Search arXiv⌕ Search

SEARCH · Search arXiv

Search Search arXiv

Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,171 records · Page 65Linked to original sources
arXiv2026

Location-Based Safety Management with Wearable Seamless Indoor-Outdoor Positioning

Reliable worker localization is essential for real-time industrial safety management, but seamless outdoor-indoor positioning remains challenging. GNSS performs well outdoors yet degrades near buildings and fails indoors, while many indoor systems require costly infrastructure or extensive calibration. This paper presents a practical safety-oriented positioning system that uses resource-constrained wearable tags and a device-cloud split architecture for continuous outdoor-indoor worker tracking. The tag integrates GNSS, IMU, barometric sensing, LTE Cat-M communication, and performs lightweight sensor fusion, while the cloud server executes map-aware fusion for continuous tracking. Developed as an operational platform, the system also includes a database, a map tool, and a dashboard that provide real-time monitoring, geofencing-based access alerts, and safety and productivity analysis. Experiments with multiple participants, each walking an approximately 1400 m outdoor-indoor route, using LTE Cat-M communication, achieve mean errors of 1.32 m indoors and 3.30 m outdoors, while maintaining low positioning latency and efficient power consumption. These results demonstrate the feasibility of wearable-based seamless positioning for scalable real-time safety management.

cs.RO↗
arXiv2026

Receiver-Specific Interventional Valuation of Semantic Atoms for Budget-Constrained Multi-Agent Communication

Cooperative multi-agent semantic communication requires deciding not only how a transmitter encodes its observation, but also which resulting semantic content should be delivered to each receiver under a limited link budget. We introduce an encoder-agnostic transmitter-side framework for receiver-specific valuation of independently selectable semantic units, termed semantic atoms. Each atom is assigned a signed value that quantifies whether delivering it is expected to improve or degrade the intended receiver's downstream decision. We formulate this effect through a controlled-delivery utility and identify it via prospective randomized payload delivery before the receiver acts. A shadow receiver estimates this utility using only transmitter-available information, from which receiver-specific leave-one-out (LOO) and Shapley atom values are computed. Experiments on a four-receiver intersection benchmark show 68.4% accuracy in recovering the sign of non-negligible LOO atom effects and a Spearman correlation of 0.302 between estimated and reference LOO values. Under limited budgets, we show that receiver-specific selection consistently outperforms shared and random alternatives, demonstrating the benefit of adapting semantic ranking to the intended receiver. At full affordability, exhaustive subset selection also outperforms full-payload transmission, showing that transmitting all available content need not maximize receiver utility. Overall, our results show that receiver-specific semantic valuation can support effective budget-constrained communication despite imperfect atom-value estimation.

eess.SP↗
arXiv2026

FilterFusion: Prior-Guided Adaptive Filter Fusion for Efficient Video Denoising

Real-world video denoising for communication systems must handle mixed degradations such as sensor noise, compression artifacts, and motion-induced inconsistency while remaining deployable under tight latency budgets. Purely deep models often achieve high accuracy but come with significant computational cost, whereas classical filters are robust and efficient but insufficiently adaptive. We present FilterFusion, a prior-guided multi-frame denoising framework that combines temporally-aware learned features with an interpretable bank of denoising branches. The model predicts a low-cost guidance signal and uses it to adaptively mix a compact set of denoising responses at the original resolution. The lightweight design uses only 109K parameters and explicitly injects temporal variance and spatial variance cues into the low-resolution controller. On CRVD (real sensor noise), it reaches 34.72 dB PSNR and 0.9330 SSIM with only 0.18 GFLOPs (network-only). Because the filter-bank operations are simple and regular (Gaussian blur, median, weighted sum), a straightforward compiled C implementation of the synthesis path achieves 17.9ms per frame (55.9 FPS) on a single CPU thread at 1080p---exceeding real-time and 122--252 times faster than lightweight CNN baselines that require over 2 seconds per frame. The results demonstrate a strong quality-efficiency-speed trade-off and a practical route to real-time denoising through low-resolution control and explicit filter-based synthesis.

