Papers

Agent benchmarks usually measure task completion and treat resource use as an auxiliary statistic. In deployment, however, the choice among a local lookup, broad search, composite research tool, stronger model, or human escalation is part of the task itself. We introduce EcoAgent-Bench, in which every task specifies priced actions and an explicit budget. Its 304 real-derived tasks span five families adapted from GAIA, HotpotQA, and MuSiQue, and test four decisions: avoiding unnecessary escalation, escalating when local evidence is insufficient, selecting a model tier, and stopping on unsupported premises. We evaluate seven LLM agents in tool-API and workspace-CLI settings, together with four oracle scripted controls. Micro-averaged accuracy rewards one-sided policies: always-escalate controls achieve high micro success while failing save-oriented tasks. We therefore also report an economic-consistency score (the worse of accuracy on upgrade-oriented and save-oriented family groups) which exposes this failure. Tool-API agents attain only 3.9-24.0% micro strict success (at most 7.3% economic consistency), often either stopping before warranted escalation or overspending on cheap tasks. A threshold-crossing budget sweep changes GPT-5.4's escalation rate from 0% to only 3%. These results show that completion under a budget and economical action selection are distinct properties. We release the task bundle, transformation pipeline, frozen evaluation environments, and integrity-bound result artifacts needed to study both.

In this paper, we consider a class of multiblock nonconvex nonsmooth optimization problems, which covers many applications such as the analysis of pre-earthquake anomalies and machine learning. To solve this class of problems, we propose the inertial block proximal linearized method with two-phase adaptive momentum (IBPL$^+$-TP). Compared to the current methods, our method possesses three main advantages: (1) it introduces a two-phase adaptive momentum strategy to effectively update the extrapolation parameters, (2) it allows using two different extrapolation points to accelerate the convergence, (3) it allows the extrapolation parameters of these two extrapolation points to be independent of and unconstrained by all other parameters. While maintaining the above advantages, we prove that our method ensures the monotonic convergence of the objective function of this class of problems, and we also prove that the sequence generated by our method globally converges to a critical point, as well as establish the convergence rate of our method. To demonstrate the effectiveness of our method, we apply it to solve two nonconvex and nonsmooth machine learning problems, namely sparse nonnegative matrix factorization with $\ell_0$-constraints and sparse nonnegative CP decomposition with $\ell_0$-constraints. The numerical experimental results on solving these problems show that our method outperforms several state-of-the-art methods.

Modern deep neural networks achieve strong performance, but their scale makes them costly and slow, especially on resource-constrained edge devices. Pruning and quantization address this, but rely on manual, expert choices and on algorithms that are hard to apply across architectures. Uniform settings also ignore how differently individual layers respond to compression, which costs accuracy. We introduce APQF, an agentic profiling-guided framework that combines structured pruning, mixed-precision quantization-aware training, and accuracy recovery in one automated pipeline. A profiling agent measures how cost is distributed across the model and how sensitive each part is to pruning, and this evidence drives per-layer pruning ratios, per-layer bit-widths, and the recovery strategy, all proposed by LLM planners and validated before execution. To our knowledge, APQF is the first framework to combine LLM-guided, profiling-grounded decisions with a fully training-aware pruning and quantization pipeline for both CNNs and vision transformers. We evaluate APQF on ResNet, VGG7, ViT, DeiT, and Swin using ImageNet-1k and CIFAR-10. On ImageNet it cuts compute to 5.6-7.7 percent of the original bit-operations, a 13-18x reduction, while keeping accuracy close to the baseline, and under a 200K-image budget it stays roughly 17 points higher in Top-1 than existing joint pruning and quantization methods. On CIFAR-10 it compresses further than that method on four of five architectures. On VGG7 it reaches 93.15 percent using only 0.41 percent of baseline bit-operations, the only method at that compression level to improve on its full-precision baseline. Ablations show that uniform compression loses the most accuracy at matched compute, and that withholding profiling data from the planner hurts every model. Six LLM planners, including free open-weight ones, all reach 97.4-97.9 percent on Swin-Tiny.

