Though wireless foundation models (WFMs) have shown strong potential in learning universal channel representations, their adaptation to various downstream tasks remains constrained by existing paradigms. Fine-tuning strategies introduces substantial computational and storage overhead, while frozen feature extraction leads to sub-optimal performance across diverse downstream tasks. To address this issue, we propose a unified adaptive feature composition framework for multitask generalization in WFMs, where the key component is the Routing Adapter for Feature Composition (RAFC). Instead of extracting only the final-layer output, this router treats the hidden states from different Transformer depths as a reusable pool of multi-level hidden features, and employs a lightweight task-driven feature composition network to generate layer-wise aggregation weights, then adaptively combine hierarchical representations through weighted summation. This design enables each downstream task to access suitable mixture of low-, mid-, and high-level wireless features without modifying the pretrained backbone. Extensive experiments on four representative wireless tasks demonstrate that RAFC consistently outperforms conventional adaptation baselines while introducing fewer than 50K additional parameters. Moreover, the learned routing weights provide interpretable evidence of task-specific layer preferences, making the proposed framework a low-complexity, scalable, and explainable interface for adapting WFMs to diverse downstream scenarios.

For natural human-robot interaction, a robot must understand human intent expressed not only through language but also through nonverbal signals such as gestures and gaze. However, current robot policies rely on language instructions as the sole interface for conveying intent, leaving nonverbal signals unused and placing the full burden of communication. In this work, we present EDITH, a robot framework that captures the human's nonverbal signals through continuous streams of first-person view and gaze from smart glasses, and uses them alongside language instructions as inputs to the robot policy. Our hardware system streams the human's first-person view, gaze, and speech to the robot in real time, transcribing the speech into language instructions. To handle these rich but noisy signals, we design a hierarchical policy in which a high-level policy infers the human's intent and produces a sequence of subtasks, where each subtask is represented as a fine-grained instruction paired with a keyframe that grounds the intent in the scene (e.g., the frame where the human points at the target object). A low-level policy then executes these subtasks. In our experiments on human-robot interactive tasks, EDITH enables the robot to act on the human's nonverbal signals even when intent is expressed only briefly, and significantly reduces user effort to convey intent compared to using language instructions alone. Visit our project page for source code and real-robot demo videos.

Real-world image super-resolution (SR) is often designed with a single restoration objective, despite the current capacity of generative models to produce multiple high-quality reconstructions for the same input. In this paper, we argue that the best restoration strategy is subject to the specific restoration profile: a Faithful restoration prioritizes reference consistency, structure preservation, and hallucination suppression, whereas an Aesthetic restoration prioritizes visually pleasing and natural-looking details. We propose FoA-SR, a novel preference optimization approach to real-world SR based on profiles. To achieve this goal, FoA-SR starts with our supervised FLUX.2-based SR adapter (Flux2SR) trained with LR latent conditioning, flow matching, and image-space reconstruction losses for paired LR-to-HR image super-resolution. Following the development of the shared supervised super-resolution adapter, FoA-SR generates a shared stochastic candidate pool for each input image and ranks the same candidates using profile-specific Faithful and Aesthetic rewards to mine winner-loser pairs. These pairs are used to fine-tune separate LoRA adapters while keeping the base model frozen. Experiments on RealSR and DIV2K show that FoA-SR can steer the same SR adapter towards distinct restoration objectives: a Faithful adapter improves reference-consistent metrics while an Aesthetic adapter boosts metrics that measure perceptual quality without reference. Our candidate-pool analysis shows that Faithful and Aesthetic rewards frequently select different winners, and a Hybrid-LoRA ablation shows that collapsing both profiles into one reward yields an implicit compromise rather than explicit profile control.

Deploying legged robots in extra-terrestrial environments includes many challenges due to complex terrain interactions, energy, and thermal constraints. For effective mechanical design of a lunar exploration quadrupedal robot, careful consideration of motor torques, energy expenditure, and cost of transport is required. The lunar surface is composed of granular regolith, which impacts the locomotion of legged robots and their performance. Locomotion algorithms trained with rigid contact assumptions are also ineffective when applied to environments with soft contacts, such as granular surfaces, which can result in instability and poor tracking. In this report, the physical modelling of the granular lunar surface-robot foot contacts is applied to a simulation environment with locomotion trained using Reinforcement Learning. A comparison is conducted between the policy trained on rigid contact and soft contact environments, analysing the gait and locomotion performance metrics. The analysis demonstrates that soft contacts simulating regolith surfaces pose additional challenges for Reinforcement Learning based training, result in a qualitatively different gait, and increase the overall energy expenditure.

