To lower the expertise barrier in machine learning, the AutoML community has focused on the CASH problem, which jointly automates algorithm selection and hyperparameter tuning. While traditional methods like Bayesian Optimization (BO) struggle with cold-start issues, Large Language Models (LLMs) can mitigate these through semantic priors. However, existing LLM-based optimizers generalize poorly to high-dimensional, structured CASH spaces. In this paper, we propose LB-MCTS, a trajectory-structured optimization framework that uses a Monte Carlo Tree Search tree as a shared state for algorithm selection, hyperparameter refinement, and BO-LLM proposer synergy. Within this shared state, BO provides algorithm-specific surrogate modeling for quantitative search, while the LLM exploits path-aware selective memory to generate semantic proposals and reflections. As the surrogate model improves, a reliability-aware proposer policy adaptively shifts from LLM-driven to BO-driven proposals within a unified search trajectory. Experiments on 104 AMLB datasets demonstrate that LB-MCTS consistently outperforms BO-based, LLM-based, and hybrid baselines.

In the information and communications technology (ICT) industry, training a domain-specific large language model (LLM) or constructing a retrieval-augmented generation system requires a substantial amount of high-value domain knowledge. However, the knowledge is not only hidden in the textual modality but also in the image modality. Traditional methods can parse text from domain documents but dont have image captioning ability. Multi-modal LLM (MLLM) can understand images, but they do not have sufficient domain knowledge. To address the above issues, this paper proposes a multi-stage progressive training strategy to train a Domain-specific Image Captioning Model (DICModel) in ICT, and constructs a standard evaluation system to validate the performance of DICModel. Specifically, this work first synthesizes about 7K image-text pairs by combining the Mermaid tool and LLMs, which are used for the first-stage supervised-fine-tuning (SFT) of DICModel. Then, ICT-domain experts manually annotate about 2K image-text pairs for the second-stage SFT of DICModel. Finally, experts and LLMs jointly synthesize about 1.5K visual question answering data for the instruction-based SFT. Experimental results indicate that our DICModel with only 7B parameters performs better than other state-of-the-art models with 32B parameters. Compared to the SOTA models with 7B and 32B parameters, our DICModel increases the BLEU metric by approximately 56.8% and 20.8%, respectively. On the objective questions constructed by ICT domain experts, our DICModel outperforms Qwen2.5-VL 32B by 1% in terms of accuracy rate. In summary, this work can efficiently and accurately extract the logical text from images, which is expected to promote the development of multimodal models in the ICT domain.

Large language models are increasingly trained via reinforcement learning for personalized recommendation tasks, but standard methods like GRPO rely on sparse, sequence-level rewards. These obscure which tokens actually contribute to high-quality outputs, creating a credit assignment gap. This gap is especially problematic when models must infer latent user intent from under-specified language without ground truth labels, which is a reasoning pattern rarely seen during pretraining but commonly required in deployment. We introduce Owen-Shapley Policy Optimization (OSPO), a framework that redistributes sequence-level advantages based on tokens' marginal contributions to outcomes. OSPO transforms task feedback into potential-based reward shaping via Shapley-Owen attributions to assign segment-level credit while preserving the optimal policy, all without parametric value models. By forming coalitions of semantically coherent units (e.g., phrases describing product attributes or sentences capturing preferences), OSPO identifies which response parts drive performance. Experiments on Amazon ESCI and H&M Fashion datasets including controlled generation tasks show consistent gains over baselines and notable test-time robustness to out-of-distribution retrievers unseen during training.

Inferring high-dimensional physical states from sparse, ad-hoc sensor arrays is a fundamental challenge across AI for Science and industrial IoT. Standard machine learning architectures struggle in these domains due to irregular, variable-cardinality sensor geometries and the profound sim-to-real distribution shift caused by unmodeled physical heterogeneities. To address these challenges, we propose Sensoformer, a set-attention framework integrated with Physics-Structured Domain Randomization (PSDR). By explicitly randomizing the underlying physical dynamics (e.g., propagation media, extreme noise, and network availability dropout) rather than just visual features, PSDR enforces the learning of domain-invariant physical operators. Using seismic source inversion as a rigorous real-world testbed, Sensoformer is pre-trained on 100,000 synthetics and evaluated on a highly complex real-world catalog. We demonstrate that Sensoformer achieves state-of-the-art precision and outperforms Message Passing Neural Networks (MPNNs) and Neural Operators (e.g., DeepONet) which struggle with extreme spatial sparsity and mixed-modality inputs. Furthermore, interpretability analysis reveals that the attention mechanism autonomously discovers optimal experimental design principles, dynamically prioritizing sparse orthogonal sensors to overcome information bottlenecks.

