Machine Learning
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Monday 14 September
Machine unlearning aims to remove specific knowledge from a trained large language model (LLM) without retraining from scratch. Existing methods modify model weights via gradient ascent and its advances. While effective on certain benchmarks, these weight-based approaches exhibit a sharp forget-utility trade-off, where stronger forgetting of target knowledge can degrade model utility, and unlearned knowledge may reappear under post-unlearning fine-tuning or prompt attacks. We propose ARIA (autoencoder-gated inference-time unlearning), a test-time unlearning method that leaves model weights intact and gates access to unwanted knowledge only when generation enters a forget-related state. ARIA uses sparse autoencoder (SAE) latents to train a lightweight linear detector, then applies an interpretable intervention on triggered states with negligible test-time overhead. Empirical evaluations on TOFU, R-TOFU, and WMDP show that ARIA improves the forget-retain trade-off over weight-based baselines across both a thinking model (DeepSeek-R1-Distilled-Qwen-1.5B) and an instruction model (Gemma-3-1B-it), e.g., reducing WMDP-cyber forget-set accuracy significantly while keeping MMLU within 1% of the pre-unlearning model. We further introduce three post-unlearning adversarial attacks targeting weight-space and decoding-space recovery, and find that ARIA remains robust under all three, with forgetting changing by less than 1% under attack. A feature-level case study leveraging the interpretability of ARIA suggests that some retain degradation may reflect response styles underlying the unlearning data rather than leakage of the targeted knowledge itself, highlighting a potential source of bias in unlearning task construction.
Logic in Computer Science
Teaching Vampire New Tricks: An Experimental Study of Neural Clause Selection
A neural clause-selection guidance approach in the Vampire theorem prover was recently shown to substantially improve the success rate of the prover's default strategy on the TPTP benchmark. We experimentally study the impact of the approach across several ITP-derived benchmark sets and its interaction with theorem proving strategies. We find that while the neural guidance consistently improves performance within individual benchmark domains, cross-benchmark application of guiding models underperforms the plain default strategy. This can be remedied by training a single model on all datasets at once. Such a model, although more expensive to obtain, helps Vampire almost catch up in performance across all datasets. The picture when considering combined strategies is less clear-cut, indicating persisting value of neural guidance but under diminishing returns.
Read the full abstract
Quantum Physics
Error-corrected function estimation advantage in multiparameter Hamiltonians
We establish the ultimate precision limits for estimating a function of multiple Hamiltonian parameters in the presence of Markovian noise. By reducing multiparameter function estimation to an optimization over effective single-parameter embeddings, we derive tight bounds on the quantum Fisher information. We identify a necessary and sufficient functional Hamiltonian-not-in-Lindblad-span condition for Heisenberg-limited scaling in time. When this condition holds, we construct a code that simultaneously removes the noise and nuisance Hamiltonian parameters while preserving the target signal. When it fails, we derive the optimal standard quantum-limit coefficient and show that it is asymptotically attainable using approximate quantum error correction. Finally, we demonstrate that direct function estimation can substantially outperform estimating all parameters individually and subsequently evaluating the function. This advantage can scale with both the evolution time and the number of parameters.
Read the full abstract
Statistical Methodology
Coherent and robust Bayesian inference under informative sampling via sample-level weight models
We develop two complementary Bayesian procedures for analytic inference under informative sampling. The first constructs a coherent likelihood by modeling the sample-level distribution of the design weight given the study variable and recovering the conditional inclusion probability through the identity of Sverchkov and Pfeffermann (2004); it is the Bayesian counterpart of a recently proposed conditional-likelihood estimator, attains the model-based efficiency bound under the correctly specified sample-level joint model, and can be biased otherwise. The second uses the same weight model only as input to a Neyman-orthogonalized analog of the design-weighted score and applies the loss-likelihood bootstrap; the resulting credible intervals are asymptotically sandwich-correct under Poisson sampling, and the posterior remains centered at the truth under misspecification of the weight model, although the working model can still affect the first-order variance. A beta regression on the population-level inclusion probability, inducing a beta prime sample-level weight distribution, serves as the parametric default. A data-adaptive implementation of the second procedure estimates the efficient factor $\bar\pi(x,y)=1/E_p(W\mid x,y)$ by a design-weighted Gamma/log-link regression and uses it in a joint one-step loss-likelihood bootstrap for the analytic parameter; its validity requires no rate conditions on the nuisance fit. We illustrate both procedures with simulations and Canadian Workforce data.