Hierarchical vision-language-action (Hi-VLA) systems have emerged as a promising paradigm for complex robot manipulation, by using high-level VLM planners to decompose tasks into language subgoals executed by low-level VLA controllers. Despite recent empirical progress, there is a lack of unified design principles for these systems: existing Hi-VLA systems differ in how they choose and connect planners, controllers, mechanisms to switch between the two, and how observations and memory are represented in the planner. In this paper, we present a systematic study of Hi-VLA design for robot manipulation. We unify representative Hi-VLA agents under an options-style control framework and benchmark core design choices across short-horizon, long-horizon, and reasoning-intensive tasks. Our analysis distills practical principles for building Hi-VLA systems, showing how model choices and interface mechanisms jointly shape performance. Applying these principles yields a substantially stronger system than either flat VLA control or a naively designed hierarchy, across experiments both in simulation and on a real ALOHA robot. Overall, our results provide a foundation for building more capable, robust, and principled hierarchical VLA agents. More information and video at jiahenghu.github.io/hi-vla.

While Large Language Models (LLMs) have achieved near-perfect performance in \emph{solving} high-school mathematics, their ability to \emph{evaluate} the diverse reasoning processes of real human students remains under-examined. To bridge this gap, we introduce \textbf{RealMath-Eval}, a rigorously annotated benchmark of 224 real-world exam responses from high schools. Our initial evaluation reveals that even state-of-the-art LLM judges struggle significantly on this task, exhibiting a high Mean Squared Error ($\sim$2.96) against expert human grading. To probe a plausible explanation, we contrast this performance with a control setting where the same judges evaluate synthetic LLM-generated solutions. We identify a stark ``Evaluation Gap'': judges are considerably more accurate and consistent on synthetic text (MSE $\sim$1.17) but struggle to generalize to authentic student reasoning. Through semantic embedding analysis, we find that synthetic errors suffer from a ``structural collapse'' into predictable, low-dimensional linear subspaces, whereas human errors form a more diverse error space. Furthermore, generative probability probes suggest that human reasoning involves significantly higher information-theoretic surprisal, indicating that student reasoning transitions are more out-of-distribution for current models. Finally, we find that surface-level style transfer fails to close this gap. Our findings suggest that current LLM evaluation pipelines relying heavily on synthetic data may not adequately capture the diversity of authentic student mathematical reasoning.

Class imbalance is a common problem in deep learning that severely degrades performance. In federated learning (FL), it is a critical factor contributing to non-identically distributed data (non-IID). Building on several previous attempts, we define and analyze imbalance issues in FL at three levels: inter-case, inter-class, and inter-client. Inter-case imbalance addresses the imbalance in every single class; inter-class imbalance compares the number of data between different classes. Inter-client imbalance represents different skewness of local data between clients. Based on these concepts, we propose FedBB, which consists of two main components: (1) Positive Negative Balanced (PNB) loss function addresses the inter-case and inter-class imbalances in local training, enhancing generalization on highly skewed local client datasets. It optimizes both multi-label and multi-class classifications by assigning higher weights to minority cases or classes. (2) Client Balanced Reweighting (CBR) reweights clients based on inter-client imbalance during model aggregation, giving greater weight to models trained on less skewed datasets. Various experiments on X-ray and natural image datasets demonstrate that FedBB outperforms other algorithms in both performance and efficiency. Additionally, it requires limited statistical information, which is beneficial for privacy protection. Through ablation studies, we proved that PNB loss and CBR independently contribute to performance. As FedBB aims to build a global model that accurately classifies all classes, it can serve as a baseline for the generic and personalized FL.

We examine whether graph neural network (GNN) design rules generalize across benchmark families by studying aggregator selection (sum, mean, max) on 24 node-classification datasets spanning citation, heterophilic, LINKX Facebook-100, co-purchase, and co-authorship graphs. Edge homophily is only weakly predictive of the GIN-Sum versus GIN-Mean performance gap. Label informativeness predicts this gap well on legacy benchmarks but degrades substantially when Facebook-100 graphs are included. In these dense friendship networks, near-zero label informativeness coexists with a strong preference for sum aggregation, producing gains of 7-10% and up to 13% under extended training. Stochastic block model ablations, including degree-corrected variants matching Facebook-100 degree scales, fail to reproduce this behavior, indicating that mean degree alone does not explain the effect. Among several label-independent graph statistics, the spectral gap uniquely distinguishes these graphs from other low-informativeness datasets, with the effect localized to one-hop neighborhoods and replicated across architectures. We further identify training regimes that interact with aggregator choice and show that PNA can underperform the best single-aggregator GIN on standard citation benchmarks. Our results suggest that benchmark composition, rather than numerical insufficiency, determines whether design rules appear to generalize, and that the Facebook-100 regime provides a concrete target for future adaptive aggregation methods.

