Large Language Models (LLMs) are increasingly used to model user preferences, with the typical output as a directly-generated ranked item list per user. However, this generative paradigm inherits the bias and opacity of autoregressive decoding. It over-emphasizes frequent (head) preferences and suppresses minority, long-tail ones. To address this, we propose SPECTRA (Softmax Probing for Extracted Category-level Token Readouts and Analysis), which treats the finetuned LLM as an implicit probabilistic model and probes its softmax to infer a probability distribution over semantically interpretable preference categories. We evaluate SPECTRA on MovieLens, Yelp, and a large-scale short-video platform. SPECTRA delivers (i) distributional alignment, reducing Jensen-Shannon divergence to the empirical preference distribution by 38 to 44 percent across public datasets; (ii) long-tail recovery with cross-user fairness, raising top-3 category exposure entropy by 23 percent on MovieLens and producing a larger gain on tail-preference users than on head-preference users; and (iii) downstream application value, with a 41 to 46 percent category-NDCG boost on MovieLens and Yelp, and a 7x improvement on long-tail category ranking on a large-scale deployment against a head-optimized production ranker.

Large reasoning models (LRMs) have attracted much attention due to their exceptional performance. However, their performance mainly stems from thinking, a long Chain of Thought (CoT), which significantly increase computational overhead. To address this overthinking problem, existing work focuses on using reinforcement learning (RL) to train hybrid reasoning models that automatically decide whether to engage in thinking or not based on the complexity of the query. Unfortunately, using RL will suffer the the reward hacking problem, e.g., the model engages in thinking but is judged as not doing so, resulting in incorrect rewards. To mitigate this problem, existing works either employ supervised fine-tuning (SFT), which incurs high computational costs, or enforce uniform token limits on non-thinking responses, which yields limited mitigation of the problem. In this paper, we propose Thinking-Based Non-Thinking (TNT). It does not employ SFT, and sets different maximum token usage for responses not using thinking across various queries by leveraging information from the solution component of the responses using thinking. Experiments on five mathematical benchmarks demonstrate that TNT reduces token usage by around 50% compared to DeepSeek-R1-Distill-Qwen-1.5B/7B and DeepScaleR-1.5B, while significantly improving accuracy. In fact, TNT achieves the optimal trade-off between accuracy and efficiency among all tested methods. Additionally, the probability of reward hacking problem in TNT's responses, which are classified as not using thinking, remains below 10% across all tested datasets.

Infographics are composite visual artifacts that combine data visualizations with textual and illustrative elements to communicate information. While recent text-to-image (T2I) models can generate aesthetically appealing images, their reliability in generating infographics remains unclear. Generated infographics may appear correct at first glance but contain easily overlooked issues, such as distorted data encoding or incorrect textual content. We present IGENBENCH, the first benchmark for evaluating the reliability of text-to-infographic generation, comprising 600 curated test cases spanning 30 infographic types. We design an automated evaluation framework that decomposes reliability verification into atomic yes/no questions based on a taxonomy of 10 question types. We employ multimodal large language models (MLLMs) to verify each question, yielding question-level accuracy (Q-ACC) and infographic-level accuracy (I-ACC). We comprehensively evaluate 10 state-of-the-art T2I models on IGENBENCH. Our systematic analysis reveals key insights for future model development: (i) a three-tier performance hierarchy with the top model achieving Q-ACC of 0.90 but I-ACC of only 0.49; (ii) data-related dimensions emerging as universal bottlenecks (e.g., Data Completeness: 0.21); and (iii) the challenge of achieving end-to-end correctness across all models. We release IGENBENCH at https://igen-bench.vercel.app/.

