Papers

Drug repurposing aims to identify new therapeutic uses for existing compounds and, compared with de novo drug discovery, offers a faster and more cost-effective path to clinical translation. However, the space of candidate drug-disease pairs is enormous and their underlying relationships often depend on complex multi-hop biological mechanisms, making it difficult to reliably predict which pairs represent true therapeutic relationships. Existing approaches tackle this from two directions: knowledge graph-based methods organize curated biomedical evidence into structured relational networks for grounded multi-hop reasoning, while LLM-based methods leverage pretrained knowledge to generate flexible mechanistic rationales. Yet neither is sufficient alone - KGs are confined to observed graph structure while LLMs lack factual grounding and risk hallucination. To address this gap, we propose DrugReason, a multi-view reasoning framework that integrates grounded KG reasoning with LLM-generated mechanistic inference for drug repurposing. DrugReason adaptively routes diverse reasoning paths to specialized experts conditioned on the query context, while a cross-expert distillation objective enables knowledge sharing without sacrificing expert specialization. Experiments on PharmaDB, DDInter, and DrugBank show that DrugReason improves average performance over strong single-view reasoning baselines and achieves competitive or superior results compared with graph-based alternatives, while providing interpretable routing-based predictions.

Ensemble fuzzing runs multiple fuzzers on a target while a scheduler allocates CPU time among them. Existing schedulers base these decisions on compact summaries of past performance and rules fixed before a campaign. Our measurements reveal two limitations. First, past-reward summaries do not reliably capture performance evolution: after accounting for estimation noise, agreement between consecutive-window rankings is statistically indistinguishable from within-window self-agreement. Second, predictive signals vary across targets: on eight of nine targets, a weighting learned from the other eight predicts reward worse than one learned on the current target. We introduce CAROL, an online scheduler that uses each fuzzer's current context. Already available to the dispatch loop, this context describes reward trends, waiting and plateau time, reached code, and estimation uncertainty. CAROL uses context in two ways: a domain-guided method detects whether a fuzzer is rising or rotting and applies a phase-specific learning rule, while a learned method predicts reward from 15 context signals and uses predictive uncertainty for online selection. Across nine Magma targets, CAROL triggers more unique bugs than each of three ensemble-scheduling baselines whenever their results differ, and fewer on none. Compared with the strongest baseline for each target, CAROL gains 11.8% and surpasses an oracle that retrospectively selects the best single fuzzer per target. Removing context eliminates the gain, and the additional bugs are concentrated among those the baselines trigger rarely or never. Run unchanged on five widely used C++ programs, CAROL finds 120 previously unknown crashing defects, deduplicated by site, fault, and entry point; all were reported to maintainers through the projects' stated disclosure channels.

While diffusion models excel at capturing complex data distributions, scientific discovery often requires steering generation toward specific, uncharacterized regions that maximize a target objective. These high-utility modes frequently reside in low-likelihood tail regions and are only revealed sequentially through interactive feedback. Existing diffusion samplers fail in this regime: they inherit the pre-trained model's bias toward high-density regions, leaving rare yet promising phenomena underexplored. Conversely, exploration-heavy samplers ensure broad coverage but fail to efficiently exploit high-utility modes when constrained by a strict sampling budget. To resolve this dilemma, we introduce Levy Adaptive Tree Search (LATS), a principled sampling framework for online feedback-driven search. LATS leverages heavy-tailed exploration coupled with tree-based value backpropagation to progressively uncover preferred modes. By maintaining broad distributional coverage, LATS successfully discovers low-likelihood, high-utility regions while preserving sample fidelity and structural diversity. Experiments across diverse benchmarks, including materials science, demonstrate that LATS significantly outperforms baselines in target discovery efficiency.

