Science & Mathematics

Spoken interaction is moving from alternating turns to full-duplex conversation, where an agent listens while speaking. Learning this coordination requires natural dialogue with persistent speaker-resolved tracks, a data requirement that remains difficult to satisfy at scale. Generative recovery offers a route from ordinary recordings, but recovering two voices is not enough: their track assignments can drift as the conversation unfolds. We address this missing link with speaker-anchored routing before diffusion: a 1.84M-parameter module matches the separator's coarse streams to fixed enrollment identities and orders them before a frozen generator reconstructs the speech, without retraining the generator or altering the reconstructed audio. On 25-minute remote conversations, routing cuts wrong-track time from 43.9% to 4.1%, below the 6.2% of matched-enrollment waveform re-identification, with its largest advantage when one participant holds the floor (17.8% versus 28.6%), and raises fixed-track activity F1 from 0.46 to 0.84 while per-chunk reconstruction quality is unchanged. The same router transfers to unseen face-to-face and telephone recordings, and a margin-gated verification step brings their wrong-track time to 0.9-1.7% while embedding 20-36% of the output, where full re-identification embeds all of it. Recovered voices thus become the participant-consistent tracks that full-duplex dialogue modeling needs. Audio examples: https://elpsykongloo.github.io/beyond-two-tracks/

Autonomous Unmanned Aerial Vehicle (UAV) navigation in urban environments requires precise pose estimation. Because Global Navigation Satellite System (GNSS) signals are often unreliable in cities, visual localization serves as a crucial complement. Within this domain, textureless Level-of-Detail (LoD) city models provide a lightweight and privacy-preserving substitute for heavy 3D textured meshes. However, localizing against abstract geometry is challenging. Existing pipelines typically score 2D silhouette overlaps, which can suffer from severe geometric ambiguities. We present MatchCityLoc, a framework that shifts this paradigm to direct feature matching from a noisy 6-DoF pose prior. By extracting surface normals from the query image using a zero-shot foundation model and matching them against rendered LoD normals, we establish explicit 3D-to-2D correspondences. This enables a Perspective-n-Point (PnP) pose solve, which we iteratively refine through a render-and-match loop. Experiments demonstrate strong pose recall and accurate translation and rotation among the evaluated off-the-shelf LoD-based methods with very few renderings. By bridging the modality gap between aerial imagery and texture-free geometry, MatchCityLoc enables accurate localization using lightweight, privacy-preserving city models.

With Q-day approaching, the transition to post-quantum cryptography (PQC) has begun, with governments across the US, UK and EU mandating migration to quantum-resistant standards. This paper benchmarks the three NIST-standardised PQC algorithms - FIPS 203 (ML-KEM), FIPS 204 (ML-DSA) and FIPS 205 (SLH-DSA) - across key generation, signing and verification operations, measuring both computational performance and energy consumption on a range of constrained IoT-class devices, from the 32-bit Raspberry Pi 1 and Zero to the 64-bit Raspberry Pi 4, as well as commodity laptop hardware. Using the FNIRSI FNB58 USB power meter and OpenSSL~3.5, results show that ML-KEM and ML-DSA achieve energy and speed efficiency comparable to classical elliptic-curve cryptography across all tested architectures. However, SLH-DSA signing is inadvisable on constrained hardware: energy costs for SLH-DSA signing on 32-bit devices were up to 243\% higher than on equivalent 64-bit hardware, making it impractical for resource-limited IoT deployments. These findings have direct implications for IoT security practitioners planning PQC migration.

Retrieval-augmented QA can fail even when the correct answer is in the retrieved evidence. This can happen when a competing answer satisfies some, but not all, of the question's requirements and is mentioned frequently in the retrieved documents, which can make it appear to be the best-supported answer. To test whether this retrieval skew actually affects the decision, we manipulate it directly: increasing the prevalence of passages that favor the competing answer makes the model more likely to select it. This points to a decision problem that candidate-level support alone cannot resolve. We therefore introduce Joint Constraint Aggregation (JCA), which compares candidates against the same question-derived requirements using the retrieved evidence. Across three evaluation settings and three instruction-tuned backbones, JCA improves overall accuracy, and its advantage grows when retrieval favors a competing answer. Together, these results suggest that when partial matches and retrieval prominence pull models toward a competing answer, comparing candidates against the question's full requirements can make selection more reliable.

