Papers

Language-model agents are evolving into long-running services that interact with models, tools, computers, mobile devices, and distributed environments. Existing agent frameworks simplify reasoning and tool invocation. However, cloud-centric designs face three limitations: centralized execution increases failure impact, scaling pressure, and compute cost; extending agents across computers, mobile devices, and edge environments requires a unified execution abstraction with permission control; and long-running executions require consistent lifecycle management across failures, recovery, results, usage, and settlement. We present Nexus, a cloud-edge platform that treats each invocation as a persistent task. Nexus uses an OpenWrt-based runtime for distributed serving, run-scoped delegation for authorized access to Computer and Mobile environments, and persistent records to track execution, outputs, failures, recovery, usage, and charging across cloud and edge components. We evaluate Nexus on controlled, cross-device, and model-driven workloads. All ten Computer-Android workflows succeed, and all six revocation tests block subsequent writes while preserving prior authorized reads. Under worker loss, journaling eliminates duplicate appends (six to zero per task), adding 0.933 s mean normal-path overhead. Across 24 matched task pairs, Nexus completes 24 tasks versus Dify's 22 and is a median 3.88 s faster on jointly successful pairs. In a separate workload, Nexus operates under a smaller tested incremental-runtime memory ceiling than Dapr (16 versus 64 MiB), although Dapr achieves lower successful-call latency. These results demonstrate how locality, operation-scoped authority, and persistent result identity support cloud-edge agent services with workload-dependent costs.

We study the field concentration phenomenon between two closely spaced inclusions with imperfect bonding interfaces of low conductivity type. The inclusions are assumed to have finite conductivities. The problem is governed by a system of elliptic equations coupled with Robin-type boundary conditions. While it is known that finite-conductivity inclusions with ideal interfaces yield bounded gradients, in this paper we show that with imperfect bonding interfaces, the gradient of the solution may blow up as $\varepsilon$ (the distance between two inclusions) tends to zero when the bonding parameter $\gamma$ is large, while it is uniformly bounded independently of $\varepsilon$ when the bonding parameter $\gamma$ is small. Moreover, we identify the threshold of $\gamma$ and the optimal blow-up rates under certain symmetry assumptions. Compared to the case when the inclusions are perfect conductors, we find a novel logarithmic blow-up phenomenon at the critical value of $\gamma$.

In autonomous driving perception, visual object tracking systems must satisfy stringent latency and power constraints while remaining robust in complex and dynamic environments. Although transformer-based trackers achieve state-of-the-art accuracy, their substantial computational and memory overheads hinder deployment on real-time, resource-constrained platforms. To move toward this goal, we propose Difference Feature Map Knowledge Distillation (DFM-KD), a novel relational distillation framework tailored for transformer-based visual object tracking. Unlike conventional feature distillation methods that minimize point-wise discrepancies (e.g., mean squared error) between teacher and student feature representations, DFM-KD transfers knowledge through inter-sample feature differences, explicitly aligning the relational structure of the feature space. By distilling how the teacher models appearance variation and consistency across samples, rather than enforcing similarity in absolute activations, DFM-KD enables the student to better capture the structural dynamics of visual changes within a batch. As a result, the distilled model exhibits enhanced feature robustness and improved tracking performance. Extensive experiments demonstrate that DFM-KD consistently outperforms conventional feature-level distillation methods in both tracking precision and success rates.

We study $W$-free planar graphs for $W\in\{W_{2,4},W_{2,5},W_{3,4}\}$, where the quasi-double star $W_{h,k}$ is obtained from a three-vertex path by attaching $h$ leaves to one endpoint and $k$ leaves to the other. We prove that every $W_{2,4}$-free planar graph on $n$ vertices has at most $9n/4$ edges, and the bound is attained whenever $8\mid n$. This determines the planar Turán density of $W_{2,4}$ as $9/4$. We also establish the sharp upper bound $5n/2$ for $W_{2,5}$. Combined with known constructions of planar graphs of maximum degree five, it yields $\ex_{\PP}(n,W_{2,5})=\lfloor5n/2\rfloor$ for every $n\ge15$. These results close the two corresponding coefficient gaps in the bounds of Liu et~al. Our proofs use structural restrictions on high-degree vertices, local deletions, and degree deficits in neighborhoods of radius two. For $W_{3,4}$, we characterize the planar graphs with a dominating vertex that avoid this tree and determine their exact extremal number, $\lfloor(5n-7)/2\rfloor$, for every $n\ge10$. Finally, $W$-free planar triangulations have at most eight, twelve, and eleven vertices for $W=W_{2,4},W_{2,5},W_{3,4}$, respectively; the first two bounds are sharp.

