Papers

Muon controls weight updates through the RMS movement of a linear layer's output. This implicitly treats the linear output as a proxy for the full module, including the nonlinearity that follows. This proxy can be misleading for softmax, where the RMS shift in activations does not faithfully measure the change in the output distribution. We propose Hilbert distance as a natural metric for bounding changes in these output distributions. Hilbert distance bounds changes in cross-entropy loss and limits probability redistribution. For softmax, it takes the form of the variation seminorm of the preactivation perturbation, allowing probability constraints to be expressed as parameter constraints. Formulating the steepest descent problem with a Hilbert distance constraint brings two recently proposed methods under a unified mathematical framework. For the LM head, solving a row-wise approximation to the resulting diameter-constrained problem leads us to RowNorm \citep{pethick2025trainingdeeplearningmodels}. For scaled dot-product attention, the same principle yields the first-order query–key steepest descent problem underlying Compositional Muon \citep{keigwin2026compositional}. These connections motivate using Hilbert distance as a guiding principle for optimizer design in softmax modules.

Large language models are attractive text compressors because next-token probabilities can be converted into code lengths. In practice, however, the main bottleneck is often not modeling quality but systems cost: a compressor must obtain logits quickly enough to compete with strong classical baselines. This paper examines systems optimizations for LLM-based compression, including multistream pipelined arithmetic coding and target-interval capture, and evaluates their trade-offs across inference engines, configurations and datasets. We evaluate LLM-based text compression on three different inference engines: Transformers, vLLM, and TensorRT-LLM. Notably, we obtain a speedup of 18.84x compared to prior work's Transformers implementation and a maximum recorded inference speed of 181 kilobytes per second at a compression ratio of 7.17, when running on an NVIDIA A100.

Operational mobile networks are rich in observations but poor in interventions: they continuously generate large volumes of passive telemetry, while controlled action–response trajectories remain comparatively scarce and costly to collect. This observation–intervention asymmetry is a central obstacle to building practical world models for RAN control. We address this problem with RAN-JEPA, an action-conditioned world model that separates representation learning from intervention learning. The model first learns a predictive latent representation from observation-only network data, adapts that representation to the target deployment domain, and then learns action-conditioned latent dynamics from controlled sleep/wake trajectories while keeping the encoder fixed. We evaluate RAN-JEPA using public operator performance data for pretraining and a real-equipment four-cell 5G testbed for domain adaptation, post-training, and control. Across four future 15-minute intervals, RAN-JEPA achieves lower action-conditioned prediction error than a GPT-2 world-model baseline and a mechanistic rule, while remaining close to a strong persistence reference. What-if evaluation further shows that the model produces distinct and coherent future trajectories under matched SLEEP and KEEP plans. In closed-loop control, RAN-JEPA consistently supports higher sleep ratios than the two baselines at comparable congestion cost. These results show that observation-first, action-conditioned learning can turn abundant passive telemetry and limited intervention data into a practical predictive model for energy-aware RAN control.

Entities in real-world systems often evolve over time: users shift between information consumption patterns, researchers migrate between fields, and firms transition between financial risk states. When such systems are modeled as dynamic graphs, these transitions correspond to nodes whose class labels change between consecutive snapshots, which we call shifters. Correctly classifying shifters is often more consequential than classifying stable nodes with persistent labels, as detecting such transitions enables timely intervention in high-stakes settings. Yet, we identify a systematic failure mode across all major dynamic graph neural network architectures: shifter performance consistently lags behind stable nodes, with errors concentrated on the old label. Our structural analysis traces this failure to embedding inertia: neighborhood aggregation keeps shifter representations anchored to the old-class centroid, and this effect is strongest precisely where it is hardest to correct, namely for shifters whose neighborhoods remain aligned with the old class. Guided by this finding, we propose CHASE, a model-agnostic CHange-Aware framework for Shifting nodEs that wraps any dynamic GNN with targeted components for detecting label shifts and overriding stale neighborhood signals. CHASE consistently improves shifter performance across all tested models and datasets (up to 114.9% and 202.5% improvement on accuracy and F1 score, respectively) while preserving stable-node performance. We additionally contribute three dynamic graph benchmarks with naturally shifting labels, filling a gap in existing resources. The source code is available at https://github.com/jhkim611/CHASE