Spectral bias, the tendency of neural networks to learn low frequencies first, can be both a blessing and a curse. While it enhances the generalization capabilities by suppressing high-frequency noise, it can be a limitation in scientific tasks that require capturing fine-scale structures. The delayed generalization phenomenon known as grokking is another barrier to rapid training of neural networks. Grokking has been hypothesized to arise as learning transitions from the NTK to the feature-rich regime. This paper explores the impact of preconditioned gradient descent (PGD), such as Gauss-Newton, on spectral bias and grokking phenomena. We demonstrate through theoretical and empirical results how PGD can mitigate issues associated with spectral bias. Additionally, building on the rich learning regime grokking hypothesis, we study how PGD can be used to reduce delays associated with grokking. Our conjecture is that PGD, without the impediment of spectral bias, enables uniform exploration of the parameter space in the NTK regime. Our experimental results confirm this prediction, providing strong evidence that grokking represents a transitional behavior between the lazy regime characterized by the NTK and the rich regime. These findings deepen our understanding of the interplay between optimization dynamics, spectral bias, and the phases of neural network learning.

Current evaluations of mathematical reasoning in large language models (LLMs) are dominated by static benchmarks, either derived from competition-style problems or curated through costly expert effort, resulting in limited coverage of research-level mathematics and rapid performance saturation. We propose a fully automated, theorem-grounded pipeline for evaluating frontier mathematical reasoning, which directly transforms recent peer-reviewed mathematical literature into executable and verifiable reasoning tasks. The pipeline identifies constructive or quantitative results, instantiates them into parameterized problem templates, and generates deterministic solutions through execution-based verification, enabling scalable, reproducible, and continuously updatable evaluation without reliance on large-scale expert authoring. By design, this approach supports temporal extensibility, intrinsic correctness checking, and domain-specific customization across mathematical subfields. Applying this pipeline yields \textbf{EternalMath}, an evolving evaluation suite derived from contemporary research papers. Experiments with state-of-the-art LLMs reveal substantial performance gaps, indicating that mathematical reasoning at the research frontier remains far from saturated and underscoring the need for evaluation methodologies that evolve in step with human mathematical discovery.

This paper introduces Interpretability-Guided Bi-objective Optimization (IGBO), a framework that trains interpretable models by incorporating structured domain knowledge via a bi-objective formulation. IGBO encodes feature importance hierarchies as a Directed Acyclic Graph (DAG) via Central Limit Theorem-based construction and uses Temporal Integrated Gradients (TIG) to measure feature importance. The framework employs a novel Relative Importance Score Hk(X, θ) that quantifies the normalized cumulative attribution of each feature over time. We propose a geometric projection mapping P for combining task and interpretability gradients, and prove convergence to Pareto-stationary points. To address the Out-of-Distribution problem in TIG computation, we outline an Optimal Path Oracle architecture, which we leave for future work. Central Limit Theorem-based construction of the interpretability DAG provides statistical guarantees on acyclicity and transitivity, with an unconditional guarantee for the median threshold and conditional guarantees for higher confidence levels.

Multimodal learning has become a prominent research area, with the potential of substantial performance gains by combining information across modalities. At the same time, model development has trended toward increasingly complex deep learning architectures, motivated by the assumption that multimodal-specific methods improve performance. We challenge this assumption through a large-scale empirical study by reimplementing 19 high-impact multimodal methods across nine diverse datasets with up to 23 modalities. Under standardized experimental conditions, including hyperparameter tuning, weight initialization, cross-validation, and statistical testing, increased multimodal complexity often yields confusion rather than effective fusion of data modalities. Accordingly, complex multimodal architectures do not reliably outperform unimodal baselines and a Simple Baseline for Multimodal Learning (SimBaMM). Through a focused case study, we further demonstrate concrete methodological shortcomings even in top-tier multimodal learning publications, underscoring the need for standardized evaluation practices. In summary, we argue for a shift in focus for multimodal learning: away from the pursuit of architectural novelty and toward methodological rigor.