We study difference-in-differences (DID) estimation of average treatment effects on the treated when the researcher is uncertain about which pre-treatment periods satisfy the parallel trends assumption. Roth (2022) shows that pre-trend testing can induce bias, complementing the statistical literature on post-selection inference. We propose the model averaged difference-in-differences (MADID) estimator, a weighted average of the candidate $2\times2$ DID estimators. The weights are normalized exponential functions of the candidate residual sums of squares, so implementation requires no specification test. We derive MADID's asymptotic properties under conditions that make invalid comparisons detectably more variable than at least one valid comparison. Under these conditions, MADID is consistent when the parallel trends assumption holds for at least one pre-treatment period. Although its joint limiting distribution across post-treatment periods is generally non-Gaussian, inference can be conducted by subsampling or simulation.

This paper study observability for the Dirichlet heat equation on a bounded interval with a real bounded potential depending on space and time. We prove an observability inequality from every measurable subset of spacetime with positive measure. The constant is uniform over potentials with a prescribed $L^\infty$ bound. By duality, positive spacetime measure is equivalent to null controllability by square-integrable distributed controls. Controls can also be chosen bounded in time with values in $L^2$. If almost every spatial slice of the observation set has a fixed positive lower bound on its measure, the observability constant is bounded above by $C_0e^{C_1/T}$, uniformly in the locations of the slices. The proof combines observation estimates on finite-dimensional spaces transported by the evolution with decay estimates for the distance to those spaces.

Constant curvature is one of the most fundamental intrinsic conditions in submanifold geometry. Investigating the existence and classification of biharmonic submanifolds under intrinsic curvature constraints is generally difficult, since the biharmonic equation is expressed in terms of extrinsic data. In this paper, we obtain a complete classification of proper, that is non-minimal, biharmonic Lagrangian surfaces with constant Gaussian curvature in 2-dimensional complex space forms. As a consequence, we classify the proper biharmonic Lagrangian surfaces with flat normal bundle and prove that pseudo-umbilical biharmonic Lagrangian surfaces are necessarily minimal.

A quantum-classical correspondence for non-Hermitian symmetry breaking has recently been established using coordinate-space path integrals, providing a semiclassical understanding of spectral transitions at the level of individual eigenstates. Here we develop its dual formulation in momentum space by constructing the corresponding trace formula and quantization condition. We show that the real or complex nature of individual eigenvalues is determined by the symmetry properties of the associated semiclassical orbits, as in the coordinate-space path-integral approach. Moreover, we demonstrate that the topology of semiclassical orbits determines the natural formulation of the quantization condition: the coordinate- and momentum-space formulations are equivalent for contractible periodic orbits in phase space, whereas for noncontractible orbits, the quantization condition along the winding direction remains valid, but its dual form must be corrected by a boundary term. In particular, real-space-winding and Brillouin-zone-winding orbits naturally select coordinate- and momentum-space quantization, respectively. As a nontrivial application, we investigate the boundary-induced spectral transition of Bloch oscillations in a finite non-Hermitian lattice, where Bloch-oscillation orbits winding across the Brillouin zone preserve the relevant symmetry and yield real energy levels, whereas boundary-reflected orbits form symmetry-related pairs and give rise to complex-conjugate eigenvalues. Our work completes the quantum-classical correspondence framework for non-Hermitian symmetry breaking, extending its applicability to a broader class of non-Hermitian problems.

