Computer Science & Mathematics

We study sampling and recovery for tensor-valued dynamical systems evolving under the t-product, using partial spatial observations over time. Although the t-product Fourier transform yields slicewise matrix dynamics, recovery does not generally decouple across slices. Eigenvalues shared across Fourier-domain frontal slices couple the corresponding eigenspaces into aggregated spectral blocks, which determine the tensor-specific recovery geometry and space--time tradeoff. We establish deterministic identifiability conditions for the initial tensor state. In the diagonalizable case, these are frame conditions on the aggregated blocks. We extend them to non-diagonalizable systems through generalized eigenspaces and Jordan structure. A temporal--spectral factorization of the finite-horizon stacked recovery system underlying Tensor Dynamical Sampling (DS) yields a decoupled two-stage formulation once the temporal threshold is reached. Under independent spatial subsampling, we prove high-probability recovery guarantees using empirical Gram matrices on the aggregated blocks. The bounds depend on aggregated multiplicities, incoherence, and population conditioning. Stability estimates separate the spatial frame factor from the temporal Vandermonde factor and show that increasing temporal depth may worsen numerical conditioning even after identifiability is ensured. We implement Tensor DS and evaluate it on synthetic and real data. Phase transitions track the predicted blockwise sampling thresholds, and stability experiments illustrate Vandermonde conditioning. Experiments on sea surface temperature and turbulence velocity data show that a learned t-product system operator generally yields more accurate and structurally faithful reconstructions than vectorized low-rank, Tucker-type, graph-regularized, graph-spectral, and vectorized dynamical sampling baselines.

Text injection trains speech recognition models on unpaired text alongside paired audio. We feed the speech encoder synthetic input derived from text so the whole model learns from it. Previously, we synthesized this input with a frozen text-to-speech (TTS) model. Here, a simple lookup table replaces the TTS model: a mean log-mel vector (or one-hot encoding) per phoneme, repeated for a random duration drawn from alignment statistics, with silence between words. Training is 1.8 times faster than with the neural TTS, achieving comparable word error rates. With an external language model, our system reaches 1.36 / 2.90 / 1.52 / 3.24 % WER on LibriSpeech {dev,test}-{clean,other}, to our knowledge state-of-the-art using only LibriSpeech paired data, text, and lexicon. We analyze components of synthetic input (durations, acoustic features), compare different HMM aligners, evaluate scaling on the 25 000 h Loquacious dataset, and show that full-encoder injection yields strong relative improvements compared to partial-encoder and decoder methods. The code of the whole pipeline behind every number in this paper is public: github.com/rwth-i6/returnn-experiments/tree/master/2026-text-injection

Electrical detection of antiferromagnetic order is a critical challenge for antiferromagnetic spintronics owing to their vanishing macroscopic magnetisation. While magnetic resonance has demonstrated sub-terahertz antiferromagnetic spin pumping and nonlinear magnon dynamics, its application to electrically contacted on-chip devices has room for further exploration. Here we show that magnetic resonance enables direct electrical detection of Néel vector states. A $180^\circ$ reversal in synthetic antiferromagnets is realised by current-driven effective fields with opposite polarisation acting on each magnetic sublayer, which we term synthetic staggered fields. The switching induces a $π$ phase shift of precessional motion in time, producing hysteretic rectified voltages in our phase-sensitive detection scheme. Angular-dependent resonance analysis offers further insight into the Néel vector directions, pointing to a model beyond the macrospin formalism. We address this by defect-incorporated micromagnetic simulations that reveal fragmented domain states after current excitation. An extracted current density of $7.4\times10^{9}$ A/m$^2$ is, to our knowledge, the lowest reported value for the room-temperature Néel vector switching and we show that it scales with the in-plane uniaxial anisotropy both experimentally and numerically. Synthetic staggered fields and the resonance-based methodology provide a research platform for nanoscale characterisation of electrically driven antiferromagnetic textures, guiding the development of high-frequency on-chip spintronic architectures.

We study online fair division of indivisible goods among $n$ agents with frequency predictions: before any good arrives, the algorithm learns the multiset of each agent's values for the goods, but neither the arrival order nor which values of different agents belong to the same good. Neoh et al. (ICML 2026) showed that such predictions allow every share-based guarantee that is achievable offline, and left open whether a positive approximation of envy-freeness up to one good (EF1) is possible. We show that deterministic online algorithms can guarantee $1/2$-EF1 for three agents and $(3-\sqrt5)/2$-EF1 for four agents, whereas for five or more agents, no online algorithm can guarantee any positive approximation of EF1, even with $n+1$ goods and identical predictions. Together with the known result for two agents, a positive approximation of EF1 is therefore possible if and only if $n\leq4$. We also rule out any positive approximation of EFX for three or more agents, even when the valuations take only two distinct positive values. Next, we use the meta-algorithm of Neoh et al. (ICML 2026) to compute an epistemic EFX (EEFX) allocation in polynomial time for any $n$, and hence a $1/(n-1)$-EF$(n-2)$ allocation for $n\geq3$, whereas algorithms that fix a picking sequence in advance can guarantee a positive approximation of EF$k$ if and only if $n\leq k+2$. Finally, for noisy predictions, we give algorithms that return EEFX allocations when the predictions are exact, and guarantee $(1-2\varepsilon+O(\varepsilon^2))$-EEFX under underestimates and, for $n\geq3$, $(1-2(n-1)\varepsilon+O(\varepsilon^2))$-EEFX under overestimates; the linear terms are the best possible, and larger overestimates rule out any positive approximation for such algorithms.