In realistic retrieval settings with large and evolving knowledge bases, the total number of documents relevant to a query is typically unknown, and recall cannot be computed. In this paper, we evaluate several established strategies for handling this limitation by measuring the correlation between retrieval quality metrics and LLM-based judgments of response quality, where responses are generated from the retrieved documents. We conduct experiments across multiple datasets with a relatively low number of relevant documents (2-15). We also introduce a simple retrieval quality measure that performs well without requiring knowledge of the total number of relevant documents.

With the rise of online dance-video platforms and rapid advances in AI-generated content (AIGC), music-driven dance generation has emerged as a compelling research direction. Despite substantial progress in related domains such as music-driven 3D dance generation, pose-driven image animation, and audio-driven talking-head synthesis, existing methods cannot be directly adapted to this task. Moreover, the limited studies in this area still struggle to jointly achieve high-quality visual appearance and realistic human motion. Accordingly, we present MACE-Dance, a music-driven dance video generation framework with cascaded Mixture-of-Experts (MoE). The Motion Expert performs music-to-3D motion generation while enforcing kinematic plausibility and artistic expressiveness, whereas the Appearance Expert carries out motion- and reference-conditioned video synthesis, preserving visual identity with spatiotemporal coherence. Specifically, the Motion Expert adopts a diffusion model with a BiMamba-Transformer hybrid architecture and a Guidance-Free Training (GFT) strategy, achieving state-of-the-art (SOTA) performance in 3D dance generation. The Appearance Expert employs a decoupled kinematic-aesthetic fine-tuning strategy, achieving state-of-the-art (SOTA) performance in pose-driven image animation. To better benchmark this task, we curate a large-scale and diverse dataset and design a motion-appearance evaluation protocol. Based on this protocol, MACE-Dance also achieves state-of-the-art performance. Code is available at https://github.com/AMAP-ML/MACE-Dance.

Discrete facility layout design involves placing physical entities to minimize handling costs while adhering to strict safety and spatial constraints. This combinatorial problem is typically addressed using Mixed Integer Linear Programming (MILP) or Constraint Programming (CP), though these methods often face scalability challenges as constraint density increases. This study systematically evaluates the potential of Conflict-Driven Clause Learning (CDCL) with VSIDS heuristics as an alternative computational engine for discrete layout problems. Using a unified benchmarking harness, we conducted a controlled comparison of CDCL, CP-SAT, and MILP across varying grid sizes and constraint densities. Experimental results reveal a distinct performance dichotomy: while CDCL struggles with optimization objectives due to cost-blind branching, it demonstrates unrivaled dominance in feasibility detection, solving highly constrained instances orders of magnitude faster than competing paradigms. Leveraging this finding, we developed a novel "Warm-Start" hybrid architecture that utilizes CDCL to rapidly generate valid feasibility hints, which are then injected into a CP-SAT optimizer. Our results confirm that this layered approach successfully accelerates exact optimization, using SAT-driven pruning to bridge the gap between rapid satisfiability and proven optimality.

With AI-generated videos increasingly indistinguishable from reality, current benchmarks primarily focus on broad semantic alignment and basic physical consistency, offering limited discriminative power for evaluating them. To address this, we introduce VideoASMR-Bench, a benchmark based on Autonomous Sensory Meridian Response (ASMR) videos that emphasizes fine-grained audio-visual perception and sensory immersion. This benchmark aims to answer two key questions: (i) Are today's video understanding models (VLMs) sensitive enough to detect AI-generated ASMR videos by recognizing minor visual, physical, or auditory artifacts? (ii) Can today's video generation models (VGMs) produce convincing ASMR videos with immersive experiences? This benchmark comprises a diverse set of 1,500 high-quality real ASMR videos curated from social media, alongside 2,235 synthetic counterparts generated by nine VGMs. Additionally, we open-source an extensible suite of prompts and reference images, enabling the benchmark to scale dynamically with future video models. Moreover, we design an automatic understanding-generation evaluation framework between VGMs and VLMs, where VGMs aim to produce realistic fake videos to fool the VLMs, while the VLMs seek to detect them, forming an adversarial game between the two parties. Our evaluation on VideoASMR-Bench reveals that even state-of-the-art VLMs, such as Gemini-3-Pro, fail to reliably detect AI-generated ASMR videos. Meanwhile, current frontier video generation models can produce ASMR videos that are difficult for VLMs to distinguish from real ones, while humans can still identify them relatively easily.