Autonomous agents now carry out entire data analyses, selecting cohorts, joining tables, and fitting models with little step-by-step supervision. When such an analysis turns out to be wrong, someone must determine which operation caused it. A recent approach does this without any labelled mistakes, learning instead from analyses known to be sound and flagging operations that depart from what that model predicts; how reliable such audits are has not been studied. This paper supplies that analysis. The choice of score determines whether an error can be localized at all. If each operation is scored by how surprising it is given the operation immediately preceding it, then operations that merely inherit an earlier error are indistinguishable from correct ones, so one mistake produces one flag; scores computed against a longer reconstruction of the intended analysis instead spread a single mistake across many operations. We quantify how far they spread, and how to choose the comparison length when an error accumulates gradually rather than at once. We then give procedures that control the proportion of falsely flagged operations within a single audited analysis, requiring only that sound analyses be exchangeable rather than that the fitted model be correct, and we quantify how much the guarantees weaken when the model is imperfect or when the analysis was selected for review in a way that depends on its content. Finally we establish a limit on what any such audit can report: errors below a certain magnitude cannot be attributed at all, being indistinguishable from ordinary variation among sound analyses. This limit falls so slowly as more sound analyses are collected that at the representation sizes now in use a hundredfold increase reduces it by under two percent, so the dimension of the representation rather than the volume of training data is the binding constraint.

Objective: Emergency department (ED) overcrowding threatens care quality, yet Full Capacity Protocols (FCPs) are typically activated only after crowding thresholds have been exceeded. This study evaluated whether machine learning-guided proactive FCP activation improves ED crowding control compared with conventional reactive activation. Materials and Methods: We developed and validated an integrated machine learning and discrete-event simulation framework representing ED patient flow and capacity management. An embedded machine learning model predicted hourly waiting-room crowding six hours ahead and triggered proactive FCP activation. Using empirically calibrated operational parameters from a large US hospital, we compared proactive and reactive deployment strategies over a 365-day simulation horizon with 10 independent replications. Outcomes included ED length of stay, wait time, percentage of time above the crowding threshold, number of FCP activations, and protocol duration. Results: Proactive deployment produced statistically significant reductions in the percentage of time above the crowding threshold and average protocol duration across all threshold settings. It also reduced crowding variability and extreme congestion events, resulting in smoother ED operations. Although proactive deployment generally improved ED length of stay and wait time, these differences were not statistically significant. Conclusion: Embedding machine learning predictions within an operational simulation enables proactive, prediction-guided capacity management before severe congestion develops. These findings support the feasibility and potential value of proactive FCP deployment as a decision-support strategy for improving ED crowding control and provide evidence to support prospective implementation in clinical practice.

Electroencephalography (EEG) is a low-cost and non-invasive signal source for dementia screening, yet existing EEG-based studies remain difficult to compare because preprocessing, EEG segmentation, feature design, classifier choice, and validation protocols vary across studies and are often evaluated in isolation. This variability limits the derivation of robust pipeline recommendations. This paper presents a leakage-safe empirical benchmark for resting-state EEG-based dementia classification and uses it to identify a practical best-practice pipeline. Using the public OpenNeuro ds004504 dataset, the benchmark evaluates artifact correction, fixed-length EEG segmentation, training-only augmentation, multi-domain feature extraction, fold-internal feature selection, classical machine-learning (ML) classifiers, subject-level aggregation, and interpretation under leave-one-subject-out (LOSO) validation. The best-performing pipeline in this benchmark combines Artifact Subspace Reconstruction (ASR) followed by Independent Component Analysis (ICA), 10 s EEG epochs with training-only amplitude scaling and Gaussian-noise augmentation, spectral, complexity, and pairwise connectivity features, mutual-information top-100 selection, linear support vector machine classification, and mean-probability aggregation. It achieves 87.69 percent accuracy, 88.89 percent F1-score, and 91.20 percent AUC on Alzheimer disease (AD) versus cognitively normal controls (CN) classification while retaining 100 features from 1596 raw descriptors. The results support compact multi-domain EEG features as an interpretable and leakage-safe baseline for subject-level dementia classification.

While modern large reasoning models (LRMs) excel at providing correct answers in many tasks, we provide additional evidence that they often struggle with a critical capability: knowing when to abstain from answering. We analyze this gap by comparing LRM behavior to results from an exploratory human study, revealing that human reasoning effort on unanswerable tasks isn't different from answerable tasks, whereas LRMs waste computational resources by generating longer Chains of Thought (CoTs) on unanswerable than on answerable prompts. To overcome this inefficiency, we take inspiration from a resource-rational perspective on human cognition and introduce a novel GRPO reward that encourages efficient reasoning about whether the task contains all the information needed to solve it. Fine-tuning several 4B and an 8B LRM with this reward leads to human-like abstention performance gains (+14.7% on average) while retaining answering capabilities and boosting the models' efficiency (44% shorter CoTs on average).