Rubinstein et al. introduce the study of connectivity and minimum cut problems in the cut query model, where an algorithm can only access the edges of a graph through CUT queries. In this work, we examine the query complexity of densest subgraph problems. Our first contribution is to generalize the uniform sparsification approach for Densest Subgraph and other heavy subgraph problems to weighted graphs, achieving the following results. >A $(1+\varepsilon)$-approximation for weighted Densest Subgraph in $O(n\log^2 n/\varepsilon^2)$ queries, roughly matching the $\Omega(n/\log n)$ lower bound we prove for randomized algorithms. >A $(1+\varepsilon)$-approximation for weighted Max Cut in $O(n\log^2 n/\varepsilon^2)$ queries. This improves upon the bound of $O(n\log^3 n(\log n+1/\varepsilon^2))$ queries achieved by Plevrakis et al., making significant progress towards their lower bound of $\Omega(n/\log n)$ queries. In contrast, we rule out approximation schemes with truly subquadratic query complexity for Densest $k$-Subgraph and settle its exact query complexity in both the unweighted and deterministic settings. For hypergraphs, we show CUT queries alone cannot achieve $o(\sqrt{n})$-approximation for Densest Subhypergraph or $o(n)$-approximation for Densest $k$-Subhypergraph, regardless of the number of queries. Augmenting CUT queries with the EDGE oracle of Chen et al., we overcome this barrier for Densest Subhypergraph and obtain a $(1+\varepsilon)$-approximation on rank-$r$ hypergraphs using $O(nr\log n/\varepsilon^2)$ queries. We also extend our Densest $k$-Subgraph hardness results to hypergraphs and prove exponential query lower bounds for exact Densest $k$-Subhypergraph. Lastly, confirming a conjecture of Chakrabarty and Liao, we show that simulating an ADD query using CUT and IS-EDGE$^?$ queries requires exponentially-many queries.

Inference-time alignment steers diffusion models toward downstream objectives without costly fine-tuning, but existing methods typically optimize expected cost (or reward) without explicitly controlling tail risk. In this work, we propose a risk-averse framework for inference-time alignment of diffusion models that balances expected cost and Conditional Value-at-Risk (CVaR) under trajectory-level Kullback-Leibler regularization. We characterize the optimal path distribution and reverse transitions as functions of the CVaR threshold, reducing the alignment problem to a one-dimensional optimization problem over the threshold. We show that this one-dimensional optimization problem can be nonconvex in general, establish sufficient conditions for convexity, and introduce quadratic regularization to enable efficient derivative-free ternary search. Under suitable assumptions, we provide finite-sample guarantees for threshold optimization and value estimation, and combine them into a terminal total-variation bound relative to the regularized target under exact resampling. Experiments on synthetic data and image generation using Stable Diffusion demonstrate improved tail performance over risk-neutral baselines while maintaining competitive average cost.

Bug localization models require substantial within-project training data that newly active or small open-source projects rarely have. Cross-project learning (CPL) promises to close this gap by reusing labeled bugs from mature projects, but prior CPL studies evaluate a single model against ample-data within-project baselines that data-scarce projects cannot reach. We ask whether CPL's effectiveness is a property of the transfer setting or depends on model architecture. We evaluate three architecturally distinct models, embedding-based BLAZE, convolutional TRANP-CNN, and GNN-based adversarial COOBA, on 63 source to target pairs from 13 open-source Python projects under identical bug corpora, data splits, retrieval candidates, and baselines. The primary baseline is within-project limited training (WP-small, 20% of target labels, the same target-side share as CP-transfer). CPL outperforms WP-small most clearly for TRANP-CNN (Cohen's d = 0.95, win rate 81.0%, p < 0.001, robust to clustering pairs by target project), and less so for BLAZE (d = 0.49, win rate 60.3%, p = 0.002; not significant once pairs are clustered by target). COOBA, under our re-implementation, shows a small gain that does not survive multiple-comparison correction (d = 0.27, p = 0.027). Negative transfer affects 9.5% of TRANP-CNN pairs and 30.2% of BLAZE's and COOBA's. Cross-model direction agreement is only 44-54%: the same pair can help one architecture and harm another, although the models' gains are positively correlated (rho = 0.24-0.42). CPL outcomes are therefore architecture-dependent: practitioners should validate CPL for their chosen model, and CPL studies should report results per architecture.