Latent visual reasoning aims to perform multimodal reasoning through hidden-state computation rather than explicit textual chains of thought. However, visual information being present in a latent state does not imply that the model actually relies on that state when producing its answer, especially when alternative image-conditioned paths remain available. We introduce Causal Visual Recurrent Reasoning (CVRR), which preserves pretrained visual competence while making recurrent computation the required image-conditioned path to prediction. CVRR initializes recurrence from the question hidden state after the pretrained vision-language model has incorporated the image, then repeatedly updates this state while re-reading the same fixed visual evidence. Before decoding, visual states and the original multimodal KV cache are removed so that only the final recurrent state carries image-conditioned information to the answer. Across the $V^*$, MMVP, BLINK, and MME-RealWorld-Lite benchmarks, CVRR retains strong performance under this strict interface, while compatible latent reasoners fail to recover comparable visual competence even when retrained under the same constraint. Causal interventions further show that predictions remain sensitive to recurrent content when the question is held fixed, and that persistent visual evidence causally revises the recurrent trajectory. These results distinguish latent informativeness from latent computation that is actually used for prediction.

Glaucoma is a progressive optic neuropathy characterized by irreversible damage to the optic nerve, making timely diagnosis critical to prevent permanent vision loss. Although deep learning has demonstrated promising performance in automated glaucoma detection, existing approaches often overlook feature refinement, suffer from class imbalance, and rely on individual classifiers that limit prediction robustness. To address these challenges, this paper proposes a hybrid glaucoma detection framework that integrates attention-enhanced deep feature extraction with heterogeneous ensemble learning. Specifically, deep representations are extracted using InceptionV3 and subsequently refined by incorporating the Convolutional Block Attention Module (CBAM) to enhance discriminative retinal features. To improve classification robustness, the extracted features are classified using multiple machine learning models together with Single-Level Ensemble (SLE) and Double-Level Ensemble (DLE) strategies, while SMOTE combined with Tomek Links (SMOTE+TL) is employed to alleviate class imbalance. Furthermore, a systematic comparison of handcrafted, deep, and attention-enhanced deep feature representations is conducted. Experimental evaluation on two public retinal fundus datasets demonstrates that deep feature-based methods consistently outperform handcrafted feature-based methods, while the proposed attention-enhanced framework achieves the best overall performance. Furthermore, Grad-CAM visualizations confirm that the proposed model focuses on clinically relevant retinal regions, providing interpretable evidence on the model's prediction process.

We present an unsupervised search for anomalous dijet events in proton–proton collision data using neural spline flow density estimation. A normalizing flow model is trained on a high-dimensional feature space comprising jet, dijet, and event-level observables to learn the dominant Standard Model background directly from data, without assuming a specific signal hypothesis. Events assigned low likelihood under the learned density are identified as potential anomalous events. Using this approach on a CMS Open Data dijet sample, we investigate extreme events in the tail of the anomaly-score distribution and perform an extensive validation and robustness study. This study includes feature-level statistical comparisons, mass decorrelation tests, permutation-based null tests, and evaluations of training stability. The selected anomalous events exhibit notable departures from the background-only expectation, primarily in jet-substructure observables, while remaining stable under several known sources of bias in unsupervised learning. The identified anomalies are distributed across the kinematic phase space and do not exhibit a narrow structure in the dijet invariant-mass spectrum. Instead, they show correlated deviations across multiple observables, consistent with a multivariate difference in jet substructure and event topology rather than a localized resonance. While no claim of new physics is made, this study demonstrates that neural spline flow-based density estimation can be sensitive to rare, structured deviations in collider data and may provide a model-independent exploratory tool for searches for physics beyond the Standard Model at the LHC.

We derive an exact gradient-step representation of the RoPE-softmax forward pass. For every deterministic RoPE-softmax attention head with arbitrary affine projection weights, we construct a query-dependent effective matrix $\Delta M_i$ satisfying $y_i = \mu_i + u_i^\top \Delta M_i$, where $\mu_i$ is the uniform mean of the attended values and $u_i$ is the augmented query input. The construction applies the classical exponential divided difference $\rho = \phi_1$ to retain the softmax exactly. Its positive coefficients give a unit gradient-step representation on a query-conditioned quadratic objective. The same function connects the RoPE generator to exact positional finite differences. We derive a tokenwise formula for the error of reusing one query's matrix and prove that a nonconstant finite-cache head cannot admit a globally exact affine query readout. Reconstruction checks and frozen-reuse calibration on one pretrained Qwen2.5-0.5B layer verify the representation and quantify the correction required when one query's matrix is reused.