A decoder that turns video or audio into text conventionally consumes the entire input before emitting a word. Offline this is merely more than the task requires; live it is impossible, since a caption cannot wait for a match to end. Streaming systems bolt on a fixed rule such as wait-$k$, which waits for the same number of input tokens before every word, regardless of the input's length or pace. We replace the fixed offset with ZENDAYA, a schedule governed by a single continuous parameter $\gamma$. It makes the visible source prefix a closed-form function of generation progress, scaled by the input's own predicted length, so an ordinary offline decoder and a real-time streaming decoder become two endpoints of one family rather than separate models. The same scalar fixes, in closed form, the mean fraction of source consumed per emitted word, $\bar{E}(\gamma) \approx 1/(1+\gamma)$, making it at once a latency dial and an interpretable budget. We prove a structural dependency theorem: under any schedule fixed in advance and non-decreasing, no emitted token can depend on input that has not yet arrived. The guarantee holds for trained and untrained weights alike, and extends to unbounded streams under arbitrary asynchronous arrival. The empirical result is counterintuitive: seeing less can produce better text, because a flood of source dilutes attention exactly when the model has the least of its own output to anchor on. Trained from scratch across two modalities and three public corpora (Charades-STA, ActivityNet Captions, LibriHeavy), a compact 29M-parameter decoder matches or beats the fixed schedule while reading less of the source, with the sharpest gains at the lowest latencies, where a fixed offset collapses. Streaming METEOR gains are statistically significant on all three corpora.

Oligonucleotide melting temperature is a fundamental determinant of nucleic acid hybridization and underpins the design of molecular diagnostics, polymerase chain reaction assays, and many other biotechnology applications. However, accurately predicting melting behavior remains difficult because it depends not only on sequence composition, but also on experimental conditions and chemical modifications commonly used in modern assay design. Existing thermodynamic models rely on fixed parameterizations that are often difficult to extend across diverse reaction environments and nucleotide chemistries. Here we show that a condition-aware nucleotide language model can accurately predict oligonucleotide melting behavior across diverse experimental conditions and both unmodified and chemically modified oligonucleotides. By combining contextual sequence representations with explicit information describing the reaction environment, the framework achieves sub-degree prediction accuracy and reduces prediction error for locked nucleic acid-modified oligonucleotides by up to 25% relative to nearest-neighbor thermodynamic approaches. The model also more accurately captures the thermal effects introduced by nucleotide modification and maintains strong performance on independent benchmark datasets spanning experimental conditions substantially different from those represented during training. Our results demonstrate that learned sequence representations can complement classical thermodynamic models by capturing context-dependent effects that are difficult to encode using fixed parameter tables alone. More broadly, this work provides a scalable framework for predicting oligonucleotide melting behavior across diverse chemistries and assay conditions, supporting more reliable molecular assay design.

Large language model agents can discover alphas, yet current methods have three weaknesses. The search cannot adapt during the run, automation usually ends at alpha generation while library selection and model choice stay manual, and alpha discovery can read the test window through loop feedback or code problems. We present AutoScientist-Quant, a self evolving search process that regards quantitative research as one budgeted search problem. A single controller conditions every decision on the remaining budget, choosing at each round whether to improve, combine, pivot, or stop, which node to expand, how many alphas to generate, and how to retrieve past trajectories from the shared memory. The same core then selects from the library and tunes the model, closing the loop from hypothesis to deployable strategy. We also review the evaluation pipeline reused from prior work, fix two lookahead problems, and keep the feedback window disjoint from the held out test window, so every comparison tests true generalization. On CSI universes, the framework attains the best value of nearly every metric in every setting, and these conclusions hold across several backbones and markets.

Supply chain forecasting systems increasingly operate under market shocks, non-identically distributed regional demand, and limited willingness to centralize commercial data. This work proposes Federated Ensemble Forecasting with Negative-Correlation Learning (FEF NCL), a distributed method that trains specialized forecasting experts across client nodes while discouraging redundant model errors. The framework combines temporal feature encoders, client level drift scoring, reliability-weighted aggregation, and an explain ability layer that exposes the market and supplier variables most responsible for each forecast. A single synthetic dataset is used to evaluate the design. It contains 124,800 weekly SKU region observations from ten regional client nodes, 60 product families, 40 suppliers, five commodity groups, and a 2021-2024 volatility profile with explicit price-shock regimes. Because the dataset is synthetic, the reported results should be interpreted as controlled evidence of internal consistency rather than real-world validation. Across the synthetic test split, FEF NCL reduces weighted mean absolute percentage error from 13.9% for the best federated baseline to 12.4%, improves delay-risk macro-F1 from 0.755 to 0.801, and lowers the high volatility quintile error by 2.1 percentage points relative to SCAFFOLD. The analysis suggests that negative-correlation specialization is useful when clients face different supplier, freight, and commodity conditions, although deployment would require stronger privacy analysis, live drift monitoring, and operational calibration. Index Terms federated learning, ensemble learning, negative correlation learning, supply chain forecasting, market volatility, data drift, demand planning, risk governance

