Papers

In on-device NLP tasks, limited resources of embedded hardware, such as the Raspberry Pi 5, require efficient inference strategies. This paper introduces FrugalSOT (Frugal Search Over The Models), a resource-aware model selection architecture for on-device NLP inference. FrugalSOT estimates each request's complexity by extracting features such as prompt length, named entity density, and syntactic complexity. The request is first made to the least complex model that is likely to pass a relevance threshold. If the output of that model falls short of the threshold, the request is made to a more complex model. It is important to note that the relevance threshold undergoes continuous updates in the background. using past validation outcomes in an adaptation process using a low-pass filtering mechanism, thus imparting adaptation to changing input patterns. Experimental results achieved on a Raspberry Pi 5 show that FrugalSOT reduces average inference time and overall computational resource use to a significant extent compared to a single-model baseline approach, without compromising output relevance to the same extent as the most sophisticated model. These results confirm that adaptive model selection can enable efficient, high-quality natural language processing inference on limited devices.

This paper investigates wearable-based recognition of human activities and gestures to support Human-Robot Interaction (HRI) in object-handover and assembly-like scenarios. Electromyography (EMG) and Inertial Measurement Unit (IMU) signals were collected using a Myo armband, culminating in a novel dataset introduced as MAGIC-HRI (Multimodal Activity, Gesture and Intention Collection) with a large taxonomy of 53 movement classes, including Brazilian Sign Language (LIBRAS) numbers (0-9), hand gestures, object/tool handover actions (pick up/give/hold), tool-manipulation tasks, and generic assembly/idle motions, collected from 11 participants with 10 samples per class (530 samples per participant). Signals are segmented by detecting muscle activation via an EMG energy envelope, then processed using sliding windows; time- and frequency-domain features are extracted. Multiple classical classifiers are tuned via cross-validated grid search, with Random Forest as the strongest baseline. A Leave-One-Subject-Out (LOSO) protocol reveals a large generalization gap, indicating substantial subject dependence. A personalized adaptation experiment suggests that injecting a small number of samples from a new user can markedly improve recognition. Overall, the study contributes a broad, HRI-driven multimodal dataset, a rigorous evaluation emphasizing generalization, and practical evidence that personalization is likely required for robust deployment in practical HRI.

Generative artificial intelligence (genAI) systems are increasingly integral to epistemic processes such as hypothesis generation, explanation construction, and decision-making. Although they reliably enhance performance, emerging evidence reveals a metacognitive dilemma: as external generative capacity increases, internal monitoring, calibration, and cognitive engagement may decline. This reflects a redistribution of cognitive control within distributed human-AI systems that cannot be explained by automation bias or reliance on algorithms alone. We propose the AIRIS (AI-Augmented Inquiry and Regulation in Hybrid Systems) framework to analyze this dilemma and specify where regulatory intervention can counteract it. AIRIS is a multi-level control allocation architecture specifying the conditions under which epistemic agency can be preserved in hybrid generative systems. Drawing on distributed cognition, cognitive load theory, multimedia learning, and self-regulated learning, it identifies seven interacting mechanisms through which hybrid cognition may become destabilized, from delegation and calibration drift to motivational-affective drift. Five regulatory operators (Anticipate, Interrogate, Reflect, Integrate, and Synthesize) target internal generative engagement at points of emerging instability. The architecture does not itself improve learning; it specifies what must remain in place for genAI-supported work to sustain understanding, whether through instructional design, teacher guidance, or learners' own regulation. We derive testable propositions concerning the seven mechanisms and the five operators, reframing AI augmentation as a problem of control allocation in distributed generative systems. Beyond theory, AIRIS offers a research agenda, a design framework for genAI-integrated learning environments, and a conceptual toolkit for the governance of hybrid human-AI cognition.

TEE-shielded inference keeps sensitive state in a trusted execution environment (TEE) while offloading linear algebra to an untrusted accelerator. Wang et al., in Game of Arrows (USENIX Security 2025), showed that five widely adopted lightweight defenses preserve vector directions and introduced ArrowMatch to exploit this leakage. They then proposed ArrowCloak, which adds a different multiple of one shared mask direction to each vector and bases its weight-recovery hardness argument on Learning with Errors (LWE). ArrowCloak successfully reduces ArrowMatch to near-black-box levels. In this paper, we revisit ArrowCloak from cryptographic and structural perspectives. Its LWE formulation does not by itself establish standard LWE hardness: the reduction direction, quantized arithmetic, and joint instance distribution do not meet the required conditions. Reusing one mask direction leaves a recoverable rank-one component across the released matrix. We exploit this structure with our proposed attack, an end-to-end, query-free recovery attack. Given a public checkpoint and the obfuscated weights, the attack removes the masking subspace, recovers the hidden one-to-one correspondence, and reconstructs protected weights without transformation secrets, victim queries, or fine-tuning data. Across six model-task pairs spanning classification, segmentation, and diffusion, the attack recovers 99.92%-100% of hidden vector correspondences. Reconstructed classification models achieve 94.39%-99.54% victim agreement and differ by at most 1.59 percentage points in accuracy; the recovered segmentation model achieves 98.35% output agreement. These findings suggest that lightweight protection should address both per-vector geometry and joint structure across released weights.