cs.CV↗
arXiv2026

Strong coloring numbers of grids

For positive integers $n$ and $d$, the $d$-dimensional grid of side length $n$ is the graph $G(n,d)$ with vertex set $[n]^d$, in which two vertices are adjacent if they differ by $1$ in exactly one coordinate and agree in all the other coordinates. We determine, for every fixed $d$, the asymptotic growth in $r$ of the $r$th strong coloring numbers of these grids. More precisely for $n$ large enough, ${\rm scol}_r(G(n,d)) = Ω_d\bigl(r^{d-1}\bigr)$ as $r\to\infty$. This matches an upper bound given by a volume packing argument, which holds more generally for touching graphs of unit balls in $\mathbb R^d$. For the lower bound, we construct in $G(n,d)$ a bramble of order $Ω_d(n^{d-1})$ with each element of radius at most $dn$. This construction was found by OpenAI's GPT-5.6 Sol.

math.CO↗
arXiv2026

A Krause-type recollement for Q-shaped derived categories

Let k be a commutative noetherian hereditary ring and let Q be a small, Hom-finite, locally bounded k-linear category admitting a Serre functor and satisfying the strong retraction property, with nilpotent pseudo-radical. For every k-algebra A, we construct a recollement of the Q-shaped homo?topy category KQ(A Inj) of objectwise injective Q-shaped modules by its full subcategory Kac Q (A Inj) of exact objects and the Q-shaped derived category DQ(A). The construction uses complete hereditary cotorsion pairs and does not require A to be noetherian. If A is left noetherian, both KQ(A Inj) and Kac Q (A Inj) are compactly generated by special preenvelopes of finitely generat?ed stalk modules. We identify (Kac Q (A Inj))c with the idempotent completion of (KQ(A Inj))c/LI ((DQ(A))c ), where LI is the fully faithful left adjoint of the localization. This paper builds on and generalizes works of Krause and Liu

math.RT↗
arXiv2026

Decomposing Brain-LLM Alignment with Interpretable Semantic Representations

Large language models (LLMs) predict neural responses during naturalistic language comprehension, but the linguistic information underlying this brain alignment remains difficult to interpret. We use SCRIBE, an interpretable contextual semantic representation spanning 65 neurocognitively grounded semantic properties, as a lens for decomposing the brain alignment of one LLM, LLaMA2-7B. Using variance partitioning between LLaMA2-7B and SCRIBE on naturalistic fMRI data from seven participants listening to narrative stories (Moth Radio Hour), we separate brain-predictive variance shared between the two representations from variance unique to each. We find that shared variance is the dominant component across the whole brain and language-related regions, while both representations also retain significant unique contributions. We then examine whether the concreteness of individual semantic properties relates to shared and SCRIBE-unique brain-predictive variance. Concrete properties show greater shared variance with LLaMA2-7B than abstract properties across the whole brain and most examined regions, whereas abstract properties do not show greater SCRIBE-unique variance. Finally, property ablations reveal that semantic contributions extend across multiple dimensions, with cumulative effects on both shared and SCRIBE-unique variance. Together, these results show that, for LLaMA2-7B on this dataset, a substantial portion of brain-LLM alignment is associated with interpretable semantic information, while revealing complementary brain-predictive semantic information beyond the LLaMA2-7B representation.

cs.AI↗
arXiv2026

On the structure of polynomial ideal codes and additive Reed--Solomon codes

We study structural properties of polynomial ideal (PI) codes and additive Reed--Solomon (aRS) codes in the folded Hamming metric (equivalently, the Hamming metric over a larger alphabet). We first determine the maximum possible length of PI codes, prove that the dual of every PI code is a generalized PI code, and deduce that PI codes are dually QMDS together with an explicit formula for their weight distribution. We then show that PI codes defined by irreducible polynomials correspond to aRS codes with pairwise nonconjugate evaluation points. For these aRS codes, the previous results on PI codes apply (duality, QMDS property and weight distribution). We show that, asymptotically, aRS codes with nonconjugate evaluation points achieve the same length as PI codes. We also compute the permutation group of such aRS codes. We also extend the family of PI codes to consider arbitrary polynomial modules (not necessarily pairwise coprime), and study when these and general aRS codes satisfy the QMDS and dually QMDS properties.

