heterogeneity and analytical complexity pose challenges for retrieval-augmented question answering. We introduce a benchmark for grounded question answering over 415 open-source geopolitical and strategic intelligence reports, comprising 5,525 document-grounded question-answer pairs spanning factual, inferential, comparative, and risk-oriented questions. We propose a document-aware generation pipeline that profiles reports before generating questions and validate dataset quality through expert annotation and an LLM-as-judge protocol. Using the benchmark, we evaluate 15 retrieval systems and six open-weight LLMs for RAG. We find that models perform similarly on document-level retrieval, while at the harder chunk level retrieval quality is not driven by embedding-model scale: bge-m3, a mid-size open encoder matches an order of magnitude larger ones and outperforms commercial embedding APIs. In contrast, answer generation remains challenging even given the gold evidence, particularly for inferential and risk-oriented questions.

Social influence, the tendency of our social environment to sway our beliefs and behaviors, is often cited as a driver of population-level trends in opinions or beliefs. Many quantitative theories of social influence assume that an agent interacts with its entire social environment. This assumption is often unrealistic. For example, users of online platforms typically see posts from only a small subset of the accounts they follow at any given time. We introduce a simple probabilistic agent-based model of social influence that explicitly models unobserved attention mechanisms. The first of these latent mechanisms is selection: each agent interacts with only one other agent at a time from its social environment. The second is retention: the agent chooses whether or not to ignore the opinion of the other. Our model admits an efficient maximum-likelihood algorithm which allows us to fit it to opinion traces even when these mechanisms are unobserved. On synthetic data, omitting either latent mechanism reduces the overall strength of social influence governed by the inferred interaction kernel, which hides the evidence for the underlying nonlinearity of the social influence. We study a selection of previously analyzed empirical datasets and curate several new ones. We show that models that include the latent mechanisms are often more predictive on these datasets; indeed, they typically improve prediction far more than the choice of interaction kernel does. Modeling latent attention increases held-out predictive likelihood on 14 of these 15 datasets. Each mechanism changes the size of the inferred kernel in a predictable direction, while its effect on the kernel's shape depends on the data.

The S-matrix bootstrap restricts low-energy interactions through unitarity, analyticity and crossing symmetry. For effective field theory couplings, positivity bounds have a natural geometric formulation, the EFThedron, in which the allowed couplings form the classical positive moment problem. We show that this geometry extends to full nonlinear unitarity. Multiplying the amplitude by fractional powers of the energy turns its real part into a moment as well, so that unitarity becomes a bounded complex moment problem. Analyticity becomes a set of linear constraints requiring different kernels to reconstruct the same couplings, analogous to the null constraints imposed by crossing symmetry. In two dimensions the problem can be solved exactly, leading to Castillejo–Dalitz–Dyson amplitudes as extremal measures of the moment problem. In four dimensions we obtain closed-form and semi-analytic bounds for the massless EFThedron model and for massive scalar scattering. For massive scalars, the analyticity constraints improve on published dual bounds, reducing the gap to the primal region. By placing the real and imaginary parts of amplitudes on an equal footing, the Real EFThedron provides a powerful framework combining unitarity, analyticity and crossing, opening a new route to mapping the space of consistent quantum field theories.

Students often perceive scientific experiments as clean, straightforward procedures with predictable outcomes. Students' beliefs about the nature of experimental physics typically remain inconsistent through the initial stages of undergraduate education. In reality, experimental work in physics involves hands-on activities such as soldering circuits, assembling provisional setups, and building functional mock-ups from simple materials before reliable measurements can be achieved. To address this gap, we designed a laboratory activity in which students were completely involved in the experiment design, construction and operation with the final goal of measuring the speed of light with a low-cost and compact setup. The educational objective of this activity was double: firstly, to make students aware of the importance of experimental choices and constraints in the setup design and secondly, to discover the potentiality of semiconductor technologies. As far back as the 1990s, with the spread of compact disk players, laser diodes and photodiodes capable of fast timing performance in the order of nanoseconds became largely available. In the last decades, consumer electronic devices made high performance technologies ubiquitous. Nevertheless, the idea of measuring ultrashort time intervals and the relevance of such measurements in science and technology remain unknown to most students. Nowadays, digital instruments with significant performance such as waveform generators and oscilloscopes are affordable for high school physics labs, paving the way to a closer contact with the world of experimental physics.