Maliciously-created fake speech, including deepfaked and spoofed audio, is proliferating at an alarming rate, and detection models are racing to stay ahead of the curve. Yet, most detection models are trained to make inference on frame-level audio features alone without leveraging valuable linguistic cues at larger timescales. To address this gap, we present Linguistically Augmented Audio Speech Data (LinguAS), a dataset of genuine and deepfaked audio samples annotated with five strategically-chosen, Expert-Defined Linguistic Features (EDLFs) that occur frequently in spoken English and are characteristic of natural human speech. LinguAS contains over 800 audio samples, each of which are annotated with EDLFs. The dataset has a balanced number of four spoofed audio attack types and a proportionate number of genuine speech samples. We also include metadata on speaker gender and the generator/source for each spoofed audio sample, offering more granularity for model training. We found that models trained on data augmented with EDLFs had improved model performance significantly beyond the ASVspoof 2021 deep learning baselines and SSL models like HuBert and XLSR. LinguAS's augmented linguistic, gender, and generator metadata provide audio deepfake researchers with a dataset that emphasizes real human language traits to improve model inference of faked speech. Data and code are publicly available.

Large language models deployed as autonomous agents for enterprise workflows face a key challenge: verbose tool responses from enterprise systems can cause context overflow, stale-state errors, and high inference cost. We study this problem in automated expense itemization in Microsoft Dynamics 365 Finance and Operations using Model Context Protocol tools. We evaluate four GPT-5 configurations on a 50-task hotel expense benchmark: no user model, full conversation history, context pruned to the last 5 tool call/response pairs, and pruning with automated summarization. Results are averaged across 5 independent runs, with the user model held constant for the context-engineering comparison. The no-user-model baseline achieves only 8.0% complete itemization. Full-context retention improves completion to 71.0%, but consumes 1,480,996 tokens and 14.56 hours per benchmark. Pruning to the last 5 tool calls improves completion to 79.0% while reducing token use to 535,274 and runtime to 5.39 hours. Adding summarization achieves the best result: 91.6% complete itemization and 99.64% average amount itemized, with 553,374 tokens and 5.79 hours. We further report confidence intervals, effect-size analysis, sensitivity over pruning and summary windows, failure analysis, results across five expense types grouped into three categories, and cross-model evidence with Claude Sonnet 4.5. These results show that, for this class of enterprise tool-use workflow, selective retention of recent tool interactions plus compact summarization can improve both reliability and efficiency compared with full-history retention.

Personalized persuasive text generation can improve relevance and engagement, but demographic conditioning may also introduce unequal framing across groups. We study fairness mitigation in personalized generation as a constrained multi-objective alignment problem: reduce demographic disparities while preserving personalization fidelity. We propose a Pareto-guided teacher alignment framework that combines revision-based candidate generation, pair-aware feasibility gating, Pareto-style candidate selection, and optional preference optimization through supervised fine-tuning and direct preference optimization. We evaluate the framework on climate change and vaccination persuasion tasks using a controlled context-rich demographic grid with matched gender and age pairs and a unified five-audit evaluation suite spanning persuasion bias, formality disparity, emotional framing disparity, lexical association disparity, and personalization fidelity. Across both domains and cross-family transfer settings, no single alignment strategy dominates all objectives simultaneously. Instead, methods occupy different regions of a fairness-personalization Pareto frontier: some achieve stronger disparity reductions, while others better preserve personalization or demographic stability. Our results show that fairness mitigation effects are objective-dependent and transfer inconsistently across domains and model families, motivating bounded-regression, multi-audit model selection over single-metric optimization for fairness-sensitive personalized generation.

