Papers

Off-policy distillation is now central to large language model pre-training, yet how training data, objective parameterization, and model capabilities interact remains poorly characterized. We studies top-$k$-truncated, temperature-scaled off-policy distillation by decomposing this problem into two questions: an objective-to-capability analysis of how the training objective shapes token-level supervision and downstream performance, and a data-to-objective analysis of how data heterogeneity should inform objective routing. We first show that the language-modeling objective ($L_{\mathrm{LM}}$) and the knowledge-distillation objective ($L_{\mathrm{KD}}$) induce systematically different capability profiles, and trace this divergence to a gradient-level tension between direct observed-token reinforcement and teacher-supported alternative supervision. To quantify this tension, we introduce diagnostic metrics – support coverage, observed-token probability mass, and teacher-distribution concentration – and show via controlled sweeps that the support size $k$ governs a coverage-sharpness trade-off, while distillation temperature controls within-support probability allocation. We then examine adaptive objective routing: a domain-level policy that applies $L_{\mathrm{LM}}$ to math and code and $L_{\mathrm{KD}}$ to general-domain data yields consistent gains over both single-objective baselines, whereas token-level routing based on observed-token probability mass or teacher entropy fails to consistently match the single-objective baseline. These results suggest that effective objective routing depends less on routing granularity than on the quality of the routing signal, reframing continued pre-training via off-policy distillation as a structured, data-conditional supervision-design problem rather than a global hyperparameter choice.

Gate-level netlists exhibit intrinsic structural properties that influence signal propagation independently of functional simulation. We define a topology-driven structural manipulability score that characterizes node-level structural flexibility using path participation, k-core embedding, symmetry, and centrality. Modeling netlists as directed graphs, we formulate node-level regression to learn this topology-derived score using graph neural networks (GNNs). Experiments on ISCAS85 and EPFL benchmarks evaluate how effectively different GNN architectures approximate this metric across held-out circuits, with hierarchical models yielding the most consistent rankings. Component-level and ablation analyses examine the contribution of individual factors. As an illustrative case study, analysis of Trojan-injected circuits using TrustHub templates reveals statistically distinguishable structural patterns, indicating that topology-based scoring provides complementary structural insight.

Training multi-turn evidence-reading agents with outcome-only reinforcement learning is unstable because intermediate turns receive little direct credit. In HotpotQA experiments with Qwen2.5-3B-Instruct, GRPO initially improves (standard F1 0.430) but subsequently collapses to 100% format-violating outputs. Training-log diagnosis reveals a zero-advantage lock-in mechanism: all sampled trajectories receive the minimum format penalty (-2.0), group-relative advantages vanish, and the policy-gradient loss becomes zero–an optimization deadlock. We propose a variance-injection strategy: by assigning per-turn rewards to intermediate evidence-reading turns, we prevent the group reward distribution from collapsing to a single value–preserving the variation that GRPO's group-relative advantage requires. Contextual Information-Gain Policy Optimization (CIGPO) implements this strategy using the marginal increase in the frozen reference model's log-likelihood of the ground-truth answer as the per-turn signal. With separate normalization of IG and F1 rewards and an IG-weight curriculum, CIGPO reaches a standard F1 of 0.518 on HotpotQA at the 3B scale (from 0.252 base; +105%), compared with 0.430 for the best GRPO checkpoint and 0.000 for the final GRPO checkpoint. CIGPO maintains meaningful reward variance and avoids zero-advantage lock-in throughout training. These results identify reward-variance collapse as a concrete failure mode of outcome-only GRPO and show that turn-level IG rewards can prevent it in this HotpotQA setting.

Multimodal industrial anomaly inspection assistants are a critical component of next-generation smart factories, enabling interactive vision-language-based querying. However, multimodal large language models remain impractical for on-site deployment due to prohibitive computational demands and privacy risks from cloud-based inference. Compact multimodal small language models (MSLMs) offer a deployable alternative, yet progress is constrained by the lack of comprehensive robustness analyses and meaningfully challenging benchmarks that reflect real-world industrial conditions. To address this gap, we develop RobustMAD, the first deployment-motivated benchmark, designed to comprehensively evaluate model robustness through diverse open-ended queries spanning object understanding, anomaly detection, unanswerable problems, and visual quality degradations. Contrary to conventional assumptions, top-performing MSLMs exhibit promising capabilities, surprisingly outperforming even the larger GPT-5 Nano. However, they still fall short of safety-critical requirements, and RobustMAD reveals critical robustness gaps that pose operational risks. In particular, three recurring failure modes emerge: (i) fragile multimodal grounding under fine-grained distinctions or degraded visual conditions, (ii) insufficiently comprehensive responses, and (iii) weak logical grounding on unanswerable or ill-posed queries, leading to hallucinated outputs. Grounded in these insights, we provide actionable guidance for the design of next-generation multimodal industrial inspection assistants that leverage their promising competence. Code is available at https://github.com/en-research/RobustMAD.