Answering questions from radiology reports involves two decisions: whether the report contains an answer and which span gives it. A false answer presents an unsupported claim as fact; an over-abstention hides what the report says. Yet RadQA results are reported as a single F1 that merges both. We propose an evaluation method that separates them, reporting false-answer and over-abstention rates beside F1 and splitting F1 changes into extraction and answerability terms. We apply it to 13 LLMs on the RadQA test set, zero-shot and with three cumulative demonstration sets, the last adding an unanswerable example. Models with similar F1 differed more than twofold in false-answer rate. The unanswerable example changed when models answered, not how well they extracted: 12 of 13 models abstained more often, and three-quarters of the answers they withdrew had F1 $\geq$ 0.5. Within models, false answers showed no clear association with extraction quality, but acting on the decision itself helped: keeping an answer only when enough other models also answered lowered the zero-shot false-answer rate by 4 points without a detectable F1 change. Because the two decisions can change independently, we suggest radiology QA evaluations report them separately, not through F1 alone.

Tracking object motion within a gripper provides essential feedback for robotic manipulation. However, limited contact coverage and uneven deformation make tactile estimation challenging. In this study, we present TacSpin, a method that tracks in-gripper object rotation on SO(3) using only two vision-based tactile sensors. The key idea is to estimate the axis and magnitude of the rotation response separately, rather than fitting a rigid-body model by least squares. Each finger's response is further refined with contact-boundary cues. The two fingers are then fused according to their agreement. On standard objects, TacSpin achieves lower mean checkpoint errors. Qualitative demonstrations on real-world objects, using calibration transferred from a standard sphere, illustrate tracking under different contact geometries.

Self-supervised pretraining has expanded the pool of 3D medical foundation models, but published evaluations use different datasets, endpoints and fine-tuning protocols. Models also differ in architecture, pretraining objective and data, while downstream targets range from voxel-level masks to patient-level diagnoses and outcomes. As a result, published findings provide little directly comparable evidence for choosing a model for a new task. Direct comparison of adaptation performance requires fine-tuning candidates on labelled target data, at substantial computational cost. We benchmark 23 candidates across the same 14 segmentation, classification and regression tasks spanning multiple anatomies and modalities. No candidate ranked among the top three across all segmentation tasks or across all classification and regression tasks; a convolutional encoder could outperform a transformer trained with the same objective on one task but underperform it on another. We benchmark transferability estimators against the fine-tuned rankings and find that their agreement varies across dense segmentation, classification and regression. InsightRank uses task-appropriate frozen-feature statistics and architecture-aware pairwise comparisons to rank candidates before fine-tuning. Across the 23-model pool, mean Kendall tau-b was 0.431 for segmentation and 0.352 for classification and regression. The segmentation Top 3 contained a task winner on four of nine tasks; the image-level Top 3 contained none on the five tasks, despite better full-ranking agreement than the matched conventional estimators. A three-model shortlist reduced recorded extraction-plus-training job time by 85.8% across seven segmentation tasks. Together, the benchmark provides a common reference for model selection and InsightRank offers a lower-cost way to prioritize fine-tuning, with shortlist performance that differs by target type.

While expert demonstrations on real robots are difficult to collect, human ego-centric videos are easy to collect and abundant on the internet. However, learning dexterous manipulation policies from human videos remains challenging—one of the most important challenges being the difficulty of reconstructing physically grounded robot trajectories due to the large embodiment gap. In this paper, we propose EgoCPT, a novel paradigm for learning high-quality retargeted dexterous hand trajectories directly from an ego-centric video, which can then be deployed directly on real robots. The only inputs to our method are an ego-centric RGB video and a reference hand pose trajectory, which can be easily obtained through off-the-shelf software or specialized cameras during recording. Our method is particularly robust to noisy hand poses attributed to our novel contact-aware optimization framework, which iteratively optimizes the contact points between the robot hands and objects. Across five manipulation tasks, our method achieves overall success rates of 67.3% on noisy EgoDex hand-pose reconstructions and 92.0% on high-quality TACO reconstructions collected using motion-capture hardware, outperforming multiple baselines on both benchmarks.