The proliferation of AI-generated video models poses new challenges to information integrity and digital trust. A key confound, however, remains unaddressed: commercial generators embed visible overlay watermarks for provenance tracking, yet no existing benchmark controls for this variable, leaving open whether detectors learn genuine generation artefacts or merely associate watermark patterns with AI-generated labels. We present RobustSora, a benchmark of 6,500 manually verified videos in four categories: Authentic-Clean (A-C), Generated-Watermarked (G-W), Generated-DeWatermarked (G-DeW), and Authentic-Spoofed (A-S), sourced from Vript, DVF, and UltraVideo (authentic) and from Sora, Sora 2, Pika, Open-Sora 2, and KLing (generated). Two evaluation tasks isolate watermark effects: Task-I (Watermark Erasure Robustness) tests detection on watermark-removed AI videos; Task-II (Watermark Spoofing Robustness) measures false-alarm rates on authentic videos injected with fake watermarks. Across ten models spanning specialized detectors, transformer classifiers, and MLLMs, watermark manipulation induces accuracy changes of $-9.4$ to $+1.6$ pp (mean 6.6 pp; $p{<}0.01$ for 7/10 models on each task). A placebo control bounds inpainting-artefact confounds at $\le$2 pp, and a watermark-aware training augmentation recovers 3-4 pp on both tasks, together providing causal evidence that detectors actively rely on watermark cues. Per-generator breakdown shows that Sora 2 induces drops of $-11$ to $-14$ pp versus $-3$ to $-6$ pp for Pika and Open-Sora 2, indicating that watermark prominence, rather than detector architecture, is the principal driver of dependency. These results argue for watermark-aware evaluation and training in AIGC video detection. Dataset, evaluation code, and pretrained checkpoints will be released.

Recent works have shown that gradient-update alignment is a powerful signal for modulating optimizer updates, often leading to faster training. We promote this update-wise heuristic as a mathematically grounded principle for selecting and tuning optimizer hyperparameters. By treating gradients and updates as signals and an optimizer as a causal filter that maps between them, we formulate optimizer selection as maximizing the expected drop rate in loss over a prescribed family of optimizers. We show that this objective is exactly the inner product between the optimizer filter and the gradient autocorrelation, and prove that a greedy optimum exists and has a stability bound under perturbations of the estimated gradient statistics. Specializing in momentum-based optimizers, the theory yields simple dynamic momentum selection rules for both SGD+Momentum and Adam/AdamW. Experiments across image classification, language model fine-tuning, and vision transformer fine-tuning show that the resulting dynamic momentum rules match or improve upon the best fixed hyperparameters found via manual sweeps, reducing the need for exhaustive momentum sweeps. Code is available at https://github.com/ironjr/gap

Land-cover underpins ecosystem services, hydrologic regulation, disaster-risk reduction, and evidence-based land planning; timely, accurate land-cover maps are therefore critical for environmental stewardship. Remote sensing-based land-cover classification offers a scalable route to such maps but is hindered by scarce and imbalanced annotations and by geometric distortions in high-resolution scenes. We propose LC4-DViT (Land-cover Creation for Land-cover Classification with Deformable Vision Transformer), a framework that combines generative data creation with a deformation-aware Vision Transformer. A text-guided diffusion pipeline uses GPT-4o-generated scene descriptions and super-resolved exemplars to synthesize class-balanced, high-fidelity training images, while DViT couples a DCNv4 deformable convolutional backbone with a Vision Transformer encoder to jointly capture fine-scale geometry and global context. On eight classes from the Aerial Image Dataset (AID)-Beach, Bridge, Desert, Forest, Mountain, Pond, Port, and River-DViT achieves 0.9572 overall accuracy, 0.9576 macro F1-score, and 0.9510 Cohen' s Kappa, improving over a vanilla ViT baseline (0.9274 OA, 0.9300 macro F1, 0.9169 Kappa) and outperforming ResNet50, MobileNetV2, and FlashInternImage. Cross-dataset experiments on a three-class SIRI-WHU subset (Harbor, Pond, River) yield 0.9333 overall accuracy, 0.9316 macro F1, and 0.8989 Kappa, indicating good transferability. An LLM-based judge using GPT-4o to score Grad-CAM heatmaps further shows that DViT' s attention aligns best with hydrologically meaningful structures. These results suggest that description-driven generative augmentation combined with deformation-aware transformers is a promising approach for high-resolution land-cover mapping.