A trained reinforcement learning policy does not determine the complete behavior that users encounter: deployment code still schedules, admits, suppresses, or replaces its proposed actions. We contribute runtime action interference (RAI), an AI control mechanism that preserves policy parameters while regulating action pacing and filtering configured action patterns after inference. RAI releases a proposed action only when its cooldown condition is satisfied and its content detector does not flag the action; otherwise, it dispatches a no-op. The detector covers specified toxic behaviors, including worker-unit harassment, while the cooldown controls action rate. We implement RAI in a replication of AlphaStar actor.py and make the implementation and reproducibility materials available through an open source code repository. We deployed RAI in a StarCraft~II human participant study that compared two presentations of the same opponent with high capability and rate limited actions; we withheld its capability claim in one presentation and disclosed it in the other. On response scales from 1 to 5, we observed pooled fairness, trust, and toxicity means of 3.90, 3.50, and 2.00 under claim withholding, compared with 2.62, 4.31, and 2.85 under disclosure. Disclosure corresponded with lower perceived fairness and higher perceived toxicity across every expertise group, whereas trust increased among novices and experts but decreased among intermediate participants. Our human evaluation therefore shows that perceptions of an opponent controlled through RAI can vary substantially with the capability information presented to users, even when the configured control remains constant. We conclude that human-computer evaluations must separate control within the execution stack from capability disclosure and assess fairness, trust, and toxicity as distinct dimensions of human experience.

Predicting underwater acoustic transmission loss rapidly and accurately is crucial for real-time ocean acoustic applications. While Fourier Neural Operators (FNO) have emerged as powerful surrogate models due to their global receptive fields, they suffer from spectral bias. The frequency truncation mechanism in FNO filters out high-frequency components, resulting in over-smoothed predictions that fail to capture fine-grained interference patterns. To overcome this limitation, this paper proposes a Spectral-Spatial Residual Learning (S2RL) framework. S2RL decomposes the prediction task into a coarse-to-fine process: a spectral Global Propagator first generates a globally consistent prediction, and a spatial Local Refiner subsequently recovers the high-frequency residuals. Experimental results on a South China Sea dataset show that the proposed method significantly outperforms FNO baselines while maintaining millisecond-level inference speeds.

Multi-layer locate-then-edit methods for knowledge editing first optimize target residual-stream activations (anchors) at selected layers, then realize them layer by layer as weight updates. This pipeline optimizes an intermediate representation but deploys multi-layer weight updates whose joint effect through the true forward pass is never itself optimized: regardless of how anchors are set or propagated, each update comes from a local solve, so propagation-induced attenuation and distortion go uncorrected, leaving a closure gap between anchor targets and realized edits. We propose DOW-KE, an anchor-free method built on a single principle: what is optimized must be exactly what is deployed. DOW-KE backpropagates the final editing objective through the complete model, jointly optimizing the updates of all edited layers so cross-layer propagation and coupling enter every gradient step. The same principle dictates where preservation resides: embedding the preservation projection in the update parameterization, inside the computation graph, makes every gradient act on the deployed update; post-hoc constraints would reopen the gap, and the constrained search keeps edits clear of protected knowledge. In large-scale sequential editing on two datasets and three models, DOW-KE achieves the highest overall Score and neighborhood Specificity in five of six model-dataset settings among the evaluated baselines.

Electroencephalography (EEG) foundation models have recently emerged as a promising paradigm for EEG decoding by learning reusable representations from large-scale heterogeneous neural recordings. However, the open release of EEG foundation encoders, while facilitating downstream developments, also introduces a previously unexplored security risk: publicly available representations may make private downstream models vulnerable. This paper investigates adversarial transfer attacks in EEG foundation model deployment in a public-encoder and private-downstream setting, where attackers have white-box access to a released encoder and a small task-matched labeled reference set, but no access or query to victim parameters, outputs, or gradients. We propose Shrinkage-Whitened Proxy Cross-Entropy (SW-ProxyCE), a query-free task-aware attack framework that recovers task-level decision geometry from a small labeled reference set through shrinkage-whitened class prototypes, enabling transferable adversarial generation without training an additional surrogate classifier. We evaluated SW-ProxyCE across three EEG tasks using three general-purpose foundation encoders and a paradigm-specific pre-trained encoder, covering both linear-probing and full-fine-tuning downstream models in cross-subject and within-subject scenarios. Results demonstrated that adversarial examples generated from the public encoder and limited labeled references can effectively transfer to inaccessible downstream models. SW-ProxyCE consistently outperformed task-agnostic representation-shift attacks, revealing that the strong transferability of EEG foundation models does not necessarily lead to adversarial robustness. Our code will be available on GitHub.

