Parametric computer-aided design records both final geometry and the ordered construction history that determines how a part can be edited. Datasets for editable CAD research should therefore expose modeling operations, parameters, and feature dependencies together with validated geometry. We introduce FllumaOne, a code-native multimodal CAD dataset whose models are generated by executable Python programs in Flluma, a Qt/C++ OpenCASCADE-based CAD system. Each sample aligns its program with a structured feature tree, a training-oriented intermediate representation, STEP geometry, a surface point cloud, natural-language descriptions, metadata, and eight canonical visible-edge renderings. The primary release, FllumaOne-100K, contains 100,000 accepted samples across four template-level complexity regimes. Programs are executed and retained only after kernel geometry, solid validity, and export checks; release reports also record modality completeness and split-level duplicate tests. A Qwen2.5-Coder-1.5B LoRA baseline trained on 80,000 samples achieves 99.98% Python syntax validity, 99.97% Flluma build success, and 99.14% STEP-export validity on the held-out 10,000-sample test split. For the 9,909 predictions converted to surface point clouds, the mean normalized Chamfer Distance is 0.002124. The dataset supports conditioned CAD reconstruction, executable program synthesis, feature-tree prediction, B-Rep analysis, retrieval, design completion, and editable reverse engineering.

Recent work has shown that LLMs overrepresent dominant cultures, particularly Western ones, while marginalizing others. We investigate whether this affects models' ability to generate culturally adapted responses by evaluating their use of local measurement units based on the user's perceived cultural background. We introduce Cultural and Pragmatic Response Inference (CAPRI), a dataset of conversations with varying levels of cultural cues. Experiments with state-of-the-art LLMs show that models can infer cultural background and recall relevant conventions, but often fail to utilize the information to adapt their answers to the relevant cultural conventions, unless explicitly prompted to perform the tasks sequentially. We further evaluate adaptation to the interpretation of time and quantity expressions, two subjective language grounding dimensions that are affected by culture. We find that models increasingly adapt their answers as cultural cues accumulate, but their priors are not culture-neutral, sometimes aligning with the model's country of origin. Overall, CAPRI provides a resource for future research aimed at narrowing the gap between cultural knowledge and culturally adaptive language generation.

Despite remarkable performance on complex tasks, Large Reasoning Models (LRMs) often generate excessively long Chain-of-Thoughts (CoT), inflating computational costs even for simple queries. Existing efforts to mitigate this inefficiency typically rely on discrete reasoning modes or fixed budget tiers, lacking a principled criterion of when reasoning is sufficient. In this work, we introduce Minimal Sufficient CoT (MSC), defined as the shortest prefix of a CoT trajectory which is adequate for producing the correct answer. We empirically show that MSC not only reduces reasoning tokens, but also improves accuracy across difficulty levels. Building on MSC, we propose Sufficiency-guided Continuous Adaptive Reasoning (SuCo), a two-stage training framework for autonomous reasoning control along a continuous spectrum. In stage 1, MSC-Aligned Fine-Tuning (MFT) constructs MSC data using problem-adaptive sufficiency thresholds that naturally scale with question difficulty, then fine-tunes the model to internalize concise yet sufficient reasoning patterns. In stage 2, Sufficiency-Aware Policy Optimization (SAPO) further optimizes the model through reinforcement learning with dynamic complexity tracking and sufficiency-aware rewards that penalize both over- and under-thinking. Extensive experiments across mathematics, code, and science benchmarks show that SuCo consistently achieves improvements in both accuracy and reasoning efficiency.

While large language models (LLMs) have greatly advanced the functional correctness of automated code translation systems, the runtime efficiency of translated programs has received comparatively little attention. With the waning of Moore's law, runtime efficiency has become increasingly important for program quality, alongside functional correctness. Our preliminary study reveals that LLM-translated programs often run slower than human-written ones, and this issue cannot be remedied through prompt engineering alone. Therefore, our work proposes SwiftTrans, a code translation framework comprising two key stages: (1) Multi-Perspective Exploration, where MpTranslator leverages parallel in-context learning (ICL) to generate diverse translation candidates; and (2) Difference-Aware Selection, where DiffSelector identifies the optimal candidate by explicitly comparing differences between translations. We further introduce Hierarchical Guidance for MpTranslator and Ordinal Guidance for DiffSelector, enabling LLMs to better adapt to these two core components. To support the evaluation of runtime efficiency in translated programs, we extend existing benchmarks, CodeNet and F2SBench, and introduce a new benchmark, SwiftBench. Experimental results across all three benchmarks show that SwiftTrans achieves consistent improvements in both correctness and runtime efficiency.