We present Muses, the first training-free method for fantastic 3D creature generation in a feed-forward paradigm. Previous methods, which rely on part-aware optimization, manual assembly, or 2D image generation, often produce unrealistic or incoherent 3D assets due to the challenges of intricate part-level manipulation and limited out-of-domain generation. In contrast, Muses leverages the 3D skeleton, a fundamental representation of biological forms, to explicitly and rationally compose diverse elements. This skeletal foundation formalizes 3D content creation as a structure-aware pipeline of design, composition, and generation. Muses begins by constructing a creatively composed 3D skeleton with coherent layout and scale through graph-constrained reasoning. This skeleton then guides a voxel-based assembly process within a structured latent space, integrating regions from different objects. Finally, image-guided appearance modeling under skeletal conditions is applied to generate a style-consistent and harmonious texture for the assembled shape. Extensive experiments establish Muses' state-of-the-art performance in terms of visual fidelity and alignment with textual descriptions, and potential on flexible 3D object editing. Project page: https://luhexiao.github.io/Muses.github.io/.

Every existing inference-time reasoning framework discards all failure context at problem boundaries, leaving a model solving problem 500 no wiser than it was on problem 1. We present ReTreVal (Reasoning Tree with Validation), a training-free framework that closes this gap through adaptive tree exploration with tool-augmented node refinement, typed-failure backtracking that injects categorized error context into the recovered branch, and a self-rewriting memory that accumulates and revises strategy entries across problems, enabling inference-time cross-problem learning on any fixed, unmodified LLM without fine-tuning. ReTreVal achieves 85.8% pass@1 on MATH-500 (+8.6 pp over Zero-Shot CoT, +8.6 pp over the strongest baseline Self-Refine) and 54.4% on MMLU-Pro (+15.3 pp over Self-Refine), with a 3.4:1 win-to-regression ratio confirming genuine error recovery rather than noise. These capabilities, previously requiring gradient updates, allow a 32B model to compete with much larger single-pass systems.

Deep reinforcement learning (DRL) has shown great promise in addressing multi-objective combinatorial optimization problems (MOCOPs). Nevertheless, the robustness of these learning-based solvers has remained insufficiently explored, especially across diverse and complex problem distributions. In this paper, we propose a unified robustness-oriented framework for preference-conditioned DRL solvers for MOCOPs. Within this framework, we develop a preference-based adversarial attack to generate hard instances that expose solver weaknesses, and quantify the attack impact by the resulting degradation on Pareto-front quality. We further introduce a defense strategy that integrates hardness-aware preference selection into adversarial training to reduce overfitting to restricted preference regions and improve out-of-distribution performance. The experimental results on multi-objective traveling salesman problem (MOTSP), multi-objective capacitated vehicle routing problem (MOCVRP), and multi-objective knapsack problem (MOKP) verify that our attack method successfully learns hard instances for different solvers. Furthermore, our defense method significantly strengthens the robustness and generalizability of neural solvers, delivering superior performance on hard or out-of-distribution instances.

Anomaly detection is a core capability for robotic perception and industrial inspection, yet most existing benchmarks are collected under controlled conditions with fixed viewpoints and stable illumination, failing to reflect real deployment scenarios. We introduce RAD (Realistic Anomaly Detection), a robot-captured, multi-view dataset designed to stress pose variation, reflective materials, and viewpoint-dependent defect visibility. RAD covers 13 everyday object categories and four realistic defect types--scratched, missing, stained, and squeezed--captured from over 60 robot viewpoints per object under uncontrolled lighting. We benchmark a wide range of state-of-the-art approaches, including 2D feature-based methods, 3D reconstruction pipelines, and vision-language models (VLMs), under a pose-agnostic setting. Surprisingly, we find that mature 2D feature-embedding methods consistently outperform recent 3D and VLM-based approaches at the image level, while the performance gap narrows for pixel-level localization. Our analysis reveals that reflective surfaces, geometric symmetry, and sparse viewpoint coverage fundamentally limit current geometry-based and zero-shot methods. RAD establishes a challenging and realistic benchmark for robotic anomaly detection, highlighting critical open problems beyond controlled laboratory settings.