Chain-of-thought (CoT) prompting improves LLM reasoning by decomposing complex problems into intermediate steps, but its sequential nature increases decoding latency and memory usage. Mixture-of-Experts (MoE) models scale capacity through sparse expert activation, yet their full expert weights often exceed GPU memory and require costly GPU-CPU transfers. Existing runtimes treat all tokens uniformly, overlooking a key structural property of CoT traces: consecutive reasoning stages exhibit coherent and predictable expert activation patterns. Ignoring this stage-level regularity leads to inefficient caching and unnecessary data movement. We propose SAEM, a stage-aware MoE inference runtime that detects reasoning stage boundaries and exploits stage-level activation coherence to guide expert placement. SAEM combines stage-aware caching, expert-aligned token repacking, and in-situ CPU execution to reduce data transfer and kernel fragmentation. On mathematical and scientific reasoning workloads, SAEM achieves an average 1.33x throughput improvement over the strongest state-of-the-art caching and offloading baselines under constrained GPU memory, rising to 1.54x when calibration data matches the workload, demonstrating the effectiveness of stage-aware, locality-driven MoE inference for CoT reasoning.

Retrieval-augmented generation relies mostly on flat, fixed-granularity indexes: documents are cut into uniform chunks and retrieved by similarity, discarding the hierarchical structure of the source. We introduce Semantic Compression Trees (SCT), a hierarchical index in which each node stores only its semantic residual – the information it adds beyond its parent – and retrieval proceeds by progressive descent from the root, so that per-query cost is governed by tree depth rather than collection size. We evaluate on QASPER (50 papers, 173 questions) under two protocols differing only in whether the benchmark supplies the relevant document, with bootstrap confidence intervals and paired significance tests throughout. The results are mixed and we report them as such. When the document is given, SCT with a zero-LLM extractive compressor matches dense retrieval on answer quality (0.274 vs. 0.277 F1, $p = 0.37$) using 30% fewer context tokens and no LLM calls to build the index, and residual storage beats storing full summaries at each node (0.274 vs. 0.205, $p < 0.001$). Increasing the collection fifty-fold multiplies flat retrieval's per-query scoring work by 48.9x and SCT's by 6.4x. Progressive descent itself is not supported. Retrieving the same residuals without the tree performs identically when the document is given ($p = 0.27$), and descent is substantially worse when the system must select the document (0.122 vs. 0.165, $p < 0.001$). Routing accuracy localises the cause: descent selects the correct paper 20.2% of the time against 39.3% for flat retrieval, because that choice is made from the root residual, the most compressed node in the tree. We conclude that the residual representation is worth keeping and top-down routing is not.

We investigate when a sparse nonnegative matrix can be recovered from a real-valued matrix of much lower rank by zeroing out its negative elements. The potential for such decompositions suggests a mathematical connection between sparsity and rank; we analyze a number of sparse matrices with this latent low-rank structure and use them to illustrate the geometric origins of this connection. Previous algorithms have discovered these decompositions via an alternating minimization over the factors of a low-rank matrix, but to do so, they have also needed to compute and store another matrix, neither sparse nor low-rank, that is the size of their product. We develop a stochastic, alternating least-squares algorithm that operates on smaller blocks of this dense matrix and scales as a result to much larger problems. We also show how to further accelerate this algorithm with sparse optimizations and customized CUDA kernels. As one example, we use the algorithm to analyze the sparse matrix of synaptic weights for the recently published $\textit{Drosphilia}$ connectome. The nonzero elements of this matrix, with 139,255 rows and columns, record the number of synapses between cells in the nervous system of a female fruit fly. Despite a slowly decaying spectrum of singular values, this matrix exhibits a latent low-rank structure that is predictive of cell categories across multiple levels of specificity.