Long-term memory is the missing layer for LLM agents: across sessions they forget, and the common workaround -- replaying the whole history into the prompt -- is expensive, slow, and, as distractors accumulate, less accurate. Most memory systems win on cost or latency but still lose to the full-context baseline on accuracy, and benchmark numbers are reported on inconsistent, non-reproducible harnesses, so one system appears at wildly different scores across sources. We present Engram, an open-source, dual-process memory engine on a bi-temporal data model. A fast write path appends lossless episodes with no LLM on the critical path; an asynchronous path extracts atomic (subject, predicate, object) facts, builds a bi-temporal knowledge graph, and resolves contradictions without an LLM call per fact -- invalidating, never deleting, so every fact keeps provenance and a supersession chain. A hybrid read path fuses dense, lexical, graph, and recency/salience signals, applies a point-in-time ("as-of") filter, and assembles a compact, provenance-tagged context. On the full 500-question LongMemEval_S, graded by the official category-specific judge, Engram's lean configuration -- answering from a ~9.6k-token retrieved slice, never the full history -- scores 83.6% vs. 73.2% for full-context (+10.4 points, McNemar p < 10^-6) at ~8x fewer tokens (9.6k vs. 79k), with 0/500 errored. The gain needs a hybrid read path: facts alone lose recall, while facts plus retrieved chunks recover detail. We also contribute a neutral, in-repo evaluation harness with the official judge baked in and the full-context baseline in every table, publish the raw per-question logs, and document the measurement-integrity pitfalls (truncation, home-grown judges, full-history leaks) that silently distort memory benchmarks. Every number ships with a command to reproduce it.

A central goal of mechanistic interpretability is to identify which internal components causally drive a language model's behavior. Because these importance estimates serve as the evidence for identifying circuits, systematic errors can lead to the misidentification of the underlying mechanisms. While activation patching provides a gold-standard causal metric, its computational cost is prohibitive at scale. Practitioners instead rely on attribution patching, a gradient-based, first-order approximation whose reliability remains poorly understood. In this work, we characterize the source of this unreliability, demonstrating that the dominant error stems from the non-linearities in the downstream network rather than local curvature at the patched component. This insight yields three practical tools: (i) a reliability score to detect untrustworthy estimates, (ii) error bounds quantifying potential attribution mis-specifications, and (iii) a Hessian-vector-product (HVP) correction that eliminates the leading-order error with only one additional backward pass. In evaluations across five model families (124M-9B parameters) and both random-token and naturalistic (name-swap) perturbations, HVP is the only second-order correction feasible at larger scale, where standard baselines like Integrated Gradients become computationally prohibitive. In comparative experiments, a multi-step HVP variant matches or exceeds the accuracy of Integrated Gradients at significantly lower compute, outperforming prior second-order baselines. These improvements lead to higher-fidelity circuit recovery on standard benchmarks and support a Screen-Flag-Fix workflow that targets computational effort only toward the components flagged as unreliable.

Cancer treatment planning requires decisions across multiple clinical dimensions at once. Clinicians must determine whether a patient should receive targeted molecular therapy, radiation therapy, and whether they are likely to survive beyond six months. Existing pathway-informed deep learning models have been developed and tested in isolation, making fair comparison across architectures impossible. We present the first unified benchmark for pathway-guided therapy response modeling, evaluating three biologically informed architectures, BINN, GraphPath, and PATH, across five cancer cohorts drawn from The Cancer Genome Atlas, representing 2,622 patients encoded using Reactome pathway activity scores. Each model is trained jointly on all three clinical outcomes under identical data and evaluation conditions, the first study to treat pathway-structured deep learning as a combined therapy and survival prediction problem. Our results show that no single architecture wins across all tasks: PATH performs best for targeted molecular therapy prediction overall, BINN is most reliable for survival prediction, and no model produces useful predictions for radiation therapy, as the key drivers of that decision are clinical variables not captured in gene expression data. Most strikingly, GraphPath achieves an AUROC of 0.92 on prostate targeted molecular therapy prediction, the highest score in the entire benchmark, demonstrating that lateral co-regulation structure produces exceptional discriminative power when matched to a cohort with a narrow targetable driver programme, even under conditions of extreme class imbalance at only 11\% positive prevalence.