Read the full abstract
Earth and Planetary Astrophysics
What the Solar System Can Teach Us About Rocky Exoplanets
The number of rocky extrasolar planets being discovered continues to increase, but so too does the apparent diversity of such worlds. As we work to understand the likely bulk properties and thermal, geological, and climatological attributes we might expect of rocky exoplanets, we can look to the Solar System for guidance. Here, we review the interior, surface, and atmospheric characteristics of modern Earth, and discuss how our homeworld has changed through Solar System history. We then visit in turn Venus, Mars, Mercury and Earth's Moon, and Io, noting how these terrestrial planets are alike, how they are different, and how they have evolved through time. Finally, we consider some examples of the types of rocky worlds known or suspected to exist without direct Solar System counterparts — but we argue that even then, for all the variety we might expect, they are still variations on a common theme. Perhaps the most important lesson the Solar System can teach us is that terrestrial bodies change through time, sometimes dramatically, and that rocky worlds of similar size and mass can have vastly different planetary outcomes.
Read the full abstract
Machine Learning
How I learned to stop worrying and love StopGrads: Stationarity, Convergence, and a case study on Flow Map Learning
Stopgrads are widely used in training machine learning models, but stopgrads can alter the gradient, stationary points and convergence guarantees of the original objective, which can make stopgrad training theoretically ungrounded. We introduce a stopgrad regression principle, which identifies a general template for stopgrad objectives with a closed-form characterization of stationary points and their uniqueness, unifying stopgrad objectives for flow maps, reinforcement learning, and diffusion samplers. We provide theoretical grounding for optimizing stopgrad flow map objectives by showing their unique stationary point is the true flow map, and showing positive convergence results for Eulerian and Lagrangian objectives, including MeanFlow and improved MeanFlow. Remarkably, we show that under functional semi-gradient flow, the learned flow map has a closed-form expression composing the initial flow map and the true flow map. We additionally use our stopgrad regression principle to propose modified stopgrad placements for flow map objectives which reduce training memory by 2x.
Read the full abstract
Dynamical Systems
Pressure at infinity and Large Deviations
We establish level-1 large deviation principles for continuous time suspension flows over countable Markov shifts, a class of non compact dynamical systems. Our results both quantify deviations of Birkhoff averages from their typical values and, importantly, provide a large deviation principle for orbits that at a finite time are close to escaping to infinity. To overcome the challenges posed by non compactness, we introduce the notion of topological pressure at infinity and prove a variational principle linking it to its measure theoretic counterpart. This framework allows the study of a natural class of potentials, known as strongly positive recurrent, which exhibit robust thermodynamic properties. Moreover, we show that compactly supported ergodic measures are pressure dense in this setting. Applications to pressure gaps and examples are given.
Read the full abstract
High Energy Physics (Phenomenology)
Enhanced Standard Model bound on charm CP violation from QCD penguins
We investigate the impact of QCD penguin operators on direct CP violation in singly Cabibbo-suppressed $D^0$ decays within the framework of light-cone sum rules (LCSR) using pion and kaon light-cone distribution amplitudes. We determine, for the first time, the hadronic matrix elements of the QCD penguin operators for $D^0\to K^+K^-$ and $D^0\to\pi^+\pi^-$ at leading order in $\alpha_s$, employing a three-point correlation function with an artificial momentum to avoid unphysical parasitic cuts. We also compute the matrix elements of the current-current operators using the same framework and compare them with previous LCSR results. We find that the scalar penguin operators $Q_5$ and $Q_6$ can have substantially enhanced matrix elements. We identify two sources of enhancement: quark-condensate contributions, which arise already at tree level and are therefore enhanced by a relative factor of $16 \pi^2$ compared to the remaining, loop-induced terms in the sum rule; and an annihilation-type contribution from the $\bar u u$ component of these operators, which enters already through twist-three light-cone distribution amplitudes and is of comparable magnitude. While the impact of QCD penguin operators on the branching ratios is negligible, they can enhance the magnitude of the “penguin-to-tree” amplitude ratios relevant for direct CP violation. The relevant strong phases, however, remain largely unconstrained within the accuracy of our framework. Assuming maximal relative strong phases, we find that the corresponding upper bound on the Standard Model contribution to charm CP violation can be enhanced compared to estimates based on the current-current operators alone, reducing the gap with the experimental measurement.