Machine unlearning aims to remove the influence of targeted training data from a model while preserving its remaining capabilities, but evaluating whether such information has truly become inaccessible remains challenging. Existing benchmarks primarily assess unlearning under clean, non-adversarial queries, leaving open whether information that appears forgotten can still be recovered through strategic prompting. We address this gap through a unified evaluation of prompt-based and fine-tuning-based unlearning methods on TOFU using Llama-3.2-3B-Instruct, followed by an adversarial robustness evaluation of methods that perform strongly under standard metrics. We introduce Attack Success Rate (ASR), an LLM-as-judge metric that measures the fraction of adversarial responses whose leakage score exceeds $0.2$, and evaluate recovery across eight attack suites. Our results reveal a substantial gap between clean-query forgetting and adversarial robustness. Although several fine-tuning-based methods achieve Forget Quality above $0.91$, targeted information remains recoverable with ASRs between $72.8\%$ and $84.3\%$, close to the $87.5\%$ ASR of the unprotected base model. In contrast, clean multilingual reformulations yield only $2.95\%$ measured leakage. A manual audit further finds agreement between binary ASR decisions and human factual assessments in seven of ten cases, indicating that ASR provides a useful, though imperfect, signal of behavioral recoverability. These findings show that strong standard-metric performance alone is insufficient to establish robustness after unlearning and motivate adversarial stress-testing as a complementary component of unlearning evaluation.

Logic Tensor Network-Enhanced Generative Adversarial Networks (LTN-GANs) inject background knowledge by grounding each logical axiom as a predicate and training the generator to raise its satisfaction, a fuzzy truth value in $[0,1]$. Previous LTN-GAN work grounded every constraint this way, at the predicate level, and improved constraint satisfaction. A predicate, however, only scores a sample, so it cannot embed hard structural constraints, rules such as orderings, positivity, and definitional identities that must hold in every generated sample. In this work, we investigate grounding each axiom as a function symbol inside the LTN framework. We compare against the state-of-the-art alternative, a constraint layer that clamps each violating sample onto the feasible boundary and so produces outputs that are always valid. Our investigation shows that a valid sample is not always a realistic one. An inequality is not merely satisfied or violated. It holds by a margin, and a faithful generator should also reproduce the margin's real distribution. We find that the resolution ratio $R$, the data's scale over the margin's spread, is a diagnostic, computable before training, of which constraints a chosen grounding can learn. When $R$ is large, the predicate receives no learning signal, the clamp pushes every sample onto the boundary, and the margin distribution is lost while every standard metric still looks fine. A function symbol avoids both failures, computing the constrained variable rather than scoring it. Together the function symbols form a chart, a coordinate system inside the feasible region, where every sample is valid by construction and the margin is learned like any other quantity.

AI is increasingly lowering the barrier to data analysis and creating visualization scripts. However, a key obstacle in AI-assisted visualization is that certain data issues can lead to visualizations that are plausible, but misrepresent the underlying data. These visualization defects are elusive and difficult to fix, particularly for non-experts who may not know what data issues cause them or how to guide AI systems to resolve them. We present findings of a preliminary empirical investigation of how commercial LLMs identify and repair defect-inducing data issues. Using a curated subset of the 911 emergency-call dataset with five injected data issues, we evaluated GPT-5, GPT-4o, GPT-4, and Claude Sonnet 4.6 under a three-stage prompting protocol, including zero-shot, guided issue-identification, and guided issue-repair. We executed this protocol under two conditions: single-agent and a multi-agent orchestration that separates data issue detection, review, repair planning, data repair, and repair quality assurance. We observed that across both conditions, LLMs identified and repaired single-field issues (e.g., missing values) but struggled to identify and repair temporal, geographic, and semantic issues. Based on these observations, we discuss design implications for agentic visualization systems, including explicit representation of data assumptions, selective human intervention for ambiguous decisions, and evidence-based repair.

Benchmarks for scientific artificial intelligence are mostly written to be scored: multiple-choice questions, curated agent tasks with reference solutions, or simulators with a known generative structure. Real scientific requests arrive differently. They are underspecified, they carry attachments, and they lack ground truth. We report K-Bench 01, an evaluation built from first-turn requests sampled from live user traffic on K-Dense Web and run end to end by nine frontier models in identical sandboxes, yielding 1,602 completed agent runs. Three blinded language-model judges scored every run against an eight-dimension rubric. On a rubric whose 8-anchor is defined as work a domain scientist would accept with minor edits, no model clears the line under all three judges. gpt-5.6-sol has the highest pooled mean, 8.04, but its 95% interval [7.80, 8.23] spans the threshold, and two of the three judges rank claude-opus-5 first instead. We therefore report the ordering of systems as the reproducible quantity, the absolute level as an attribute of the instrument, and the top of the table as unresolved. Across all 39,934 scored judgments – the eight dimension scores plus a holistic overall for each assessment, excluding not-applicable cells – 47.6% fall below the 8-point threshold. Difficulty is not uniform across the rubric: scientific accuracy averages 6.22 against 7.33 for communication, on identical denominators and in the same direction within every one of the nine models. The single leading failure tag is overclaiming, on 31.4% of assessments. We argue that the informative quantity for scientific agents is not a leaderboard position but the joint distribution of what was delivered, what was claimed, and what artifacts were produced.