Existing multi-task learning methods rely on hard sharing, multiple paths or experts, adaptive sharing, and dynamic expansion. However, their capacity changes are usually constrained by predefined structures or triggered by task boundaries and conflict signals. This raises a fundamental question: can a network start from exact single-path computation and grow a new independent path only when persistent optimization evidence appears? We propose the Emergent Modular Atomic Network (EMAN), an optimization-driven framework for exposing an antisymmetric growth direction through latent relative phases without instantiating a second path, and for monitoring multiple decision signals during training to transform local optimization evidence into a structural decision. EMAN materializes two equal-capacity independent paths only after certification. EMAN adaptively allocates shared and task-specific representation capacity to accommodate varying task requirements. Extensive experiments on controlled rank settings, PASCAL-Context, and NYUv2 validate its effectiveness, achieving improved performance at a competitive computational cost.

Predicting 30-day hospital readmission is essential for assessing patient stability and optimizing healthcare resources. As clinical risk evolves with the accumulation of evidence during hospitalization, capturing these dynamic trajectories is essential. However, many existing approaches compress the complex longitudinal history into fixed representations, often losing the granular, day-level clinical signals that reflect a patient's evolving physiological state. To address this, we propose Mr.Dec (Multimodal Readmission-risk prediction Decoder), which models each admission as a natural chronological sequence of daily multimodal events. By leveraging a Transformer Decoder, Mr.Dec integrates daily Electronic Health Record(EHR) updates and intermittent Chest X-ray(CXR) findings in a time-aligned stream, reflecting the actual clinical workflow. To ensure robustness, we utilize Disease-Specific Supervised Contrastive Learning as an auxiliary regularization to induce a diagnosis-aware structure in the latent space. Evaluations on the MIMIC-IV and MIMIC-CXR datasets show that Mr.Dec achieves state-of-the-art performance by preserving the integrity of the clinical sequence. Furthermore, our model identifies "Critical Days" within an admission, providing actionable and clinically grounded interpretations for real-time risk stratification. Code is available at: https://github.com/yejix-ai/MR.DEC

Automatic sensitivity classification of organizational documents is a critical yet underserved problem, where the consequences of misclassification range from regulatory violations to security breaches. While AI-based approaches offer a scalable alternative to manual review, their reliability depends fundamentally on the integrity of training data. A pervasive but underreported problem in this domain is label leakage: residual classification markers embedded within document bodies that allow models to exploit surface shortcuts rather than learning genuine content-based sensitivity signals, producing performance estimates that are inflated and unreliable. This paper addresses this problem by introducing Strategic 16K, a carefully constructed, leakage-controlled corpus of 16,000 diplomatic cables sourced from the WikiLeaks Public Library of US Diplomacy (PlusD), and presents a systematic benchmark evaluating six model architectures spanning classical machine learning and transformer-based approaches. We document an extended leakage removal protocol that identifies and eliminates three categories of residual classification markers embedded within document bodies. On the clean benchmark, BERT achieves the strongest performance (Accuracy = 89.14%, F1 = 89.33%), followed by ELECTRA (Accuracy = 88.57%, F1 = 88.90%). Among classical models, TF-IDF with Logistic Regression achieves the strongest performance at significantly lower computational cost. These results constitute the first fully reproducible sensitivity classification benchmark constructed under explicit leakage-controlled conditions from WikiLeaks PlusD.

As supervised fine-tuning data continues to scale, selecting high-value subsets from large candidate pools is crucial for reducing training cost and improving model performance. Existing methods often measure diversity directly in the original embedding space, where geometric metrics entangle dominant semantic directions, fine-grained supervision differences, and local noise. We address this limitation by formulating data selection as a coarse-to-fine hierarchical coverage problem and propose MASS. MASS learns low-dimensional principal manifold coordinates with a dense autoencoder for coarse semantic grouping, and then performs quality-aware sparse feature coverage within each group using a TopK sparse autoencoder. Experiments on Vision Flan and LLaVA-CoT show that MASS consistently outperforms strong data selection baselines across multiple budgets, and in several settings matches or surpasses full data training with only a small subset of data.