Entity matching is the task of deciding whether two records describe the same real-world entity. Current entity matching methods rely on pre-trained language models (PLMs) which require task specific fine-tuning or generative large language models (LLMs) which do not require task-specific fine-tuning but are slow and expensive at matching large record sets. Decision models are a new class of LLM-based models that return a probability for each predefined answer option rather than generate text. This paper investigates decision models as an alternative to generative LLMs and fine-tuned PLMs for entity matching. We evaluate six decision models in a zero-shot setting across eight benchmarks and compare their performance with that of the generative LLM GPT-6 Luna and the PLM-based matcher Ditto. Our experiments show that, without task-specific training data, the four strongest decision models achieve mean F1 scores close to that of GPT-6 Luna, with the best model slightly surpassing it. This model also outperforms Ditto on six of the eight benchmarks. All decision models are faster than GPT-6 Luna, and most are less expensive. For three of the four strongest decision models, the predicted match probabilities are well calibrated without requiring any task-specific recalibration. A model cascade that routes the least confident record pairs to an ensemble of generative LLMs improves the mean F1 scores of the four strongest decision models by 1.9-4.0 percentage points. Using the best decision model and its best-performing prompt for each benchmark, the cascade surpasses the highest published F1 scores from literature on four of the eight benchmarks.

Electro-optic materials, whose refractive index changes linearly with an applied electric field (the Pockels effect), set the speed and energy of optical modulators and microwave-to-optical links for quantum computing. Designing better materials requires optimizing composition alongside temperature, strain, and bias field. However, general repositories lack electro-optic tensor data. We present a pipeline that builds such a database from literature following the rule: "no value is better than a wrong value." The pipeline has five stages: physics-aware acquisition, layout-aware extraction via LLM and deterministic table parsers, a three-tier verification gate, cross-paper consensus using median and median absolute deviation, and a noise-robust surrogate model audited by leave-one-material-out (LOMO) testing. Physics enters directly: notation is mapped, point-group symmetry-forbidden Pockels components are rejected, equivalent labels are merged, and values carry provenance tags. In a run over 49 publications, the gate processed 816 candidates into 414 accepted, 215 held for review, and 187 rejected. Against a hand-certified gold standard of 303 values from 16 papers, the run achieved a precision of 0.51 and recall of 0.48 for exact record matches. Across four database snapshots, held-out materials where the surrogate model beat a predict-the-mean baseline grew from 1 of 4 to 9 of 13 as the database expanded from 10 to 22 materials. While not a controlled comparison, this trend supports the premise that a precision-oriented, physics-checked database can power forward models beyond its training set and clarifies requirements for inverse material design.

We study the vulnerability of beam allocation approaches in large low-Earth-orbit (LEO) satellite constellations to adversarial demand inflation resulting in denial-of-service attacks. We consider an attack model in which an adversary places (injects) additional terminals and evaluate these attacks against greedy and max-min-fair beam allocation algorithms, each combined with either equal or demand-proportional intra-cell capacity sharing. Vulnerability is measured using mean demand satisfaction, coverage, and the fraction of legitimate users receiving at least a threshold fraction of their requested capacity under several attack heuristics. There is a trade-off for the system between robustness and performance, meaning the algorithm with the most vulnerability to the attacks (greedy-proportional) has the highest performance before the attacks. Demand-proportional sharing increases the fraction of under-served users up to 34% under both allocation algorithms, whereas equal sharing limits the increase up to 13%. The sharing rule also changes the most damaging attack allocation in nearly all cases: demand-proportional sharing rewards concentrating injected terminals in the most loaded cells, while equal sharing performs better when distributing these terminals across the least loaded cells. Furthermore, the max-min-fair beam allocation algorithm forces the attacker to use a higher number of targeted cells to result in extensive damage to the system.