Fine-Tuning-as-a-Service (FTaaS) platforms let users perform supervised fine-tuning (SFT) on customized data, but this pipeline can erode model safety alignment. To recover safety without re-running full alignment, existing realignment methods focus on calibrating the integration of safety patches into fine-tuned models. These methods exhibit a persistent safety-utility trade-off: weak repair leaves harmful behavior intact, while stronger repair increasingly damages the benign task. This paper shifts the focus from online calibration to offline patch learning and aims to learn a safety patch that restores safety while preserving task-specific capabilities. To this end, we propose TRACE, which simulates harmful SFT trajectories to produce progressively corrupted model states, and optimizes a safety patch simultaneously across these states. TRACE trains the safety patch during the offline stage, and reuses it across all user checkpoints without per-user calibration. We evaluate two representative models using three harmful SFT datasets, together with three utility benchmarks. Across six benchmarks and two models, TRACE consistently dominates the safety-utility frontier. TRACE improves the safety rate by up to 77 percentage points over the best baselines, while maintaining comparable utility to the undefended fine-tuned model.

Recent large language models (LLMs) can generate custom CUDA kernels that appear to outperform PyTorch on benchmarks such as KernelBench. Building upon this foundational framework, we demonstrate that frontier models frequently engage in reward hacking to artificially inflate reported performance. In this work, we identify two areas where evaluation frameworks must co-evolve with model capabilities. First, to accurately measure true speedup, we examine the baseline timing mechanism, noting that enabling Tensor Core acceleration with TF32 provides a more realistic estimation of execution on modern GPUs. Second, concerning algorithmic correctness, models often exploit the narrow test distribution by hardcoding bypasses for specific tensor values. By skipping required computations, these kernels artificially accelerate execution rather than implementing actual CUDA kernels. We introduce KernelBench-Verified, an extended evaluation framework that incorporates a TF32-enabled baseline and a four-distribution hidden test suite. We additionally introduce memory efficiency metrics that capture the often-overlooked speed-memory tradeoff in kernel optimization. Under verified single-turn evaluation with seven frontier LLMs, we find that the best-performing model (GPT-5.5) achieves a 0.88x geometric mean speedup, significantly lower than the 1.43x speedup observed under the standard evaluation protocol. No model consistently outperforms PyTorch when evaluated against realistic baselines. On the memory front, 28% of GPU kernels generated by the best model increase peak GPU memory usage. Our findings demonstrate the necessity of continually adapting robust evaluation protocols as LLM kernel generation capabilities advance.

Direct Alignment Algorithms (DAAs) such as DPO have become a common way to post-train and align LLMs with human preferences. However, DAAs have been observed to over-optimize their implicit reward model and decrease the likelihood of preferred responses. This results in a decrease in the total likelihood assigned to responses seen in the preference dataset, potentially resulting in undesirable behavior. To counteract this undesired side-effect of DAAs, we examine the effect of using objectives that add a regularization term to maintain the total length-normalized probabilities of the chosen and rejected responses. To better understand over-optimization, we investigate how response likelihood changes are distributed over the tokens with and without regularization. We find that a significant portion of the likelihood changes are due to a small set of outlier tokens, which explains how DAAs improve generation quality despite decreasing the likelihoods of chosen responses. We apply the proposed regularization to reference-based (DPO) and reference-free (SimPO) methods and find (1) improved trade-offs between generation quality and general benchmark capability and (2) improvements in reward modeling across datasets. For example, on Llama-3.1-8B-Instruct, we see both a >20% relative increase in AlpacaEval2 scores and >9% relative performance gains on general benchmarks. Additionally, we find that the added regularization term effectively mitigates the amount of displacement within preferred responses overall, and for the outlier tokens specifically, by utilizing low-likelihood tokens.