Developing artificial agents that unify representation, memory, adaptation, and prediction remains a fundamental challenge in artificial intelligence. Here we introduce a geometric framework in which cognitive computation emerges from Riemannian gradient flow on a learned latent manifold. The learned metric encodes representational constraints and computational preferences, while anisotropies in the geometry naturally generate multiple timescales of behaviour, yielding both rapid reactive responses and slower adaptive dynamics without explicit memory modules or recurrent mechanisms. We instantiate this framework through Riemannian representation and dynamics models and evaluate them in partially observable reinforcement-learning environments. Across observation masking, sensory blackouts, dynamics perturbations, and predictive latent-modelling tasks, the proposed approach consistently outperforms feedforward baselines, achieves robustness comparable to recurrent architectures, and produces highly predictable latent trajectories with low long-horizon rollout error. These results suggest that learned latent geometry can serve simultaneously as a substrate for representation, memory, adaptation, and prediction. More broadly, the framework provides a principled connection between dynamical systems, representation learning, and world-model-based intelligence.

Most federated learning (FL) approaches assume a fixed device set. However, real-world scenarios often involve devices dynamically joining or leaving the system, driven by, e.g., user mobility patterns or handovers across cell boundaries. This dynamic setting introduces unique challenges: (1) the optimization objective evolves with the active device set, unlike traditional FL's static objective; and (2) the current global model may no longer serve as an effective initialization for subsequent rounds, potentially hindering adaptation, delaying convergence, and reducing resource efficiency. To address these challenges, we first provide a convergence analysis for FL under a dynamic device set, accounting for factors such as gradient noise, local training iterations, and data heterogeneity in this practical setting. Motivated by this analysis, we propose a model initialization algorithm that enables rapid adaptation whenever devices join or leave the network. Our key idea is to compute a weighted average of previous global models, guided by gradient similarity, to prioritize models trained on data distributions that closely align with the current device set, thereby accelerating recovery from distribution shifts in fewer training rounds. This plug-and-play algorithm is designed to integrate seamlessly with existing FL methods, offering broad applicability. Experiments demonstrate that our approach achieves convergence speedups typically an order of magnitude or more compared to baselines, which we show drastically reduces energy consumption to reach a target accuracy.

LLMs have shown the capacity to improve their performance on reasoning tasks through reflecting on their mistakes, and acting with these reflections in mind. However, continual reflections of the same LLM onto itself exhibit degeneration of thought, where the LLM continues to repeat the same errors again and again even with the knowledge that its wrong. To address this problem, we instead introduce multi-agent with multi-persona debators as the method to generate reflections. Through out extensive experimentation, we've found that the leads to better diversity of in the reflections generated by the llm agent. We demonstrate an accuracy of 47% EM HotPot QA (question answering) and 82.7% on HumanEval (programming), both performances surpassing reflection with a single llm.

The attention mechanism in a Transformer architecture matches key to query based on both content -- the what -- and position in a sequence -- the where. We present an analysis indicating that what and where are entangled in the popular RoPE rotary position embedding. This entanglement can impair performance particularly when decisions require independent matches on these two factors. We propose an improvement to RoPE, which we call Polar Coordinate Position Embeddings or PoPE, that eliminates the what-where confound. PoPE is far superior on a diagnostic task requiring indexing solely by position or by content. On autoregressive sequence modeling in music, genomic, and natural language domains, Transformers using PoPE as the positional encoding scheme outperform baselines using RoPE with respect to evaluation loss (perplexity) and downstream task performance. On language modeling, these gains persist across model scale, from 124M to 774M parameters. Crucially, PoPE shows strong zero-shot length extrapolation capabilities compared not only to RoPE but even a method designed for extrapolation, YaRN, which requires additional fine tuning and frequency interpolation.