LLM agents act on the external world through tools and third-party skills, so a single executed action can cause irreversible harm. Existing defenses judge a proposed action without tracking whether its content comes from the user or from external sources such as tool outputs, and they judge it in isolation rather than against an available alternative, which forces a trade-off between security and utility. We present PROVISO, a controller that decides which action to commit after the agent proposes an action and before it is executed. PROVISO encodes trusted observations, external observations, the proposed action, and an alternative into provenance-typed memory, relates them through role-specific queries, and commits the action with the highest calibrated value learned from outcomes. On four agent security tasks spanning the agent lifecycle, no baseline exceeds PROVISO in both security and utility. PROVISO improves F1 over the strongest baseline by 13.84 points on SkillsBench and by 7.07 points on ToolSafe, and it reaches 88-99 F1 on three additional skill detection benchmarks.

LLM agents perform tasks over multiple steps and revise their actions based on feedback. As the total reasoning budget is often limited and must support future steps, the agent must choose between short and long reasoning before each action. Longer reasoning may improve the final task outcome by preventing later errors, but it also uses budget capacity that may be needed for future decisions. This creates two challenges: estimating the downstream value of long reasoning and deciding when that value is worth using the remaining budget. We propose Paired Outcome-guided Reasoning Allocation (PORA), which learns this decision by comparing short and long reasoning from the same interaction state. During training, \method{} completes the task twice from each selected state, once with a short reasoning allowance and once with a long allowance, and compares their final quality and total processing to estimate when longer reasoning is valuable. During inference, a controller weighs this estimate against learned prices on remaining-budget pressure and chooses either short or long reasoning without running both alternatives. We evaluate PORA across application, memory, code, retail, and mathematics tasks, achieving strong results across all settings. On BFCL memory, it exceeds the strongest existing method by 17.4--32.9 percentage points. Within the same agent, it matches or exceeds always using long reasoning on code and mathematics with 50.6% fewer actor tokens and 31.2% fewer model requests, respectively.

Multimodal question answering (MMQA) is based on integrating heterogeneous data sources, selectively leveraging relevant modalities, and ignoring distractors. Existing approaches based on Multimodal Large Language Models (MLLMs) improve performance by splitting the task into intermediate steps, but do not quantify how each modality contributes to reaching the final answer. We argue that quantifying the effect that each modality has on the answering process improves process traceability and final performance. Thus, we propose Entropy-Driven Multimodal Answering (EDMA), which uses logical entropy to formalise multimodal QA as an iterative process of entropy reduction and to quantify the contribution of each modality in distinguishing correct from incorrect candidate answers. This enables the construction of quantifiable answering trajectories, where each step is associated with a measurable entropy reduction. EDMA introduces (i) modality-conditioned partitioning to estimate unimodal entropy reduction, (ii) entropy-driven multimodal fusion to capture complementary information across modalities, and (iii) entropy-based selection of the answering trajectory that minimises the most logical entropy. Evaluated across standard MMQA benchmarks, EDMA outperforms competitive prompting baselines, with average gains of 12.8 F1 and 4.6 LLM-J points over Planning Prompting and 18.6 F1 and 8.3 LLM-J points over Chain-of-Thought. Moreover, EDMA outperforms self-consistency baselines and is competitive with the multi-agent system baseline. Moreover, we show that higher entropy correlates with more false positives and, through controlled trajectory selection, demonstrate that logical entropy serves as an effective, monotonic selector for correct reasoning paths, validating entropy level as a reliable proxy for performance.

Mixture-of-Experts (MoE) models reduce computation through sparse expert activation, but deployment still requires storing a large number of expert parameters. We study post-training expert pruning using task-domain calibration data as a subset-selection problem: given a pretrained MoE layer, select a fixed number of experts that best reconstruct its original output. Exact subset search is combinatorial and becomes intractable for modern MoE models with many experts per layer. We develop OptEP, a scalable continuous relaxation that directly optimizes differentiable expert-selection variables, followed by a lightweight reconstruction-based update of the retained expert weights. Across multiple MoE architectures and tasks spanning mathematical reasoning and code generation, our approach improves over the strongest existing expert-pruning baselines, with absolute accuracy gains of up to 15.05% on Qwen3-30B-A3B, 14.55% on Qwen3-235B-A22B, 11.06% on GLM-4.5-Air, and 16.20% on Mixtral-8x7B. The improvements are particularly pronounced under aggressive pruning, demonstrating that reconstruction-driven expert selection and weight adaptation can preserve substantially more of the original model's performance while reducing the number of retained experts.