Voice has been proposed as a low-cost screening signal for chronic obstructive pulmonary disease (COPD). COPD is strongly age-associated and voice changes with age, thus such results admit a trivial alternative explanation. We re-evaluate a public sustained-phonation corpus ($1246$ recordings, $68$ participants) under a strictly participant-level protocol. We therefore evaluate on repeatedly drawn age-matched cohorts and report the discrimination achieved by the confounders themselves on those same cohorts. Where raw (unmodelled) age ($0.510$ $[0.469, 0.551]$) and raw gender ($0.479$) are both measured at chance, acoustic models excluding age retain ROC-AUC $0.717$ $[0.552, 0.859]$ and average precision $0.747$ $[0.581, 0.892]$ against a one-to-one baseline of $0.5$, whereas models containing age fall to $0.531$–$0.679$. The separation is reproduced by two further learners with fixed hyperparameters. Two findings have broader methodological implications: models trained with age transfer less effectively to an age-balanced target cohort than otherwise identical models trained without age, and fourteen classical voice-quality and perturbation measures achieve comparable discrimination to a $55$-dimensional combined representation. We conclude that a non-age acoustic signal is present, that confounding by recording conditions cannot be excluded from the released features, and that the evaluation protocol in standard use cannot distinguish these possibilities.

Clinical data sources such as electronic health records and wearable sensors record patient status repeatedly over follow-up, often at irregular times and on different schedules for different measurements. These data create opportunities for continuously updated, individualized risk prediction. Existing approaches, however, often simplify the temporal structure for model fitting. We introduce Random Hazard Forests (RHF), a survival tree ensemble that estimates how a patient's hazard changes in continuous time as new measurements become available. The method formulates the estimation problem directly through a nonparametric hazard likelihood for predictable covariate processes. An efficient working model guides tree construction, after which flexible time-varying hazards are estimated for each terminal node. Given any predictable covariate path, each tree follows the path through its terminal nodes over time and assembles the corresponding node-level hazards into a trajectory. Averaging these trajectories across trees yields the pathwise hazard estimate. Because routing at each time uses only the covariate state available immediately beforehand, the construction accommodates internal longitudinal covariates without lookahead. Simulations and an intensive care application show that the forest accurately estimates changing risk under irregular and asynchronous covariate updates.

Target Wake Time (TWT), introduced in IEEE 802.11ax, lets an access point schedule exactly when each station wakes, transmits, and dozes. Existing TWT schedulers optimize energy or throughput, treating information freshness at best as a constraint and offering no performance guarantees. We design the wake schedule itself for freshness: minimize the weighted average Age of Information (AoI) over stations subject to per-station energy budgets, where the decision variables are the TWT triples (wake interval, offset, service period duration). We derive a renewal-exact AoI model for TWT under per-SP block fading and validate it against packet-level 802.11ax simulation with ~1% mean error. We show that, unlike preemptive scheduling, non-preemptive TWT packing can be infeasible at schedule density 1, and identify the granularity condition under which a small-first best-fit packer provably succeeds. Around this we build Harmonic-Greedy, a scheduler combining a convex relaxation, anchor-optimized power-of-two rounding, and a best-of-uniform safeguard, and prove it is a constant-factor approximation: 4/ln 2 ~= 5.77 under a mild granularity assumption and 6/ln 2 ~= 8.66 unconditionally. We implement the complete system in ns-3 – a TWT wake/doze mechanism integrated with the power-save architecture, plus the scheduler – and show that it is the only scheduler that stays near a relaxation lower bound across all regimes: against a strong energy-greedy baseline it ties when per-station energy floors already pin the periods, and wins by 4-36% exactly where the schedule density is binding and must be redistributed by AoI weight or channel quality rather than by energy budget – the regime our analysis identifies.