The linear arrangement of a simple graph $G = (V,E)$ is a bijective labeling $\pi$ of the vertices of $G$ to the set $\{1, \ldots, n\}$. The maximum linear arrangement ($\texttt{maxLA}$) problem is to determine whether $G$ has a labeling such that the sum of labels across $G$'s edges, $\sum_{uv \in E} |\pi(u) - \pi(v)|$, is at least a given value $k$. Minimum linear arrangement ($\texttt{minLA}$) is defined analogously. Both $\texttt{maxLA}$ and $\texttt{minLA}$ are known to be NP complete for general graphs. $\texttt{minLA}$ has shown to be NP complete for bipartite graphs by Evans and Shiloah in 1975. It has been an open problem whether the same holds for $\texttt{maxLA}$. In this paper we prove that $\texttt{maxLA}$ is NP complete for bipartite graphs.

We completely resolve Question~5.3 of Kania's paper \cite{Kania} and establish a more general realization theorem. We show that for infinite cardinals $\mu\leqslant\nu<\kappa<\lambda$, a space with \[ \pc(X)=\mu,\qquad\pcl(X)=\nu,\qquad \pcc(X)=\kappa,\qquad\pcd(X)=\lambda \] exists if and only if $\kappa\leqslant2^\mu$. The witnesses may be zero-dimensional and Tychonoff spaces, of the least possible cardinality $\lambda$. They may have the least possible number of non-isolated points, or alternatively be crowded. The main ingredient is an exact computation of the centered local $\pi$-character of star filter spaces in terms of ultrafilter character, obtained from a finite selection lemma for filter pushforwards. We also construct a single space with countable $r$-linked local $\pi$-bases for every finite $r$, centered value $\omega_1$, and any prescribed decreasing value greater than $\omega_1$.

Computer-use agents (CUAs) are rapidly improving, yet synthesizing diverse and reliable software environments at scale remains challenging. Native application environments require substantial construction and integration effort, which limits application coverage and increases the risk of application-specific overfitting. Browser-hosted environments provide limited interaction depth and workflow complexity. These limitations motivate SoftwareGym, an automated pipeline for synthesizing GUI software environments with diverse interfaces, executable functionality, and runtime-verified behavior. It generates structurally diverse designs, realizes them as runnable applications, and uses GUI evaluation with iterative repair to ensure functional reliability. Using SoftwareGym, we synthesize 182 applications across twelve categories and achieve a mean GUI task pass rate of 91.5%. We construct SoftwareGym-Bench, a 120-task evaluation suite spanning the same twelve categories, and use trajectories collected from these environments for CUA training. The resulting model improves Pass@3 from 41.39% to 56.97% on the strictly out-of-domain OSWorld benchmark and achieves 40.0% on SoftwareGym-Bench. These results demonstrate the potential of synthesized software environments as scalable and transferable supervision for CUA training.

Ethnic and cultural identities rarely align with political borders, a tension made sharper by migration and long-standing cross-border ties. Communities such as the Sami, the Rohingya, and the Kurds have long lived across more than one state. This tension is especially clear in mainland Southeast Asia, where many communities share history, language, and culture across state lines. We study nine such transborder ethnic communities along China's southwestern border with mainland Southeast Asia, looking at how large language models (LLMs) represent them. We define and study identity ossification: a pattern where LLMs turn a fluid, cross-border ethnic identity into a single, fixed nation-state identity. We introduce the Transborder Identity Ossification Diagnostic Framework (TIODF) that compares direct factual questions with open-ended narrative questions to test whether ossification comes from missing knowledge or from how models frame their answers. We test GPT-5.1 and DeepSeek-V3.2 in English and Chinese. Ossification varies substantially across models and query languages. It appears in only 10.1% of GPT-5.1's English answers, but in 62.6% of DeepSeek-V3.2's Chinese answers. Models often still tell a single-nation story even when they answer the factual question correctly. This decoupling suggests that ossification reflects narrative framing rather than simple knowledge absence, which matters for how we test cultural fairness in multilingual AI systems.