Current jailbreak attacks on large language models (LLMs) predominantly rely on LLMs themselves to generate adversarial prompts, creating a critical efficiency bottleneck: each attack requires substantial computational resources and API queries, limiting scalability and practical deployment. To overcome this limitation, we propose Adversarial Prompt Distillation (APD), a novel framework that transfers jailbreaking capabilities from LLMs to small language models (SLMs) for efficient, low-resource attacks. APD integrates three key components: (1) masked adversarial knowledge pre-training via LoRA fine-tuning, (2) dynamic temperature-controlled knowledge distillation to bridge architectural gaps, and (3) reinforcement learning-based template optimization for adaptive refinement. Extensive experiments across 12 models show that APD achieves state-of-the-art attack success rates (e.g., 96.4% ASR_k on GPT-4) while dramatically improving efficiency - generating prompts 3.7x faster with 11.3x fewer parameters than teacher models. Our work establishes the first practical framework for lightweight jailbreak attacks, exposes new vulnerabilities in LLM defenses, and provides a scalable testbed for advancing AI safety research. Our code is available at: https://github.com/lxgem/Efficient_and_Stealthy_Jailbreak_Attacks_via_Adversarial_Prompt.

Development of reliable and physically interpretable probabilistic frameworks for industrial prognostics remain nascent, and existing literature is often insensitive as inputs move away from the training manifold. In this paper, we develop two sampling-free, distance-aware physics-constrained probabilistic frameworks: (i) PC-SNGP and (ii) PC-SNER. Both apply spectral normalization to hidden layer weights, enforcing bi-Lipschitz distance-preserving representation from the input to the latent space. PC-SNGP replaces the dense output with Gaussian process whose posterior variance increases with input distance from the training manifold. PC-SNER modifies the output layer to predict Normal-Inverse-Gamma~(NIG) parameters for distance preserving estimation. To maintain balance between data fidelity and physical consistency during training, we introduce a dynamic weighting strategy for the physics-constrained loss. We also introduce a distance-aware-coefficient~(DAC) metric to quantify sensitivity to distributional shifts. Empirically, we validate both frameworks on rolling-element-bearings (REBs) prognostics using the PRONOSTIA, XJTU-SY, and HUST benchmark datasets. Experimental results demonstrate improved prediction accuracy and well-calibrated uncertainty estimates relative to competing baselines, while maintaining auditable performance in cross-validation and robustness under extreme adversarial perturbations.

We propose a collaborative edge-to-server inference framework for vision-language models (VLMs) that reduces communication cost while maintaining inference accuracy. In typical deployments, visual data captured at edge devices (clients) is transmitted to the server for VLM inference. However, transmitting full-resolution images incurs high communication cost. Conversely, aggressive downsizing or excessive compression to mitigate communication overhead can discard fine-grained details, leading to accuracy degradation. To overcome this limitation, we design a communication-efficient two-stage framework. In the first stage, the server performs inference on the downsized thumbnail (global image) and quantifies the min-entropy of the output tokens. If the min-entropy exceeds a predefined threshold, the server identifies a region of interest (RoI) using the VLM's internal attention and requests the edge device to send a detail-preserved local image of the RoI. The server then refines its inference by jointly leveraging the global and local images. This selective retransmission strategy ensures that only essential visual content is additionally transmitted. Experimental results consistently confirm that the proposed framework substantially reduces communication overhead while maintaining inference accuracy across diverse VQA benchmarks.