Read the full abstract
Databases
How Can We Shrink the Family of Test Databases? Query Containment with Nulls and Comparisons
Query containment and equivalence drive database query optimization and rewriting. For plain conjunctive queries, both are decided by evaluating one query over a single canonical database of the other. This classical test breaks down in two settings that pervade real queries: databases with null values under SQL's three-valued semantics, and queries with order comparisons. In both, deciding containment is $\Pi_2^p$-complete, and the known characterizations replace it by an exponential family of test databases, leaving no practical route to certifying equivalence. We ask how the family of test databases can be shrunk. For conjunctive queries over databases with nulls, we shrink the family to one that is exponential only in a special set of variables, and place containment in NP when that set has constant size. For queries with comparisons, we construct canonical values that decide containment, and then shrink the family by decomposing the test into independent components and by splitting it along the order conflicts. Finally, we combine the two features, and prove that the number of test databases is fixed-parameter tractable in three local parameters of the two queries, with the null-only and comparison-only tests as special cases. Each test evaluates the containing query over a family of databases, an operation native to any database system.
Read the full abstract
Neurons and Cognition
A neural-astrocyte architecture implements a hybrid automaton for evidence accumulation
Astrocytes are non-neuronal glial cells that are receiving widespread attention due to their emerging role in neural computation. In this paper, we propose and study dynamical mechanisms by which astrocytes may augment the ability of neural networks to infer context in reinforcement learning (RL) settings. We construct a biologically inspired, two-level dynamical neural-astrocyte network with distinct spatial and temporal organization. We train this model on a hierarchical multi-context task that requires the agent to infer changes in latent task rules based on derived rewards. We find that in this setting, astrocytes enable evidence accumulation of changes in context and subsequent context-specific modulation of neural dynamics. We show that these functions are implemented via two dynamical mechanisms: (i) reward-induced bifurcations that relocate an asymptotically stable attractor into different, context-specific regions of state space, and (ii) the relative shallowness of these attractors, mediated by the entropy of the environment, giving rise to behavioral stickiness. Together, these mechanisms amount to a hybrid automaton, in which uncertainty accumulates until, eventually, the neural dynamics are switched to a new context. This model provides a neuro-dynamic schema, compatible with neural-astrocyte biology and prior empirical observations, for how astrocytes may integrate information from the periphery and drive contextual changes in neural circuits.
Read the full abstract
Quantum Physics
$d$-pod realization of nonadiabatic holonomic quantum computation
Holonomic quantum computation (HQC) realizes quantum gates through non-Abelian geometric phases, providing an experimentally accessible approach to quantum control. While the nonadiabatic HQC framework has been extensively developed for three-level $\Lambda$ systems encoding qubits, its systematic extension to higher-dimensional qudits remains largely unexplored. In this work, we generalize nonadiabatic HQC to a $d$-pod configuration, where a single excited state is coupled to $d$ ground states, the latter forming the computational subspace. This scheme enables universal holonomic single- and two-qudit gates using only optical or microwave pulses on trapped atoms or ions, offering an efficient route to implement a discrete universal gate set with minimal pulse coordination. As an explicit example, we analyze in detail the qutrit ($d=3$) case, demonstrating compact realizations of single- and two-qutrit holonomic gates, each gate requiring at most two loops in the Grassmannian generated by at most three pulses.