Clinical notes in Electronic Health Records (EHRs) capture rich temporal information on events, clinician reasoning, and lifestyle factors often missing from structured data. Leveraging them for predictive modeling can be impactful for timely identification of chronic diseases. However, they present core natural language processing (NLP) challenges: long text, irregular event distribution, complex temporal dependencies, privacy constraints, and resource limitations. We present two complementary methods for temporally and contextually grounded risk prediction from longitudinal notes. First, we introduce HiTGNN, a hierarchical temporal graph neural network that integrates intra-note temporal event structures, inter-visit dynamics, and medical knowledge to model patient trajectories with fine-grained temporal granularity. Second, we propose ReVeAL, a lightweight test-time framework that distills LLMs' reasoning into smaller verifier models. Applied to opportunistic screening for Type 2 Diabetes (T2D) using temporally realistic cohorts curated from private and public hospital corpora, HiTGNN achieves the highest predictive accuracy, especially for near-term risk, while preserving privacy and limiting reliance on large proprietary models. ReVeAL enhances sensitivity to true T2D cases and retains explanatory reasoning. Our ablations confirm the value of temporal structure and knowledge augmentation, and fairness analysis shows HiTGNN performs more equitably across subgroups.

Spatial cognition is fundamental to real-world multimodal intelligence, allowing models to effectively interact with the physical environment. While multimodal large language models (MLLMs) have made significant strides, existing benchmarks often oversimplify spatial cognition, reducing it to a single-dimensional metric, which fails to capture the hierarchical structure and interdependence of spatial abilities. To address this gap, we propose a hierarchical spatial cognition framework that decomposes spatial intelligence into five progressively complex levels from basic observation to high-level planning. Building upon this taxonomy, we construct SpatialBench, a large-scale, fine-grained benchmark covering 15 tasks aligned with these cognitive levels. To provide a unified evaluation across heterogeneous tasks, we further introduce a high-level capability-oriented metric that reliably assesses a model's overall spatial reasoning ability. Extensive experiments over massive MLLMs reveal distinct performance stratification across cognitive levels: models exhibit strong perceptual grounding yet remain limited in symbolic reasoning, causal inference, and planning. Additional human tests demonstrate that humans perform selective, goal-directed abstraction, while MLLMs tend to over-attend to surface details without coherent spatial intent. Our work establishes the first systematic framework for measuring hierarchical spatial cognition in MLLMs, laying the foundation for future spatially intelligent systems.

Recently, Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as an effective approach to incentivizing reasoning capability in Large Multimodal Models (LMMs), while the underlying mechanisms behind this post-training paradigm are poorly understood. We begin by exploring how input activations are affected by RLVR through the perspective of logit lens. Our systematic investigations across multiple post-trained LMMs suggest that RLVR shifts low-entropy activations unexpectedly, while high-entropy ones are less affected. We further demonstrate that such phenomena are associated with LMM reasoning by controlled experiments, suggesting a potentially beneficial role of modulating low-entropy activations. To this end, we propose Activation Replay, a novel simple yet effective training-free approach that boosts multimodal reasoning of post-trained LMMs without requiring expensive policy optimization. Our design involves manipulation of visual tokens at test time, replaying low-entropy activations from the input context of base LMMs to regulating the RLVR counterparts. Activation Replay triggers better reasoning across diverse scenarios, including mathematics, o3-like visual agents, and video reasoning. We further show that Activation Replay boosts Pass@K and mitigates narrower reasoning coverage of RLVR. Our design is compared against alternative choices, such as replaying high-entropy activations instead of low-entropy ones, or direct cross-model intervention instead of manipulating input tokens, demonstrating the superiority of our implementation. Code is publicly available at https://github.com/latentcraft/replay.