Humanoid and service robots operating in clut- tered human environments may encounter motion planning problems that are infeasible when all surrounding objects are treated as fixed obstacles. The Minimum Constraint Removal problem, which we refer to as MCR, addresses this situation by identifying the minimum number of constraints, such as movable objects, whose removal enables a collision-free path. However, computing an optimal MCR solution is NP-hard. In this work, we propose a graph-based heuristic pipeline that combines bidirectional feasible-set generation with fixed- length searches over reduced MCR graphs. We further derive an extended feasible-set generator for instances in which the candidates produced by the first generator do not contain a complete optimal removal set. Across six benchmarks, the evaluated pipeline matches the exact solution in three cases and, in the remaining cases, selects at most 31.25% more obstacles than the optimum, while being between 3.4 and 33.7 times faster than exact search. In one of the benchmarks, involving a manipulation task amidst clutter, selecting two rather than seven obstructing objects reduces the combined MCR, inverse- kinematics, and motion planning time from 84.72 s to 30.59 s. These results indicate that the proposed approach provides a practical compromise between computational efficiency and solution quality for motion planning among movable obstacles.

We present a Fourier-based method for evaluating two-electron four-center Coulomb repulsion integrals over Slater-type orbitals. Spherical Fourier-Bessel coefficients of the convolution factors are precomputed and reused. Since the spherical part of the integral can be evaluated directly, there remains only a one-dimensional quadrature problem in the radial direction. For the ten reference integrals reported here, the largest absolute discrepancy is $1.7\times10^{-9}$. A preliminary resident-cache benchmark computes approximately $4.2\times10^5$ unique integrals per second on an NVIDIA V100 at the contraction stage.

We study symmetries and anomalies of dissipative Majorana chains described by Lindblad dynamics. Using the operator-state mapping, we represent the Lindbladian as a non-Hermitian operator on a doubled Hilbert space. In this formulation, the ket and bra contours admit strong fermion parities and strong translations, while their diagonal combinations define weak symmetries. For even and odd chain lengths, we diagnose anomalies by comparing symmetry algebras with and without weak fermion-parity defects. These algebras retain the lattice precursor of the modulo-$8$ anomaly of the closed Majorana chain and exhibit projective relations between strong fermion parity and weak translation, as well as between weak fermion parity and strong translation, which are intrinsic to open quantum systems. We establish the IR correspondence between the lattice symmetry algebras and those of the doubled Majorana CFT in four spin sectors, identifying chiral fermion parities emanated from the lattice translations. Finally, we develop the lattice bosonization of the dissipative chain. For even length with collective energy dephasing, combining the defect and no-defect sectors and imposing GSO projections gives rise to an Ising Lindbladian with non-invertible Kramers–Wannier symmetries descending from the anomalous strong translation and fermion parity of the Majorana chain. For odd length, constructing mutually commuting ket- and bra-sector spin operators instead yields an Ising Lindbladian with a Kramers–Wannier duality defect. Our results provide a controlled framework for extending lattice anomalies and non-invertible symmetries to open quantum systems.