Read the full abstract
Artificial Intelligence
Where Should the KV Cache Live? Placement Policies Across GPU, CPU, and SSD for Long-Lived Sessions
GPU high bandwidth memory is scarce and expensive, and KV caches consume much of it as chats, agent loops, and document question answering accumulate state. Systems such as Mooncake, LMCache, FlexGen, InfiniGen, and AttentionStore extend GPU memory with CPU DRAM and SSD. The harder question is which blocks belong in each tier, when to move or evict them, and whether prefetching helps. We study these choices in a discrete event simulator spanning GPU HBM, CPU DRAM, and SSD, calibrated against a random forest execution time predictor. We compare recency, reuse frequency, predicted reuse, and an EWMA predictor with prefetch lookahead across chat, agent, and document question answering workloads. Tiering supports 73.02 times more concurrent sessions per GPU and lowers cost per session by 62.04 times. These gains come from tier capacities of 1 plus 8 plus 64, not placement policy. Decode is compute bound at batch size one in our setup, so placement barely affects throughput. It mainly changes PCIe migration traffic and time to first token. Recency produces 2.30 times less migration traffic than reuse frequency for chat. Reuse frequency performs best for agents and document question answering. The existing predicted reuse policy is byte identical to recency, making its agent recommendation effectively recency. A genuine EWMA predictor changes behavior but still ranks behind reuse frequency on the workloads prediction was expected to help. Prefetching does not justify its bandwidth cost. Across the policy and cache size grid, even an oracle with knowledge of future requests never beats no prefetch on migration traffic. Workload specific placement can reduce data movement, but the predicted reuse and prefetch recommendations are not supported as implemented.
Read the full abstract
Cryptography and Security
Analyzing Multi-Factor Authentication Through Cryptographic Security Properties
Modern authentication systems use cybersecurity techniques to validate the identity of the person (or applications acting on behalf of the person) as a primary defense against unauthorized access. While initially built around single mechanisms such as usernames/passwords, physical tokens, or biometrics, current systems have evolved into multi-factor authentication (MFA) platforms that combine multiple mechanisms. Among them, several focus on strategies that prevent replay attacks (i.e. the reuse of a component that could have been potentially compromised).
Read the full abstract
Artificial Intelligence
Artificial intelligence and biosecurity: capabilities, threat pathways, and defense-in-depth governance
Artificial intelligence is reshaping biological research across an increasingly connected digital-to-physical workflow. General-purpose large language models can retrieve and integrate scientific information, support experimental planning, and computational analysis; biological foundation models can predict, optimize, and generate proteins, genes, and genome-scale sequences; agentic systems can coordinate multistep research tasks; automated laboratories can partially close the design-build-test-learn cycle. These technologies could greatly benefit medicine, public health, and biotechnology. However, their biosecurity risk depends not only on what the AI can do, but also on who uses it, their expertise and intent, their access to laboratory tools and materials, and the safeguards in place. Current evidence shows that AI uplift exists but primarily affects digital rather than physical tasks. Frontier systems have exceeded expert baselines on in-silico, and screening-evasion benchmarks, whereas controlled wet-laboratory studies find that tacit knowledge and physical execution remain substantial barriers. This review describes the different biological threats from AI tool use, from information gathering and biological design to procurement, synthesis, testing, scale-up, and potential release. We further examine why alignment techniques for general-purpose models transfer poorly to biological ones, and the emerging role of interpretability in auditing whether hazardous capabilities are genuinely removed. We argue for defense-in-depth governance that links capability thresholds to proportionate responsibilities across the biological AI ecosystem, reducing high-consequence risk while preserving beneficial use.
Read the full abstract
Geometric Topology
The underlying manifold of a low-volume hyperbolic 3-orbifold
This paper proves strong restrictions on the underlying topological space of a closed, orientable hyperbolic 3-orbifold \orbM whose volume satisfies a certain upper bound. For instance, if vol(\orbM) < 0.1491, then the underlying space of \orbM must be either a small Seifert fibered space, or the connected sum of two lens spaces (or of a lens space with S^2 x S^1), or the gluing of one or two highly restricted Seifert fibered spaces along a single incompressible torus. If vol(\orbM) < 0.1571 and the singular locus of \orbM is a link, then the topology of the underlying space is restricted even further. Our methods are primarily topological, and involve studying the underlying topology of orbifold books of I-bundles. Along the way, we provide an exposition of some foundational material about orbifolds that was previously absent from the literature.