Reinforcement learning pipelines for Large Language Model (LLM) training often rely on manually redesigned environments between stages, requiring practitioners to heuristically infer which configuration will best improve the current policy. To automate this process, we propose the LLM-as-Environment-Engineer framework in which the current policy model analyzes failure trajectories together with contextual information and proposes modifications to the next-stage training environment configuration. We also introduce MAPF-FrozenLake, a controllable testbed whose generator exposes multi-dimensional environment configurations, making it suitable for studying and benchmarking environment redesign. On this testbed, we condition the environment engineer on structured summaries of policy behavior, failure cases, and environment statistics, from which it produces the configuration for the next training stage. With Qwen3-4B as the backbone, our framework achieves the strongest aggregate performance on our benchmarks, outperforming larger proprietary LLMs (e.g., GPT, Gemini) and fixed-environment training baselines. We further analyze which forms of context are most effective, finding that successful environment updates rely on failure evidence and preserve configurations that already work. Interestingly, the current RL checkpoint serves as a better environment engineer than the original base model, suggesting that policy learning improves the model's ability to diagnose its remaining weaknesses.

Reinforcement learning (RL) has emerged as a powerful paradigm for training Large Language Models (LLMs) as agents. However, conventional RL methods for long-horizon agentic tasks often struggle with sparse outcome rewards. Intuitively, this overlooks the rich environment dynamics information contained in rollout interaction trajectories. We argue that the interaction experience inherently serves as an implicit supervision signal, reveals the underlying transition mechanisms of the environment, and enables the agent to construct a more accurate internal model of the environment.. Therefore, in this work, we investigate how to leverage this additional signal to improve policy learning. Specifically, we propose EnvRL, a framework that incorporates environment dynamics learning into agentic RL via two auxiliary objectives: state prediction and inverse dynamics. By jointly optimizing with the primary RL objective, we encourage the agent to internalize environment dynamics from its own interaction experience. Extensive experiments on two long-horizon agentic benchmarks demonstrate that EnvRL achieves significant improvements on success-rates over RL-only baselines, e.g., when trained with GRPO, lifting Qwen-2.5-1.5B-Instruct from 72.8% to 77.4% on ALFWorld, and from 56.8% to 67.0% on WebShop.

Multimodal large language models (MLLMs) integrate strong text reasoning with visual inputs, yet their responses can be inconsistent with the underlying images, indicating ineffective utilization of visual evidence during inference. The prevailing training paradigm relies on large-scale caption-based pretraining for general alignment, followed by supervised fine-tuning and reinforcement learning to enable instruction following and complex reasoning. However, such pretraining provides only weak visual grounding: short, coarse captions bias models toward salient objects while neglecting fine-grained visual evidence. In this paper, we introduce Visual Evidence Pre-Alignment (VEPA), an intermediate stage between pretraining and post-training that explores a novel sufficiency-driven objective with Group Relative Policy Optimization (GRPO) to optimize question-conditioned visual evidence descriptions. Extensive experiments across diverse benchmarks show that our VEPA consistently enhances performance on visually demanding evaluations and complements standard supervised post-training. Further analyses show that the income stems from strengthened, transferable visual grounding, rather than from additional task-specific training.

Magnetic resonance imaging-signal fat fraction (MRI-SFF) quantifies tissue fat and serves as an established biomarker for metabolic and musculoskeletal disorders. The acquisition requires, however, specialized MRI sequences, which are not available routinely. We investigate whether SFF can be estimated from widely available T2-weighted (T2w) MRI via image-to-image translation (I2I). We further compare a lightweight 4-level U-Net to a state-of-the-art Denoising Diffusion Probabilistic Model (DDPM) using a dataset of 230 048 paired 2D images (183 517 train, 23 621 val, 22 910 test) from the German National Cohort (NAKO). Both models clearly outperform the identity baseline (Pearson correlation r = 0.769, mean absolute error MAE = 0.070 +/- 0.054), which confirms that the models learn a non-trivial cross-modal mapping. Interestingly, the lightweight U-Net outperforms the DDPM in both correlation (r = 0.975 vs. 0.962) and error (MAE = 0.014 +/- 0.015 vs. 0.019 +/- 0.019), while reducing inference time by a factor of 208 (25.2 ms vs. 5 227.2 ms per image using 50 Denoising Diffusion Implicit Model (DDIM) steps). The strong clinical performance at substantially reduced computational cost enables real-time clinical use.