Recursive self-improvement (RSI) enables agents to iteratively optimize their workflows via execution feedback. However, standard RSI typically operates as a first-order optimizer: it repeatedly patches surface-level parameters in response to immediate failure symptoms, often leading to trial-and-error thrashing without resolving underlying mechanisms. To address this limitation, we introduce SO-RSI, a framework that elevates workflow optimization to a second-order diagnostic inquiry, investigating why failures occur before committing to structural interventions. SO-RSI passively monitors execution traces for three structural anomalies (recurrence, opposing edits, and expectation mismatch) to trigger targeted mechanism investigations. By executing lightweight diagnostic probes and maintaining persistent inquiry memory across RSI rounds, SO-RSI accumulates causal evidence to guide systematic workflow edits rather than parameter patches. Across Lean 4 proof generation and Verus-based verifiable code generation, SO-RSI improves final held-out pass rates over Naive RSI by 21.8 and 25.8 percentage points under matched 24-hour search budgets. Behavioral analyses further confirm that SO-RSI substantially suppresses failure recurrence and eliminates unproductive zero-progress optimization loops.

Recurrent sequence models must decide how strongly to overwrite their memory at each token. Read as Bayesian filtering, this write is the gain of a Kalman update, set by uncertainty from two sources that pull it in opposite directions: volatility, how quickly the underlying associations change, and stochasticity, how noisy each observation of them is. First, we show that the update of gated delta-rule memories is the form this filter takes under isotropic uncertainty. Next, we introduce Voltic, a recurrent memory that keeps the covariance anisotropic and makes both noise variances input-dependent, so the write is vector-valued and carries uncertainty accumulated over the sequence. A dense covariance would have to be propagated token by token, ruling out the parallel training these models depend on. We therefore give two assumed-density approximations, diagonal and quasi-diagonal, both of which leave the memory update in delta-rule form and reuse its chunked kernels. On controlled recall tasks in which associations change and observations are corrupted, Voltic leads all baselines. On the task combining volatility and stochasticity, its margin over the strongest baseline is larger at both extrapolation sizes than at the training sizes. In 45M-parameter language models it leads an eight-task reasoning average and achieves higher retrieval accuracy beyond the training context length than gated baselines, at throughput close to those baselines. Deriving the write from an uncertainty recursion therefore makes memory more responsive to change.

The efficacy of traditional disease prediction is limited by spatial gaps and temporal lags, which impact the timing and targets of resource deployments. Outbreaks escalate undetected, chronic disease burdens are quantified years later, and at-risk populations in data-sparse regions remain unaddressed. Planetary geospatial foundation models complement existing epidemiological workflows to provide operational improvements, encoding multimodal search, mobility, and environmental signals into generalizable place representations. As illustrations of this complementarity, we present independent global health case studies of Google Earth AI's Population Dynamics Foundation Model (PDFM) – a foundation model for geospatial inference – across four domains (vaccine-preventable, communicable, noncommunicable, maternal mental health), five tasks (spatial extrapolation, interpolation/nowcasting, probabilistic forecasting, prospective forecasting, risk stratification), and four countries (USA, Canada, Mexico, and the Democratic Republic of the Congo). Across these case studies, PDFM addresses critical surveillance gaps across domains: improving US-Canada border MMR vaccination coverage predictions by capturing cross-border behavioral spillovers domestic models miss; nowcasting cardiovascular disease to accelerate data availability; enhancing short-term municipal Mexican dengue forecasts for timely outbreak vector control; improving forecasts of cholera hotspots; and adding a transferable signal to individual-level postpartum-depression risk prediction in US states the model had never seen, while not replacing individual socioeconomic data or closing demographic screening gaps. Together, these results showcase capabilities of geospatial foundation models for public health surveillance.

We study the large time asymptotics of the Feynman-Kac semigroups of the symmetric Lévy process on unbounded open sets. Our main result proves the exact exponential asymptotic decay rate for the survival probability given in terms of the bottom of the spectrum of the associated nonlocal Schrödinger operator. We also prove quantitative upper bounds with an explicit polynomial correction. The proof is probabilistic and is done by decomposing the Lévy process into a finite-range jump process collecting the small jumps and an independent compound Poisson process describing the large jumps. Our approach gives a direct link between the spectral properties of nonlocal Schrödinger operators and pointwise decay of survival probabilities, which was previously unknown for jump processes.