A major challenge in calculating electron-phonon coupling at interfaces from first principles is the prohibitive computational cost due to the large supercells involved. Here, we present an efficient approach, which builds the electron-phonon matrix of weakly bound heterostructures from the properties of the isolated layers. The two key ingredients, the dynamical matrices and the potential response, are constructed in two steps. In the expansion step, these quantities are folded from the unit cells onto the supercell through an exact Fourier mapping. In the addition step, the interatomic force constants and the potential responses of the heterostructure are assembled blockwise from the corresponding quantities of the isolated layers, neglecting the weak interlayer contribution. The method is implemented in the electron-phonon module of the all-electron full-potential code exciting, enabling efficient calculations of the electron-phonon self-energy. We demonstrate the approach for the phonon dispersion and spectral functions of bilayer WSe$_2$ and pyridine/MoS$_2$, in comparison with exact reference calculations.

Internet path diversity is usually verified at the logical layer: two alternatives may share no AS or IP link yet traverse the same facility, conduit, or submarine landing zone. We call this gap illusory resilience and define the Illusory Resilience Index (IRI) as the latent hazard-weighted physical overlap of a logically disjoint path pair. To detect such hidden exposure without carrier underlay maps, we audit three public signals: correlated BGP withdrawals during documented physical incidents, operator interconnection-footprint overlap, and shared submarine landing zones. Two signals survive calibration; the 60-second BGP signal does not. Across 47 operators in ten regions, same-region pairs have substantially greater listed facility overlap than cross-region pairs, and one Marseille campus colocates 153 operator pairs. Across 68 cable systems in twelve corridors, 31-100% of candidate within-corridor pairs share a 25 km coastal hazard zone; two corridors are fully saturated, and the same rule recovers ten of eleven cable pairs documented as co-damaged in 2024. The rejected temporal predicate has a 0.444 null firing rate and no hits among twelve monitored positive pairs. Public data therefore expose substantial shared-risk structure behind nominal diversity and identify where stronger physical-path verification is warranted.

We give a short-description decomposition of bounded-width CNFs. For every fixed $\varepsilon>0$, every width-$b$ CNF on $m$ variables is an exact disjoint union of at most $2^{\varepsilon m}$ width-$b$ CNFs, each determined by $2^{O_\varepsilon(b)}\log(m+1)$ satisfying assignments. For every fixed $d\ge3$, we prove a worst-case size lower bound of $2^{Ω_d((n\log n)^{1/(d-1)})}$ against depth-at-most-$d$ unbounded-fan-in AND/OR circuits with input negations for three polynomial-time Boolean families: inverse trace over binary fields, the Goldreich--Tal trilinear function, and OpenAI's polynomial evaluator with its original parameters. All parameter bits count as inputs. For inverse trace, every such circuit of size at most $S$, where $n\le S\le2^{c_dn^{1/(d-1)}}$, has uniform sign correlation at most $2^{-η_dn\log\log S/(\log S)^{d-1}}$. Here $c_d,η_d>0$ depend only on $d$. This improves on the classical parity correlation bound. The lower bounds combine the decomposition with restriction arguments; the correlation bound additionally uses common partial decision trees. We also construct exact nonuniform AND/OR circuits for inverse trace of size $2^{O(n^{2/3})}$ at depth eight and $2^{O(\sqrt n\log n)}$ at depth nine. For every fixed integer $h\ge2$, the construction gives depth-$(4h+1)$ circuits of size $2^{O_h(n^{1/h}\log n)}$. The upper bounds implement finite-field evaluation and interpolation using circuits that enumerate intermediate states.