Ferrers diagram rank-metric codes are essential components of the multilevel construction of constant dimension codes for random network coding. This paper aims to construct optimal Ferrers diagram rank-metric codes, focusing on the diagram \([3,3,3,5]\) with rank distance \(3\). We construct an optimal \([[3,3,3,5],6,3]_q\) Ferrers diagram rank-metric code for every prime power \(q\), thereby answering an open problem posed by Etzion et al. in 2016~\cite{Etzion2016}. The construction first builds a \([3\times4,4,3]_q\) MRD code and then proves that the Delsarte dual of its puncturing on the last column contains a \([3\times3,3,3]_q\) MRD code, which comes from semifield codes. We further construct optimal Ferrers diagram rank-metric codes for two infinite families of Ferrers diagrams: the first relies on the \([[3,3,3,5],6,3]_q\) construction and holds for every prime power \(q\), while the second is binary and obtained directly by the puncturing-inclusion construction.

We construct a structure-preserving effective linear approximation of the quantum Boltzmann collision operator for Bose and Fermi gases in the normal phase. For variable hard-sphere kernels, the resulting Fourier spectral evaluation costs $O(N_v^4\log N_v)$, with $N_v$ velocity nodes per direction and no angular quadrature factor. The continuous operator conserves mass, momentum and energy and satisfies a linearized $H$-theorem. We also use the operator at local quantum equilibrium as an approximate collision Jacobian in a modified Newton solver that retains the full nonlinear residual. Numerical tests show smaller errors than the quantum BGK model in homogeneous relaxation and fewer full collision evaluations for steady Couette and Fourier flows than the same Newton solver using a quantum BGK Jacobian.

Artificial dissipation improves the robustness of high-order methods for under-resolved flows, but can compromise accuracy and suppress physically relevant structures if its strength and spatial distribution are not appropriately controlled. We present a sensor-activated nonlinear artificial-dissipation framework, inspired by the residual correction method, for continuous and discontinuous summation-by-parts discretizations. In contrast to many existing spectral-element dissipation operators that either reduce the formal order of accuracy or target only the highest polynomial modes, the proposed framework permits design- and higher-order dissipation with explicit control over its magnitude, spatial distribution, activation, and convergence rate under mesh refinement. This modular construction also recovers several existing methods through appropriate choices of its components. The dissipation preserves discrete conservation and semidiscrete entropy stability whenever the base method satisfies these properties. It is dimensionally consistent, element local, and discretization agnostic, as the added term is constructed independently of the base spatial residual. For smooth solutions under suitable mesh assumptions, we establish parameter choices that retain the formal accuracy of the base discretization. Numerical experiments with summation-by-parts discretizations of the Burgers and Euler equations demonstrate substantially improved robustness for under-resolved flows and flows with shock waves while retaining high-order accuracy for smooth solutions, using the same parameter choices across several test problems. Overall, these results show that targeted dissipation can improve the robustness of high-order methods without excessive damping of resolved flow structures.

Multi-robot exploration under communication and mobility constraints requires balancing information gain with reachability. We present a distributed semantic exploration framework that combines Dirichlet mapping, budget-constrained trajectory selection, and Resource-Aware distributed Greedy (RAG) coordination. Each robot clusters frontiers, computes terrain- and heading-aware least-cost paths with a single budget-limited Dijkstra search, and selects feasible paths by marginal information gain, represented by the expected reduction in Dirichlet entropy. Each robot's per-round selection is exact, the team's joint per-round information gain is at least half the centralized optimum, and the bound degrades gracefully under limited communication. Ten-run gridworld ablations show improved coverage, semantic mIoU, and map certainty with lower entropy and traversal cost per unit distance than shortest-distance frontier selection. Budget-sensitivity experiments show that an intermediate budget best balances path reach and receding-horizon execution. In photorealistic full-system simulation against five coordination rules run in their native published forms, the proposed method attains the highest mean coverage and semantic accuracy while being the only one that respects the per-trajectory mobility budget.