Read the full abstract
Cryptography and Security
Feasibility of Homomorphic Inference for a Genomic Foundation Model
Human genomic sequences can identify individuals, cannot be replaced after disclosure, and are the inputs that genomic foundation models are designed to interpret. We assess whether a compute provider can execute a released genomic foundation model without receiving query-derived genomic values in plaintext and whether correctness, memory, or cost prevents complete encrypted inference. We first reproduce the released model on three genomic task families and freeze an independently validated numerical reference. We then implement a client-assisted approximate homomorphic encryption protocol: the provider evaluates linear algebra on ciphertexts, while the key-holding data owner evaluates exact normalization, causal softmax, and activation functions at fixed boundaries. A noninteractive configuration completes one released-weight block but exceeds the tested accelerator-memory envelope when configured for composition. The client-assisted configuration executes all released transformer blocks and the task head for one heldout genomic-signal input at its full prompt length. It matches the frozen final label, peaks at 9,839 mebibytes of accelerator memory, and completes in 6,683 seconds on one accelerator. These results establish arithmetic feasibility for a complete classifier, while repeatability, network transport, and private token-index lookup remain unresolved. The biomedical significance is that, under the stated threat model, a served genomic model can process an encoded sequence without exposing plaintext queryderived activations to the compute provider.
Read the full abstract
Atmospheric and Oceanic Physics
A Self-Diagnosing Structural Error-Aware Parameter Estimation Method for Earth System Models
We propose a fully automated, structural error-aware, interpretable climate model parameter estimation method that leverages Perturbed Parameter Ensembles (PPEs). It is based on history matching and aligns with an increasingly-used iterative simulation-emulation-calibration methodology. The method is motivated by the negative impacts of structural error and emulator and observational uncertainties on climate model parameter estimation efforts, as well as the problems associated with sparsely-sampled PPEs. To address these challenges, the method explicitly builds simpler emulators that avoid overfitting, detect structural error, avoids compensating for structural error through inflated mismatch tolerances, and sequentially excludes structurally inconsistent variables for parameter estimation. The method decomposes the high-dimensional calibration problem into linked low-dimensional subproblems, and integrates their constraints to reconstruct the jointly plausible region of the full parameter space. The method is applied to a 100-member PPE with 34 perturbed parameters generated by a version of CAM6 with machine learning-based warm rain microphysics parameterization. Through iterative application, the method greatly reduces the ensemble spread and improves the matching between simulated and observed zonal climatologies. The method also finds ensemble members that outperform the default CAM6 configuration in root mean square error across multiple diagnostics. Controlled experiments demonstrate that overly-conservative emulator uncertainty could lead to neglect of informative observations, and tolerance of the structural error, in the context of this method, biases the estimated parameters toward compensating for structural error. Our work also emphasizes the value of interpretability for diagnosing structural error and informing parameter estimation in PPE-based calibration.
Read the full abstract
Astrophysics of Galaxies
Gravitational Wave Bias in IllustrisTNG300 from Machine-Learned Population-Synthesis Calibrated Merger Rates
The large scale clustering of gravitational wave (GW) sources offers an independent probe of cosmology and compact-binary astrophysics. Interpreting future GW clustering measurements requires understanding how merger rates depend on host-galaxy properties and environment. We develop a simulation-based framework to model the clustering of binary black hole (BBH) mergers and estimate the corresponding GW bias. We populate IllustrisTNG300 with GW sources using a machine-learning emulator trained on the GALAXYRATE data set, which provides population-synthesis calibrated merger rates linked to galaxy formation histories. This enables physically motivated mock GW catalogs with galaxy-specific BBH merger rates that depend on stellar mass, star formation rate, metallicity, and redshift. We compare this framework to a simpler phenomenological model in which the merger rate depends only on stellar mass. In both approaches, GW sources are more strongly biased than the overall galaxy population, reflecting the preference for BBH mergers to occur in more massive, more strongly clustered halos. The machine-learned model also produces stronger scale dependence in the GW bias than stellar-mass weighting alone, indicating that additional host properties affect GW clustering. We further find that metallicity correlates positively with both GW and galaxy bias. These results highlight the importance of physically motivated merger rate models for GW large scale structure analyses and provide a framework for interpreting future clustering measurements from experiments such as the Einstein Telescope and Cosmic Explorer.
Read the full abstract
Quantum Physics
Link-middle-cut lower bounds for Clifford circuit synthesis
Minimizing the number of CNOT gates required to synthesize a Clifford operator is a central problem in quantum circuit optimization. We extend the link–middle–cut (LMC) framework for linear reversible circuits to Clifford operators represented by binary symplectic tableaux. By introducing block-support connectivity graphs and determinantal invariants, we obtain efficiently computable lower bounds on ancilla-free CNOT complexity. We prove that these bounds are tight for qubit permutations: a permutation of $n$ qubits with $k$ cycles requires exactly $3(n-k)$ CNOT gates, even when arbitrary one-qubit Clifford gates are available. We also derive efficiently computable bounds for synthesis up to a permutation of the output qubits, corresponding to free qubit relabeling. Finally, we further demonstrate the utility of the bound by using it as an admissible heuristic for $A^*$-based Clifford synthesis and show that, on benchmark instances, the resulting search can prove optimality beyond what was possible with earlier SAT-based methods, and that it may be used in combination with earlier $A^*$ heuristics to improve the CNOT counts obtained through heuristic search.