N-1-1 contingency analysis studies the loss of one element followed by system adjustments and the loss of a second. The number of contingency pairs increases quadratically with the number of applicable elements, and traditional ranking-based screening methods cannot guarantee that omitted pairs are bounded. This paper determines the minimum bounding set of two-branch-outage pairs: the smallest set that can reproduce the worst value of every violated thermal and voltage limit of a linear network model. In the DC model, effects of an outage pair can be quantified by the extent of the coupling between the outages. The paper measures this by the cosine between the two outages in the network's cycle space, with a bound that splits into one term per outage and is within a near-ideal factor of $\sqrt{1+\varepsilon}$. This argument extends to every linear model in which an outage subtracts a positive semi-definite low-rank term from a symmetric positive definite matrix, with a tight bound and a collapse test. A model with voltage magnitudes is extracted by performing one Newton step on the energy function of the lossless network in rectangular coordinates. A best-first search built on these bounds reproduced brute force on public systems of up to 2{,}869 buses with 130 to 3{,}800 times fewer pair evaluations for flows and 170 to 23{,}700 times fewer for voltages. Gaps in translation to nonlinear models are explored on the IEEE 118-bus system.

We prove that for any prime $p$ and tame level $N$, the weight map from the cuspidal Hilbert eigenvariety constructed by Andreatta-Iovita-Pilloni for a totally real field $F$ to the weight space is proper in the sense of Buzzard-Calegari, under a non-vanishing hypothesis on certain Sen polynomials of Galois representations at the primes above $p$. In particular, this yields the properness of the eigenvariety away from non-positive integral Sen weights. Our approach extends the $p$-adic Hodge-theoretic argument of Diao-Liu.

This paper presents a low-complexity tensor-based channel estimation framework for downlink mmWave massive MIMO systems using passive Butler-matrix beamforming. By modeling received signals as a third-order PARAFAC tensor, we exploit known pilot structures via bilinear alternating least-squares (BALS) optimization to reduce computational overhead. The decoupled spatial factors are then processed via 1D and 2D ESPRIT to extract angles of arrival and departure with automatic pairing. Additionally, we derive the Cramer–Rao lower bound (CRLB) for this architecture. Simulations confirm the proposed estimator outperforms conventional least-squares, achieving accurate channel reconstruction and approaching the CRLB at moderate-to-high signal-to-noise ratios.

Broadband soliton microcombs near 1 $\mu$m offer coherent, evenly spaced optical lines for measurements in the biological window. However, combining broad spectral coverage with closely spaced comb lines increases the pump-power requirement and makes thermally stable soliton access more difficult. Here, we demonstrate broadband silicon nitride (Si3N4) soliton microcombs enabled by a resonant coupler that enhances pump delivery to the nonlinear resonator. The device generates a spectrum spanning 930-1380 nm with an approximately 98-GHz comb-line spacing at on-chip pump powers of tens of milliwatts, while providing reliable thermally stable access to the single-soliton state. These results provide a route to compact, power-efficient microcomb sources that combine broad spectral coverage with closely spaced comb lines for spectroscopy and imaging in the biological window.

Continual learning (CL) seeks to adapt pretrained models while preserving prior capabilities. Yet the tasks supplied for preservation cover only part of what a model knows, while forgetting can extend beyond them. Recent work shows that language models can supplement this limited evidence by generating useful replay without access to their original training data. This raises a question about self-preservation: can a model help choose preservation targets that protect its own capabilities more broadly? To study this question, we organize task descriptions and their examples into Task Banks, whose entries can be selected and combined to guide replay generation by a frozen model. Continual instruction tuning with Qwen models of varying sizes yields three findings. First, the new task and learning rate shape which prior capabilities degrade, while tracking previously correct answers reveals losses that average accuracy can conceal. Second, generated replay reduces forgetting on tasks outside the bank. Combining partial banks strengthens this protection, improving mean retention of previously correct answers over either bank alone across the tested sizes. Third, the model can also help choose replay targets: selection informed by the new task improves mean held-out retention over random selection at 4B, with similar new-task gains, although smaller models show no corresponding advantage. Under a fixed example budget, however, replacing new-task examples with replay improves preservation at the cost of smaller new-task gains. Together, these findings show that models can support broader self-preservation by generating replay data and, in some settings, choosing which tasks to replay.