Self-consistency -- sampling multiple reasoning paths and selecting the most frequent answer -- was designed for an era when language models made frequent, unpredictable errors. This study argues that the technique has become increasingly wasteful as models grow stronger, and may degrade performance on problems that modern models already solve reliably. Using Gemini 2.5 models on HotpotQA and MATH-500, we show that accuracy gains from increasing the number of sampled reasoning paths are minimal -- 0.4% on HotpotQA across 20 samples, and 1.6% on MATH-500 -- while token costs scale nearly linearly with sample count. Critically, performance plateaued early and in some configurations declined at high sample counts, suggesting that additional paths introduce noise rather than signal when models already solve problems reliably. As inference costs rise with model scale, indiscriminate self-consistency is difficult to justify. We recommend reserving multi-path sampling for problems that demonstrably exceed a model's single-pass reliability.

Computer-assisted surgery research requires large, deeply annotated video datasets that capture clinical and technical variability. Existing cataract surgery resources lack the diversity and annotation depth required to train generalizable deep-learning models. To address this gap, we present a dataset of 3,000 phacoemulsification cataract surgery videos acquired at two surgical centers from surgeons with varying expertise. The dataset provides four annotation layers: temporal surgical phases, instance segmentation of instruments and anatomical structures, instrument-tissue interaction tracking, and quantitative skill scores based on competency rubrics adapted from ICO-OSCAR and GRASIS. We demonstrate the technical utility of the dataset through benchmarking deep learning models across four tasks: workflow recognition, scene segmentation, instrument-tissue interaction tracking, and automated skill assessment. Furthermore, we establish a domain-adaptation baseline for phase recognition and instance segmentation by training on one surgical center and evaluating on a held-out center. Ultimately, these multi-source acquisitions, multi-layer annotations, and paired skill-kinematic labels facilitate the development of generalizable multi-task models for surgical workflow analysis, scene understanding, and competency-based training research.

Real-world deployment often exposes models to distribution shifts, making test-time adaptation (TTA) critical for robustness. Yet most TTA methods are unfriendly to edge deployment, as they rely on backpropagation, activation buffering, or test-time mini-batches, leading to high latency and memory overhead. We propose \textbf{ELaTTA} (\textit{Efficient Latent Test-Time Adaptation}), a gradient-free framework for single-instance TTA under strict on-device constraints. ELaTTA freezes model weights and adapts each test sample by optimizing a low-dimensional coefficient vector in a source-induced principal latent subspace, pre-computed offline via truncated SVD and stored with negligible overhead. At inference, ELaTTA encourages prediction confidence by optimizing the $k$-D coefficients with CMA-ES, effectively optimizing a Gaussian-smoothed objective and improving stability near decision boundaries. Across six benchmarks and multiple architectures, ELaTTA achieves state-of-the-art accuracy under both strict and continual single-instance protocols, while reducing compute by up to \emph{63$\times$} and peak memory by up to \emph{11$\times$}. We further demonstrate on-device deployment on a ZYNQ-7020 platform.

Coreference Resolution (CR) is a critical task in Natural Language Processing (NLP). Current research faces a key dilemma: whether to further explore the potential of supervised neural methods based on small language models, whose detect-then-cluster pipeline still delivers top performance, or embrace the powerful capabilities of Large Language Models (LLMs). However, effectively combining their strengths remains underexplored. To this end, we propose \textbf{ImCoref-CeS}, a novel framework that integrates an enhanced supervised model with LLM-based reasoning. First, we present an improved CR method (\textbf{ImCoref}) to push the performance boundaries of the supervised neural method by introducing a lightweight bridging module to enhance long-text encoding capability, devising a biaffine scorer to comprehensively capture positional information, and invoking a hybrid mention regularization to improve training efficiency. Importantly, we employ an LLM acting as a multi-role Checker-Splitter agent to validate candidate mentions (filtering out invalid ones) and coreference results (splitting erroneous clusters) predicted by ImCoref. Extensive experiments demonstrate the effectiveness of ImCoref-CeS, which achieves superior performance compared to existing state-of-the-art (SOTA) methods.