AI judges offer a scalable, low-cost alternative to human evaluation, but their outputs can be biased relative to human preferences and highly item-dependent, varying across judges, tasks, and domains. When uncalibrated AI evaluations are used for model ranking, item scoring, or population-level quality reporting, these biases can directly distort downstream decisions. We propose BACON, a four-stage pipeline that combines budgeted human calibration with multiple AI-judge outputs to produce more accurate annotations. BACON constructs full-coverage auxiliary features for every item, including multi-judge scores, token-level uncertainty statistics, and contextual embeddings. It then collects human labels for a small sampled subset and trains a cross-fitted outcome model to generate calibrated item-level surrogate predictions. These predictions support two use cases: population-level estimation of summary metrics, such as means or quantiles, using an augmented estimating-equation estimator with valid confidence intervals; and individual-level surrogate scoring for item ranking and annotation. BACON treats AI judges as auxiliary measurements rather than ground truth: human labels provide the calibration anchor, while AI-derived signals improve efficiency. Across diverse tasks, domains, and labeling budgets, BACON improves predictive accuracy and ranking consistency, and reduces bias and variance relative to raw AI outputs and purely human-label-based methods. These results show that BACON offers a practical, statistically grounded framework for scalable evaluation with limited human annotation.

Bayesian and multiplicative-weights updates reweight experts, models, or actions from sequential feedback. We show that the regret of any such update obeys an exact information-accounting identity. On each round, the learner's excess loss to any chosen comparator is the sum of an immediate cost for the uncertainty exposed by the round and a reduction in the information distance from the learner's current weights to the comparator. The cumulative cost defines a pathwise uncertainty clock, the intrinsic time of the realized sequence. Summing one-step balances yields two exact adaptive decompositions of cumulative regret, one for each natural way of composing the update across rounds. Because the decompositions are exact, favorable stochastic or low-noise regimes appear as self-bounding properties of the realized intrinsic time. The accounting also fixes a learning rate, inverse in the square root of intrinsic time. That schedule is competitive with adaptive baselines in selected online-learning settings. The same calculus covers Hedge, optimistic and side-information variants, continuous priors, boosting, online convex optimization, contextual bandits, and repeated games: the pathwise account is the same in every case.

Value-of-information (VOI) analysis is usually conducted under a single probability measure. However, in practice, the available evidence often pins the measure down only to a set. Consequently, under a set of probability measures, VOI requires different formulations. First, we explicate a rule-specific VOI that fixes a decision rule for acting under imprecision (such as Gamma-maximin) and measures what the information is worth to a decision maker who uses that rule. Second, we derive a fixed-measure envelope that evaluates the classical VOI functional over all admissible precise measures. We formalize this distinction and explicate its consequences for the expected perfect, partial, and sample information. The expected value of perfect information is concave over the credal set. Hence, when the set is generated by finitely many measures, its lower envelope endpoint is obtained exactly from the generators, while its upper endpoint may be interior and is computed by a finite linear program. The Gamma-maximin value, in contrast, can exceed the entire envelope, so a rule-specific value is not recovered from the envelope's endpoints. A continuity bound limits how much the VOI can change as the measure varies, and we identify when the partial- and sample-information endpoints can still be obtained from the generators. Because the single-measure VOI must itself be estimated, the procedure we give combines standard estimators for it with a search over the credal set. By using a worked decision problem, we show how the two quantities separate conclusions that hold across every admissible measure from conclusions that depend on one unidentified choice of measure.

Introduction: Objective neuroimaging biomarkers may improve Parkinson's disease motor assessment by capturing brain variation not directly observable from clinical examination. We used interpretable machine learning to predict current motor severity, measured by MDS-UPDRS Part III, from QSM and multiband multi-echo resting-state fMRI-derived ReHo features. Methods: Regional QSM and ReHo features were extracted from 28 participants, including 24 individuals with Parkinson's disease and 4 controls. Thirteen feature-set experiments evaluated imaging-only, clinical-only, imaging-plus-clinical, full, reduced, and multimodal inputs. Support vector regression, Elastic Net, Random Forest, and XGBoost models were trained using nested cross-validation. Performance was assessed using pooled held-out R^2, RMSE, MAE, Pearson correlation, permutation testing, and the proportion of participants predicted within +/-5 MDS-UPDRS Part III points. Results: Imaging-only models carried meaningful predictive signal, whereas the clinical-only model performed weakly. Full fMRI, full QSM, and clinical variables provided the strongest global fit, explaining 45.4% of variance in motor severity. Selected QSM plus clinical variables produced the most clinically close predictions, with 75.0% of participants predicted within +/-5 points and the lowest MAE among top-performing models. SHAP highlighted cerebellar, thalamic, striatal, insular, and motor cortical features. Conclusion: QSM and multiband multi-echo fMRI-derived ReHo capture distinct, interpretable dimensions of Parkinson's disease motor severity. These findings show that structural and functional imaging contribute differently depending on the clinical prediction goal.