Generative artificial intelligence (AI) models can perform increasingly complex tasks, yet greater AI usage does not necessarily translate into proportional productivity gains. We identify token-max as one source of this inefficiency: when token consumption is treated as productive effort, agents are encouraged to over-exert and expend computation beyond what is necessary. We instead propose outcome-max, which rewards independently verified task completion per unit cost and induces a principled stopping rule. Then, to study these objectives, we develop a three-level simulation framework spanning immediate interaction, long-run behavioral adaptation, and organizational collaboration. Across all three levels, outcome-max improves the efficiency of AI-assisted production while largely preserving verified task performance. To further align these incentives with outcome-max, we introduce OutcomeShare, an incentive mechanism. Theory and simulation show that OutcomeShare can induce participation while generating shared gains for employees, firms, and LLM providers. Together, our results suggest that AI productivity not only depends on model capability, but also on how to construct the objectives governing AI use.

A $k$-substring edit in a sequence first deletes a substring of length at most $k$, and then inserts a sequence of length at most $k$ at the same position. A code that can correct a $k$-substring edit is called a $k$-substring edit code. In this paper, we develop a new localization method and use it to construct a $q$-ary $k$-substring edit correcting code with $\log n+8\log\log n+o(\log\log n)$ bits of redundancy for any fixed $q\ge2$ and $k\ge1$, where $n$ is the code length. For the binary alphabet, this improves upon the redundancy $\log n+16k\log\log n+o(\log\log n)$ obtained by Li et al. When the deleted substring and the inserted sequence have different lengths, we further construct codes with redundancy $\log n+O_{q,k}(1)$, which is optimal up to an additive constant. As corollaries, for all fixed $q\ge2$ and $1\le t\le T$, we obtain $q$-ary $(\le t)$-burst-deletion correcting codes and $(t,T)$-localized deletion correcting codes with redundancies $\log n+O_{q,t}(1)$ and $\log n+O_{q,T}(1)$, respectively. To the best of our knowledge, these are the first constructions attaining optimal redundancy up to an additive constant for these two deletion models over the full range of fixed parameters. For $(\le t)$-burst-deletion correction with $t\ge2$, such redundancy had previously been achieved only for $q=t=2$ by Levenshtein in 1967.

Virtual Model Control (VMC) is an approach to design a controller for force-controlled robots in complex uncertain environments. While this method was primarily investigated for legged robot locomotion in the past, it can be more generally applicable to other types of robotic systems. This paper investigates the VMC framework for reaching tasks in a force-controlled robotic arm. We propose six different approaches to designing virtual models in order to achieve reaching tasks in environments with obstacles and uncertainties. A force-controlled 8 degree-of-freedom humanoid robot was used to validate the proposed approach in the real world. We conducted three experiments to test the performance of VMC controllers in terms of predictability, sensitivity to external force, and adaptability against known and unknown obstacles. Experimental analyses show that, even though the proposed approach needs to sacrifice accuracy and trajectory optimality, it enables us to design complex reaching motions under uncertainties, in an intuitive and extendable manner.

As mobile communications cross the threshold from 5G to 6G, the 3GPP has entered a decisive phase. Following the Technical Specification Group (TSG) plenary meetings of June 2026 and the associated 6G workshop in Singapore, the Release-20 study phase is now well underway across all three TSGs: Radio Access Network (RAN), Service and System Aspects (SA), and Core Network and Terminals (CT). Concurrently, the timeline for the first normative 6G specifications in Release 21 has been established. Rather than offering another technology vision, or a chronological account of committee progress, this article asks what the first body of agreed 6G material tells us. We review service requirements, system architecture, the emerging radio interface, native AI, ISAC, non-terrestrial networks (NTN), and core-network protocols, tracing how each moves from study item toward specification. Read across the three TSGs, this material also supports seven cross-cutting lessons: AI has moved from a feature to a primitive spanning every stage, non-terrestrial access has crossed from overlay to substrate, and connectivity is now one objective among three alongside computing and sensing; the industry favors deployability over redesign, with spectrum access and site reuse, rather than waveform innovation, appearing to be the tighter constraint; and standardization targets AI workflows rather than AI models, with testability across vendors recurring as a decisive criterion for which techniques survive. Because the material spans very different stages of maturity, every item cited is tagged with its status at a single snapshot date (approved, draft, individually proposed, or vendor-demonstrated), and the lessons are offered as the authors' synthesis rather than as 3GPP positions. We close with the road toward trials and commercial launch around 2030, and with the open questions these lessons help prioritize.