cs.IT↗
arXiv2026

How Much Screening Buys a Second? Speed, Selection, and the Maker's Spread

We study the spread a liquidity provider needs to break even against counterparties that can exploit a stale quote, and we place two defenses on a single scale: repricing sooner and screening informed flow. A break-even condition pins down the exact screening rate at which a slower maker matches a faster unscreened venue: the selection that buys back the extra staleness. The result requires only an integrable price innovation and therefore holds under jumps, atoms, and infinite variance. Predictable drift is a centering problem: it widens a spot-centered quote, but its cost vanishes once the quote is recentered on the expected reference. In four USDC-WETH pools benchmarked against Binance ETH-USDT from June to August 2026, the one-second benchmark innovation is sharply non-Gaussian: 73% of seconds show no price change. Cutting cadence from twelve seconds to one sheds 84.5% of the expected shock, so a twelve-second maker must screen at least 84.5% of informed flow to match a one-second venue, far above the 71.1% that a Gaussian model implies. In trending markets, the stale pool is picked off at execution: with-trend fills obtain an effective spread 11.1 bps lower than against-trend fills, with the same sign in all four pools.

q-fin.TR↗
arXiv2026

Imaginary quadratic fields of class number at most 2000

We determine all imaginary quadratic fields with class number at most 2000, unconditionally: there are 16,453,493 of them, and the largest $|d_K|$ among them is 1,062,350,083. To our knowledge, such a list was previously known only for class number at most 100. We make explicit a recent argument, released by OpenAI and not refereed, that excludes Landau-Siegel zeros: every real zero $β$ of $L(s,χ)$, for a primitive real character $χ$ of conductor $q\ge3$, satisfies $(1-β)\log q\ge5.8\cdot10^{-4}$. An elementary lower bound for $L(1,χ)$ then shows that a field with class number at most 2000 has $|d_K|<1.61\cdot10^{15}$. The fields with $6\cdot2^{28}\le|d_K|\le1.65\cdot10^{15}$ are excluded by Watkins' sieve, and those below are read from the class group tables of Mosunov and Jacobson. As a consequence, the lists of Cremona and Sutherland of imaginary quadratic orders with class number at most 1000, and the counts of Holmin, Jones, Kurlberg, McLeman and Petersen for odd class numbers up to 1999, previously known under the generalized Riemann hypothesis, hold unconditionally. The two steps, the bound for $β$ and the bound for $|d_K|$, are proved in Lean, building on the existing formalization of that argument, and the proofs are checked by two independent kernels. The contributions of AI systems to this work are stated in the paper.

math.NT↗
arXiv2026

QREF: Improving DeepSearch Agents by Aligning Search with Reasoning via Query Refinement and Evidence Filtering

DeepSearch agents achieve strong performance on complex information-seeking tasks through iterative search and reasoning. However, they often issue intent-misaligned, under-constrained, or redundant search queries, resulting in poor evidence coverage and accumulating retrieval noise. To mitigate these limitations, we propose QREF, a plug-and-play Query Refiner and Evidence Filter. Specifically, the Query Refiner rewrites search queries according to the current reasoning intent while ensuring complementary coverage across each search batch. The Evidence Filter retains evidence relevant to both the current search batch and the reasoning objective, reducing distracting retrieval results. We initialize the Query Refiner with supervised fine-tuning and optimize it through reinforcement learning based on interactions with a real-world web search environment, reducing ineffective queries. Comprehensive experiments demonstrate that our approach improves the performance of DeepSearch agents across four challenging benchmarks, achieving average relative Pass@1 improvements of 33.55% on WebSailor and 11.04% on TongyiDeepResearch without modifying architectures or retraining their reasoning policies. Compared with the original DeepSearch backbone without QREF, the RL-trained refiner reduces the fraction of query batches that lack discriminative retrieval anchors by 95%. Compared with the SFT-trained refiner, it also lowers the filter-based ineffective query ratio.