We compile 129 heterogeneous LLM prompt datasets (>1.22 TB, >673M instances) into a structured taxonomy and conduct a multi-level linguistic analysis (lexical, syntactic, and semantic) on seven representative corpora, surfacing systematic patterns that distinguish prompts from general text. Three downstream experiments validate practical utility: prompt filtering (F1 = 0.90), domain classification (Macro-F1 = 0.975), and prompt quality prediction (AUC = 0.792), all without invoking any additional model. A central finding is that 62-d syntactic features (POS + dependency distributions) serve as a uniquely efficient routing primitive, recovering >93% of GPU-embedding accuracy at 1.9 $\times$ lower single-request latency (3.0 ms vs. 5.7 ms) with no GPU and no corpus vocabulary. A complementary discriminative--predictive divergence shows that features most useful for routing are precisely those most negatively correlated with response quality, while lexical diversity (Cohen's $d$ = 0.71) dominates the quality signal but carries minimal routing weight, directly motivating two-stage pipeline design. Our datasets and code are available.

Federated learning (FL) enables collaborative training without raw data sharing, but still risks training data memorization. Existing FL memorization detection techniques focus on one sample at a time, underestimating more subtle risks of cross-sample memorization. In contrast, recent work on centralized learning (CL) has introduced fine-grained methods to assess memorization across all samples in training data, but these assume centralized access to data and cannot be applied directly to FL. We bridge this gap by proposing a framework that quantifies both intra- and inter-client memorization in FL using fine-grained cross-sample memorization measurement across all clients. Based on this framework, we conduct two studies: (1) measuring subtle memorization across clients and (2) examining key factors that influence memorization, including decoding strategies, prefix length, and FL algorithms. Our findings reveal that FL models do memorize client data, particularly intra-client data, more than inter-client data, with memorization influenced by training and inferencing factors.

The popular path query - identifying the most frequented routes between locations from historical trajectory data - has important applications in urban planning, navigation optimization, and travel recommendations. While traditional algorithms and machine learning approaches have achieved success in this domain, they typically require model training, parameter tuning, and retraining when accommodating data updates. As Large Language Models (LLMs) demonstrate increasing capabilities in spatial and graph-based reasoning, there is growing interest in exploring how these models can be applied to geo-spatial problems. We introduce CompassLLM, a novel multi-agent framework that intelligently leverages the reasoning capabilities of LLMs into the geo-spatial domain to solve the popular path query. CompassLLM employs its agents in a two-stage pipeline: the SEARCH stage that identifies popular paths, and a GENERATE stage that synthesizes novel paths in the absence of an existing one in the historical trajectory data. Experiments on real and synthetic datasets show that CompassLLM demonstrates superior accuracy in SEARCH and competitive performance in GENERATE while being cost-effective.

Evaluation of multimodal reasoning models is typically reduced to a single accuracy score, implicitly treating reasoning as a unitary capability. We introduce MathLens, a benchmark of textbook-style geometry problems that exposes this assumption by operationally decomposing performance into perception, reasoning, and multimodal-specific components. Each problem is derived from a symbolic specification and accompanied by visual diagrams, text-only variants, multimodal questions, and targeted perceptual probes, enabling controlled measurement of each component. Using this decomposition, we show that common training strategies induce systematically different capability profiles that are invisible under aggregate accuracy. Reinforcement learning primarily improves perceptual grounding and robustness to diagram variation, while textual SFT yields gains through reflective reasoning. In contrast, as perception and reasoning improve, a growing fraction of remaining errors fall outside these components and are categorized as multimodal-specific. These results suggest that apparent progress in multimodal reasoning reflects shifting balances among subskills rather than uniform advancement, motivating evaluation beyond scalar accuracy.