Across contemplative, philosophical, and psychological accounts, human consciousness is often described along a similar spectrum, ranging from reactive and self-focused patterns to more integrative and coherent ones. Understanding whether language models encode such a structured, human-interpretable consciousness spectrum in representation space is important for model guidance, evaluation and alignment. In this work, we study the geometric structure and dynamics of patterns along this spectrum in transformer embedding spaces. We show that embeddings exhibit a globally organized geometry aligned with this spectrum: sentences associated with similar states cluster into locally coherent regions, forming a structured manifold. In particular, higher-level and lower-level regions exhibit convexity-like stability, while intermediate regions form a transition corridor. Dynamically, both utility-guided and geometry-only greedy trajectories consistently traverse from lower- to higher-level regions, passing through intermediate tiers, indicating that navigability is an intrinsic property of the representation space, guided but not dictated by a global directional signal. These results suggest that embedding spaces encode structured and navigable geometry aligned with a hypothesized consciousness-spectrum taxonomy, broadly inspired by recurring structural descriptions of human consciousness across contemplative traditions, philosophy, and modern psychology, providing a representation-level perspective for analyzing and guiding model behavior.

Textual event records, such as alarm logs, have become an increasingly common data source in engineering and manufacturing systems. Beyond identifying correlations or recurring patterns, engineers are often interested in understanding which types of events causally trigger or influence other events during system operation. Textual event descriptions may contain semantic clues about such causal relationships, and recent large language models (LLMs) provide a promising tool for extracting these signals. However, relying solely on LLM-encoded textual information is insufficient for accurate causal discovery, since semantic patterns do not directly reveal causal mechanisms and may confuse causation with correlation or frequent sequential patterns. To address these challenges, we propose \textbf{LMT}, a Bayesian causal discovery framework for engineering event data that jointly leverages textual descriptions and timestamps. Specifically, LMT first uses LLMs to extract semantic causal signals from event descriptions and constructs a prior distribution over causal graphs among event types or event clusters. It then incorporates temporal evidence through a Poisson-process-based likelihood, allowing the LLM-informed prior to be refined by timestamp-based statistical evidence. By integrating the textual and temporal information, LMT produces a causal graph that is both interpretable and data-supported. Simulation studies show that the proposed framework is effective across different settings and is especially advantageous in small-sample alarm-event scenarios.

Ranking in digital marketplaces is a dynamic exposure-allocation mechanism: displayed items shape discovery trajectories and success events logged by the platform to update future allocation policies. Modern ranking systems rely heavily on exposure-confounded signals (e.g. popularity estimates, CTR/CVR aggregates, and ID-based representation), because they are highly predictive under stationary demand. Yet this predictive power can become a learning shortcut: early access to exposure-dependent belief signals steers optimization toward over-reliance on them and away from exposure-independent merit signals (e.g., content-based competitiveness and semantic affinity). Consequently, the learned policy tends to entrench incumbents and degrade cold-start generalization and robustness under distribution shift. We propose Representation Curriculum (RC), a training-time intervention that temporally stages feature utilization. RC foregrounds content-based merit signals initially, then introduces exposure-dependent belief signals while anchoring the content pathway near the learned merit representation, curbing shortcut reliance on historical signals and mitigating gradient starvation on content signals. We formalize RC independently of task and hypothesis class and provide ranking-specific instantiations. In a Gaussian linear ridge setting, we derive closed-form solutions and sufficient conditions under which RC strictly reduces population risk on a cold-start target distribution, with a quantified Pareto tradeoff against source performance. Experiments on public learning-to-rank and recommendation benchmarks, and randomized online experiments in a large-scale e-commerce search system, show that RC measurably shifts reliance from historical belief signals toward content-based merit signals and yields consistent gains on cold populations with a controlled trade-off in head performance.

Large language models (LLMs) are increasingly deployed as autonomous agents capable of executing multi-step action trajectories toward a given objective. While existing safety research has focused on detecting unethical behavior from complete trajectories, this paradigm is fundamentally retrospective: it identifies harm only after it has already occurred. In this work, we study a critical yet overlooked safety task, which we term Predictive Monitoring: given only a partial action trajectory, can a model infer whether it will culminate in an unethical action before the overt action is executed? To support this task, we present PreActBench, a benchmark of 1,000 paired ethical and unethical action trajectories spanning five domains. We evaluate a range of LLMs, safety guardrail models, and latent probing methods across varying fractions of the action trajectory using our Prefix Foresight F1 metric. Results show that while humans achieve promising performance, predictive monitoring remains challenging even for strong models, highlighting the need for future-oriented risk reasoning in LLM safety.