Data selection in supervised fine-tuning aims to select a small set of effective samples from large-scale candidate data, reducing training cost while preserving model performance. However, existing methods usually treat data value as a relatively static property, and pay limited attention to the compatibility between data and the capability distribution of the target model. To address this issue, we propose Data-DPO, a target model-oriented SFT data selection method. Data-DPO observes the local training feedback of the target model on different samples through one-step probing, transforms activation differences among samples into pairwise data preferences, and trains a lightweight reward model to learn target-model-aware data preferences. In the final selection stage, Data-DPO further combines target model preference, external quality scores, and marginal diversity to construct a more stable and effective training subset. Experimental results on Vision-Flan and LLaVA-CoT show that Data-DPO consistently outperforms existing data selection baselines under multiple data budgets and stably surpasses full data training performance.

Knowledge distillation (KD) is widely used to transfer the capabilities of large language models (LLMs) to smaller students, but existing objectives often struggle to balance faithful imitation and robust generation. In particular, existing methods mainly combine FKL and RKL, overlooking that RKL itself provides a mechanism for adjusting the student's imitation strength. Motivated by this, we revisit on-policy Reverse Kullback-Leibler (RKL) distillation and decompose its objective into a teacher-fitting term and a student-entropy term, without introducing an explicit FKL branch. We show theoretically that the token-level optimal student distribution corresponds to a tempered variant of the teacher distribution, where the adaptive weight controls the trade-off between mode-seeking and uncertainty preservation. Guided by this insight, we propose Adaptive Entropy Distillation (AED), which uses the teacher's entropy to dynamically calibrate token-level imitation strength. Experiments on instruction-following and mathematical reasoning benchmarks demonstrate that AED achieves superior overall performance and generally improves teacher–student distributional and entropy alignment.

LLM judges increasingly evaluate responses against fine-grained rubric checklists. When a sample requires multiple rubrics, current methods typically assess each in a separate inference call. Evaluating all rubrics in a single pass is a natural alternative with greater efficiency, but we find that it introduces rubric interference: the verdict on one rubric shifts depending on which other rubrics are co-present. In a preliminary study, only one-third of samples receive fully consistent verdicts when evaluated under rubric sets of varying composition. We develop a measurement framework that probes interference through four controlled operations: rubric set expansion, subsetting, reordering, and noise injection. To mitigate interference without external supervision, we propose Self-Anchored Rubric Alignment (SARA). SARA uses a model's own single-rubric judgments as stable anchors and aligns multi-rubric reasoning with these anchors through on-policy self-distillation. We validate SARA on three datasets (HealthBench, FLASK, ResearchQA) and two model families (Qwen3, Llama-3.1). SARA consistently improves evaluation consistency while maintaining agreement with both base models and GPT-4.1 as a reference judge. Furthermore, the learned consistency transfers across datasets, confirming that SARA teaches a general capability rather than fitting dataset-specific patterns.

Given a patient's clinical findings, a diagnostic system ranks possible diseases and must decide when to endorse its first prediction or defer it for review. This decision is usually made by thresholding the top score. Selective prediction over ranked outputs begins with two checks. First, the ranker must produce enough correct top-ranked predictions to make the target feasible. Across 2,000 patient records stratified by disease prevalence, eight small open-weight LLMs achieve at most 4.6% Recall@1 on ultra-rare diseases. At 10% coverage, even a perfect confidence ranking of their existing predictions therefore cannot reach 50% selective accuracy. More accurate models pass the same check, showing that the limit is regime-specific. Second, the confidence signal must match the decision being made. For fixed-candidate rankers, the top-two margin cancels components shared across candidates. On phenotype-only Exomiser, it selects 10% of cases at 29.0% accuracy, compared with 13.3% overall, while the top score provides no reliable gate. Yet that cancellation can remove information needed to detect whether the candidate list contains an answer. SciFact retrieval and biomedical entity linking confirm this distinction. Finally, we prove that unlabelled scores alone cannot determine whether switching to the margin will help.