cs.AI↗
arXiv2026

Probing cold gas and dust in quiescent galaxies across cosmic time

The discovery of massive quiescent galaxies at high redshift by JWST has revealed that star formation can cease rapidly in the early Universe, yet the physical mechanisms responsible for quenching remain unknown. A key missing element is the systematic measurement in statistically representative quiescent samples of the cold gas and dust content, which are crucial for distinguishing quenching mechanisms and evolutionary pathways. Such observations are currently not feasible because state-of-the-art sub-millimeter facilities (e.g., ALMA) cannot routinely detect the faint molecular gas and dust reservoirs of quiescent galaxies, nor pinpoint outflowing molecular gas. A next-generation facility (ALMA2040) capable of providing transformative access to the cold interstellar medium of quiescent galaxies across redshift is therefore essential. In particular, it will allow the community to probe the molecular gas content of quiescent galaxies at cosmic noon and beyond by targeting the CO, [CI], and [CII] emission lines, as well as unveiling the physical conditions of the cold interstellar medium by constraining the CO spectral line energy distribution. It will also reveal the dust content of the passive population, a crucial ingredient to constrain the different quenching mechanisms. Finally, it will enable probing the molecular gas kinematics and identify molecular outflows driven by supermassive black holes. When combined with the rich legacy of JWST, ELT, and SKAO surveys, ALMA2040 will deliver a comprehensive, multi-phase view of feedback and quenching, allowing us to establish when, how, and why galaxies shut down their star formation across cosmic time.

astro-ph.IM↗
arXiv2026

Poison Valleys: Diagnosing Path-Dependent Failure in LLM Code Repair

When code-repair agents fail, final tests reveal that a patch is wrong but not which earlier commitment caused it or which software obligation failed. Poison Valleys traces this path in 1,844 model requests on two synthetic Python cases: PersistQueue adds atomic batch claiming to a disk-backed queue, while Dijkstra adds source-sensitive filtering to shortest-path search. We define plausible but incorrect software mechanisms, place each in a model's hypothesis, implementation plan, or both, and evaluate the generated code with an Oracle and targeted diagnostic coordinates. Moving the misconception from hypothesis to plan increased failure by 32.2% in PersistQueue and 28.9% in Dijkstra after matching the experimental mix. Different misconceptions produced distinct, model-dependent patterns of violated obligations, although several could converge on the same behavior. This creates a measured connection from upstream misconception to code mechanism and failure form. Code, data, and demo: https://poisonvalleys.org.

cs.SE↗
arXiv2026

TRACE: Typed Reference-Aware Conversational Editing for Humanoid Motion Diffusion

Language-conditioned humanoid motion generation enables users to create and refine robot movements through natural language. The central challenge is to connect an underspecified request to concrete motion structure that can subsequently be revised relative to an accepted reference. We present TRACE, a generation and editing framework for the Unitree G1 humanoid robot built around typed, measurable kinematic quantities. Our insight is to ground generation conditions in paired training motions and editing goals in the current motion. A language model learns to predict support, root, and hand events from motion-derived supervision; their time-aligned encodings complement the original text in a diffusion generator. Gravity-referenced states and trajectory--contact objectives improve kinematic consistency. For editing, typed requests define body--time scopes and source-relative goals. A separate reference-conditioned diffusion editor uses these goals and revised events to realize changes, while sampling-time projection preserves protected state coordinates. Quantitative evaluation and human assessment show improved semantic alignment and motion quality over the compared methods. Controlled ablations support the value of motion-grounded conditioning and kinematic consistency, while multi-turn editing evaluations demonstrate more reliable execution of successive relative requests under shared preservation constraints. Together, these results support typed kinematic structure as a common basis for motion generation and reference-aware revision.

cs.RO↗
arXiv2026

Space Sustainability: The Global Commons

Space is no longer a distant, barely reachable frontier of scientific exploration; it has become a critical infrastructure for life on Earth and a rapidly expanding zone of intense commercial activity. Yet this expansion is unfolding in ways that replicate, and amplify, the unsustainable dynamics already driving ecological and social crises on our planet. This chapter explains the concept of space sustainability as a multidimensional challenge; it raises the questions of governance, knowledge, power, justice, and global responsibility. It highlights both the profound benefits derived from space activities and the uneven distribution of costs and risks, from orbital debris and atmospheric impacts to privatization of essential services. We therefore argue that it frames space as a global commons that demands stewardship rather than pure exploitation, cooperation rather than competition, and governance that prioritizes long-term wellbeing rather than laissez-faire approach. Sustainability in space is inseparable from sustainability on Earth. This chapter invites readers, whether they are space administration and industry actors, policymakers, scientists, or simply engaged citizens, to see space not as a limitless frontier, but as a shared domain whose future depends on collective choices made today.