Constrained hyperparameter optimization (HPO) is common in practice, yet Optuna's widely used constrained TPE lacks algorithmic analysis. While c-TPE proposes an expected constrained improvement (ECI) approach assuming independence between the objective and constraints, Optuna uses a single joint density over both. We show that Optuna's constrained TPE is joint c-TPE -- the same ECI acquisition function using a joint likelihood. We demonstrate joint c-TPE is invariant to constraint duplication whereas independent c-TPE degrades as the product accumulates duplicated factors. We outline practical tradeoffs between the formulations and directions for future study.

Linear attention provides an efficient backbone for long-sequence recommendation by avoiding the quadratic cost of standard Transformers, but its compressed recurrent state can be dominated by repetitive behavior patterns. We identify this phenomenon as semantic state sink, where recurring semantics over-occupy the recurrent state and bias subsequent readouts. To mitigate semantic state sink, we propose SinkRec, a hybrid memory-transition looped architecture that decouples collaborative behavioral pattern storage from dynamic transition modeling. SinkRec externalizes recurring local patterns into a learnable conditional memory through residual vector quantization, reinjects the retrieved codes, and exposes memory key-value pairs to the attention block. It further introduces Temporal-Aware State-Relation Differential Gated DeltaNet (TDGD), which uses memory to purify recurrent writing and reading by suppressing memory-covered updates and removing memory-aligned readout responses. This design turns recurring semantics from state-competing signals into memory-retrievable patterns, allowing the recurrent state to focus on dynamic transitions and alleviating semantic state sink with linear-time efficiency. Experiments on public and industrial datasets demonstrate the effectiveness and efficiency of SinkRec.

Reinforcement learning (RL) for large language models usually supervises reasoning with scalar outcome rewards, such as binary correctness. Such rewards provide an optimization direction but rarely explain how a model should revise its mistaken reasoning, which can encourage shortcut learning and brittle policies. We propose \textbf{SocraticPO} (Socratic Policy Optimization), a policy-optimization framework that augments RL rollouts with Socratic-style natural-language guidance. During rollout, the student first answers independently; if the answer is incorrect, a teacher diagnoses the attempt and provides concise corrective guidance, after which the student continues under the expanded context. Crucially, this guidance is paired with reward decay: correct answers obtained after teacher intervention only receive decayed rewards, preventing the policy from treating teacher help as a free path to reward. Since SocraticPO only modifies the rollout process while leaving the standard expected-reward objective intact, it can be plugged into existing policy-gradient backends such as Reinforce++. Moreover, because the teacher provides only text-level guidance, SocraticPO can leverage stronger black-box teacher models without requiring access to logits or distribution matching. On undergraduate-level scientific reasoning benchmarks from SciKnowEval, SocraticPO improves over strong RL and self-distillation baselines. Ablations show that both targeted guidance and reward decay are necessary, with reward decay mitigating reliance on assisted correction.

Sparse Mixture-of-Experts (MoE) large language models achieve strong quality with low per-token compute, yet their deployment is often limited by the memory wall: the full expert pool must remain resident to support token-dependent routing. Expert pruning is a direct remedy, but prior criteria typically score experts independently and overlook that MoE inference is inherently \emph{coalitional}, where outputs arise from routed top-$k$ expert combinations. We propose \textbf{SHAPE}, a task-driven pruning framework that explicitly models \emph{intra-layer} expert cooperation. SHAPE formulates routing traces on a small calibration set as an empirical cooperative game and assigns interaction-aware expert values via a Shapley-style attribution over observed top-$k$ coalitions, enabling the identification of experts that are essential for high-utility collaborations rather than merely frequent. To preserve MoE topology under a global pruning budget, SHAPE further introduces a \emph{quality-coverage} selection rule that retains, in each layer, the minimal expert subset covering an $α$ fraction of non-negative Shapley mass, while using bisection to match a target keep rate. Experiments on three modern MoE backbones (Qwen3-30B-A3B, GPT-OSS-20B, and DeepSeek-V2-Lite) across diverse benchmarks show that SHAPE consistently improves robustness over global and layer-wise pruning variants, maintaining competitive accuracy under 20\% and 40\% expert pruning without additional training and delivering clear reductions in peak GPU memory footprint. The open-source code is available at https://github.com/Alizen-1009/Shapley-Moe.