Real-world time series inherently intertwine global periodic structures with localized non-stationary variations. Existing approaches process these heterogeneous dynamics within a single computational domain, incurring fundamental limitations: time-domain models suffer from periodic misalignment over long horizons, while frequency-domain models over-smooth transient spikes. We argue that the optimal computational domain is not a property of the model, but of the data itself. Based on this principle, we propose RouteTS, a unified forecasting framework that partitions the frequency spectrum via amplitude routing and delegates components to their mathematically optimal domains. Dominant frequencies are processed by a complex-valued linear predictor in the frequency domain to preserve periodic structure, while residual spectral energy is reverted to the time domain and modeled by a lightweight MLP for local variations. Extensive experiments demonstrate that RouteTS achieves competitive prediction accuracy across diverse real-world datasets, with routing decisions guided by the underlying spectral signature. Furthermore, the lightweight design of RouteTS provides significant computational efficiency advantages, offering a principled solution to the longstanding dilemma between global periodicity and local transience.

Active learning can reduce labeling cost by selecting informative examples, but the most uncertain examples may also be the hardest to label correctly. This study tests whether uncertainty sampling fails because it acquires more corrupted labels or because errors concentrated in difficult regions are especially harmful. Margin-based uncertainty sampling is compared with random sampling under clean labels, random classification noise (RCN), and bounded difficulty-dependent noise on three public binary tabular datasets. The design uses 100 paired seeds, nine expected noise rates from 0 to 0.30, annotation budgets from 20 to 120, and logistic regression with regularization re-selected by cross-validation at every budget. An exposure-matched RCN control aligns mean final acquired corruption, while a clean-label extension reaches budget 400. Under clean labels, uncertainty sampling improved normalized balanced-accuracy area under the learning curve by 1.09 to 1.77 percentage points on all datasets. Difficulty-dependent noise reduced this advantage more than RCN at six of eight rates on Breast Cancer Wisconsin, but at no tested rate on Banknote Authentication or MAGIC Gamma Telescope. Exposure-matched analyses found no corrected evidence for a universal additional penalty from structured error location. On clean MAGIC data, uncertainty sampling improved balanced accuracy while reducing average precision and true-positive rate at fixed false-positive rates. Thus, uncertainty sampling was label-efficient, but its apparent robustness depended on dataset, budget, noise structure, and evaluation metric.

Predicting fatigue failure in steel components experimentally is costly, requiring testing across multiple compositions and processing conditions, spurring research on data-driven prediction models. Studies using the NIMS MatNavi steel fatigue dataset often report high point-prediction accuracy but rely on aggregate error metrics, leaving uncertainty about the reliability of individual predictions and whether accuracy is consistent across the fatigue-strength spectrum. This paper is the first to apply conformal prediction to steel fatigue strength, comparing seven interval-construction methods across 50 independent data splits and distinguishing marginal coverage from coverage within specific sub-regions of the predicted property. A gradient-boosting point model achieves an R^2 of 0.976 +/- 0.009 and a mean absolute error of 18.3 +/- 2.3 MPa. Split-conformal prediction provides valid marginal coverage (0.918) but drops to 0.758 in the highest-strength quartile, where design margins are most critical, a pattern also observed with a Gaussian process baseline. Two locally-adaptive methods correct this: a cross-fitted normalised conformal method holds 0.872-0.940 across quartiles at no cost in average width, and Mondrian group-conditional conformal prediction holds the tightest band of any method (0.917-0.939) at a 12% width premium, part of which traces to the more conservative finite-sample quantile level implied by per-group calibration at this sample size. Conformalized quantile regression, by contrast, restores marginal validity but inflates intervals in every quartile without closing the conditional gap. Marginal coverage claims for ML-based fatigue-strength predictions can conceal systematic unreliability precisely where engineering decisions are most risky; therefore, conditional coverage should be routinely assessed alongside marginal coverage.

Semi-supervised Class Incremental Learning (SSCIL) is a severe challenge for neural networks, and it is hardest in the exemplar-free setting where no past data may be stored. Existing methods forget catastrophically due to feature drift, and their pseudo-labels become increasingly unreliable as the label space grows. In this paper, we propose MAGIC (Manifold Anchoring and Geometric Incremental Calibration), a framework that stabilizes plasticity without storing exemplars. MAGIC's design centers on two components. The first is Soft-Weighted Geometry Calibration (SWGC), which uses graph-based label propagation on the learner's plastic feature space to weight and calibrate class means and variances computed on the frozen backbone; from these calibrated Gaussians, we sample phantom features that stand in for data from previous tasks. The second is a Geometric Structural Alignment (GSA) objective that preserves representation topology by matching the relational structure of student and teacher heads and aligning feature anchors with the fixed classifier prototypes, locking the orientation of the feature space. Together, these constraints keep the adapter from drifting, so geometric relations between classes remain stable as new classes arrive. We implement MAGIC with a frozen ResNet-18 backbone and a learnable plastic adapter. Across CIFAR-100, CUB-200, and ImageNet-R, at label ratios of 1%, 5%, and 10%, MAGIC improves average incremental accuracy over most of the supervised CIL methods equipped with FixMatch and native SSCIL baselines; the largest gains occur in the fine-grained, low-label setting, where confidence thresholding fails most clearly.