Group-based RLVR methods estimate advantages by repeatedly sampling multiple trajectories for each prompt, making long-horizon agent training expensive and discarding useful experience accumulated across iterations. We ask whether historical experience can replace these repeated within-iteration comparisons without directly optimizing on stale trajectories. We introduce Long-Short Term Advantage Estimator (LSTAE), a single-stream RL algorithm that uses history for advantage estimation while updating the policy only with the current rollout. LSTAE maintains a persistent tracker for each task anchor. At the trajectory level (long term), a drift-aware historical baseline tracks the anchor's moving success frontier and measures the relative contribution of each new trajectory. At the step level (short term), a recent state-experience buffer exploits recurrent states to estimate localized action advantages. This two-timescale design converts accumulated experience into multi-granular credit signals, requiring only one rollout per anchor. Across agentic and mathematical reasoning benchmarks, LSTAE matches or improves upon strong group-based baselines while substantially reducing rollout cost.

The In-context learning (ICL) paradigm aids large language models (LLMs) to adapt to new tasks without need for fine-tuning. However, selecting an optimal combination of demonstration examples from a large pool of example subsets is a challenging problem. Existing approaches for selection do not model the complex relationship between ICL samples and downstream LLM performance. They typically perform static task-level selection, choosing subsets once offline, which can fail to generalize to unseen queries. We introduce DearICL (Data Efficient Algorithm for Ranking) ICL samples, a new framework that models demonstration example selection as a subset ranking problem. DearICL employs a non-linear surrogate employing a differentiable sorting objective within a gap-index bandit algorithm. The gap-index based approach enables fine-grained separation of good arms and borderline arms, which is used as an auxiliary objective to train the non-linear surrogate through sufficient sampling of borderline arms, supporting instance-level subset ranking. On exemplar selection benchmarks with open-source LLMs, DearICL achieves 8.08-15.9% accuracy gains over strong linear bandit baselines, with low sample complexity. Code and data: https://github.com/VenkteshV/DearICL.

Multimodal graphs couple node attributes in different modalities, such as text and images, with relational structure, enabling topological structure and cross-modality attributes to be modeled jointly. Multimodal graph foundation models seek unified representations from such data that transfer across different graph domains and downstream tasks. However, existing methods exhibit two fundamental limitations. (1) Cross-Scope Context Entanglement. They merge scope-specific graph contexts into a unified representation, obscuring their distinctions during multimodal construction. (2) Scope-Ignorant Modality Routing. They route modalities within a fixed graph scope, overlooking how modality relevance varies across neighborhood ranges. To address these challenges, we propose BRAIN, a unified model that focuses on graph context that combines neighborhood scope with modality composition. BRAIN comprises a scope-conditioned Bridge that combines structural information spanning local-to-global neighborhood scopes with different modality compositions; a hierarchical Router that estimates the relevance between the scope and the task, and selects compositions separately within each scope, allowing modality utility to vary with graph range; and a lightweight residual Adapter that further specializes the routed embedding for downstream prediction. BRAIN is trained through multi-graph pretraining followed by task-specific adaptation. Experiments across nine datasets and four task families demonstrate its broad effectiveness, improving node-classification and link-prediction performance by up to 4.73% relative to the strongest baseline, while achieving an average relative improvement of 14.72% across four graph-to-text and two graph-to-image metrics.

Adaptive dosing requires policies that reduce tumor burden without excessive toxicity. Learned dosing policies are typically judged against historical or heuristic comparators, which cannot show whether a policy has found the best behavior available. We instead study a three-population tumor-control ODE in which optimal-control analysis fixes the form of a good schedule – bang-bang dosing punctuated by a singular arc – and construct a numerical controller of that form as a proxy for near-optimal behavior. Judged against this reference under a sustained-cure criterion – 200 consecutive days below 5% carrying capacity – Soft Actor-Critic (SAC) trained from scratch never reaches cure. Behavioral cloning (BC) of the reference reproduces it (100% sustained cure, 30/30 seeds), but SAC fine-tuning of the cloned policy destroys it across five entropy coefficients, and TD3 and BC-regularized SAC fail identically; the pattern persists under multiplicative pharmacokinetic action noise. Along curative trajectories the post-collapse critic ranks the collapsed-policy action above the reference action in 96% of states, concentrated in the maintenance phase, and the policy settles into a non-curative adaptive-therapy equilibrium. The reference is what makes this legible: against a heuristic comparator the fine-tuned policy would read as a competent controller rather than a failure.