Read the full abstract
Computer Vision
Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics
Spatial Transcriptomics (ST) has transformed biomedical research by enabling the spatial mapping of gene expression across tissue sections. However, high operational costs, specialized equipment requirements, and sensitivity to experimental noise limit the accessibility and scalability of ST. Recent computer vision approaches aim to overcome these limitations by predicting spatial gene expression directly from histopathology images. While effective, current approaches often suffer from gene expression over-smoothing and overly uniform predictions across tissue regions, suggesting that further progress depends on learning representations that reflect the hierarchical and asymmetric structure of gene regulation and tissue morphology. To address these issues, we propose Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics (HyCLoST), a hyperbolic contrastive learning model that captures the intrinsic hierarchical relationships within ST data. By leveraging hyperbolic geometry and a gene-to-image entailment loss, HyCLoST learns structured, biologically grounded representations that improve gene expression prediction accuracy, achieving a 6% reduction in MSE and an 8% increase in PCC across 26 ST datasets, over previous methods. Our source code is publicly available at https://github.com/BCV-Uniandes/HyCLoST
Read the full abstract
Artificial Intelligence
Calibrate, Then Route: A Measured Study of Learned Request Routing for Disaggregated LLM Serving
Disaggregated LLM serving places compute heavy prefill and memory heavy decode on separate GPU pools. Systems such as DistServe, Splitwise, and Mooncake make this separation fast, but routing still determines which instances handle each request. We study a router that estimates the additional completion time on each instance using exact prompt length, predicted output length, post admission KV cache pressure, and SLO class. We develop the policy in a discrete event simulator and validate it on eight NVIDIA A40 GPUs, each running a vLLM engine, with NIXL transferring KV caches between pools. All workloads run at measured saturation. Across three mixed, bursty arrival traces, the calibrated router achieves the highest mean goodput at 0.864, compared with 0.835 to 0.847 for round robin, least loaded, and a length heuristic. It also shows the lowest variance across traces. It beats round robin and the length heuristic on all three traces and least loaded on two. On the third, it trails by 0.003, within run to run noise. Hardware calibration matters: simulator derived constants cost 4.5 goodput points and roughly 40 percent of the tail latency advantage, reducing the scorer to little more than queue counting. Benefits grow with decode pool size and traffic heterogeneity but disappear in pools with three instances, where queue counts are often enough. Under extreme scarcity, greedy cost minimization concentrates requests on the cheapest scored instance, and blind spreading performs better. With calibrated costs, the learned router matches the goodput of round robin using six GPUs instead of seven.
Read the full abstract
Astrophysics of Galaxies
Filament Formation via Collision-induced Magnetic Reconnection – Kinematic Features
The Collision-induced Magnetic Reconnection (CMR) mechanism has recently been proposed as a filament formation model that emphasizes the active role of magnetic fields. To enable better observational tests with data from modern telescopes, we provide more specific observable predictions for filaments formed via CMR. Two types of CMR kinematics are revealed in position-velocity (PV) diagrams. First, due to the nature of CMR, the midplane contains a converging motion close to the filament and two diverging motions farther from the filament. They exhibit a blueshift, redshift, blueshift, redshift (BRBR) velocity pattern along the line-of-sight when the collision midplane is inclined, giving rise to a special pattern in the PV-diagram. Second, the longitudinal PV-diagram along the filament spine exhibits a velocity oscillating pattern, which is currently attributed to the magnetic transport of gas clumps in the converging motion. The first type of pattern involves emission primarily outside the filament, which can be confused with the environment. The second type, however, mainly emerges from gas inside the filament, thus more robust for observational tests. Both the integral-shaped filament and the Stick filament in the Orion A cloud show the velocity oscillation in PV-diagrams, although the spatial oscillation frequency and their velocity spread differ.