Speaker tracking is the task of separating and following multiple speakers over time. It must address overlapped speech, speech onset and offset, talker identity, and time-varying speaker count. We propose a segment-online, modular framework for single- and multi-channel speaker tracking. The proposed system first performs speaker separation in each segment and computes speech activity via voice activity detection (VAD). To generate speaker tracks over time, we introduce a two-stage sequential organization strategy: Overlap-based stitching for continuous grouping and memory-based speaker verification for discontinuous grouping. For speaker separation, we employ complex spectral mapping to estimate the real and imaginary spectrograms of underlying speakers. The proposed system achieves state-of-the-art segment-online tracking performance on the LibriCSS and AMI datasets. Our framework significantly reduces diarization error rate (DER) and concatenated minimum-permutation word error rate (cpWER) compared to other methods.

Learning from trajectory demonstrations offers a route to active-inference control of complex systems whose dynamics are difficult to model explicitly. We introduce generative active-inference control (GenAIF), in which one generative trajectory model learns from demonstrations and measured action interventions to supply a goal-conditioned policy distribution and a state-to-observation likelihood mapping. From this control design, we derive three model requirements: (i) useful action proposals, (ii) accurate prediction under imposed actions, and (iii) probabilistic observation evidence for belief updating and expected information gain. We benchmark diffusion, autoregressive Transformers, conditional variational autoencoders (CVAEs), and flow matching in a MuJoCo manipulation task with multiple physical conditions. Diffusion delivers the strongest control across the tested dynamics, while CVAE combines comparable short-horizon prediction with much faster inference. Correct conditioning is decisive, and trajectory reuse offers further computational savings. In replay after an unannounced tilt change, pretrained diffusion updates belief fastest among the original models; fine-tuning on recovery demonstrations further accelerates identification and sustains accurate tracking. These findings support the use of shared generative trajectory models to connect action proposal, controlled prediction, and observation evidence within GenAIF.

Latent audio generative models are typically trained in two stages: an audio codec is learned first, followed by a latent generative model. This decomposition leads to a decoder train-generation mismatch: the codec decoder is trained on encoder-induced latents but deployed on generator-produced latents at inference time. Across diverse datasets and latent generative models, we observe clear reconstruction-generation gaps under both FD and FAD, showing that strong reconstruction quality does not necessarily translate into strong end-to-end generation quality. A natural remedy is to adapt the decoder on generation-produced latents, but generated latents lack correspondence with source audio and therefore cannot directly provide the paired supervision used for decoder fine-tuning. We introduce AudioGAR, which constructs intermediate latents by perturbing encoder latents and denoising them through the frozen latent diffusion model. These latents form a trajectory from reconstruction toward generation, with lower-noise latents retaining source correspondence and supporting paired decoder fine-tuning. We fine-tune only the codec decoder on these latents, while keeping the codec encoder and latent generative model frozen. When applied to AudioX, AudioGAR substantially improves generative performance. It requires only 1.5% of the original training audio hours and 0.26% of the original training cost.

Large language models (LLMs) are increasingly used for literature search and synthesis. However, it is unclear whether they retrieve accurate bibliographic information in environmental science. Therefore, we quantitatively compared the errors of widely used LLM platforms in retrieving references related to original articles from five leading environmental science journals (Energy and Environmental Science, Nature Sustainability, Nature Climate Change, Lancet Planetary Health, and Environmental Science and Technology) published in 2024 to 2025. Claude, ChatGPT, Grok, DeepSeek, Perplexity, and Gemini were used as the LLM platforms. LLMs retrieved 10 references for each of the 50 randomly selected original article using either the article's abstract or its full-text as prompt. The retrieved references were subject to a multimetric score ratio combining validity of bibliographic data, Google Scholar link, digital object identifier, Scopus Electronic Identifier and relevance score (cited by or being the index paper), and the proportion of complete fabrication that failed all metrics. Abstract-only prompt yielded significantly higher accuracy than full-text one. This advantage was confirmed in multilevel mixed-effect multivariable regression after adjusting for journal, platform, and output order. Source journal and the position of a reference within the output list were also independently associated with retrieval accuracy, with lower-listed references associated with lower accuracy. These findings suggest that LLM assisted literature retrieval in environmental science remains moderately accurate and overall inconsistent, varying significantly by platform, journal, prompt type, and output position. Abstract-based prompting, as task-aligned information compression, may outperform full-text one in literature retrieval. Caution should be used when generalizing our findings.