Long-video understanding requires models to efficiently acquire and reuse sparse visual evidence from long and redundant video streams. Recent video tool-use agents address this challenge by iteratively invoking visual Tools at different temporal scales, but their Tool-Planner communication typically relies on textual observations. Such text-only interfaces provide lossy summaries of Tool computations, causing previously computed visual evidence not verbalized to be discarded and unavailable for subsequent planning. We identify this limitation as the Tool observation bottleneck and propose Latent Visual Evidence-Enhanced Planning (LAVE), a training-free framework for reusing latent visual evidence from completed Tool calls. LAVE introduces a dual-channel observation interface: the visible channel preserves the original textual trajectory, while the latent channel stores pre-verbal visual updates with their Tool roles, source-frame timestamps, and visual locations. During planning, LAVE retrieves evidence relevant to the current Planner state but not covered by textual observations, and integrates it through bounded timestamp-aligned latent updates with entropy-constrained frame-time routing. This enables video agents to reuse existing visual computation without additional training, frame replay, or modifications to the original orchestration. Extensive experiments on Video-MME, LongVideoBench, and CG-Bench show that LAVE consistently improves video tool-use agents across backbones. Under a comparable frame budget, LAVE improves the Video-MME overall score by 3.76 points over the strongest baseline, demonstrating the effectiveness of latent visual evidence reuse for multi-step video-agent planning.

Multi-agent LLM systems route among model-backed advisors, yet a deployer rarely knows before shipping whether routing will help at all. Prevailing routers optimize a gate's AUC and presume that advisor complementarity suffices. We show that neither determines the deployable gain. We introduce RouteGuard, a deployment-certification framework. Routing gain decomposes as $G = \pi \Delta_E$, and the achievable gain is governed by a conditional-regret functional $\Phi$, not by AUC. A finite-sample certification bracket comes with a matching Le Cam lower bound, constant-sharp over the fixed-activity class, and a robustness phase transition. On two benchmarks the framework acts as a guardrail. On RouterBench (11 cross-family models) the verdict depends on the sampling unit: the protocol certifies a gain over GPT-4 under prompt-level sampling and withholds it under workload-cluster resampling, because the gain rests on 3 of 86 workload cells. On OpenRCA (three Gemini advisors) the advisors are statistically redundant: the realized oracle sits at or below the independence baseline in all pools we tested (221 RouterBench pools and three OpenRCA distributions), so the protocol correctly refuses to certify. A pre-registered semi-synthetic control confirms calibration: the protocol certifies a genuine gain once $m \ge m^\star$ and does not certify a true null. Code and frozen artifacts will be released with the published version.

Learning from Electronic Health Records (EHRs) has gained significant attention due to its potential to improve clinical prediction. However, effective learning remains challenging because EHRs encode heterogeneous, temporally ordered clinical interactions. In particular, EHRs contain: (i) heterogeneous clinical entities, including patients, visits, diagnoses, prescriptions, and procedures, together with their heterogeneous interactions, (ii) longitudinal patient trajectories across hospital visits and (iii) shared statistical dependencies across related clinical prediction tasks. Existing EHR learning methods capture only a subset of these properties. To bridge this gap, we propose Multi-task Graph transformer for Heterogeneous Temporal EHRs (MiGHT-EHR), which jointly models all three within a unified representation learning method. MiGHT-EHR constructs a heterogeneous graph from EHRs in which nodes represent clinical entities and edges connect statistically associated entities identified via normalized point-wise mutual information. Across MIMIC-III and MIMIC-IV datasets, MiGHT-EHR outperforms state-of-the-art methods on average across four tasks: drug recommendation, prediction of length-of-stay, mortality, and readmission, with particularly strong improvements in mortality and readmission prediction. Furthermore, a post-hoc analysis of the learned representations reveals that patient neighborhoods are organized by clinical outcomes, salient medical concepts are recoverable as linear directions in the representation space, and task probabilities are well calibrated. Collectively, these findings demonstrate that MiGHT-EHR representations support diverse prediction tasks while preserving clinically interpretable structure.