Read the full abstract
Machine Learning
Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration
Safety guardrails in open-weight language models can be readily bypassed using Refusal Feature Ablation (RFA), a technique that identifies and projects out a linear refusal direction from the residual stream, often achieving a high attack success rate (ASR) while preserving model capability. Defending against these attacks typically requires computationally expensive safety finetuning for every new checkpoint. We introduce Decoy Direction Optimization (DDO), a fast, post-hoc weight-editing defense that requires no base-model finetuning. Our approach is based on a simple mechanistic insight: ablation attacks rely on contrastive estimators to find the refusal direction. Rather than trying to hide the true refusal circuitry, DDO actively injects a high-magnitude, nonlinear decoy signal into the network's MLP neurons. When an attacker attempts to locate the refusal direction, the decoy corrupts their estimator, tricking them into ablating a harmless orthogonal feature while the actual safety mechanism remains intact. We prove a spectral bound formalizing this effect and evaluate DDO across six model families, achieving <10% ASR under standard RFA. On Llama-3-8B-Instruct, DDO remains comparable to trained defenses under adaptive multi-phase attacks (65% vs. 58% worst-case ASR) and reduces Heretic weight-level attack ASR from 88.7% to 18%, all at 30 to 450 times lower optimization cost per configuration than the trained baselines.
Read the full abstract
General Economics
Global Poverty Beyond the Official Line: A bounded estimate of material insufficiency
Numbers this large invite a defensive reflex: reach for the reassuring figure and move on. By the most widely cited measure, the World Bank's extreme-poverty line of \$3.00 per day (2021 PPP), approximately 847 million people, or 10.4% of the world's population, lived in poverty in 2024. That figure is accurate as measured. It is also, by design, a floor: a threshold built to mark bare survival in the poorest economies, not to describe what it takes for a family anywhere to live with basic security. The central argument of this report is that the reassurance offered by that single line is largely an artifact of how we chose to measure. Held to the \$3.00 floor, roughly 847 million people are poor; held to the standards of their own societies, the number is several times larger, and even the most cautious, honest count exceeds a billion. Measured against income standards appropriate to each country's level of development, some 3.7 billion people fall below the World Bank's \$8.30/day upper-middle-income reference line, roughly 1.1 billion are multidimensionally poor, and around 2.3 billion are food insecure. These are not competing errors but answers to different questions. This report does not dispute the official data. It takes those figures as accurate and asks what they do and do not capture. It disaggregates by geography, age, and gender and supplements the monetary count with multidimensional measures of food security, working poverty, social protection, and relative poverty. It also documents a data-quality problem that runs in one direction: survey coverage is weaker in regions with the worst poverty, so official figures are far likelier to understate deprivation than to overstate it.
Read the full abstract
Earth and Planetary Astrophysics
Rescuing Kepler's False Positives: A Possible Habitable Zone Exoplanet in the Triple Star System KOI-1623
The Kepler mission has significantly advanced our understanding of exoplanet demographics. However, Kepler planet candidates showing centroid shifts are often classified as false-positives and thus excluded from demographic analyses. Some of these systems may be erroneous false positives, resulting from unresolved binary stars with planets orbiting one of the stars, thereby introducing biases into past exoplanet demographic studies. Here, we present updated Kepler centroid offset values derived from Gaia Data Release 3 (DR3) and analyze KOI-1623.01, a Kepler centroid offset false positive (COFP) with a clear transit signal and an orbital period of 111 days. High-contrast imaging with Keck/NIRC2 reveals that KOI-1623A has a close companion with a separation of $0.704 \pm 0.002$ arcsec, consisting of a slightly evolved G-type primary and a K-dwarf secondary, while Gaia detects a tertiary companion with common proper motion at a separation of $\sim 8.4$ arcsec. By combining transit photometry, Kepler pixel level data, direct imaging, and Gaia DR3, we find that the transit signal is consistent with a $10.02 ^{+0.44}_{-0.31} R_{\oplus}$ giant planet located in the habitable zone of the secondary. We furthermore use resolved companions from Gaia to estimate that another $\approx 70$ out of $515$ of Kepler's centroid false positives were misclassified and instead are real planets that belong to binary systems.