Driving video captured through a windshield violates the world-space consistency assumed by 3D Gaussian Splatting (3DGS): the road scene exhibits parallax while cabin reflections remain attached to the camera, causing standard 3DGS to embed reflected content into the reconstructed scene. We introduce RIDE-Splat, a pipeline for reflection-free 3D reconstruction from posed driving video without clean images, reflection masks, or pretrained networks. RIDE-Splat assigns world-consistent transmission through the windshield to a 3DGS scene and the dominant camera-attached reflection to one shared, signed RGB image. It builds the clean reconstruction in two passes. First, it learns the 3D geometry and the signed reflection image, representing both added radiance and transmission attenuation. Then, it uses the reflection-removed training images to fit a clean 3DGS model. At inference, the reflection image is discarded, preserving the standard 3DGS rendering path. We also introduce the RIDE-Splat Dataset, the first 3D reconstruction benchmark for reflection removal from driving video -- with real driving scenes, physical windshield reflections in training views, and separately captured non-reflective evaluation images. Across the RIDE-Splat Dataset, XLD, and Wayve, RIDE-Splat outperforms 3DGS, DC-Gaussian, and RobustSplat on all nine aggregate metrics, improving PSNR by up to 3.59 dB and reducing LPIPS by up to 0.0876. We achieve the highest metrics in 38 out of 42 per-scene comparisons and removes windshield reflections that remain visible in the baselines while preserving the transmitted scene.

Repository-level code migration rewrites an entire project in a new programming language while preserving its observable behavior. It is challenging because behavior spans interdependent modules, dependencies and build infrastructure must be adapted to the target ecosystem, and the source code is often the only specification. Large language model agents can now attempt these migrations end to end, but three challenges remain. First, an agent that writes both the translation and its acceptance tests may overlook the same requirements in both, leaving missing or incorrect behavior undetected. Second, large repositories can exceed a model's context window, and dividing the work into smaller tasks does not ensure that all required behavior is migrated. Third, even behaviorally correct translations may be slower than the original implementation. We present CodeMeepo, a multi-agent pipeline for complete and efficient migration of large repositories. Each milestone is evaluated against a fixed external oracle that the agents neither author nor modify. A separate parity assessment compares source and target components in scope and adds identified gaps back into the plan as required work. Ordered milestones limit the scope of each task, while cumulative regression checks guard against breaking already-migrated behavior. A benchmark-guided optimization phase addresses performance, followed by full-suite conformance testing and repair. Eight specialized agent roles communicate through file artifacts, and a deterministic state machine governs repair, deferral, scope changes, and termination. On nine subjects in three language pairs, CodeMeepo achieves the best pass rate on every subject and passes 18% more tests than a state-of-the-art multi-agent baseline using the same model. On a production daemon, its optimization phase improves half of the benchmarks, by 25% on average.

Reusable Web-agent verification requires a rule to follow changing requirements and support subsequent execution. We introduce AgentProofBench, which connects paired tests of a frozen acceptance rule to execution from identical saved openings. The resource spans 97 tasks, 79 families, and four Web services. Experiments with DeepSeek Flash/Pro establish two linked measurements. Across 48 families, complete-panel and decision-replay tests retain 82/96 Flash and 79/96 Pro source programs; paired reuse-rebuild BA intervals are [-5.47, 1.57] and [-10.16, 0.00] percentage points. Across 22 execution families, deployment-state panel correctness separates program assignments whose two saved openings are both delivered correctly: 32/33 assignments with error-free panels, versus 1/11 with panel errors. All reference-defined program-gate regressions and withheld deliveries occur in the latter group. Neutral continuation attains the highest correct delivery under the tested generic feedback, providing a concrete baseline for measuring the value added by a gate. Released instructions, states, programs, reference rules, and trajectories connect each test failure to its execution consequences. AgentProofBench makes reusable verification assessable at both the requirement and policy levels.

Animals have a remarkable capacity for autonomy---driven by intrinsic motivation---without external sensory cues or extrinsic objectives, yet the computational mechanisms behind this ability largely remain a mystery. We combine whole-brain recordings in the fruit fly (Drosophila melanogaster), which reveal neural representations of spontaneous exploratory decisions more than 10 seconds in advance, with a neuromechanical digital twin to test whether optimizing intrinsic goals yields realistically embodied artificial agents that quantitatively match fly behavior and internal representations of future volitional action during exploration. Our hierarchical policy selects from a discrete library of low-level motor primitives, consistent with ethological theories of behavioral syllables. We train agents without external reward using 20 intrinsic motivation algorithms spanning curiosity, entropy, and competency methods. The agents are evaluated in a virtual tethered-ball assay modeled on the fly experiment, comparing spontaneous locomotor statistics and internal representations and decoding future turn direction from kinematics and neural activity on the same timescales observed in flies. DIAYN, a hierarchical competency algorithm learning discrete policies that maximize behavioral entropy and the separability of induced sensory observations, achieves the strongest joint behavioral and neural prospective coding performance relative to flies. Its learned policies closely capture individual variability in fly exploration strategies, suggesting that this diversity serves to maximize state-space coverage at the population level. DIAYN's latent activity also closely captures the sum-difference rule for interhemispheric encoding of kinematics in biological flies. These results establish hierarchical, intrinsically motivated agents as a powerful framework for reverse-engineering animal autonomy.