physics.soc-ph↗
arXiv2026

SCRIBE: Scalable Contextual Representations of Interpretable Brain-Grounded Semantics

Understanding how the human brain constructs meaning from real-world language remains a fundamental challenge, as semantics depend on context that varies in natural settings. To enable the study of real-world linguistic meaning in the brain, we introduce SCRIBE, a framework grounded in a neuroscience-informed semantic space that leverages an ensemble of open-source large language models to generate interpretable, word-level semantic representations of arbitrary text. In a behavioral study with 921 participants, we show that human semantic feature judgments vary systematically with context, and that SCRIBE captures these shifts, reaching 75-81% of the estimated human noise ceiling. Using brain activity recorded while participants listen to naturalistic stories, we further demonstrate that SCRIBE predicts neural responses in language-selective brain regions, significantly outperforming non-contextual semantic representations. Variance partitioning reveals that SCRIBE explains unique brain-predictive information beyond syntactic features, suggesting that its representations capture aspects of linguistic meaning not accounted for by the measured syntactic features. Together, these findings establish SCRIBE as a scalable and interpretable foundation for studying natural language semantics in the brain.

cs.AI↗
arXiv2026

Hierarchical Interpretable Multitask Audio Representation Learning: An Analysis of Event-Scene Relationships through the Information Bottleneck

Joint Acoustic Scene Classification (ASC) and Audio Tagging (AT) models traditionally rely on unconstrained multihead architectures. However, this design ignores the natural hierarchical dependency between sounds and acoustic scenes. In this work, we investigate a hierarchical (sequential) conditioning paradigm that predicts scene categories directly from intermediate event probabilities. Through a large-scale empirical benchmark with 66 full training runs across three state-of-the-art backbones (PaSST, PANNs, and BEATs) and different loss-weighting settings (beta in [0.0, 1.0]), we show that hierarchical conditioning achieves competitive classification performance while altering multitask optimization dynamics. Unlike multihead baselines, which sequentially solve the easier ASC task before optimizing AT, the hierarchical architecture enables synchronized dual-task learning for specific loss-weighting settings (beta in [0.6, 0.9]). Furthermore, we propose a multi-perspective interpretability framework that combines an analysis of the information bottleneck, manifold geometry, and Formal Concept Analysis (FCA) on a Term-Frequency Inverse-Document-Frequency (TF-IDF) event representation, providing complementary insights into event-scene relationships. Our findings reveal that hierarchical conditioning systematically yields lower conditional entropy, higher mutual information, and superior latent cluster separability (higher Silhouette scores), establishing a more structured and semantically faithful intermediate feature space.

eess.AS↗
arXiv2026

Wavefront Engineering for Physical Layer Security in Near-Field Wireless Communications

With wireless transmission exposing information to potential interception by unintended receivers, physical layer security (PLS) offers a promising approach to securing wireless communications. Beamforming can mitigate this vulnerability by exploiting the spatial characteristics of wireless propagation to concentrate power toward the intended receiver while suppressing it at potential eavesdroppers. However, as wireless systems enter the radiating near field of large-aperture antenna arrays, the resulting wavefront curvature provides additional opportunities for spatial wavefront engineering to enhance PLS. In this work, we investigate the use of different beam types for PLS, including focused, curved, abruptly autofocusing, and cosine beams, generated using uniform linear and planar arrays (ULAs/UPAs). To quantify their performance, we introduce power-based metrics that characterize the power received by the legitimate user and a potential eavesdropper. For all beam types we introduce focusing along the vertical UPA direction, to independently control the phase distributions along the two UPA dimensions and enhance the power discrimination between the legitimate receiver and the eavesdropper. We also propose hollow-focused beams, in which an inactive strip is introduced in the transmitting aperture to create a protected low-power region between the transmitter and the legitimate receiver. The results demonstrate that vertical focusing dramatically enhances the PLS performance and that hollow-focused beams achieve superior PLS performance in both line-of-sight (LoS) and vicinity-of-user (VoU) scenarios.

eess.SP↗