Neural-network (NN) controllers are increasingly used in nonlinear control systems, but their highly nonlinear behavior makes them difficult to explain and verify, raising trustworthiness concerns in safety- and mission-critical applications. A key step toward certifiable trustworthiness is to find a Forward-Invariant Set (FIS): a state-space region such that any trajectory starting inside remains inside. If the FIS excludes unsafe states, safety can be guaranteed for initial states within it. Finding an analytically proven FIS for a given AI-controlled system with a fixed controller is difficult. We propose a framework that uses an Invertible Neural Network (INN) to transform the original state space into a latent space where a regular-shaped FIS is more likely to exist. We train the INN so that a preferred hyper-rectangular candidate becomes invariant in the latent space, then formally verify it. We prove that, whenever verification succeeds, both the latent-space candidate and its inverse-transformed counterpart in the original state space are analytically proven FISs. We evaluate the approach on 45 AI-controlled systems across three representative control testbeds. Our method finds certified FISs for all 45 systems, whereas an adapted state-of-the-art baseline finds none. It is also faster on 40 of the 45 systems, and the centers of the resulting FISs roughly match domain-expert preferences.

We study cooperative multiplayer bandits with adversarial rewards and no collision information. In each of $T$ rounds, each of $m$ labeled players picks one of $K>m$ arms. A player alone on its arm receives that arm's reward in $[0,1]$; players who share an arm receive $0$. Each player observes only its own reward, so it cannot distinguish a collision from an arm with zero reward, and there is neither shared randomness nor a communication channel. Bubeck, Li, Peres and Sellke (2020) obtained regret $\widetilde O(KT^{3/4})$ for two players and $\widetilde O(mK^{3/2}T^{1-1/(2m)})$ for $m$ players, and conjectured that regret $\Omega(T^{2/3})$ is unavoidable already for two players. We refute this conjecture. We give a strategy whose expected regret against every oblivious adversary is $O(m^{3/2}\sqrt{KT}\log^2(KmT))$, which is optimal in $K$ and $T$ up to logarithmic factors. Three ideas make this possible. (i) Bits are sent through deliberate collisions as random parity checks, and the regret spent while a message is delayed is charged to the number of arms with positive reward rather than to elapsed time: an adversary that jams the channel with zero rewards forfeits the regret it could inflict. (ii) An exponential-weights learner over sets of arms combines a frequently resampled coordinator arm with a follower assignment that changes only $O(\sqrt{mT/K})$ times in expectation; delayed updates are reconstructed exactly. (iii) We teach the followers a privately chosen row index in a fixed public table using reward observations alone, allowing them to coordinate without shared randomness. The bound holds simultaneously for all reward matrices, even ones chosen with knowledge of the table, and a random table works with probability $1-O(1/(KmT))$.

We study state-constrained stochastic control problems with a scalar surplus $R \geq 0$. Controls for which the drift and volatility of $R$ vanish at zero define a lower-dimensional Hamilton-Jacobi-Bellman equation. Starting from control-wise state-constraint viscosity inequalities, we prove that the upper and lower limits of the interior value are a subsolution and a supersolution of this boundary equation. Terminal compatibility and comparison identify their common limit. We require one-sided bounds on positive surplus drift and generator growth, together with domination by boundary generators. For compact or coercive controls, these follow from a local lower bound and continuity of the boundary control set. For a smooth stochastic target value $w$, the transformation $R = Y - w(t,X)$ flattens the viable epigraph, and, under the stated hypotheses, the boundary control value supplies the Dirichlet datum. Applications include set-valued boundary controls, unbounded drift with quadratic costs, and redundant hedging instruments with a nonconstant target boundary.