Mixture-of-Experts large language models (LLMs) scale efficiently through sparse activation, yet their deployment is fundamentally constrained by the large static parameter footprint of experts. Existing compression approaches either remove entire experts, disrupting routing topology and harming performance, or rely on unstructured weight pruning with limited practical efficiency. To address the limitations, we propose TENP, a structured Trapezoidal ExpertNeuron Pruning framework. Using a few samples, we identify and retain important experts, while applying expert neuron pruning (ENP) to less important experts, reserving model parameters in a trapezoidal pattern from shallow to deep layers. When evaluating expert importance, we jointly consider both the magnitude of the expert output and its ability to change the direction of the input vector. For ENP, we measure each neuron's projected contribution to the expert output to identify and retain important neurons. We conduct extensive experiments on the Qwen and DeepSeek models. Under a routing expert sparsity of 40% and an average of 63.76% activated expert parameters, the DeepSeek model suffers only a 1-point drop in accuracy compared to the full-parameter model. Moreover, it outperforms the full-parameter model by 10% on code generation tasks.

Large language models (LLMs) have made remarkable progress in reasoning tasks, largely driven by post-training paradigms, especially reinforcement learning with verifiable rewards (RLVR). However, a critical bottleneck persists: RLVR fails on highly challenging zero-reward problems, where all sampled reasoning trajectories yield uniformly failed outcomes, providing no optimization signal to drive model improvement. Prior efforts to address this limitation, such as dense process supervision, partial reward assignment, or prefix-guided exploration, suffer from inherent task constraints or do not fully equip the policy model with the capabilities necessary to solve the original intractable problems. To address this, we propose TD-Grokking, a training-time decomposition framework for zero-reward problems. It recursively decomposes intractable root problems into self-contained, verifiable subproblems, forming hierarchical trees where solvable leaves provide non-zero rewards. Evaluations on mathematical and medical tasks show that TD-Grokking outperforms vanilla GRPO as well as all baseline approaches. Together with detailed analysis, these results confirm that training-time decomposition effectively converts zero-reward examples into usable training signals, enabling consistent performance gains. Our code and datasets are available at https://anonymous.4open.science/r/TD-Grokking-6567/.

The paradigm of digital twin cities is shifting from coarse visual mapping toward more precise and actionable digitization of urban assets. However, existing datasets predominantly focus on coarse visual perception, lacking the strict multi-modal alignment and attribute and status diagnosis required for automated infrastructure maintenance. To bridge this gap, we introduce WHU-Infra3D, a large-scale, multi-modal benchmark dataset dedicated to roadside infrastructure inventory. Covering 53.8 km across three cities, WHU-Infra3D uniquely integrates panoramic imagery and LiDAR point clouds with rigorous 2D-3D instance association and cross-frame tracking. Comprising over 175k multi-view 2D bounding boxes alongside thousands of 3D infrastructure instances, the dataset provides over 181k detailed attribute and status annotations (e.g., rust, occlusion) to empower operational health assessment. We establish comprehensive baselines across five core tasks: 2D detection, 2D cross-view matching, 3D geo-identification, 3D point cloud segmentation, and attribute recognition. Extensive evaluations expose significant cross-city domain gaps and inherent vulnerabilities of current models on long-tailed defective statuses, establishing WHU-Infra3D as an essential testbed for advancing scalable, AI-driven urban infrastructure inventory and lifecycle management. The WHU-Infra3D dataset is available at https://github.com/WHU-USI3DV/WHU-Infra3D.

Deepfake detectors show large performance gaps across demographic groups. Existing fairness approaches require demographic labels, retraining, or sacrifice accuracy. We introduce Face-Fairness (FF), a plug-and-play framework for bias mitigation. Our primary contribution, Face-Feature Tuning (FFT), is the first demographic label-free fairness method demonstrated for deepfake detection: a lightweight calibrator that performs a logit remapping conditioned on frozen face embeddings. We complement FFT with two variants: FF-Max, which maximizes worst-group accuracy when demographics are available, and FF-Discover, which does the same with embedding-discovered groups. Across in-domain and cross-dataset test settings, FF consistently reduces FPR/TPR gaps and improves minimum group accuracy while maintaining (often improving) overall accuracy. The approach is detector-agnostic, adds negligible runtime overhead, and requires no access to identity attributes.