Reinforcement learning with verifiable rewards (RLVR) improves the reasoning ability of vision-language models (VLMs), and diversifying the rollouts within each optimization group amplifies its gains. Existing approaches diversify through decoding temperature or pixel-space image distortion; we ask whether the perturbation belongs in the model's latent space instead. We introduce Noise-Contrastive GRPO (NC-GRPO), which injects scale-calibrated Gaussian noise into the last hidden layer of the prompt-encoding pass for half of each rollout group, branching those rollouts from a displaced departure state. Branches that reach the answer despite the displacement are reinforced over those derailed by it, converting sensitivity at the branch point into policy-gradient signal; the objective, reward, and inference protocol are untouched. On Qwen2.5-VL-7B trained on Geometry3K, NC-GRPO significantly improves out-of-domain mathematical reasoning over vanilla GRPO across five held-out benchmarks (pooled McNemar $p \le 0.001$) while also improving in-domain accuracy and hallucination robustness – the latter an axis on which image-space noise regresses even while posting a larger OOD average on perception-heavy benchmarks. Mechanism ablations indicate that independent stochastic diversity, not noise budget or direction, is the active ingredient, and a noise-scale study exposes a dial between reasoning specialization and general capability. NC-GRPO is designed to be modality-agnostic and integrates into a standard RLVR pipeline as a ~50-line change to the inference engine.

We give a catalytic logspace algorithm for the problem of graph spectral sparsification. Given an undirected graph $G$ on $n$ vertices and $\varepsilon>0$, our algorithm outputs an $\varepsilon$-spectral sparsifier of $G$ with $O(n\varepsilon^{-2}\log n)$ edges, matching the effective resistance sampling of Spielman and Srivastava (STOC 2008). This gives a new, natural problem in catalytic logspace that is not known to be in deterministic $\mathbf{NC}$ or $\mathbf{SC}$. Our main contribution is an entirely new technique in the compress–or–random paradigm for catalytic logspace that we believe will have further applications. We first analyze effective-resistance sparsification using a pessimistic estimator that can itself be computed in catalytic logspace. The estimator is motivated by the viewpoint of graph quasirandomness and immediately gives a simple, deterministic greedy algorithm for graph sparsification. Subsequently, we show that such a pessimistic estimator can be transformed into an algorithm that performs an in-place compression of a string with bad potential. Our algorithm is based on using the potential function to define a measure over strings, and implementing arithmetic coding using this measure in-place. This compression technique is substantially distinct from all prior tools in the field of catalytic computation.

Contact-rich robotic manipulation requires an accurate model of the kinematic relationship between a robot's joints and the task features it senses. This relationship is rarely known exactly: it changes with each tool the robot picks up and shifts, sometimes almost instantaneously, as contact modes change — especially for multi-fingered hands that make and break contact at points that are not exactly prescribed, as in full-hand grasping. This paper develops an adaptive scheme that estimates that relationship online, using only joint-angle sensing and a wrist-mounted force/torque sensor, with no exteroceptive measurement of the tool tip. We derive a provably stable kinematic update law that identifies the kinematics of an unknown tool from force/torque feedback alone, and prove stability of both the rigid case and the case with a compliance controller as an inner loop. We show that identification is confined to the directions the motion excites — so that, for example, a tool's length is unobservable under a rigid insertion push, while a compliant loop's passive yielding partially excites it; and that with a second-order admittance the compliant certificate holds unconditionally in continuous time. We also pose the combined control and estimation problem as a Quadratic Program (QP): the formulation yields the prediction term of the update law exactly but, instructively, cannot reproduce the tracking adaptation term. We validate the scheme in simulation on a peg-in-hole insertion. This work is the first step in a research program aimed at factoring manipulation learning into a task policy which can be learned in isolation of the robot, for instance by reinforcement learning, and an adaptive kinematic component that adapts online to the particular robot, hand, or tool in use.

Multi-horizon rare-event forecasting is hard under long macroeconomic series' data constraints: labeled events are scarce, and standard uncertainty quantification assumes an exchangeability that autocorrelation violates. A controlled ablation shows an apparent rare-event threshold for Adaptive Conformal Inference instead reflects calibration-set size. Across 200 random calibration sets, support width of the nonconformity-score distribution explains up to 85% of coverage variance versus 2% for rare-event count; the same, not the same magnitude, replicates across synthetic conditions and five countries (five-country Spearman $\rho$ 0.45-0.66 vs. 0.02-0.23). A diversity-maximizing selector built on this is the only strategy tested that improves long-horizon coverage (67.8% to 81.4% at six months); Mondrian, shift-robust, and extreme-value alternatives fail to close it. Mondrian even worsens coverage under oracle labels. A compact proposition explains why: coverage deficit reflects how closely the calibration set's upper quantile reaches the test distribution's. Diversity is necessary, not sufficient. Demonstrated on a two-stage U.S. recession-forecasting framework with RegressorChain, whether six-month coverage reaches 90% under honest scoring remains open, a question this paper quantifies rather than resolves.