Multivariate time series imputation is fundamental in applications such as healthcare, traffic forecasting, and biological modeling, where sensor failures and irregular sampling lead to pervasive missing values. However, existing Transformer- and diffusion-based models lack explicit inductive biases and frequency awareness, limiting their generalization under structured missing patterns and distribution shifts. We propose FADTI, a diffusion-based framework that injects frequency-informed feature modulation via a learnable Fourier Bias Projection (FBP) module and combines it with temporal modeling through self-attention and gated convolution. FBP supports multiple spectral bases, enabling adaptive encoding of both stationary and non-stationary patterns. This design injects frequency-domain inductive bias into the generative imputation process. Experiments on multiple benchmarks, including a newly introduced biological time series dataset, show that FADTI consistently outperforms state-of-the-art methods, particularly under high missing rates. Code is available at https://anonymous.4open.science/r/TimeSeriesImputation-52BF

Conventional soft pneumatic actuators, typically based on hollow elastomeric chambers, often suffer from small structural support and require costly geometry-specific redesigns for multimodal functionality. Porous materials such as foam, filled into chambers, can provide structural stability for the actuators. However, methods to achieve programmable deformation by tailoring the porous body itself remain underexplored. In this paper, a novel design method is presented to realize soft porous actuators with programmable deformation by incising specific patterns into the porous foam body. This approach introduces localized structural anisotropy of the foam guiding the material's deformation under a global vacuum input. Furthermore, three fundamental patterns on a cylindrical foam substrate are discussed: transverse for bending, longitudinal for tilting, and diagonal for twisting. A computational model is built with Finite Element Analysis (FEA), to investigate the mechanism of the incision-patterning method. Experiments demonstrate that with a potential optimal design of the pattern array number N, actuators can achieve bending up to $80^{\circ}$ (N=2), tilting of $18^{\circ}$ (N=1), and twisting of $115^{\circ}$ (N=8). The versatility of our approach is demonstrated via pattern transferability, scalability, and mold-less rapid prototyping of complex designs. As a comprehensive application, we translate the human hand crease map into a functional incision pattern, creating a bio-inspired soft robot hand capable of human-like adaptive grasping. Our work provides a new, efficient, and scalable paradigm for the design of multi-functional soft porous robots.

We present DIJIT, a novel binocular robotic head expressly designed for mobile agents that behave as active observers. DIJIT's unique breadth of functionality enables active vision research and the study of human-like eye and head-neck motions, their interrelationships, and how each contributes to visual ability. DIJIT is also being used to explore the differences between how human vision employs eye/head movements to solve visual tasks and current computer vision methods. DIJIT's design features nine mechanical degrees of freedom, while the cameras and lenses provide an additional four optical degrees of freedom. The ranges and speeds of the mechanical design are comparable to human performance. DIJIT attains 85\% of the peak human saccade speed. Our design includes the ranges of motion required for convergent stereo, namely, vergence, version, and cyclotorsion. Here, we present DIJIT and some aspects of its performance. We also present a novel method for saccadic camera movements, using a direct relationship between camera orientation and motor values. The resulting saccadic camera movements are close to human movements in terms of their accuracy, with 1.17$^\circ$ and 1.14$^\circ$ mean error for the left and right cameras, respectively.

Two traditions of interpretability have evolved side by side but seldom spoken to each other: Concept Bottleneck Models (CBMs), which prescribe what a concept should be, and Sparse Autoencoders (SAEs), which discover what concepts emerge. While CBMs use supervision to align activations with human-labeled concepts, SAEs rely on sparse coding to uncover emergent ones. We show that both paradigms instantiate the same geometric structure: each learns a set of linear directions in activation space whose nonnegative combinations form a concept cone. Supervised and unsupervised methods thus differ not in kind but in how they select this cone. Building on this view, we propose an operational bridge between the two paradigms. CBMs provide human-defined reference geometries, while SAEs can be evaluated by how well their learned cones approximate or contain those of CBMs. This containment framework yields quantitative metrics linking inductive biases -- such as SAE type, sparsity, or expansion ratio -- to emergence of plausible\footnote{We adopt the terminology of \citet{jacovi2020towards}, who distinguish between faithful explanations (accurately reflecting model computations) and plausible explanations (aligning with human intuition and domain knowledge). CBM concepts are plausible by construction -- selected or annotated by humans -- though not necessarily faithful to the true latent factors that organise the data manifold.} concepts. Using these metrics, we uncover a ``sweet spot'' in both sparsity and expansion factor that maximizes both geometric and semantic alignment with CBM concepts. Overall, our work unifies supervised and unsupervised concept discovery through a shared geometric framework, providing principled metrics to measure SAE progress and assess how well discovered concept align with plausible human concepts.