Truck-drone delivery is an emerging last-mile logistics mode combining the long-haul capacity of trucks with the flexible service capability of drones. In locker-based operations, smart lockers serve not only as temporary parcel storage facilities but also as automated drone docking and service nodes. These automated nodes support drone takeoff, landing, parcel handover, and battery replacement, thereby significantly extending the service range and operational flexibility of drone-assisted delivery networks. However, practical locker-based delivery systems face complex real-world challenges, requiring the integrated coordination of not only parcel delivery, return pickup, battery-constrained and load-dependent drone flights, but also necessary detours around restricted airspace. To address this practical and multifaceted challenge, this paper introduces a locker-based truck-drone routing problem with integrated considerations of pickups, deliveries, and no-fly zones (LTDRP-PDNF), with the objective of minimizing the total operational cost of a fleet of drone-equipped trucks. We formulate the route construction process as a Markov Decision Process and develop a two-stage deep reinforcement learning-based neural heuristic. The first stage utilizes an attention-based encoder and a Bidirectional Gated Recurrent Unit decoder to solve the truck-only routing problem, formulated as a capacitated vehicle routing problem. The second stage combines a policy-transfer strategy with a hybrid dispatch assignment heuristic to construct fully coordinated truck and drone routes for LTDRP-PDNF. Experiments on instances of different scales demonstrate that the proposed method outperforms metaheuristic and neural heuristic baselines in most cases while maintaining exceptionally short computation times, offering an effective, scalable solution framework under practical operational constraints.

Deploying spiking neural networks (SNNs) on neuromorphic hardware demands aggressive synaptic pruning while preserving temporal computation integrity. Existing strategies either neglect neuronal criticality or rely on convex relaxations of the inherently combinatorial pruning problem whose fractional masks, upon binarisation, destroy accuracy at moderate-to-high sparsity. We present Criticality-Constrained Quadratic Pruning (CQP), a native PyTorch pipeline that fuses weight magnitude with surrogate-gradient criticality into an analytically exact importance metric, eliminating the rounding artefacts endemic to solver-based approaches. We formally characterise a continuous-relaxation trap wherein OSQP-solver fractional masks overshoot the intended sparsity by up to 12 percentage points (pp), precipitating a 44 pp accuracy collapse. We identify and remediate a zombie-weight failure mode in which Adam's first-moment tensors resurrect pruned synapses, violating the binary sparsity guarantee. An iterative schedule - prune, fine-tune with gradient masking, recompute criticality, and repeat - eliminates gradient staleness at high sparsity. A KL-divergence temporal analysis identifies a redundant simulation timestep, enabling a free 10% theoretical energy reduction without weight modification. On MNIST (60,000 training examples), CQP yields 95.6% accuracy at 90% sparsity versus 93.4% for magnitude pruning (+2.2 pp). A criticality-threshold sweep reveals an empirical criticality cliff: accuracy falls from 87.0% to 14.4% as the threshold reaches tau = 0.9, constituting a quantitative SNN-level analogue of the Critical Brain Hypothesis. Combined weight sparsification and temporal truncation yield a compound 73% reduction in per-inference energy at 70% sparsity, confirming the practical value of the proposed pipeline for neuromorphic deployment.

Early detection of dementia through speech analysis offers a non-invasive screening alternative, but capturing both acoustic and linguistic biomarkers remains challenging. We propose a multimodal framework leveraging Whisper for dual-purpose extraction: acoustic representations from encoder outputs and transcripts via automatic speech recognition (ASR). For the acoustic pathway, temporal networks with attention pooling aggregate variable-length sequences into fixed-dimensional embeddings. For the linguistic pathway, we prompt a large language model (LLM) to extract interpretable features spanning lexical diversity, syntactic complexity, semantic coherence, and discourse patterns. A gated fusion network integrates both modalities. On ADReSS and ADReSSo, our method achieves F1-scores of 89.47% and 90.14%, demonstrating effective integration of acoustic and LLM-augmented linguistic features. Ablation shows that multimodal fusion consistently outperforms either modality alone.