Functional decline in older adults is typically recognized only after falls or observable gait impairment, closing the window for prevention. We investigated whether temporal trajectories of routine biomarkers, already recorded but rarely analyzed longitudinally, can identify patients in the pre-clinical phase of mobility decline. Using the All of Us Research Program (N = 297,861; 11.1% cases), we derived trajectory features (slope, variability, delta, mean) for twelve biomarkers over a three-year pre-index window. LightGBM models incorporating trajectories significantly outperformed static laboratory summaries (AUROC 0.797 vs. 0.755; DeLong p < 0.001; AUPRC 0.380 vs. 0.304). A 1:1 age- and sex-matched analysis confirmed an independent trajectory signal (AUROC 0.727 vs. demographics-only 0.680). A horizon analysis demonstrated sustained prediction 3-12 months before decline onset (AUROC 0.768-0.740). Because the model uses only measurements already ordered in routine care, it supports passive, zero-burden EHR integration for early detection of pre-clinical functional decline.

Under mild assumptions, any quantum mechanical (QM) model or quantum field theory (QFT) admits a representation in terms of an ensemble of neural networks with countably many random parameters. We investigate the features of NN-QM and NN-FT models with finitely many parameters, such as a feedforward network of width $N < \infty$. We find that, generically, such models must violate one of the properties of conventional Euclidean QFTs, such as reflection positivity or cluster decomposition. We present several complementary ways of understanding which features can and cannot be preserved at finite $N$, both in QM and in QFT.

We introduce a large language model (LLM)-driven approach for generating dynamic algorithmic dispatch heuristics in high-performance linear algebra. By combining prompt engineering with LLaMA 3 and a curated performance database, the model learns to synthesize selection heuristics that exploit structural patterns to identify fast algorithmic choices. A case study on LU factorization demonstrates the model's ability to replicate expert-designed strategies. This work, developed as part of the DARPA-MIT SmartSolve project, highlights the promise of LLMs for algorithmic discovery and the development of more adaptive, fast linear algebra software.

Oral cancer is a leading cause of mortality in low-to-middle-income countries, where a shortage of specialists delays diagnosis. While point-of-care screening via smartphones offers a scalable solution, developing robust AI for resource-constrained settings poses significant challenges, including class imbalance in training data, variable data quality, and computational constraints on edge devices. In this paper, we present the optimisation of lightweight deep learning models for smartphone-based oral cancer screening. Using a diverse, multi-centre retrospective dataset of approximately 30,000 images acquired over a decade, we systematically evaluate state-of-the-art convolutional, transformer, and hybrid architectures. Through rigorous pipeline ablation, we demonstrate that directly optimising hybrid architectures for the edge strictly outperforms computationally heavy paradigms, such as large models or knowledge distillation. Furthermore, interpretability analysis and simulated noise-stress tests revealed that the system anchors on clinical features and remains robust to unstructured sensor noise, despite vulnerabilities to impulse bit errors. In the held-out test set, our optimised MobileViTv2 models achieved an average sensitivity of 83.2 $\pm$ 1.5% and an average specificity of 86.0 $\pm$ 0.8%, with the best model exhibiting 87.4% sensitivity, 86.5% specificity, and a critical negative predictive value of 97.2% with reference to specialist labels. These results confirm that with targeted architectural selection and streamlined optimisation, interpretable and robust lightweight AI models exhibit high potential for edge deployment to enable automated triage in primary care settings.

World models built on the RSSM architecture, such as DreamerV3, keep a recurrent hidden state $h_t$ trained only to reduce prediction error. We show this state also tracks its own confusion, hiding in plain sight: nearly orthogonal to $h_t$'s directions of greatest variance, invisible to any variance-based method. It is functionally distinct from ensemble disagreement, which flags new inputs, and reconstruction error, which flags bad predictions right now. On a test holding prediction error fixed while confusion varies, a linear probe on $h_t$ finds the signal (AUROC 0.72, 5 runs), while an ensemble baseline scores below chance. A discounted count of recent high-error steps explains 80% of the probe's output ($R^2=0.80$). We confirm the signal is causally used, not merely present, by editing $h_t$ directly and watching behaviour change, including a check using real values from other trajectories instead of synthetic edits. Its geometry and closed form generalize across three control tasks; the decisive dissociation test itself holds cleanly on only one, and its practical use, deciding when to check reality instead of trusting imagination, generalizes to only two of the three tasks.