Modern power systems are growing increasingly complex as they integrate diverse generation sources to meet rising demand, making accurate load forecasting challenging. Recent advances in time-series foundation models (TSFMs) resulted in promising performance in zero-shot univariate load forecasting tasks. However, real-world load forecasting often involves multiple target variables and requires the integration of exogenous variables, raising important questions about the utility of TSFMs in realistic settings. In this study, we position Chronos-2, a recently developed model by Amazon, as a representative multi-channel TSFM that supports univariate, multivariate, and covariate-informed forecasting, and conduct a systematic investigation of how such models can be used for real-world load forecasting. While prior work has evaluated Chronos-2 on a limited number of energy-related tasks in a zero-shot setting, its performance relative to established task-specific deep learning models and its behavior when adapted using task-specific historical data remains insufficiently understood. In this work, we evaluate Chronos-2 on two real-world utility datasets, ISO New England and ENTSO-E, and benchmark it against widely used task-specific deep learning models. Our results show that Chronos-2 benefits substantially from task-specific fine-tuning and achieves strong short-horizon forecasting performance, but its zero-shot accuracy lags behind task-specific models and its forecasting error grows more rapidly with increasing forecast steps. Overall, this study provides a detailed characterization of the strengths and limitations of TSFMs such as Chronos-2 in grid load forecasting and offers practical insights into how a pretrained TSFM can be effectively adapted for operational load forecasting applications.

Diffusion models have general generative abilities but struggle to align with specific objectives. Fine-tuning can improve alignment, yet its training cost is often prohibitive. This led to training-free methods that apply objective-guided terms in sampling to bias the generation distribution toward designated regions, e.g., high-reward areas. However, these methods face two issues: (1) the strong directional bias narrows the pretrained distribution and generation diversity, and (2) indiscriminate constant guidance fails to prune redundant signals, hurting both quality and efficiency. To address the above challenges, we propose SwiftExplorer, a plugin that mitigates distribution collapse caused by excessive diversity loss and reduces compute costs. First, we adopt an Inheritance-Restart exploration mechanism to avoid early convergence, while exploration also increases the likelihood of high-reward trajectories. Additionally, it balances diversity and fidelity, adding diversity without causing a distribution over-shift. Second, our Quality-Efficiency arbitration mechanism improves guidance by removing incorrect signals, and it reduces computation by dynamically stopping generation when completeness and marginal reward gain are optimal. In an extensive number of experiments and different types of evaluation metrics, the proposed SwiftExplorer achieves excellent performance on all metrics, including preference, fidelity, diversity, and richness.

Reward hacking during reinforcement learning from verifiable rewards (RLVR) can induce reward seeking and broad misalignment in language models. Studying this misgeneralization is important for developing better threat models and countermeasures, but is often infeasible due to the cost of RL on large models. As an alternative, we propose studying emergent misalignment from iterative DPO, which preserves important properties of RLVR while reducing costs and enabling training on popular finetuning APIs. In practice, we find that training GPT-4.1 with iterative DPO on a single-turn reward hacking environment induces covert misaligned power-seeking and alignment faking, the first openly available (semi)-online training pipeline to induce these concerning forms of misalignment. We also find that training Qwen2.5-32B-Instruct with the same pipeline induces both misalignment and improved instruction following accuracy, showing that iterative DPO can be used as a testbed for selective generalization. Overall, we think iterative DPO can help democratize and accelerate the study of emergent misalignment from RLVR.

Fine-grained mobile traffic demand forecasting is essential for long-term planning of 5G and future 6G networks, including radio upgrades, site densification, backhaul expansion, and spectrum activation. This paper proposes the Traffic Demand Spatio-Temporal Graph Transformer (TD-STGT), a graph neural forecasting framework for predicting changes in wireless mobile traffic demand across fine geographic grids. The framework uses a population-scaled demand proxy developed from crowdsourced mobile measurements and daytime population information. Experiments across five Canadian metropolitan regions show that TD-STGT achieves the best performance in forecasting grid-level demand changes, reaching a $\Delta R^2$ of 0.462 and reducing $\Delta$RMSE by 5.7% relative to the strongest baseline. The proposed model provides a practical tool for identifying areas with increasing demand pressure and prioritizing future mobile-network capacity upgrades.