Complex and long-horizon agent tasks often rely on frontier models deployed in the cloud to achieve strong performance. However, using these models for every step of a long-horizon trajectory incurs substantial inference cost, even though not all steps may require the same level of model capability. In this paper, we find that allocating the LLM to only a small subset of key steps can achieve performance comparable to full LLM execution while significantly reducing inference cost. Motivated by this observation, we propose EdgeWise, an edge-cloud collaboration framework that deploys an SLM on the local device as the primary solver and selectively delegates critical steps to a cloud LLM when stronger model capacity is needed. To enable autonomous delegation without additional models at inference time, we introduce a step-level RL method that jointly optimizes task solving and delegation, with fine-grained credit assignment for individual delegation decisions. We also introduce an external judge-based delegation strategy and use it to systematically investigate key design choices in edge-cloud collaboration, including when to delegate, how delegation decisions should be made, what context should be shared, and how frequently the cloud LLM should be invoked. Experiments show that EdgeWise matches or outperforms pure-LLM execution while reducing inference cost up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1. We further deploy EdgeWise on a local workstation, demonstrating comparable cost savings in a practical edge-cloud setting.

We propose ATOM (Abstract Transition and Option Modeling), a hierarchical navigation framework that combines higher-level graph planning with reusable lower-level option policies for efficient repeated navigation in large-scale 3D environments represented by hierarchical structures such as 3D scene graphs (3DSGs). ATOM operates on a hierarchical graph representation, implemented using, for instance, the Rooms and Places layers of a 3DSG, where each directed abstract transition is associated with a navigation option. For each transition, ATOM computes and stores a reusable option policy along with its state-dependent cost-to-go. Across repeated navigation tasks, ATOM maintains a smoothed empirical distribution over option initiation states and uses it to progressively refine the corresponding abstract transition cost. Ablation studies show that ATOM reduces cumulative repeated-planning time by 97.5% relative to a monolithic MDP baseline while maintaining trajectories within approximately 7% of a Global Places A* shortest-path reference. Experience-informed refinement further reduces abstract transition-cost estimation error by 10.3%. Physical experiments validate ATOM on an F1Tenth vehicle, successfully completing all five navigation queries with 43% of inter-region option executions reusing previously generated solutions.

An efficient shallow Ritz method was developed in [8, 9] for one-dimensional linear diffusion and diffusion-reaction problems. In this paper, the method is extended to a nonlinear diffusion-reaction equation with a cubic reaction term: (1) a basic error estimate of the shallow Ritz discretization is established and (2) a reduced block Newton (rBN) method for solving the resulting non-convex optimization problem is developed. Again, the computational cost of the rBN method is $\mathcal{O}(n)$ for one-dimensional problems, where $n$ is the number of neurons. Combining a convex splitting scheme for time discretization with the efficient shallow Ritz method, we develop a time-stepping neural network method for solving the Allen-Cahn equation. The unconditional energy stability of the time discretization and an error estimate in the $L^2$ norm are established. Numerical experiments demonstrate that the neural network approach yields accurate solutions for singularly perturbed nonlinear diffusion-reaction equations and the Allen-Cahn equation. The method effectively resolves interior and boundary layers without oscillations and significantly outperforms the finite element method on fixed uniform meshes.

Recent video super-resolution (VSR) models achieve efficient one-step generation through diffusion distillation. However, high-resolution inference remains costly, and distilled models are difficult to align with human preferences. These challenges call for improvements in both model architecture and training strategy. We present FlashVSR+ to address both limitations. On the architecture side, we introduce hybrid attention based on the strong locality observed in pretrained VSR models, while retaining sparse global interaction for long-range consistency. It achieves up to $2\times$ DiT speedup and 45% lower peak memory, enabling spatially untiled 4K VSR on a single 24GB RTX 4090. On the training side, we propose Reward-Yield Distribution Matching Distillation (Ry-DMD) for preference-aligned one-step generation. Ry-DMD uses reward-weighted teacher rollouts to incorporate black-box rewards into distribution matching, without backpropagating through the reward model. We further show that Ry-DMD generalizes beyond VSR to T2I generation. Together, FlashVSR+ delivers stronger restoration quality and substantially improves high-resolution efficiency, with consistent gains in user preference. We release our code, models, and training data at \href{https://www.flashvsr.tech}{\textcolor{FlashBlue}{www.flashvsr.tech}}.