While Large Language Models (LLMs) have revolutionized text-based role-playing, creating immersive Speech Role-Playing Agents (SRPAs) requires a seamless bridge between cognitive reasoning and paralinguistic nuances. Current SRPAs primarily rely on end-to-end (E2E) fine-tuning. However, this paradigm suffers from poor generalization to unseen characters due to its reliance on role-specific data, while imposing a "modality alignment tax" that degrades intrinsic LLM reasoning capabilities. We propose DeSRPA, an agentic framework for character role play via inference-time intervention on frozen backbones. DeSRPA employs a dual-level control vector mechanism, Internal Cognitive Steering and External Expressive Rendering, to synchronize "mind" and "voice". Experiments on SpeechRole and OmniCharacter benchmarks demonstrate that DeSRPA significantly outperforms E2E baselines in personality and emotional consistency. It achieves high speech naturalness, narrowing the gap with proprietary models like GPT-4o Audio, while remaining a scalable and training-free paradigm.

Molecular dynamics (MD) simulation is computationally demanding, particularly for large-scale systems requiring long-term analysis. Accurate forecast of the outcomes of a MD simulation is not only an attractive scientific challenge but also has substantial practical value. In this work, we developed a data-driven framework, termed ASTEROID (Advanced Spatiotemporal TransformER fOr Inferring Dynamics), that can directly predict multi-step atomic coordinates, avoiding conventional iterative integration. For this purpose, our ASTEROID reformulates MD trajectories as high-dimensional spatiotemporal sequences and integrates the Spatiotemporal Information (STI) Transformation equation into a Transformer architecture. The core innovation of ASTEROID lies in its ability to model multiscale spatiotemporal dependencies. In particular, for spatial dependencies, a local-global self-attention mechanism captures both short- and long-range interactions. For temporal dependencies, an encoder-decoder structure integrates global context with autoregressive forecasting. ASTEROID was evaluated on several quantum-mechanics derived molecular datasets. Our results indicate that ASTEROID achieved not only a higher level of accuracy in multi-step prediction than existing methods on various benchmarks, but also significantly reduced computational cost of conventional MD simulation. Moreover, the model supports iterative multi-step forecasting over an extended time scale. This work establishes a robust and generalizable data-driven paradigm for accelerating MD simulations.

In recent years, the rapid development of foundation models and graph pre-training technologies has spurred increasing interest in constructing a universal pre-trained graph model or Graph Foundation Model (GFM). However, a significant challenge is that existing models are unable to address feature heterogeneity in graph data without textual information, which hinders the transferability of graph models across different datasets. To bridge this gap, we propose the concept of learnable graph patches, which we regard as the smallest semantic units of any graph data. We decompose the graph into learnable graph patches by unfolding the node features and constructing corresponding patch structures separately. We then design a framework that mines transferable information from graph data across domains. Specifically, after extracting graph patches, we propose a patch encoder to extract knowledge from each unit and a patch aggregator to learn how the units are combined into a whole. Due to its domain-agnostic nature, the model can be applied to downstream data across different domains. Furthermore, we analyze the connection between our method and existing graph models, as well as the transferability of the node embeddings it generates. Empirically, our method not only achieves the capability to use multi-domain graphs for pre-training, but also shows enhanced performance across various downstream datasets and tasks. Moreover, we observe consistent improvement in downstream performance as the volume of pre-training data increases.

Fine-tuning large language models (LLMs) is compute-intensive and error-prone: model performance depends sensitively on data quality and hyperparameter choices, and naïve runs can even degrade model performance. This raises a practical question:can we predict fine-tuning performance before committing to a full training run? We present TUNEAHEAD, a lightweight framework for pre-hoc prediction of fine-tuning performance. TUNEAHEAD encodes each candidate run as a meta-feature vector that combines static dataset descriptors with dynamic probe features from a short standardized probe. A predictor maps these features to performance estimates, while SHAP-based attributions provide interpretable diagnostics that reveal which specific features drive the prediction. Across 1,300+ fine-tuning runs on Qwen2.5-7B-Instruct, TUNEAHEAD consistently outperforms strong baselines such as Early-Stop Extrapolation and ProxyLM. On a held-out test set of 370 runs, TUNEAHEAD achieves an RMSE of 1.47 percentage points and places 95.1% of predictions within +3/-3 percentage points of the true score. These accurate continuous predictions support practical go/no-go screening policies that can reduce unnecessary full fine-tuning while retaining most promising runs.