Consider the following problem of learning an unknown linear order on $n$ items. In each round, the learner guesses a complete ordering of the items and receives either confirmation that the guess is correct or a counterexample: a pair of items in the wrong order. The goal is to identify the unknown order using as few queries as possible. We study this problem when up to $k$ of the returned counterexamples may be untruthful, where $k$ is not known in advance. We determine the optimal query complexity up to constant factors: \[ \Theta(n\log n + nk). \] Thus, while the noiseless complexity matches the classical complexity of sorting, each untruthful counterexample incurs an additional cost of order $n$. The upper bound is based on a geometric representation of permutations and Grünbaum's theorem, while the lower bound combines sorting arguments with a Condorcet-type construction. We also study the case where the target ranking has a low-dimensional geometric representation: each item is represented by a point in $\mathbb{R}^d$, and the ranking is obtained by projecting the points onto an unknown direction. For these classes we give an upper bound of $O(d^2\log n+dk)$ and a lower bound of $\Omega(d\log n+dk)$, leaving a factor of $d$ gap in the noiseless term.

Multimodal Large Language Models (MLLMs) are increasingly deployed in high-stakes domains where fairness is a critical safety requirement. In practice, these models are continually updated through continual learning (CL) to adapt to evolving tasks and data distributions. Prior work has shown that backdoor attacks can manipulate MLLM responses through hidden triggers, but naively implanted backdoors degrade as models undergo subsequent updates of CL. Although fairness has emerged as a central concern for MLLM deployment, whether backdoor-induced fairness violations can survive CL remains unexplored, leaving two critical questions unanswered: (1) whether a backdoor can reliably induce fairness violations in MLLMs, and (2) whether such fairness-targeted backdoors can persist through continual learning. We bridge this gap by proposing Persistent Fairness Backdoor Attack (PFBA) to inject persistent and group-specific discrimination into MLLMs. Specifically, PFBA achieves this through two novel mechanisms. The Latent Space Fairness Reinforcement reshapes the model's deep feature geometry by anchoring privileged-group representations to preserve utility while repelling and clustering targeted-group representations to sustain discrimination, and the Continual Learning Simulation iteratively optimizes the trigger against simulated parameter drift to ensure backdoor persistence across future updates. Extensive experiments demonstrate that PFBA induces severe fairness disparities that persist across continual learning rounds, evading standard backdoor defenses. The data and code are publicly available at https://github.com/lyygua/PFBA.

The classical procedure for solving hydraulic networks relies on the assumption of constitutive equations, which state the relationship between the pressure gradient and fluxes along an edge. In this paper, we propose a data-driven framework that bypasses these constitutive models, formulating the incompressible flow problem directly on the graph topology. By assigning discrete measured data points to network edges, the problem is cast as a mixed-integer quadratic optimization over nodal pressures, edgewise states, and data assignments, accommodating both laminar (convex) and turbulent (non-convex) regimes. To solve this, we evaluate three algorithms: a GPU-accelerated Brute Force method, the Alternating Direction Method (ADM), and Deterministic Annealing (DA). The Brute Force method certifies global optima for small networks, establishing a good baseline for the iterative solvers. We demonstrate that ADM is highly sensitive to its initialization, requiring a faithful surrogate model to avoid local minima. In contrast, DA eliminates this dependence through unsupervised clustering. By annealing the data assignment from the centroid to strict nearest-neighbor projections, DA consistently reaches the global optimum without prior manifold reconstruction. Furthermore, numerical experiments reveal that DA is robust to noisy data, remains thermodynamically admissible on all but the coarsest and noisiest datasets, and sustains its convergence rate on larger networks where Brute Force is intractable and ADM degrades. Finally, the framework is successfully validated on complex configurations, including mixed-component networks and a $958$-edge arteriovenous bed featuring a non-Newtonian Carreau–Yasuda model, demonstrating its scalability and practical applicability.

Consider a typical test of a robot system: one observes a sequence of outcomes concerning some aspect of interest (crash or no crash, tracking error, time to completion), and reports a mean (crash risk, average error, mean time to completion) and, more importantly, an interval guaranteed to contain that mean at a prescribed confidence, referred to as a performance certificate. Given expensive real-world trials, the sample size is therefore small, and the certificate is often loose. Now consider the same procedure, except that before each real outcome is revealed, the operator “peeks” at a large bank of simulated results, and places a bet on where the real outcome will land. As the real outcomes settle the bets, the operator gains or loses wealth. One's “trust” over simulators also shifts within the portfolio. This paper develops that idea into a sim-to-real betting certificate framework with three contributions: (i) An algorithm that links a scalable bank of simulators to effective bets, and the accumulated betting wealth to the certificate. (ii) A proof that the returned certificate is anytime valid, covering the true mean with the prescribed probability, using any simulator bank. (iii) The guaranteed wealth-regret bounds yield configuration principles for the proposed algorithm and simulator bank design to deliver tight certificates. Experiments across synthetic distributions and real-world robot tests, covering both replayed standardized testing outcomes and online runtime evaluation, show the proposed method narrows the certificate by $51.6\%\pm16\%$ against classic and state-of-the-art baselines, and by $32.26\%\pm8\%$ in the extremely limited-sample regime ($\leq30$ samples).