Can exploratory UAV waypoint sequences be generated from multimodal onboard observations and a fixed-dimensional recurrent internal state without maintaining a persistent global map in the deployed policy? We investigate this question through CAVEAT, a diffusion policy conditioned on a recurrent internal state updated from fused LiDAR, visual, and pose features and trained from trajectories generated by the map-based FUELv2 expert. Rolling inference partially warm-starts consecutive predictions, while a temporary local signed distance field provides heuristic obstacle guidance. Simulation results evaluate both inference mechanisms and compare CAVEAT with its demonstration-generating expert. Proof-of-concept experiments on a Flyability Elios 3 demonstrate partial exploration of a previously unseen indoor environment and target-directed visual servoing using a separately trained policy.

Professional box-lacrosse statistics summarize outcomes but provide limited information about shot quality or the roles behind scoring opportunities. This study develops a documented framework for estimating expected goals (xG) and attributing recorded offensive involvement using 1,006 manually annotated Rochester Knighthawks shot attempts, including 151 goals, from 13 consecutive 2025-2026 National Lacrosse League games. Logistic regression, random forest, and extremely randomized trees were evaluated across three nested feature sets using Leave-One-Game-Out cross-validation and a training-fold base-rate benchmark. The contextual baseline random forest had the lowest observed pooled log loss (0.4189) and Brier score (0.1260), improving on the benchmark by 1.22% and 1.50%; five of nine specifications did not beat the benchmark. Adding two-man-action and pick-type fields did not improve the primary metrics. Core Offensive Impact attributes recorded involvement through shooter xG and shot-based expected assists for final passers. Expected Pick Value (xPV) compares a qualifying pick's observed-state probability with a no-pick counterfactual. Its magnitude was indistinguishable from model noise. Its directional pattern exceeded 200 row-permutation replicates, but limited tail resolution and failure to preserve game-level pick composition make the diagnostic descriptive rather than inferential. Accordingly, xPV is reported only as an exploratory augmented component. Given the single-team, 13-game sample, the results are an initial case study rather than league-wide or causal estimates.

The optimal transport (OT) map provides a geometric transformation for aligning probability distributions and has become a useful tool in machine learning. However, existing estimators of the OT map still exhibit a gap between sharp statistical guarantees and practical parametric estimation based on stable training objectives. Theoretical estimators achieve minimax optimal convergence rates, but they are typically nonparametric and can incur demanding implementation design or inference costs. Practical estimators are parametric and scalable, but their statistical guarantees remain underexplored, and their min-max, adversarial-like training objectives can be sensitive to optimization algorithms. We propose BROT (Barycentric Regression for OT), a simple two-step method that first computes the unregularized OT plan and then fits a deep neural network (DNN) to the induced barycentric targets by least-squares regression. Under standard regularity conditions, we prove that the DNN estimator of BROT attains the minimax convergence rate, when the ground-truth OT map is Lipschitz. Numerical studies on synthetic datasets and an image dataset show that BROT provides accurate map estimates, strong target distribution matching, and competitive transport costs, compared to existing estimation methods. Experiments on two downstream tasks, single-cell perturbation prediction and unsupervised domain adaptation, further suggest that the accurate estimation of BROT can translate into stronger task performance.

Choosing when to translate multilingual documents is a central routing problem in text classification: translation can improve predictions for some languages while degrading others or adding unnecessary computation. Uniform translation and heuristic language tiers do not provide statistically controlled route selection. We introduce a language-level router based on paired e-processes that continuously compares direct and translation-assisted classification before freezing a routing policy. A familywise-controlled threshold of 280 bounds the probability of any false route across 14 eligible languages per dataset by 0.05. On SIB-200 and MASSIVE, the router selects translation for 4 of 15 languages and 14 of 15 locales, improving held-out accuracy over direct classification by 8.14 and 16.70 percentage points, respectively. All 28 decisions remain stable across 50 outcome-independent orderings and relative to the per-group threshold. Our results demonstrate that paired e-processes enable statistically controlled, anytime-valid, and auditable multilingual classification routing.