How do later loops in a pretrained looped language model repair two-hop retrieval? We study Ouro-1.4B and Ouro-2.6B by transplanting attention queries between loops at the same layer, head and final prompt position. On synthetic in-context questions, injecting the layer-indexed collection of loop-3 queries throughout loop 1 rescues 90% and 97% of questions answered incorrectly at loop 1 and correctly at loop 3, versus 24% and 21% for shuffled donors. The rescue also holds under greedy generation in Ouro-2.6B; in Ouro-1.4B it is limited to four-candidate ranking. Crossed donor-recipient swaps favor retrieval through the donor's bridge entity over copying its answer, although a substantial positional component remains in Ouro-2.6B. Late keys and values rescue far fewer targets. Freezing queries at their loop-1 values also damages two-hop accuracy more than a matched question that supplies the bridge. On natural questions, rescue is smaller and replicates only in Ouro-2.6B, and crossed swaps do not establish bridge redirection. These results identify a transferable collection of late-loop queries that supports bridge-conditioned retrieval on synthetic scenes; they do not locate where the bridge is computed or establish a single-vector handoff.

Large language model (LLM)-based personal agents are increasingly deployed to handle everyday tasks on users' behalf. These agents have access to private user information, while their tasks involve communicating with third parties; this raises a critical privacy concern that agents may inappropriately reveal sensitive user information to these recipients. Existing benchmarks primarily assess direct privacy leakage, where an agent explicitly states a pre-specified secret or attribute. However, they do not capture \textit{inferential leakage}, where the model leaves indirect cues in the reply that implicitly leaks sensitive attributes. We introduce \textsc{InferPriv}, a benchmark that measures whether an honest-but-curious recipient can infer a user's sensitive attributes from an LLM's reply on routine task. Evaluating five frontier models, we find that inferential leakage substantially exceeds direct leakage across models and prompts. Under a privacy-aware system prompt, GPT-6.1-Sol, Sonnet-5.5, and Gemini-3.8-Flash achieve near-zero direct leakage, yet the inferential leakage rate remains between 16.2--25.2\%. These findings suggest that evaluating explicit leakage alone can substantially understate the privacy risks.

Machine learning across materials science, chemistry, energy, and medicine increasingly relies on literature-extracted datasets due to a lack of curated databases. However, automated extraction introduces errors: text recognition misreads digits, units and notations are confused, values are misattributed, and LLMs hallucinate numbers. Drawing on our work building an electro-optic property database, we present a taxonomy of errors organized by their origin and signatures, alongside a layered framework for error detection. The framework combines schema and provenance checks, physical admissibility (bounds, symmetry, and governing equations), cross-source statistical consensus, model auditing, and suspicion-ranked review. In a production run, our verification gate halted 402 of 816 candidate values with typed reasons. Yet, an audit against curated references showed that 30 of 112 accepted values were defective; standard symmetry rules alone caught 21 of these. On measured BaTiO3 and KNbO3 tensors, constitutive identities caught 8 of 10 single-value errors and isolated the corrupted entries. In a controlled benchmark, the combined framework caught 81% of errors at 85% precision, including errors in single-report materials, bringing model accuracy on held-out data to within 3% of a model trained on clean data. Uncleaned data severely impairs discovery: 89% of the twenty largest values were errors, causing AI searches to prioritize false positives. Once bad numbers are controlled, missing measurement conditions emerge as the main limit, as distinct records become indistinguishable without full context. We outline directions for the field, including shared physics rule libraries and self-reporting databases.

Organisms must constantly weigh exploiting the resources they know against exploring for better ones. When foragers compete, the best behavior for each forager also depends on what its competitors do. How does competition change the strategic balance between exploration and exploitation? Here we address this question in a minimal model: two $ε$-greedy temporal-difference learners on a one-dimensional lattice, who start at a small reward at one end and are unaware of a larger reward at the other. For a lone forager, the optimal exploration rate increases monotonically with the value of the distant reward, and with its distance up to a critical distance. When two foragers share the lattice and split rewards, the choice of exploration rate becomes a game whose Nash equilibria have both foragers exploring, both staying, or one exploring while the other stays (anti-coordination). Under local best-response dynamics, pairs of foragers often converge on both exploring or both staying, even when anti-coordination would be socially optimal. This social dilemma arises for a wide range of parameters, and it persists even in a simplified model without explicit spatial structure. A minimal, naturalistic signal, in which one forager informs the other that it has found the distant reward, can remove many of the dilemmas in which both foragers explore, and can produce a Nash equilibrium whose collective payoff exceeds that of every outcome available without signaling.