Deep Neural Network (DNN) classifiers suffer from poor calibration when their softmax outputs (predictive confidence) deviate from the empirical likelihoods. This manifests itself as either overconfident incorrect predictions or under-confident correct predictions. Label smoothing (LS) enhances model calibration by introducing entropy regularization during training through redistributing probability mass from the ground-truth label to the remaining classes. LS, including Margin-based LS (MbLS), have restrictive assumptions: they rely on predefined, uniform smoothing rules and only tackle overconfidence. In reality, samples exhibit diverse characteristics, such as difficulty/ambiguity, that interact with the evolving nature of the model being trained. In training, samples may have various degrees of under- or overconfidence. To overcome this, a mechanism that identifies the specific confidence state of each sample and determines the appropriate degree of smoothing in each training step is needed, tailoring the adjustment to the individual sample. We propose FedLAS: Feature-Modulated Bidirectional Label Smoothing, a plug-and-play algorithm for label smoothing-based losses. In FedLAS, we introduce a Feature Norm-based Confidence Indicator (NCI) to control smoothing and a Bidirectional Calibration Gating (BCG) module to detect both over and under-confidence. Our algorithm can be integrated with LS and MbLS based losses when applied to standard DNNs, enhancing performance. Extensive experiments on standard and fine-grained high-resolution vision benchmarks show that FedLAS consistently improves calibration compared to modern baselines, reducing Expected Calibration Error (ECE) and Adaptive ECE while maintaining Top-1 accuracy. Code: github.com/nadarasarbahavan/FEDLAS

Online high-dimensional regression requires algorithms that can update sequentially while preserving structural sparsity. We propose Adaptive Iterative Hard Thresholding (AIHT), an online sparse-regression framework that alternates stochastic subgradient updates with adaptively scheduled hard-thresholding steps. The key idea is to separate support discovery from local refinement: early in the learning process, AIHT delays thresholding so that weak but informative coordinates have time to accumulate signal, while later it increases the projection frequency to stabilize the sparse estimator and exploit local curvature. We develop the theory for high-dimensional online quantile regression, a challenging setting in which the loss is nonsmooth and the data may exhibit heterogeneity or heavy-tailed noise. Under restricted curvature and gradient-leakage conditions, AIHT remains in an inflated sparse cone, exhibits a two-phase convergence behavior, and attains logarithmic regret for the sliding-window objective. Simulations for online quantile regression, together with threshold-scheduling ablations, support the proposed mechanism and illustrate its advantage over standard online sparse-learning baselines.

Parameter settings in evolutionary algorithms and metaheuristics are important because such parameter values can influence the performance of algorithms under evaluation. For a given algorithm, there are many different numerical experiments to show that the algorithm can work well in practice; however, in most cases there is no theoretical analysis of parameter settings. In this work, we show that theoretical analysis using the theory of dynamical systems and evolution of population variance can give some good results in terms of parameter ranges for the bat algorithm. We also show that results from numerical experiments are consistent with theoretical bounds. Such analyses can provide good insights from different perspectives about the algorithmic characteristics such as variance evolution, transition between exploration and exploitation as well as convergence behaviour.

Localized generative editing needs localized evaluation: full-image identity metrics are structurally confounded under hard-composited edits. We present Envisage, a FLUX.1-Fill inpainting reference pipeline for rhinoplasty goal visualization from a single frontal photograph. The pipeline combines 8 rhinoplasty clinical presets (the released framework also includes 8 blepharoplasty and 8 rhytidectomy presets), MediaPipe masks, and hard-mask compositing. The composite preserves outside-mask pixels by construction, so full-face identity scores are dominated by copied pixels rather than by the diffusion backbone. Because full-face identity metrics cannot grade localized edits, we introduce SurgicalScore, a mask-decomposed 0-1 protocol scoring edit direction, edit magnitude, masked LPIPS, realism, and outside-mask preservation; SS_raw assigns 0.919 [0.918, 0.920] to a perfect-predictor control , anchoring the ceiling. On N=211, the paired ArcFace gain (output-to-GT minus input-to-GT) is negative for all methods (Envisage -0.048 smallest, vs. ICEdit -0.139, Kontext -0.242, InstructPix2Pix -0.294; p < 1e-4), with external validation on a 457-pair ASPS/PCA corpus showing a larger negative gap. With SurgicalScore, Envisage achieves the highest score (0.599 [0.579, 0.619]) and leads on both metrics, but the all-negative ArcFace gap shows that full-face identity is poorly aligned with localized surgical accuracy under hard compositing. A 5-seed GT-oracle (an upper bound, not a deployable result) reduces the residual ArcFace gap by 73% (-0.054 to -0.015), with positive output-to-GT gain on 33.9% of cases, indicating candidate-space headroom for a learned ranker. For localized edits, progress should be measured with edit-region fidelity rather than full-face identity metrics. We release Envisage, SurgicalScore, preset definitions, and matched split manifests.