In recent years, time-series question answering (QA) systems have made significant progress. However, generating a correct answer does not show whether retaining the supplied numerical series improves task performance, nor whether the prediction is sensitive to changes in that input. While some systems provide rationales, answer accuracy also does not show whether their numerical claims are grounded in the supplied series or whether the stated inference is valid. In this work, we focus on evaluating four time-series QA systems: TimeOmni-1, ChatTS, TimeOmni-VL, and Time-MQA. First, for three systems with released evaluation data, we reproduce their reported results and compare the performance of the systems with their backbones. Then, we introduce a benchmark named COMMON-TSQA, which collects public evaluation datasets from existing time-series benchmarks and unifies their sample representation, task definitions, and answer schemas, while evaluating each system through its own interface under common evaluation criteria. The evaluation uses the original condition and six interventions while keeping the question and target fixed. Our analysis shows that aggregate performance alone can obscure how systems use numerical evidence. Similar task-level scores can arise despite substantial changes in individual predictions. Some interventions induce simple fallback behavior rather than preserved task ability. We also evaluate rationales for factual grounding, inference validity, and consistency with the final answer. We find that rationales often contain time-series claims unsupported by the input. Moreover, the rationale audit shows that agreement between a rationale and its final answer can coexist with incorrect numerical descriptions or invalid intermediate inferences.

Reasoning models write most of their KV cache while decoding long chains of thought (CoT), so the cache has to be compressed online under a fixed memory budget. Decode-time methods mostly decide which tokens to evict. We ask how a fixed byte budget should be split between the number of cached tokens and their precision. BreadthKV spends the bytes on more tokens at low precision, combining quantization with eviction, and picks the bit-width for each model and budget with a 60-problem end-to-end calibration, since offline attention error does not predict it reliably. On three reasoning models and four math and science benchmarks, it scores above eviction alone in 17 of 18 settings and produces shorter outputs. Much of what eviction loses comes from derailed runs, which keep reasoning until the length cap without reaching an answer. On Qwen3-8B at our tightest budget, eviction sends 91% of AIME samples to the cap and BreadthKV 40%. Under the same protocol, BreadthKV is statistically indistinguishable from a joint rate-distortion allocator (RDKV) that uses 27% more KV memory-time, and it outperforms our re-implementation of ThinKV.

Modern deep generative models are primarily studied for their ability to generate realistic samples, yet the generative dynamics they learn can also serve as objects of statistical inference. We develop this idea for two-sample testing, the problem of deciding whether the same distribution generated two finite datasets. Using stochastic interpolants, we connect both distributions to a shared Gaussian bottleneck, so that each half of the resulting path is a Gaussian channel acting on a single population. We prove that the null hypothesis holds if and only if the population denoiser, or equivalently, the velocity fields of the two halves, coincide at any single noise level, which amounts to a reflection symmetry of the path about the bottleneck. Deviations from this symmetry yield a continuum of two-sample witnesses, which we estimate via held-out regression risks on learned denoisers and velocities and aggregate along the path; under an information-theoretic weighting, the aggregated discrepancy equals the Jeffreys divergence between the noise-smoothed distributions. Calibrating the resulting statistics by permutation yields tests that are valid in finite samples for any trained networks and consistent when the fields are learned accurately. On a synthetic benchmark and three image benchmarks, the proposed tests improve power over the strongest baseline by up to 33 percentage points at an equal total sample budget, with the best choice of regression representation and path weighting depending on the data modality. These results show that generative paths provide a principled representation for statistical testing, extending stochastic-interpolant models beyond generation.

AI systems can now write and optimize production GPU kernels, but validating them remains an important challenge. Evaluating the kernel on a few random inputs and checking that its outputs match a trusted reference kernel within numeric tolerances is not sufficient: races can cause nondeterministic behavior that fails to manifest in tests, and numeric tolerances can hide bugs and cause false positives even after extensive calibration. To address this challenge, we present RESOLVE, which combines testing and formal verification to build a comprehensive kernel validation pipeline. It operates in three steps: First, it tests for nondeterminism using binary instrumentation that perturbs execution timing to expose races. Second, an agent rewrites the candidate and reference kernels to obtain "reduced-concurrency" versions that are simpler to analyze but still produce bitwise-identical outputs in all tests. Third, the reduced kernels are formally analyzed in the F*/Pulse framework and prove that they perform the same computation on real numbers. This sidesteps the need for numeric tolerances. We show that RESOLVE can validate a broad selection of kernels using KernelBench, and prove equivalence across fused GEMMs in three state-of-the-art frameworks and languages: CUTLASS, Triton, and Gluon. It also analyzes mega-kernels, notoriously difficult to validate, and finds four previously unreported issues, including two clear bugs. We show that agents can use RESOLVE to repair the issues, with minimal performance impact, highlighting that agents can optimize aggressively when they can rigorously check their results.