Accurate marine wind forecasts are essential for safe navigation, ship routing, and energy operations, yet they remain challenging because observations over the ocean are sparse, heterogeneous, and temporally variable. We present an observation-informed correction approach for global numerical weather prediction (NWP) of marine winds. Rather than forecasting winds directly, we learn local correction patterns by assimilating the latest in-situ observations to adjust the Global Forecast System (GFS) output. We propose ORCA (Observation-informed Real-time Correction with Attention), a transformer-based deep learning architecture that (i) handles irregular and time-varying observation sets through masking and set-based attention mechanisms, (ii) conditions predictions on recent observation--forecast pairs via cross-attention, and (iii) employs cyclical time embeddings and coordinate-aware location representations to enable single-pass inference at arbitrary spatial coordinates. We evaluate ORCA over the Atlantic Ocean using observations from the International Comprehensive Ocean-Atmosphere Data Set (ICOADS) as reference. ORCA reduces GFS 10-meter wind error at all lead times up to 48 hours, achieving 45% improvement at 1-hour lead time and 13% improvement at 48-hour lead time. Spatial analyses reveal the most persistent improvements along coastlines and shipping routes, where observations are most abundant. The tokenized architecture naturally accommodates heterogeneous observing platforms (ships, buoys, tide gauges, and coastal stations) and produces both site-specific predictions and basin-scale gridded products in a single forward pass. These results demonstrate a practical, low-latency post-processing approach that complements NWP by learning to correct systematic forecast errors.

Stochastic Variance Reduced Gradient (SVRG) and its variants aim to speed-up training by using gradient corrections. Originally proposed over a decade ago, these methods have never been connected to any Bayesian method at a fundamental level. Here, we fill this gap and derive surprising new connections of SVRG to a recently proposed Bayesian method called `posterior correction'. Our main contribution is to show that SVRG can be recovered as a special case of posterior-correction over isotropic-Gaussian posteriors. Novel extensions of SVRG are automatically obtained by using more flexible exponential-family posteriors. We derive two new such extensions by using Gaussian families: a Newton-like variant with novel Hessian corrections, and an Adam-like extension that scales to large problems. Our work is the first to connect SVRG to Bayes and use it to speed-up training.

Controlling autonomous systems under real-world conditions often requires policies that can be evaluated with low latency and minimal energy consumption. Unfortunately, these conditions are at odds with the use of high-precision deep neural networks as controllers. In this work, we introduce Differentiable Weightless Controllers (DWCs), a symbolic-differentiable architecture that learns flexible, non-linear, yet highly efficient control policies. DWCs can be trained end-to-end via gradient-based techniques, yet compile directly into FPGA-compatible circuits with few- or even single-clock-cycle latency and nanojoule-level energy cost per action. Across five MuJoCo benchmarks, including high-dimensional Humanoid, DWCs achieve returns competitive with standard deep policies (full-precision or quantized neural networks). Furthermore, DWCs exhibit structurally sparse and interpretable connectivity patterns, enabling direct inspection of which input values influence control decisions.

Assessing embryo fragmentation is crucial for predicting IVF success, yet manual grading is prone to subjectivity, and existing AI models struggle with clinical interpretability and segmentation errors. We propose AttnRegDeepLab, a Multi-Task Learning (MTL) framework designed to solve these challenges. The model enhances a DeepLabV3+ decoder with Attention Gates to filter out cytoplasmic noise and retain sharp contour details. It also introduces a Multi-Scale Regression Head with Feature Injection, guiding the segmentation process with global grading priors to eliminate systematic area estimation errors. Based on a two-stage decoupled training strategy and a range-based loss for weakly labeled data, our method resolves MTL gradient conflicts. AttnRegDeepLab yields high grading precision and excellent segmentation quality (Dice coefficient = 0.729), avoiding the trade-off between contour integrity and grading accuracy seen under standard joint optimization. This provides a reliable, clinically interpretable tool balancing visual and quantitative accuracy.