Connected Vehicles (CVs) rely extensively on communication technologies to enable data-driven predictive analyses for enhancing performance and safety. These communication channels can be exploited by adversaries to launch cyberattacks such as Sybil attacks, which could threaten both safety-critical and mobility applications, leaving CVs vulnerable and putting human lives at risk. As CV deployment continues to expand, the need to detect and mitigate cyberattacks in real-time becomes increasingly urgent. This study presents an in-vehicle Digital Twin (DT)-based collision warning framework with built-in capabilities for Sybil attacks detection. The framework integrates a Temporal Convolutional Network (TCN) for learning temporal dependencies in vehicle trajectory data and Hierarchical Navigable Small World (HNSW) algorithms for efficient similarity-based classification. Our framework is evaluated on real-world Sybil attack data, collected through field experiments. The framework achieved accuracy, recall, and F1 scores of 0.984, 1.00, and 0.944, respectively, in detecting Sybil-generated fake vehicles. During the safety evaluation, the framework reduced the mean Time Exposed Time-To-Collision (TET) and mean Time Integrated Time-To-Collision (TIT) of near-collision events by 88% and 72%, respectively. Furthermore, real-world feasibility evaluation shows that the framework conformed to the standardized maximum allowable latency for safety applications and operated well within the capacity of modern processors – demonstrating the promise of an in-vehicle DT-based framework as an attack mitigation mechanism against Sybil attacks for next-generation CVs.

Effective sub-typing (also known as grouping or clustering) of patients using their electronic health record (EHR) data can greatly inform precision medicine efforts. However, subtyping temporal EHR datasets is known to be challenging due to inherent EHR issues, including complexity and irregularity. In this study, we propose a self-supervised Mamba-based model that learns effective EHR representations and enables enhanced patient subtyping. We evaluate the proposed model on public and private real-world EHR datasets to classify the data based on the available labels and subtype patients based on the representations learned from the model. Through an extensive set of experiments, we demonstrate that our model's design choices lead to better performance compared to competitive baseline models for prediction. Moreover, we evaluate several clustering techniques to demonstrate that our findings offer valuable insights into subtyping patients based on temporal records from EHR models\footnote{Our implementations are available at https://github.com/healthylaife/triplet_mamba.

Fayyazi et al. (2025) recently proposed FACTER, a model-agnostic framework designed to jointly enforce fairness and statistical coverage in LLM-based recommendation through conformal thresholding and iterative prompt repair. In this work, we conduct a reproducibility study of the FACTER framework across diverse architectures and dataset sparsity levels, evaluating both the original open-ended generation task and a constrained re-ranking extension. Under the strict reproduction, we observe a divergence in recommendation utility, which we trace to underspecified target-set evaluation in the original study. We then use the constrained re-ranking setting to evaluate FACTER when the candidate set is fixed, and introduce a static Fair Zero-Shot baseline to isolate the contribution of the iterative prompt repair loop. Our analysis shows that FACTER consistently reduces adaptive-threshold violation counts, but that these reductions are not consistently reflected under the fixed threshold or in global fairness metrics. In the constrained ranking setting, static fairness instructions achieve comparable semantic-parity outcomes to FACTER's dynamic repair loop, suggesting that the additional online repair mechanism provides limited benefit in this formulation. All code and reproduction artifacts are available at https://github.com/oscar-omlf/facter-repr.

In stochastic linear bandits, the canonical Upper Confidence Bound (UCB) algorithm admits a simple frequentist regret analysis but can be computationally demanding, while Thompson Sampling (TS) is computationally attractive yet typically harder to analyze due to its non-optimistic nature. We propose Absolute Thompson Sampling (ATS), a simple modification of TS that ensures optimism in expectation by replacing the signed exploration noise with its absolute value. This preserves the computational efficiency of TS while avoiding the technically involved anti-concentration arguments common in TS analyses, enabling a simple UCB-style regret analysis. We show that ATS achieves $\tilde{O}(d^{3/2}\sqrt{K})$ regret, matching existing bounds for TS in linear bandits. We further introduce Ensemble Absolute Thompson Sampling (EATS), which takes the maximum over multiple absolute perturbations with normalization by the ensemble size. As the ensemble size grows, EATS converges to the UCB objective, recovering UCB behavior in the limit. Experiments show that moderate ensemble sizes already yield strong performance. Our results point to a bridge between randomized exploration and deterministic optimism both in theory and practice.