Synthetic data is central to data-efficient Dyna-style model-based reinforcement learning, but it can also degrade performance. We study this failure in Model-Based Policy Optimization (MBPO), which performs actor-critic updates using model-generated synthetic state transitions. Although MBPO reports strong sample-efficiency gains on OpenAI Gym, recent work shows that it often underperforms Soft Actor-Critic (SAC), its non-Dyna base, in the DeepMind Control Suite (DMC), despite both suites involving MuJoCo-based proprioceptive continuous control. We identify two coupled causes of this collapse: scale mismatch between dynamics and reward targets, which suppresses reward learning and induces critic underestimation, and residual next-state prediction, which inflates model variance and produces unreliable synthetic transitions. We introduce Fixing That Free Lunch (FTFL), a minimal repair that combines independent target normalization with direct next-state prediction. FTFL outperforms SAC in five of seven previously failing DMC tasks while preserving MBPO's strong Gym performance. We further show that MBPO-lineage algorithms, including uncertainty-aware variants that filter, penalize, or reject synthetic transitions based on model uncertainty, still inherit these failures unless FTFL is applied to their shared learned-model backbone. More broadly, our results show how benchmark-limited evaluation can encode environment-specific assumptions into algorithm design, motivating taxonomies that map MDP structure to algorithmic failure modes.

Reinforcement learning with verifiable rewards (RLVR) enables large language models to acquire slow, multi-step reasoning from sparse final-answer signals. We provide a statistical-physics picture of this emergence. We show that an autoregressive model's finite capacity forces it to compress its exponentially large prefix space into a Markov network of predictive states, on which slow thinking unfolds as a random walk -- the Concept Network (CoNet) picture. Within CoNet, RLVR dynamics are governed by two mechanisms: merging of compatible paths and frustrated competition among incompatible ones. Together they drive the network through nucleation, growth, and freezing into multi-input, single-output directed inverse trees. The picture reproduces the training dynamics of a 1.5-billion-parameter LLM and yields three predictions: reasoning chains lengthen as a geometric necessity of sparse topology; SFT induces catastrophic forgetting through bridge-node rupture; and frustration drives policy collapse. Building on the structural timing inherent in inverse-tree freezing, we propose Annealed-RLVR -- a brief SFT intervention at the moment of maximum frustration. It outperforms standard RLVR on both in- and out-of-distribution benchmarks, with the largest gains at high sampling budgets where standard RLVR collapses. The same SFT applied after the trees freeze instead triggers catastrophic forgetting, isolating timing as the active ingredient.

Given a set of graphs from some unknown family, we want to generate new graphs from that family. Recent methods use diffusion on either graph embeddings or the discrete space of nodes and edges. However, simple changes to embeddings (say, adding noise) can mean uninterpretable changes in the graph. In discrete-space diffusion, each step may add or remove many nodes/edges. It is hard to predict what graph patterns we will observe after many diffusion steps. Our proposed method, called GraphWeave, takes a different approach. We separate pattern generation and graph construction. To find patterns in the training graphs, we see how they transform vectors during random walks. We then generate new graphs in two steps. First, we generate realistic random walk "trajectories" which match the learned patterns. Then, we find the optimal graph that fits these trajectories. The optimization infers all edges jointly, which improves robustness to errors. On four simulated and five real-world benchmark datasets, GraphWeave outperforms existing methods. The most significant differences are on large-scale graph structures such as PageRank, cuts, communities, degree distributions, and flows. GraphWeave is also 10x faster than its closest competitor. Finally, GraphWeave is simple, needing only a transformer and standard optimizers.

Synthetic text generation with Differential Privacy (DP) guarantees emerges as a principled approach that can enable the sharing of sensitive datasets across institutional and regulatory boundaries, while bounding the risks of re-identification and membership inference. LLM-based methods deliver promising results; however, comparisons are exacerbated by differing evaluation setups and "private" datasets, potential pre-training contamination is not considered and guarantees are not verified with DP audits. To advance this field, we introduce a unified evaluation framework with standardised utility and fidelity metrics and privacy audits, encompassing nine curated datasets that capture domain-specific complexities such as technical jargon, long-context dependencies, and specialised document structures. In a large-scale empirical study, we benchmark LLM-based state-of-the-art DP text generators of varying sizes (between 1--8B). Our results indicate that DP synthetic text generation remains an unsolved challenge, with quality deteriorating more as the private datasets deviate further from the generators' pre-training corpora. Our novel synthetic text membership inference attack (MIA) explains this observation: Synthetic data quality is overestimated when LLMs have been pre-trained -- without DP -- on portions of the "private" data to be generated. Finally, our work provides the first quantitative evidence that this "public pre-training and private generation" paradigm invalidates the guaranteed privacy bounds of real-world private datasets.