Knowing domain rules does not guarantee applying them to a case. We study this distinction in traditional Chinese Bazi through 3,000 Chinese multiple-choice questions spanning 14 Theory and 11 Case categories. Six endpoint systems are evaluated, with primary results reported on a 2,492-item model-informed refinement. Theory accuracy exceeds Case accuracy for every system, and gaps of 16.60-29.56 percentage points remain when invalid responses are excluded. The contrast is more specific than a general case-reasoning deficit. Across six systems, Twelve Stages and Nayin reach mean accuracies of 89.10% and 88.62%, while Shensha Basics reaches 75.96%. Within Case, Luck Pillars averages 84.62%, but Career and Family Relations average only 36.98% and 38.19%. Overall rankings also conceal different category strengths. On the original 3,000 items, paired DeepSeek native/disabled comparisons associate native configurations with Theory gains of 6.53 and 12.20 points for Flash and Pro, respectively; Case changes are -3.67 and +1.27 points. These are provider-configuration associations, not isolated causal effects of reasoning. The results motivate task-specific evaluation of cultural-domain applications rather than reliance on aggregate knowledge scores. The benchmark measures agreement with a model-generated, model-verified answer key, not real-world predictive validity. Final-set results are post-selection descriptions, and incomplete provenance and expert validation constrain their interpretation.

Automatic speech recognition (ASR) for low-resource languages remains a major challenge. Sinhala, the primary language of Sri Lanka with about 16 million speakers, illustrates the difficulty: agglutinative morphology, a 54-phoneme inventory, subject-object-verb (SOV) syntax and scarce annotated speech data limit both conventional and modern ASR systems. This paper presents the first critical review of Sinhala ASR research, tracing its development from Hidden Markov Models (HMMs) through deep neural networks to self-supervised pre-trained models such as wav2vec 2.0, XLS-R, Whisper and Massively Multilingual Speech (MMS). We compare existing Sinhala systems with related low-resource ASR work on Tamil, Malayalam and Hindi in terms of architecture, training data, word error rate (WER) and robustness to real-world acoustic conditions, and we assess self-supervised and transfer learning as responses to scarce labeled data. We show that most reported WERs are not directly comparable because they differ in corpus, data split and scoring, and that the only controlled comparison in the literature attributes an 18.1% relative WER reduction to corpus correction alone. We also discuss context-aware ASR that draws on phonological, syntactic and semantic knowledge. We identify six research gaps: (1) the lack of large annotated corpora covering multiple dialects and acoustic conditions; (2) weak contextual modeling of Sinhala morphosyntax; (3) high WER in real-world conditions; (4) the absence of standardized benchmarks; (5) the lack of parameter-efficient fine-tuning studies; and (6) the absence of annotated code-switched Sinhala-English speech resources. We outline a research agenda to address these gaps, intended as a roadmap for researchers working on Sinhala and other morphologically rich languages.

Instrumental variable (IV) analysis in underrepresented target populations suffers from low statistical efficiency and weak IV bias due to limited sample size. Leveraging external knowledge from a source population with larger samples offers an appealing solution to this problem. We propose in this paper the KNowledge-powered IV Estimator for underrepresented Subpopulations (KNIVES), a framework that enhances the IV analysis for underrepresented subpopulations by transferring knowledge from external source population data. KNIVES constructs a knowledge-transferred IV score by combining multiple candidate models through an effective signal-to-noise ratio criterion tailored to the downstream IV estimator, rather than prediction accuracy alone. To address weak associations between the IV score and the exposure, as well as a high correlation among candidate models in the target population, KNIVES employs a bias correction technique to mitigate the resulting bias and ensure robust performance. Theoretical investigation demonstrates that our method is robust to weak IV bias and achieves improved efficiency compared to the causal effect estimator based on any single candidate model. In addition, KNIVES avoids negative transfer by guaranteeing that its asymptotic variance is not larger than that of the estimator relying solely on target population data. Extensive numerical studies demonstrate the superior finite-sample performance of our method over existing methods. Two Mendelian randomization studies on ethnic minority subgroups in UK Biobank data further illustrate the advantages of our method.