Large language models (LLMs) are increasingly deployed in high-stakes domains, where free-text explanations such as chain-of-thought and post-hoc rationales are used to justify model outputs. Yet it remains unclear whether these explanations are sufficient, i.e., if they contain enough information to explain the model's output-generating process. We generalize classical sufficiency from feature attributions to arbitrary explanations and prove that explanation sufficiency can change depending on the input distribution, which must be explicitly defined for LLM explanations. We propose using the LLM itself to generate alternative inputs conditioned on an explanation, capturing its beliefs about possible inputs. We formalize self-consistent sufficiency as a goal for free-text explanations and introduce an information-theoretic metric, SCSuff, that enables evaluation of free-text explanations without relying on predefined biases or shortcuts. Our experiments show that SCSuff agrees with targeted perturbation tests where applicable and demonstrate that explanation sufficiency can vary with the input distribution. We find LLM explanations are generally insufficient and weakly correlated with model size, accuracy, or output entropy. Analysis of final-token hidden states shows that top and bottom SCSuff scores can be predicted from internal representations, suggesting that SCSuff can guide detection and improvement of sufficient LLM explanations. The code for this paper is available at https://github.com/rajesh-lab/self-consistent-sufficiency .

Neuromorphic and edge computing research has focused on reducing the inference cost of neural network controllers, yet in physical closed-loop systems the actuator can rival or exceed an efficient controller in energy. An efficient controller is therefore necessary but not sufficient, because the actuator becomes the cost worth reducing once inference no longer dominates it. Here, we introduce energy-aware learning, an approach that incorporates actuator energy directly into the reinforcement learning reward, and demonstrate it in closed-loop deep brain stimulation (DBS) for Parkinson's disease. A deep spiking Q-network, trained in a biophysical cortico-basal ganglia-thalamic circuit model, learns to suppress pathological alpha-beta oscillations by 45.2% while reducing stimulation charge by 80.0% relative to continuous DBS. Sparsity-constrained knowledge distillation compresses the policy onto the SynSense XyloAudio 3 neuromorphic processor at 0.52 mW inference power, yielding 28.1x lower energy per inference than an equivalent artificial neural network on conventional edge hardware. By co-optimizing stimulation energy and inference efficiency, the framework addresses both major power demands in implantable neuromodulation.

Prediction sets should have high coverage to be useful, but some coverage notions are more practically relevant than others. In the classification setting, class-conditional coverage requires that the prediction set (i.e., the set of candidate labels for a new test point) must achieve the target accuracy level within each class, which may be challenging to satisfy when many classes are rare and have few calibration points. At the other extreme, marginal coverage requires only that coverage holds on average over the distribution of all classes, which can lead to low-probability labels being essentially ignored. To find a middle ground, recent work has introduced macro-coverage, defined as the unweighted average of class-conditional coverages. Macro-coverage offers a compromise between marginal coverage and class-conditional coverage that is particularly appropriate for long-tailed settings. In this work, we show that label-weighted conformal prediction can be used to produce prediction sets with a finite-sample macro-coverage guarantee, and more generally a guarantee on a family of generalized macro-coverage objectives that aggregate coverage at the level of arbitrary class groupings and take a weighted average. We further characterize the form of the smallest prediction sets satisfying a given generalized macro-coverage objective and propose a corresponding conformal score function. We validate our theoretical results on two large-scale image classification datasets.

Modern machine learning models are trained by optimizing high-dimensional non-convex empirical risk functions. Such cost functions can have a multitude of local optima and yet, gradient-based optimization appears to converge to near-global optima. Within a simple supervised learning setting, we develop a precise picture of which parts of the empirical risk landscape are accessible by polynomial-time algorithms. We are given i.i.d. pairs $\{(\boldsymbol{x}_i,y_i):\; 1 \le i\le n\}$ with $\boldsymbol{x}_i\in \mathbb{R}^d$ standard Gaussian feature vectors, and $y_i\in\mathbb{R}$ response variables that depend on $\boldsymbol{x}_i$ through their projections on an unknown $k$-dimensional subspace. We use empirical risk minimization to learn a model that depends on an $m$-dimensional projection of the data (e.g., an $m$-neurons neural network). We propose an incremental approximate message passing (IAMP) algorithm and precisely characterize the training error it achieves, as well as the relation between test and training error, in the high dimensional asymptotics $n,d\to\infty$, with $n/d\to\alpha \in (0, +\infty)$. Based on earlier work in related models, we expect that the performance achieved by our algorithm is optimal among polynomial-time algorithms.