Mixture of Prompt Experts (MoPE) adapts multimodal transformers through input-dependent prompt composition, while retaining a fixed prompt length. We investigate a layer-wise gating extension in a binary chest X-ray classification pilot study. The controller predicts a retention ratio for each sample, averages these ratios within a mini-batch, and uses the resulting integer length to truncate the static and mixed visual prompts. Retained mixed prompts are also scaled by the individual ratios. In one recorded run per configuration, the gated model reached a best validation accuracy of 0.8996, compared with 0.8969 for the fixed-length baseline; the corresponding final values were 0.8963 and 0.8802. The exported gate statistics imply a retained length of one at all recorded training points, relative to a configured maximum of six. This reduces the complete visual sequence from 210 to 200 tokens, but no direct runtime measurements establish an acceleration benefit. Report-derived labels, report text as input, sequential data partitioning, and the absence of repeated controlled experiments limit interpretation. The findings document prompt shortening under the configured gate penalty; they do not establish sample-specific length allocation, superiority over fixed short prompts, or clinical utility. Code is available at: https://github.com/jingtaolei/mope-dynamic-prompt-truncation.

A shared reward gives agents a common objective, but leaves open when, how and even whether they must cooperate to succeed. We address these questions in the Laser Learning Environment, a multi-agent path-finding environment where cooperation materializes as one agent blocking a laser to let a teammate pass safely. We represent these interactions through temporal cooperation graphs whose timed edges connect helpers to beneficiaries, define six cooperation profiles as overlapping graph predicates, and prove that every cooperative trajectory satisfies at least one. By encoding the environment dynamics and profile predicates as propositional formulae, we distinguish tasks that admit}a profile in some winning trajectory from those that require it in every winning trajectory within a specified horizon. Used as filters, these queries turn a random layout sampler into a generator of tasks with certified cooperation requirements. Experiments with five multi-agent reinforcement learning algorithms show that training diversity improves joint success on unseen tasks when cooperation-free solutions exist. When cooperation is required, greater diversity improves individual-agent exits, but joint success remains near zero. Across five profile-certified pools, final exit rates averaged over algorithms separate the pools into four statistically distinguishable levels but this ordering primarily reflects partial completion: policies collect rewards for individual exits but rarely exhibit the profile required for joint success. Our framework exposes this gap between rewarded partial completion and realized cooperation by certifying what cooperation successful completion requires and using temporal cooperation graphs to reveal what policies exhibit.

We investigate stochastic simple bilevel optimization with smooth and possibly nonconvex upper- and lower-level objectives. Existing stochastic extensions of dynamic barrier gradient descent (DBGD) either obtain fast convergence under an unverifiable trajectory-dependent “rare-visit” assumption, or remove this assumption at a substantially higher oracle cost. We show that a simple denominator-only regularization of the DBGD multiplier eliminates the need for such an assumption while preserving fast convergence rates. Specifically, our method achieves $(\varepsilon, \varepsilon)$-stationarity in $O(\varepsilon^{-2})$ iterations using $O(\varepsilon^{-4})$ upper-level and $O(\varepsilon^{-7})$ lower-level stochastic gradients, which improves upon the best assumption-free complexities. We additionally derive anytime parameter schedules.

Social deduction games (SDGs) require agents to reason under partial observability by maintaining relational beliefs about hidden roles and team alignments. While recent LLM-agent approaches improve gameplay through prompting and preference optimization, they often optimize actions and in-game speech without explicitly grounding them in such beliefs. This frequently leads to strategically inconsistent behavior, especially for compact LLM agents. We introduce Multi-Agent Relational Belief Optimization (MARBO), a belief-grounded preference optimization framework that leverages relational beliefs to guide strategic decisions and in-game speech. MARBO provides preference feedback only when behaviors are supported by reliable relational beliefs and lead to strategically favorable social outcomes, encouraging more consistent learning under uncertainty. Experiments on representative SDGs show that MARBO enables compact LLM agents to consistently outperform existing baselines. The Code is available on https://github.com/PleaseTakemeAway/MARBO.