Prompt optimization is an attractive alternative to reinforcement learning for adapting language-model agents, since it is cheap and produces human-readable artifacts. On factual tasks, however, it carries a failure mode that scalar-reward reinforcement learning does not. An optimizer that reads the training examples can write their answers into the prompt, so an in-distribution gain may be memorization rather than transferable behaviour. We apply reflective prompt evolution (GEPA) to the two system prompts of a multi-agent deep-research agent and evaluate it under a topic-disjoint split in which no topic is shared across train, validation, and test. Held-out citation reward rises by 0.079, from 0.702 to 0.781, with a bootstrap 95% confidence interval of [0.051, 0.107], and the gain replicates across three seeds (mean 0.107). Nearly half of the winning prompt is a memorized answer key, yet a controlled ablation shows this content earns no reward even on the exact training questions it encodes. The large validation-minus-held-out gap is therefore small-sample selection under a noisy judge, not memorization. Varying the verifier isolates what makes the gain transfer. Decomposed per-dimension feedback is the causal ingredient, and collapsing it to a single scalar reward makes held-out transfer negative. The gain persists under an out-of-family judge and is largely reproduced by a hand-written version of the discovered recipe. Credible evaluation of prompt optimizers therefore requires held-out topics, multiple seeds, and verifier ablations. We release the candidate prompts, lineage diffs, figures, tables, and analysis.

Empirical relations and planet formation simulations suggest an upper mass limit for rocky planets of $\sim$10M$_{\oplus}$, as cores above this will almost certainly undergo runaway gas accretion. Nevertheless, a small number of planets appear to have densities consistent with rocky Earth-like compositions above this threshold, indicating that unusual formation and evolutionary scenarios are capable of producing extremely massive rocky planets called "Mega-Earths." However, all of these planets orbit active stars with log$R'_{HK}$ $>$-4.75. We find a <1$\%$ probability that all literature ultra-massive rocky planets would coincidentally orbit such active stars, and find no astrophysical reason to expect Mega-Earths to preferentially form around active stars. We therefore conclude that Mega-Earths are artifacts of stellar activity, and we question the reliability of Gaussian Processes to enable mass measurements for rocky planets around active stars.

Sparse computation can scale the parametric capacity of LLMs without proportionally increasing active computation, but pretraining sparse architectures remains expensive and data-hungry. Existing sparse upcycling methods mainly target Mixture-of-Experts models, where expert homogeneity, sample inefficiency, and system overheads can limit practical scalability. Static sparse memory offers predictable access and token-specific updates that can help address these challenges, but naively integrating it into pretrained computation yields poor returns. In this work, we introduce STAPLE, a recipe built upon STEM for integrating static sparse memory into pretrained models. Whereas STEM replaces FFN up-projections with token-indexed embeddings, STAPLE retains them and adds the embeddings as a residual pathway, preserving pretrained computation while learning complementary token-specific information in a scalable manner. On OLMo-2 1B, STAPLE improves over continued dense training by 4.27 points on HumanEval, 2.80 on MBPP, 2.27 on GSM8K, and 1.65 on DROP. Average performance increases with static capacity, and improvements persist on the larger OLMo-2 7B backbone. STAPLE also outperforms static sparse baselines such as STEM and an Engram-style model, and achieves a higher average score than an MoE upcycling baseline while decoding 2x faster with less than half its GPU model memory and measured GPU energy per token in our throughput-oriented decoding evaluation. Importantly, despite introducing billions of parameters late in training, our host-offloaded implementation keeps generation throughput within 1.5% and training throughput within 3% of the dense baseline, while caching only the ~5% most frequent embeddings on GPU. These results show that static sparse memory can provide a sample-efficient, scalable, and sustainable path for expanding pretrained models without sparse pretraining from scratch.