Flow matching (FM) has shown promising results in data-driven planning. However, it inherently lacks formal guarantees for ensuring state and action constraints, whose satisfaction is a fundamental and crucial requirement for the safety and admissibility of planned trajectories on various systems. Moreover, existing FM planners do not ensure the dynamical consistency, which potentially renders trajectories inexecutable. We address these shortcomings by proposing SAD-Flower, a novel framework for generating Safe, Admissible, and Dynamically consistent trajectories. Our approach relies on an augmentation of the flow with a virtual control input. Thereby, principled guidance can be derived using techniques from nonlinear control theory, providing formal guarantees for state constraints, action constraints, and dynamic consistency. Crucially, SAD-Flower operates without retraining, enabling test-time satisfaction of unseen constraints. Through extensive experiments across several tasks, we demonstrate that SAD-Flower outperforms various generative-model-based baselines in ensuring constraint satisfaction.

The effectiveness of self-supervised learning (SSL) for physiological time series depends on the ability of a pretraining objective to preserve information about the underlying physiological state while filtering out unrelated noise. However, existing strategies are limited due to reliance on heuristic principles or poorly constrained generative tasks. To address this limitation, we propose a pretraining framework that exploits the information structure of a dynamical systems generative model across multiple time-series. This framework reveals our key insight that class identity can be efficiently captured by extracting information about the generative variables related to the system parameters shared across similar time series samples, while noise unique to individual samples should be discarded. Building on this insight, we propose PULSE, a cross-reconstruction-based pretraining objective for physiological time series datasets that explicitly extracts system information while discarding non-transferrable sample-specific ones. We establish theory that provides sufficient conditions for the system information to be recovered, and empirically validate it using a synthetic dynamical systems experiment. Furthermore, we apply our method to diverse real-world datasets, demonstrating that PULSE learns representations that can broadly distinguish semantic classes, increase label efficiency, and improve transfer learning.

Large Language Models have achieved impressive performance across diverse applications, yet their training typically depends on centralized data collection, raising serious privacy and governance concerns. Federated Learning offers a decentralized alternative by enabling multiple clients to collaboratively train shared models without exposing raw local data. However, integrating FL with LLMs introduces new challenges, including data heterogeneity, convergence instability, communication overhead, and computational constraints. This survey provides a comprehensive and up-to-date overview of Federated Learning for Large Language Models (FedLLM). We systematically review recent advances, with particular emphasis on federated fine-tuning and federated prompt learning, and analyze how existing methods address efficiency, personalization, and security challenges. We further summarize emerging directions such as federated pre-training and federated agents. Our goal is to offer a structured perspective on this rapidly evolving field and to highlight promising avenues for future research.

We present BLINQ, a new model-based algorithm that learns the Whittle indices of an indexable, communicating and unichain Markov Decision Process (MDP). Our approach relies on building an empirical estimate of the MDP and then computing its Whittle indices using an extended version of a state-of-the-art existing algorithm. We provide a proof of convergence to the Whittle indices we want to learn as well as a bound on the time needed to learn them with arbitrary precision. Moreover, we investigate its computational complexity. Our numerical experiments suggest that BLINQ significantly outperforms existing Q-learning approaches in terms of the number of samples needed to get an accurate approximation. In addition, it has a total computational cost even lower than Q-learning for any reasonably high number of samples. These observations persist even when the Q-learning algorithms are speeded up using neural networks to predict Q-values.