Large language models (LLMs) are often considered fragile under aggressive sparsification, and maintaining reliable performance typically requires sticking to moderate sparsity levels. However, recent studies suggest that LLMs are more resilient to high sparsity than previously thought, reframing the problem as a design challenge rather than a fundamental limitation. In this work, we challenge the perceived limits of unstructured post-training LLM pruning by revisiting elementary pruning strategies that have remained relatively underexplored at this scale. Through a progressive sparsification framework with second-order saliency and continued training coordinated with sparsity progression, we show that pretrained LLMs can retain strong performance far beyond commonly studied sparsity regimes. Across LLaMA-2 and Qwen-3 model families, our approach improves perplexity and downstream accuracy up to 99% sparsity, surpassing both the current state-of-the-art and representative baselines. Precisely, on LLaMA-2-7B, our approach achieves WikiText-2 perplexities of 13.48 and 19.67 at 95% and 99% sparsity, respectively, while delivering 3.23$\times$ decoding speedup and 6.21$\times$ memory savings at 95% sparsity. Taken together, our results show that LLMs can be pushed into extreme sparsity while retaining strong performance, providing a foundation for further improving sparse models in this regime.

Mobile vendors and application developers increasingly deploy LLMs on smartphones for diverse prefill-only services. Yet current systems rely mainly on dense models whose regular computation maps efficiently to mobile NPUs, leaving more capable MoEs underused. MoE prefill does not fit mobile NPUs: NPU graphs are fixed at compile time, yet MoE picks experts at runtime; and one request touches most experts, more than a phone can hold in memory. We present EStream, which resolves both by separating what the NPU must fix from what MoE decides at runtime. A single compiled expert graph serves every expert, with each expert's routed tokens and weight address bound at call time, so dynamic MoE execution runs entirely on the NPU without padding or CPU/GPU fallback. Expert virtualization keeps the expert pool in UFS flash storage and pages it through a fixed-size NPU-addressable arena, group by group, with loading hidden behind computation, so memory is bounded by the arena rather than by the model. It further introduces a hardware-aware configuration algorithm that automatically configures the UFS–NPU pipeline and maximizes loading–computation overlap. Across 18 comparative settings covering three 7B–16B MoEs and 256–4,096-token prompts, we evaluate EStream on a commercial Snapdragon smartphone. Compared to the fastest baseline at each setting, EStream achieves a 2.25–27.57X pure-prefill TTFT speedup and reduces peak physical memory by 1.19–12.29X. EStream further scales to MoE models with up to 46.7B parameters.

Although machine learning can be used to predict the evolution of physical systems from data, a formulation that learns only the system state at each time leaves the temporal and dynamical structure of the solution to be resolved within a broad hypothesis space. We introduce Learning$^2$, a representation-level framework that structures this space by coupling a primary representation to a second representation through a known physical transformation. The resulting cross-representation constraint restricts the effective hypothesis space and provides an ante-hoc, physically interpretable criterion for excluding solutions that satisfy the primary representation alone. We instantiate Learning$^2$ through EuLaNet, an Eulerian–Lagrangian representation for fluid dynamics. Given the velocity state $u(\mathbf{x},t)$, EuLaNet constructs its induced Lagrangian flow map $X(\mathbf{a},t)$ through $\dot{X}(\mathbf{a},t)=u(X(\mathbf{a},t),t)$, from which material transport and finite-time deformation are derived. The resulting representation couples the predicted state to the dynamical consequences it induces, providing a second consistency criterion beyond state-level agreement. We formalize this construction through an effective hypothesis space $\mathcal{H}_{L^2}\subseteq\mathcal{H}$ and define the conditions under which a consequence representation provides discriminative constraints on candidate solutions. EuLaNet is implemented as a model-independent representation module, separating the physical constraint from the downstream learning architecture. This construction provides an ante-hoc mechanism for physically interpretable constraint in scientific learning and offers a basis for developing and evaluating broader classes of Learning$^2$ architectures. The implementation is open-sourced to support the development and extension of the architecture across scientific domains.