Read the full abstract
Image and Video Processing
A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires, July 2020-August 2026
This article describes an open image dataset for developing and evaluating active-fire segmentation methods in satellite imagery. The dataset contains 2,148 image-mask pairs from 25 California wildfires, with acquisitions spanning July 2020 to August 2026. Each image is a 512x512-pixel, three-channel composite derived from Sentinel-2 Level-2A bands B12, B11 and B8A at 20 m spatial sampling. A fixed linear rendering is applied throughout the dataset. Corresponding masks distinguish background, SWIR-rule active fire and invalid observations. The masks were generated from shortwave-infrared brightness and near-infrared contrast, followed by constrained neighborhood growth. The release includes chip-level metadata and an incident-disjoint partition containing 18 training, three validation and four test fires. Among the image pairs, 841 contain active-fire labels; these labels occupy 0.0766% of all grid cells. A mask-blind analyst review covers 233 test chips and provides a separate assessment of the rule-generated labels at chip and connected-component levels. Reference training and evaluation code accompanies the data, including a ResNet-34 U-Net implementation with validation-based checkpoint and threshold selection. The archived images, masks, metadata and review annotations support research on rare-class segmentation, learning from algorithmic labels and transfer across fire incidents. The versioned dataset is deposited on Zenodo, with preparation and reuse software maintained in a public GitHub repository.
Read the full abstract
Quantum Physics
Utility-Based Path Selection and Configuration in Quantum Networks via Layered Shortest Paths
A path in a quantum network is a chain of repeaters that distributes entanglement between two users. Selecting a path requires balancing the rate and quality (e.g., fidelity) of the delivered entanglement, but these quantities, unlike standard routing metrics, compose non-additively. The problem is compounded by link-level configuration choices (e.g., distillation rounds or emitter brightness tuning), each trading rate against fidelity, so that a path's performance depends jointly on its route and its per-link settings. We cast this joint path selection and configuration problem as a shortest path computation on a layered graph whose layers track discretized end-to-end fidelity. A single run returns the full rate fidelity Pareto frontier, from which the path maximizing any nondecreasing utility function of rate and fidelity can be selected. We prove that for certain utility functions (including the secret key rate of BB84), the method is a fully polynomial time approximation scheme, returning a near-optimal path within a specified tolerance. We further characterize exactly when cheaper scalarization-based routing suffices: it is optimal for utility functions with convex fidelity profiles, but can be arbitrarily suboptimal otherwise (e.g., for step-like, sigmoidal utilities), whereas the layered method remains reliable in all cases.
Read the full abstract
Nuclear Theory
Obtaining tensor-polarized nuclei with spin $S\ge1$ via spin filtering upon nuclear beam transmission through an unpolarized target (spin dichroism effect)
Upon transmission through an unpolarized target initially unpolarized nuclei with spin $S\ge1$ acquire tensor polarization due to spin dichroism effect. For nuclei passing several nuclear lengths in the target, the acquired component $p_{zz}$ can reach substantial values of ~0.2-0.8. The beam obtained in this way can be effectively used to study reactions with tensor-polarized nuclei.
Read the full abstract
Algebraic Geometry
Filtrations on Local Cohomology and Injectivity Theorems
We prove that the Hodge filtration on local cohomology is contained in the symbolic Ext filtration, and hence in the usual Ext filtration, answering a question of Mustata and Popa. Our main observation is that the Ext filtration admits a filtered-duality interpretation via Hartshorne's algebraic de Rham complex and its infinitesimal filtration. Comparing this construction with the filtered Du Bois complex yields the containment. The same mechanism proves the higher injectivity conjecture of Popa–Shen–Vo and several variants.
Read the full abstract
Combinatorics
Combinatorial slicing problems of polytopes: How (not) to reconstruct a polytope from its slices
We study combinatorial aspects of hyperplane sections of polytopes, focusing on how much the combinatorics of the sections determines the combinatorics of the original polytope. We show that, in general, combinatorial information about the sections is not enough to determine even the $f$-vector of the polytope. In contrast, for sufficiently generic simple polytopes, the function recording the number of vertices of each central section determines the full combinatorial type. We organize different combinatorial slicing properties into hierarchies, separately for affine and central sections, according to the level of combinatorial structure they determine on the polytope. In analogy with classical metric slicing problems, we formulate combinatorial analogues of the Busemann-Petty problem and Bourgain's slicing problem by replacing volume with face numbers, and show that they fail in every dimension and every face dimension.