Prompt optimization has become a central mechanism for eliciting strong performance from LLMs, and recent work has made substantial progress by proposing diverse prompt evaluation metrics and optimization strategies. Despite these advances, prompt evaluation and prompt optimization are often developed in isolation, limiting the extent to which evaluation can effectively inform prompt refinement. In this work, we study prompt optimization as a process guided by performance-relevant evaluation signals. To address the disconnect between evaluation and optimization, we propose an evaluation-instructed prompt optimization approach that explicitly connects prompt evaluation with query-dependent optimization. Our method integrates multiple complementary prompt quality metrics into a performance-reflective evaluation framework and trains an execution-free evaluator that predicts prompt quality directly from text, avoiding repeated model executions. These evaluation signals then guide prompt refinement in a targeted and interpretable manner. Empirically, the proposed evaluator achieves 83.7% accuracy in predicting prompt performance. When incorporated into the optimization process, our approach consistently outperforms existing optimization baselines across eight benchmark datasets and three different backbone LLMs. Overall, our results demonstrate that reliable and efficient evaluation signals can serve as an effective foundation for robust and interpretable prompt optimization.

Current vision-language models (VLMs) in medicine are primarily designed for categorical question answering (e.g., "Is this normal or abnormal?") or qualitative descriptive tasks. However, clinical decision-making often relies on quantitative assessments, such as measuring the size of a tumor or the angle of a joint, from which physicians draw their own diagnostic conclusions. This quantitative reasoning capability remains underexplored and poorly supported in existing VLMs. In this work, we introduce MedVision, a large-scale dataset and benchmark specifically designed to evaluate and improve VLMs on quantitative medical image analysis. MedVision spans 22 public datasets covering diverse anatomies and modalities, with 30.8 million image-annotation pairs. We focus on three representative quantitative tasks: (1) detection of anatomical structures and abnormalities, (2) tumor/lesion (T/L) size estimation, and (3) angle/distance (A/D) measurement. We show that current off-the-shelf VLMs perform poorly on these tasks. However, supervised and reinforcement fine-tuning on MedVision significantly enhances performance across detection, T/L estimation, and A/D measurement. MedVision provides a foundation for developing VLMs with robust quantitative reasoning capabilities in medical imaging.

Existing Test-time Adaptation (TTA) studies rely heavily on static and homogeneous corruption protocols, such as ImageNet-C and CIFAR-10-C/100-C, leading to inconsistent evaluation settings and potentially inflated robustness estimates that are compared with real-world situations. TTA lacks a standardized evaluation infrastructure capable of modeling realistic heterogeneous acoustic degradation. We introduce DHAuDS, a standardized benchmark suite for evaluating audio classification TTA robustness under dynamic corruption severity and heterogeneous noise mixtures. Rather than proposing a new TTA algorithm, DHAuDS focuses on exposing robustness limitations that remain hidden under conventional fixed-noise evaluation protocols.

Muon, a recently proposed optimizer that leverages the inherent matrix structure of neural network parameters, has demonstrated strong empirical performance, indicating its potential as a successor to standard optimizers such as AdamW. This paper presents theoretical analysis to support its practical success. We provide convergence proofs for Muon across four practical settings, systematically examining its behavior with and without the inclusion of Nesterov momentum and weight decay. We then demonstrate that the addition of weight decay ensures almost-sure boundedness of the parameter and gradient norms -- without relying on the commonly imposed bounded-gradient assumption -- and clarify the interplay between the weight decay coefficient and the learning rate. Finally, we derive a lower bound on the critical batch size for Muon -- the batch size that minimizes the stochastic first-order oracle (SFO) complexity of training. Because the resulting formula involves problem-dependent quantities that are not directly observable (gradient variance, target precision, effective rank), it does not predict the critical batch size in absolute terms; rather, it reveals how the hyperparameters $β$ (momentum) and $λ$ (weight decay) govern the qualitative scaling of this value. Our experiments validate these hyperparameter-dependent predictions across workloads including image classification and language modeling.