Oversmoothing is a well-known failure mode of Graph Neural Networks (GNNs). However, most existing diagnostics rely on global aggregation measures that fail to capture the heterogeneous dynamics of message passing. Real-world graphs exhibit pronounced community structure, and message passing operates on two timescales, with representations collapsing rapidly within communities and slowly across them. This creates a critical gap in which intra-community representations can become indistinguishable while inter community separation persists, a failure mode that we refer to as the Echo Chamber Effect. To quantify this effect, we introduce the Echo Chamber Index (ECI), which stratifies pairwise distances by community membership and reveals when global energy diminishes while inter-community separation persists. ECI further shows that feature retention mechanisms can preserve the echo chamber under the conditions of our theoretical analysis. The consequences depend on label structure: when communities align with classes, the echo chamber can sharpen node classification, whereas when they do not, the same collapse makes classification provably harder. Motivated by this analysis, we propose Community-Aware Split Propagation (CASP), a lightweight plugin that decouples intra- and inter-community aggregation and learns their balance from label structure. CASP improves diverse backbone GNNs across most evaluated homophilic and heterophilic settings.

Forecasting multivariate graph signals is challenging when node-level trajectories are nonstationary but stable relations persist across nodes and features. In an error-correction representation, long-run equilibrium restoration and short-run transient propagation represent different predictive roles and need not share a common cross-feature geometry. We introduce role-specific predictive geometries in which directed Long relations act on estimated equilibrium coordinates, whereas directed Short relations act on lagged differences. Matrix-valued Long responses mix equilibrium coordinates before graph propagation, while Short responses use graph-filtered transient designs; a direct multi-horizon estimator couples forecast corrections across adjacent horizons. Temporal cross-fitting and Frisch-Waugh-Lovell partialling-out give selected edges a conditional predictive interpretation relative to a graph-temporal backbone. The Long operator remains right-factorized through the equilibrium subspace and therefore annihilates source common-trend directions. Controlled experiments recover all planted Long relations (20/20), all planted Short relations (20/20), and both role families in every Dual realization (10/10). Across four real-world benchmarks, the proposed predictor improves on the G-VARMA backbone in three datasets, with all 25 fold-horizon comparisons favorable on the five-fold financial benchmark.

Adaptive frequency hopping against predictive jamming must address both model uncertainty and policy exposure: the context-loss relationship may vary across operating regimes, while persistent hopping patterns may expose high-probability channels to attack. We propose D-PACT-AFH, a model-adaptive and risk-constrained adversarial contextual-bandit framework in which a Tsallis-FTRL master combines a global linear learner with a partitioned local learner and selects the model class online. D-PACT-Hit incorporates channel-wise marginal hit risk into model selection, while D-PACT-Safe applies a minimum-Kullback-Leibler projection to enforce a per-slot risk budget. We establish estimator validity under non-anticipating attacks, an oracle decomposition relative to the better fixed base, and an exact conditional-risk guarantee for the Safe projection. Experiments across diverse channel regimes and jammer types demonstrate effective model adaptation and a controllable goodput-risk tradeoff: D-PACT-AFH recovers 95.5% of the local learner's gain under observable switching while avoiding 77.7% of its degradation in a negative-control regime.

Physical and synthetic models may describe complementary aspects of the same PDE-governed system while receiving different, possibly fragmented, observations. We propose Bi-Objective HYCO (Bi-HYCO), a cooperative framework that retains both representations and their local observational objectives while coupling their predicted states at unlabeled interaction points. These points contain no measurements and do not augment the data; they provide a communication mechanism in the common state space. The two criteria form a vector-valued objective, and weighted scalarizations provide computational realizations. For the deterministic shared-observation algorithm with fixed interaction points, we prove sufficient decrease and finite length of the whole alternating sequence, which converges to a mixed critical point under the stated Kurdyka-Lojasiewicz-type assumptions. Elliptic transmission and two-dimensional Navier-Stokes experiments assess parameter and state reconstruction, noise and scalarization effects, and PINN/XPINN references. Ablations show that removing state interaction while retaining aggregation deteriorates parameter recovery in the tested configurations, particularly for Navier-Stokes.