Model order reduction has the potential to enable real-time patient-specific assessment of cardiovascular flows. However, it still faces significant challenges stemming from the geometric variability of vascular anatomies. Building upon the SynthAorta dataset (Bosnjak et al., IEEE Trans. Med. Imag. 45(1), 2026) of physiological aortic geometries, we generate an extensive dataset of physiologically consistent computational fluid dynamics solutions through an automatic boundary-condition adjustment procedure. A common reference geometry allows all patient-specific anatomies to be represented by displacement fields, providing point-to-point correspondence without the need for geometric registration. Based on this dataset, we develop Graph U-Net and Transformer U-Net surrogate models to predict velocity and pressure fields at peak systole. We further investigate two strategies to enhance data efficiency and prediction accuracy: active learning and differentiable physical consistency constraints based on the finite element residual. The results show that active learning is strongly regime-dependent: while it offers little benefit over random sampling during early training, it improves performance during fine-tuning through gradient- and subspace-based acquisition strategies. Similarly, physical consistency constraints prove to be substantially more effective as a fine-tuning objective than as a regularizer applied throughout training. Finally, out-of-distribution studies demonstrate that the primary limitation of surrogate models lies in the geometric support of the underlying parametrization rather than in the neural network architectures themselves. The proposed framework combines synthetic dataset generation, geometry-aware surrogate modeling, physics-informed optimization, and active learning into a scalable workflow for inter-patient hemodynamic prediction.

Deep reinforcement learning (DRL) is challenging long-held assumptions in active flow control. By learning directly from interaction with a flow, DRL agents can find actuation strategies for nonlinear, unsteady flows without explicit models, often exploiting mechanisms that classical control would not anticipate. High-speed flows are a demanding test of this promise: shocks, shock/boundary-layer interactions, transition, aerothermal loading and microsecond-scale dynamics strain both algorithms and hardware. Yet most demonstrations remain idealized, low-speed simulations with limited evidence of robustness, interpretability, transferability or experimental deployment. In this perspective, we examine what constrains DRL-based high-speed flow control, including compressible simulation cost and fidelity, data efficiency, sensing in extreme environments, reward design, explainability, and generalization across Mach and Reynolds number. We argue that these limits reflect a tension between data-driven optimization and the structured physics, constraints and uncertainties of real flows. We then highlight where DRL may go beyond conventional approaches: adaptive control of shock-induced separation and turbulent boundary layers, coordination of distributed actuators, and discovery of unconventional flow-manipulation mechanisms.The central challenge, however, is translating these capabilities into reliable technology, which will require validated compressible simulation environments, physics- and constraint-aware learning, robustness to off-design conditions and sensor degradation, real-time operation, hardware-in-the-loop and wind-tunnel validation, and credible pathways to safety and certification.

We study the bipanchromatic number $χ_B$, introduced by Lalou et al. (Discrete Appl. Math. 391 (2026), 511-522) as the maximum number of colors in a panchromatic hypergraph coloring where every color appears at least twice. They show that merging uniquely occurring colors in a maximum panchromatic coloring gives the lower bound $β=χ_P-\lceilα/2\rceil \leq χ_B$, where $χ_P$ is the panchromatic number and $α$ is the minimum number of such colors among maximum panchromatic colorings. They conjectured that $χ_B=β$ and subsequently disproved it (arXiv:2610.08042). We identify the smallest counterexample, with six vertices and three hyperedges, both minimum possible. We show that the ratio bound $χ_B/β

With the advancement of powerful AI chips, on-device machine learning is gaining traction for its low latency and privacy protection. However, when edge devices are compromised, attackers can obtain white-box access to the deployed models by reversing the apps directly to obtain the knowledge of peer devices, which is known as Cross-Node Attacks, like performing efficient adversarial attacks. Most existing defenses target known vulnerabilities, yet white-box access still enables attackers to uncover more unknown weaknesses. This creates a critical risk for edge-device deployments, where one compromised device can endanger all models distributed from the same source. Motivated by findings that the model robustness is primarily influenced by the curvature of the decision boundary, we investigate how model parameters shape its curvature. In this paper, we propose a novel method, BMO, that obfuscates model parameters to diversify the decision boundary further resisting cross-node attack while preserving its benign accuracy. Specifically, we explore decomposing model parameters into boundary-position and boundary-curvature components based on null-space constraints, then models with differentiated decision boundaries are instantiated by injecting noise into the boundary-curvature component. Across all four major types of adversarial attacks with four widely-deployed on-device models, BMO achieves an average success rate of 68.68% in defense, far exceeding the most competitive baseline, which attains only 22.32% in defending against cross-node attacks. Moreover, BMO achieves comparable performance in benign accuracy compared with baselines.