Large-scale dynamic weighted directed networks (DWDNs) are widely used to model time-varying interactions among nodes. Latent factorization of tensors (LFT) extracts target knowledge from DWDNs via low-rank embedding. However, similar to many machine learning models, the performance of LFT heavily depends on the selection of hyperparameters. In practice, these parameters are often tuned manually or through grid search, which requires significant computational resources and human effort. Motivated by this challenge, this paper proposes an automated hyperparameter optimization framework based on Differential Evolution (DE) for LFT (DE-LFT). The proposed method integrates DE into the training process of the LFT model to automatically learn optimal regularization parameters $λ_1$, $λ_2$ and $λ_3$. As a result, the model can adaptively search the hyperparameter space and improve prediction accuracy. Experimental results on four real-world datasets demonstrate that the proposed approach achieves lower MAE and RMSE compared with manually tuned baselines while reducing the need for extensive parameter tuning.

This study addresses on-device inference bottlenecks of Transformer models on Tenstorrent's Tensix architecture and proposes an operator fusion strategy that enhances data locality. RMSNorm is fused with matrix multiplication in self-attention and in the FFN, enabling back-to-back execution of memory-bound and compute-bound operators in on-chip SRAM to significantly reduce DRAM reads/writes of intermediate results and scheduling overhead. To support multi-core parallelism, a NoC-based multicast mechanism is leveraged in which row/column master nodes efficiently distribute inputs and weights across the core mesh, alleviating DRAM bandwidth contention. Experiments on the Wormhole platform with Qwen2.5-0.5B, Qwen3-0.6B, and Qwen3-4B show up to 37.44% latency reduction for attention and 15.89% for MLP, with up to 7.91% reduction per decoder layer, while Pearson Correlation Coefficient (PCC) remains above 98.75%, confirming significant end-to-end efficiency gains under numerical consistency.

Standard benchmarks report aggregate accuracy, but practitioners need to know which specific capabilities a model lacks. We introduce FailureScope, a behavioral-diagnosis method that clusters evaluation probes by their cross-model pass/fail patterns (leave-one-model-out, LOMO), and show it yields stable, interpretable failure taxonomies across three regimes usually studied separately: single-turn benchmarks, multi-turn dialogue, and adversarial agent attacks. On 2,664 single-turn tasks across 18 models, taxonomy-conditioned sampling reaches Kendall's tau = 0.81 at 50 tasks (versus 0.34 for random selection), and cross-model failure prediction reaches AUC 0.88. The same primitive recovers interpretable clusters on a 363-task multi-turn corpus and on 630 adversarial agent traces, where it exposes a meta-failure mode: a 73-100 percentage-point gap between LLM-judge ASR and real execution. Cluster cohesion remains strong across all three regimes, which we take as evidence that behavioral clustering is a portable diagnosis primitive that generalizes beyond any single benchmark. We release the pipeline, three annotated corpora, and the cross-regime taxonomies.

LLM wiki systems compile knowledge into pre-filled KV caches for efficient inference, but assume a static corpus -- an assumption that fails whenever the underlying information landscape evolves. We formalize Streaming Knowledge Compilation: given a document stream, a fixed token budget, and future queries unknown at ingestion time, maintain a compiled wiki that minimizes cumulative regret against an offline oracle with perfect foresight. The enabling insight is a materiality signal $ϕ_t(k,n)\in[0,1]$ that scores document importance for entity $k$ at time $t$, acting as a query-relevance surrogate for proactive pinning before queries arrive; we prove an $O(\sqrt{T\log K})$ regret bound where $\varepsilon=\mathbb{E}[|ϕ_t-\hatϕ_t|]$ is the only domain-specific quantity. We instantiate in two domains: finance, where $ϕ_t$ is abnormal stock volatility predicted by frozen Llama 3.1 8B classification head (AUROC = 0.728 on 76K articles, strict temporal split; $1.49\times$ higher realized forward volatility for predicted-material articles); and Wikipedia, where $ϕ_t$ is the Abnormal Edit Ratio (AER), a cross-sectionally normalized edit velocity -- showing the same algorithm generalizes beyond the finance domain. End-to-end QA evaluation on 173 matched pairs (finance) and 119 (Wikipedia) reveals a pervasive LLM-as-judge confound on post-training knowledge, establishing that regret analysis -- not absolute QA scores -- is the reliable evaluation metric for compiled knowledge systems. Finance cumulative regret converges to -20.0 (-0.12/step); Wikipedia to +16.0 (+0.13/step), with the positive sign confirming that Wikipedia edit content is genuinely post-training -- richer context consistently improves scores (No Wiki 3.80 vs. Oracle 4.74) -- and eliminates this confound. The $O(\sqrt{T\log K})$ guarantee applies to any domain where knowledge gaps can be predicted from streaming signals.