The dominant barrier to running large clinical language models at the point of care is not arithmetic throughput but graphics memory: consumer GPUs cap at 4-16 GB, while the working set of a 27-billion-parameter model in 4-bit precision is 17-25 GB. We argue that the edge node should be treated as capacity-enabled (system RAM) rather than compute-limited (GPU VRAM), and we test this on commodity hardware. Using an INT4 build of a 27B open model as an architecture-faithful stand-in for a 27B medical model, we measure the memory footprint, inference latency, energy, and federated communication cost of a fully CPU-based pipeline driven by synthetic Fast Healthcare Interoperability Resources (FHIR) prompts. On two independent consumer machines (a desktop-class machine with 67 GB of RAM and a 6 GB GPU, and a mobile workstation with 34 GB of RAM and a 4 GB GPU) the model loads entirely into system RAM (37% and 60% utilization, respectively) while both GPUs are out-of-memory before the first step. Inference runs at overnight-batch speed (0.6-2.4 tokens/s for the 27B model), and federated fine-tuning would transmit only a 22.6 MB low-rank adapter, a 99.96% reduction over full weights. The capacity result is independent of CPU microarchitecture, RAM size, operating system, and GPU. We further identify an open, CPU- and NPU-optimized inference runtime as the natural production path for this approach, indicating a privacy-preserving and broadly deployable route for on-premises clinical AI. All data is synthetic; this is a research artifact and not a medical device.

Identifying a robot task from a short instruction is sensitive to which descriptions are available for retrieval. We investigate this effect in the natural-language instructions included with the RLBench manipulation benchmark. We extract 296 fixed descriptions from 62 publicly available task files and compare five training-free lexical rankers over 80 random choices of one indexed description per task. Exact strings assigned to conflicting task labels are excluded from the primary score. Character n-gram TF-IDF achieves 61.3% mean top-1 accuracy on the resulting held-out instructions, compared with 55.6% for BM25 and 54.9% for word TF-IDF. Across gallery choices, character TF-IDF ranges from 55.2% to 66.8%; its 5th-95th percentile range is 56.9%-64.8%. Adding a second indexed description improves character TF-IDF by 7.0 percentage points on matched queries. These results quantify the influence of reference wording and representation choice on instruction-to-task retrieval under a consistent evaluation protocol.

Federated learning (FL) enables collaborative model training without data centralization but struggles with incentive alignment, accountability, and robustness under heterogeneous non-IID data. Existing decentralized FL (DeFL) systems leverage blockchains for coordination yet face high on-chain overhead, unverifiable self-reports, and weak fairness guarantees. We present the FLAI Protocol, a trusted execution environment (TEE)-backed DeFL framework that integrates enclave-validated aggregation, sparse Peer Evaluation Graph (sPEG) scoring, and on-chain settlement. sPEG incorporates peer-based quality, class coverage, promptness, effort, and reputation into composite contribution scores. The protocol formalizes contribution evaluation and tokenized incentive distribution, ensuring verifiable fairness and resilience against free-riding and collusion. We evaluate FLAI on image, text, and healthcare datasets and demonstrate parity with FedAvg in utility while improving fairness and accountability. A case study in consumer eHealth (aggregating wearable and electronic health record data for real-world evidence generation) illustrates the protocol's applicability to privacy-critical domains where equitable rewards, auditability, and data sovereignty are essential.