Lung cancer remains a leading cause of cancer-related mortality worldwide, and early diagnosis is critical for improving survival. However, early-stage malignancies can be subtle on chest X-rays, creating challenges for radiologists. This study evaluates Vision Transformers (ViTs) for predicting lung cancer one to two years before clinical diagnosis. We analyzed 259,361 chest X-rays from 91,020 imaging studies at the Jamaica Plains VA Hospital in Boston, MA. The dataset showed extreme class imbalance, approximately 1:150 cancer to non-cancer, which was addressed using hybrid under- and over-sampling and class-weighted loss optimization. Three ViT configurations were evaluated: a model trained from scratch, an ImageNet-pretrained model, and a Corona-pretrained model fine-tuned on the lung cancer dataset. Transfer learning improved performance, with pretrained models exceeding the scratch baseline by 6-10 percentage points in AUC and about 10-12 percent in balanced accuracy. ImageNet-pretrained models showed the most stable overall performance, while Corona-pretrained models achieved higher sensitivity in some settings but greater variability. Moderate resampling ratios, including 1:1 undersampling and 1.5:2 oversampling, provided favorable trade-offs between sensitivity, precision, and computational efficiency, reducing runtime by up to 70 percent without major performance loss. These findings demonstrate the potential of ViTs for early lung cancer risk prediction from routine chest X-rays. Although performance remains below clinical deployment thresholds, the results support further development of ViT-based triage systems to flag high-risk patients for earlier evaluation.

Deploying a safety layer for large language models on commodity hardware is constrained by the guards available to do it: current open guard models hold between 1 and 9 billion parameters, are oriented toward the graphics processing unit, and answer in seconds per request on a central processing unit. This paper presents a reproducible, license-aware knowledge-distillation recipe addressing that constraint. A strong open guard labels a corpus of roughly 97,000 prompts, drawn from 24 public datasets, into seven safety categories aligned to a public hazard taxonomy, and a fleet of small students spanning lexical, shallow, encoder and generative architectures is trained to reproduce that signal. The corpus is partitioned at the license boundary, so that a deployable and a research model differ only in their training data and the cost of that restriction becomes measurable. Every model is scored against an independent gold benchmark of 6,361 rows over four slices, labeled apart from the teacher and including a slice of harmless prompts that makes over-defense measurable. The distilled students match the teachers on adversarial text within overlapping confidence intervals and reduce false alarms on harmless prompts, the smallest generative student reaching 3.8% against 4.8% for the 8-billion-parameter teacher, while the encoder classifies in roughly 24 ms per request on CPU. Per-class rebalancing is the only decisive ingredient of the recipe. No superiority over the distilled guards is claimed; on the clean reference slice they remain ahead.

Human mobility serves as an essential proxy for understanding social, economic, and environmental dynamics in urban systems. Geospatial transferability, which measures a model's capability in a new location or unseen region, is a critical dimension for comparing different human mobility generation models. However, few studies have studied the intrinsic characteristics of geospatial transferability. To this end, this study systematically investigates the geospatial transferability of four representative human mobility generation models using a large-scale benchmark dataset of census tract level commuting flows across 2265 counties in the United States. Inspired by the domain adaptation theory in machine learning, we introduce geographic domain shift to describe the intrinsic differences in geographic feature distributions and spatial structures between source and target regions, which may jointly affect model transferability. Moreover, we propose two metrics, mutual information and spatial shift, to quantify the geographic domain shift. To examine their associations with model transferability, we employ linear mixed-effects regression to analyze the associations between geographic domain shifts and transferability. Our results reveal substantial spatial heterogeneity and asymmetry in transfer performance across regions. Both information shift and spatial shift exhibit statistically significant and complementary explanatory power. This indicates that geospatial transferability depends not only on model design but also on intrinsic geographic differences. These findings provide a novel methodological framework for evaluating and improving the geospatial transferability of human mobility generation models and support more robust and fair human mobility data synthesis across diverse regions. It also offers insights on spatial transferability for GeoAI model development.