The axiom of choice has divided the foundations of mathematics for over a century, but the distinction between classical and constructive proofs has remained a philosophical and methodological one. We use Lean 4's kernel-level tracking of axiom dependence to show that the axiom of choice has a measurable geometric correlate in proof space that obeys a one-parameter mixture law and has operational consequences for neural theorem provers. To do this, we partition $471{,}260$ declarations of Mathlib by transitive dependence on the axiom of choice and represent a filtered population of $42{,}355$ traced theorems by their sequences of tactic invocations. We use the constructive proofs in this dataset to train a self-supervised proof encoder and show that when using it to measure classical proofs, three complementary measurements (anomaly score, reconstruction loss, and density-superlevel containment) exhibit a common decline with the proof's distance from the axiom in the dependency graph, from sharp separation at the shallow boundary (AUC $0.847$ at distance $2$) to indistinguishability at distance~$9{+}$. Robustness controls show that the signature survives length, file, author, and topic controls, and replicates under full-source encoders trained on normalised proof source. Operationally, we show that on an evaluation sample of $251$ Mathlib theorems, Lean's aesop tactic solves constructive theorems at $13\times$ the rate of classical ones, and a neural-guided hybrid using the ReProver tactic generator compresses the gap to $5\times$. The geometric anomaly score predicts aesop failure beyond proof length, providing an operational link between the geometric signature and prover performance.

Speech-driven 3D facial animation methods face significant challenges in simultaneously achieving high-fidelity motion and precise artistic control at production quality. Existing controllable models typically learn global style control by relying on large-scale, low-quality in-the-wild datasets that compromise overall animation realism. Furthermore, these frameworks often lack the fine-grained temporal precision required for demanding tasks such as dialogue localization (e.g., dubbing), where matching specific facial expressions is as critical as lip synchronization. We present KM-Speaker (Keypoint-Matching Speaker), a novel keypoint-conditioned flow-based generative framework that provides both global style guidance and frame-level temporal control from reference performances. We propose a disentanglement strategy that separates audio-driven lip motion from keypoint-driven upper-face dynamics, together with a global style context preservation mechanism to ensure coherent full-face expressiveness. KM-Speaker advances example-based 3D facial animation by achieving high-fidelity motion and flexible controllability in a data-constrained setting, consistently outperforming state-of-the-art methods in lip-sync accuracy, style adherence, and expressive temporal control.

We study sparse self-attention in which each query attends to a dense local window plus a set of Fibonacci-spaced offsets, with a per-layer scalar alpha that compresses or expands the spacing. Across 21 language models trained under one matched recipe (60M parameters, 512 hidden, 16 layers, 426M tokens), we compare four ways of setting alpha across depth: fixed, per-layer learned, a static linear stagger, and a coprime (anti-gridding) reassignment of that stagger, together with a reach-matched power-of-2 control. Three results stand out. First, a static per-layer stagger improves perplexity over both fixed and learned alpha, and the gain is base-agnostic: applying the same stagger to a power-of-2 base lifts it above fixed Fibonacci and to parity with learned Fibonacci attention. Second, learning per layer is inert: it does not beat the static schedule and costs roughly five times the inference latency. Third, and most consequential, all sparse variants extrapolate to four times their training length with little or no degradation, whereas a recipe-matched dense baseline collapses (perplexity rises by 201% at 4x length); we attribute this to fixed-offset attention only ever querying relative positions seen during training. We also report two honest negatives: at training length the best sparse model has about 26% higher perplexity than the dense baseline, and the staggering gain is uniform across context positions rather than concentrated at long range.

In many real-world applications, such as retail sales, energy usage, and supply chain planning, forecasting is performed across hierarchical structures. These structures often represent aggregations (e.g., products to categories to regions), where forecasts must not only be accurate but also coherent, meaning that lower-level predictions sum correctly to higher-level forecasts. Traditional statistical methods, such as Bottom-Up and MinT, enforce coherence through post-processing but fail to model complex nonlinear temporal dependencies and covariate interactions. We propose Hierarchical Temporal Fusion (HTF), a novel extension of the Temporal Fusion Transformer (TFT) that integrates structured hierarchical embeddings with a coherence-aware loss function to ensure consistent forecasts across all levels of a hierarchy. Rather than applying reconciliation after forecasting, HTF embeds coherence directly into the training objective. The coherence loss penalizes the difference between aggregated child forecasts and their corresponding parent forecasts during training, enabling the model to learn both temporal dynamics and structural consistency simultaneously. We evaluate HTF on two publicly available benchmark datasets: the M5 Walmart forecasting dataset and a publicly available hierarchical energy consumption dataset. Results demonstrate that HTF substantially reduces forecast incoherence while improving forecasting accuracy compared with classical reconciliation methods and deep learning baselines. In addition, attention visualization and embedding analysis provide insight into how temporal and structural information contribute to hierarchical forecasting performance.