The randomized primal-dual framework of Devanur, Jain, and Kleinberg (SODA 2013) splits each unit matching gain between endpoint dual variables according to the matched offline rank. The gain-sharing function affects the certificate but not Ranking's decisions. Keeping the splitting rule and expected edgewise feasibility criterion fixed, we define certifying capacity as the strongest uniform guarantee obtainable by functional optimization over a prescribed function class. Its shortfall from Ranking's actual competitiveness is the certifying gap. We study unweighted online bipartite matching under adversarial arrivals on graphs with online degree at most $d$ and on $d$-regular graphs. The optimized protocol is tight on both degree-one classes and in the unrestricted setting, but has a strictly positive gap on both degree-two classes, even over all measurable nonnegative complementary splits. In both degree-two settings, we determine Ranking's exact competitiveness and the exact certifying capacity over nondecreasing gain-sharing functions. For general $d$, we optimize over nondecreasing, right-continuous functions $f:[0,1]\to[0,1]$ with $f(1)=1$, using their probability-measure representation. Both the certifying capacity and actual competitiveness converge to $1-1/e$ in both instance families as $d\to\infty$. The resulting gap is at most $1/(2ed)+O(d^{-2})$ for online degree at most $d$; on $d$-regular graphs, it vanishes but is at least $(1/(2e)-o(1))/d$. These results distinguish function-class restrictions from limitations of the fixed protocol and motivate tighter randomized primal-dual frameworks.

Thought templates, skill libraries, agent memories and caches turn many reasoning steps into one. We call this reasoning contraction. A contracted memory saves reasoning steps and errors until its facts change, then quietly answers wrong. Most systems trust memories by counting successes and failures. We prove that such counters are badly tuned change detectors: a symmetric counter notices a partial change only after a power of its false-alarm horizon, and an unclipped one may never notice it. On five LLMs answering multi-hop questions over edited facts, ExpeL- and ACE-style counters serve 7.0-13.5x more stale answers than certified memories given the same information, and on code tasks with an exact verifier 35-73x more. Safety also has a price. When facts change and each memory is judged only by its own verified outcomes, one number, the certification load, sets how much reasoning any safe memory can skip, and under Zipf demand it yields a scaling law. Our recipe weighs each memory's failures by its own likelihood ratio, certifies its utility against live reasoning, and restarts the certificate when facts change, with the likelihood-ratio detector as the safety net for unannounced changes; for large edits it also invalidates the memory.

The rapid advancement of cybersecurity agents in vulnerability discovery has increased the need for benchmarks that evaluate their capabilities accurately. However, synthetic benchmarks lack realism, real-life benchmarks are prone to contamination, and many existing benchmarks have already been saturated by current models. We therefore set out to develop a benchmark that is realistic, reproducible, resistant to contamination, and challenging. Our key insight is that real-life business-logic vulnerabilities are subtle and exercise complex application interactions, making them naturally demanding. The difficulty lies in specifying evaluation criteria precise enough to faithfully measure model capabilities while keeping the tasks free of data-leakage. We introduce AnonVul, a cross-language benchmark of 132 application-level business-logic vulnerabilities, built by a semi-automated pipeline that anonymizes each task, injects oracles from expert-written specifications, and differentially verifies proof-of-concept exploits. The benchmark spans nine languages, with two evaluation tracks on black-box and grey-box settings, where each task runs in a sandbox with specialized oracles embedded in key exploitation steps. We evaluate six state-of-the-art vulnerability-discovery agents on the AnonVul benchmark, measuring discovery success, cost-effectiveness, anonymization sensitivity, oracle validity, and failure modes. Our results show that AnonVul remains challenging, with the best agent succeeding in only 23.11% of attempts. On a matched cohort of 20 vulnerabilities, restoring software identity increases discovery from 21.25% to 40.83%, highlighting the importance of controlling recognizable software knowledge when evaluating vulnerability discovery.

Safety verification of a Markov decision process (MDP) asks whether the probabilities of reaching several failure states jointly lie in a convex safe region under every scheduler. Watanabe and Unno (2026) prove safety by refining an outer approximation of the failure probabilities, where each query bounds a nonnegative weighted sum of the probabilities. Their rule selects weights by the largest Euclidean violation. We present a unified framework to measure violation in an explicitly chosen error unit: directions are scored by the support-function gap between the outer approximation and the safe region, normalized by the support function of the unit. We show that the maximal score, when positive, equals the smallest enlargement of the safe region in multiples of that unit that contains the outer approximation. The Euclidean unit recovers the earlier rule; two new rules are generated by the unit cube and the safe region itself. We experimentally evaluate how the choice of measurement rule affects the performance of safety verification for MDPs.

Textual reasoning before acting has improved vision-language-action (VLA) models, but at the cost of latency: every action conditioned on the reasoning text must wait for autoregressive decoding. We show that actions do not need to wait for or condition on the generated reasoning to benefit from it, because the reasoning supervision already improves the backbone representation. We introduce TacitVLA, a reasoning VLA whose action expert never attends to the reasoning it generates for the current call. On long-horizon tasks where the reasoning text also serves as episodic memory, it can generate each entry during action execution rather than before acting. We evaluate it in a controlled study on four benchmarks spanning manipulation and autonomous driving, memory and non-memory tasks: RoboMME, RMBench, LIBERO, and PhysicalAI-AV. On all four, TacitVLA matches the performance of its explicit-reasoning counterpart at up to 3.4x lower latency, and matches the latency of the same architecture trained without reasoning at up to +17.1% success. In absolute terms, it surpasses the best previously published results on the RoboMME Counting suite (82.7% vs. 66.8%) and RMBench (77.5% vs. 56.5%), and reaches 72.9% on LIBERO-Plus and a 52.5% scene score on the public split of the PhysicalAI-AV closed-loop driving challenge. Project page: https://tacitvla.github.io.

Lexical representations can encode both abstract information and information specific to particular contexts. We investigate how these forms of information are organized in language models by decomposing contextual word representations into item, class, and word-class interaction components. Across six pretrained transformers and two linguistic categories, we find that the interaction is consistently present, grows across layers, and contains information associated with both item and class. Unlike the nearly orthogonal item and class effects, the interaction overlaps with both and encodes class information beyond the shared class effect. An interaction also emerges in artificial-language learners even when the input contains only additive structure. Together, these findings show that lexical representations combine abstract information with structure specific to particular word-class combinations.

Actuator sizing is an early and consequential decision: joint torque and velocity specifications fix the motors, gearboxes and brakes before any CAD model exists. Sim4RAD is a MATLAB toolkit for that phase and for the verification that follows it. It builds a rigid-body model from a handful of parameters, and four analyses supply the specification: a static analysis combining lever-arm sampling with a posture search, giving a rated torque at a chosen payload envelope and a reproducible peak torque; a dynamics pipeline that reshapes a trajectory under one of four constraints, including the actuator torque-speed envelope; a stochastic analysis over many random motions, so a specification rests on a class of motions, not one trajectory; and a workspace and dexterity analysis that separates where the arm can reach from how it can reach there. Replacing the synthesized model by a URDF file turns the same analyses to verification. A 7-DOF anthropomorphic arm is compared in three layouts differing in the shoulder axis order and in whether the last wrist joint is coaxial with the payload. The static demand at half arm length is nearly the same in all three, and it sizes the joints rather than the sampled motion; reordering the shoulder axes leaves the worst-case torque unchanged at 55 N m and the reachable volume within 8 percent. A last wrist joint coaxial with the payload removes its gravity torque, reduces the adjacent wrist joint by almost a third and cuts the last joint demand tenfold in a 5 m/s scan, at the price of dexterity (mean 0.570 to 0.348). The same arm needs 20 to 37 times more torque for a 5 m/s target than for a 2 s motion: the scenario, not the trajectory, must be fixed first. Verification should use the URDF model of the detailed design; the toolkit is available at https://github.com/LasterLi/Sim4RAD.

Coordinated online accounts can amplify messages and manipulate attention, but detecting such behavior is challenging due to confounding factors like account activity and item popularity. We propose a statistical null-model framework that assesses behavioral similarity between accounts by comparing observed coordination indicators against activity- and popularity-conditioned reference distributions. By combining evidence across multiple indicators and partners, our method can identify suspicious coordination links more accurately. Evaluations on nine labeled information-operation datasets involving over 152 thousand accounts show that aggregating p-values multiplicatively across indicators and partners yields improved detection performance. While results vary across information operations, this approach offers a generalizable and statistically principled tool for uncovering coordinated online behavior.