This study introduces enhancements to physics-constrained neural networks (PCNNs) that improve the accuracy and stability of hybrid short-term weather forecasting models. Building on the WeatherGFT architecture, three innovations are proposed. First, an upgraded numerical solver, combining a fifth-order weighted essentially non-oscillatory scheme (WENO-5), a beta-plane approximation, and subgrid-scale viscosity, permits a fourfold increase in the integration time step to 1200 s while reducing the daily mean squared error by up to 26%. Second, a unified autoregressive hybrid block replaces the original chain of 24 specialised modules, eliminating overfitting to specific lead times. Third, the physical core is integrated with two state-of-the-art neural backbones, resulting in PI-PredFormer and PI-IAM4VP. Evaluation on the WeatherBench South Pacific subset from 2000 to 2004 shows that these hybrids reduce root mean squared error at 1-12 h lead times by 8-22% compared to purely neural counterparts, while better preserving physical consistency. These results demonstrate that incremental refinement of hybrid components offers a practical route toward more accurate and efficient short-range weather forecasting.

People make decisions differently in strategic interactions. Some update beliefs like a Bayesian; others exhibit biases like motivated reasoning. Although creators of large language models use simulated humans for safety evaluations and training, they often fail to cover this breadth of human behavior. We argue that cognitive science and economics provide a convenient tool for doing so, making use of mathematical models of human decision-making. We propose an approach that we call Equation-to-Behavior Prompting for guiding large language models to match cognitive models, and evaluate this approach on persuasion games based on legal decision-making. We find that large models can approximate equation-based specifications -- Bayesian updating, affine distortion, motivated updating, and Grether's $α$-$β$ model -- using prompting, but small models fail to do so. However, training small models with reinforcement learning to adhere to mathematical rules, Equation-to-Behavior RL, reduces belief error by 26.5% in out-of-distribution parameterizations. We show that these simulations can help create diverse training environments; training small models to consider different kinds of decision-makers improves average belief change by 2.5%--12% over Bayesian-only training, even when persuading GPT-5-mini. Our work could improve human simulations for training and evaluation in increasingly realistic settings, and could also enable novel research into more complicated mathematical models of human decision-making.

Text-guided Open-vocabulary Object Counting (TOOC) aims to estimate the number of objects described by text prompts, which is particularly challenging in dense scenes with large scale variations. Existing TOOC approaches predominantly rely on Transformers, whose quadratic complexity with respect to image resolution limits their scalability. Mamba offers a promising alternative due to its linear complexity. However, previous Mamba-based methods have two main limitations. On the one hand, the inherent causal formulation of Mamba constrains the bidirectional spatial dependency modeling required by non-causal vision tasks. On the other hand, existing Mamba-based vision models often overlook the unconstrained high entropy in the spatial token responses, which can weaken local details and high-frequency cues. To address these limitations, we propose MambaCount, an efficient framework built on the Spatial Sparse State Space Duality (S^4D) block. Specifically, we analyze and reconstruct the decay dynamics of hidden states in Mamba to alleviate the dependency constraints introduced by causal modeling. Moreover, we introduce a Spatial Token Selection (STS) sub-block to reduce the unconstrained high entropy in spatial token responses within Mamba. In addition, we design Multi-Granularity Prototypes (MGP) to identify object-like regions at different semantic levels, improving cross-modal alignment and interpretability. Extensive experiments on FSC-147 demonstrate that MambaCount achieves state-of-the-art performance among methods without secondary querying, obtaining a test MAE of 12.23, while retaining linear complexity.

The high cost of fine-tuning LLMs poses a significant economic barrier; pre-hoc performance prediction offers a critical solution to substantially reduce this expense. However, the theoretical limits of pre-hoc performance prediction remain unexplored. We formulate it as a stochastic estimation problem under information constraints, decomposing prediction risk into two components: an intrinsic limit (static data-model compatibility) and a reducible optimization variance. We prove that optimization variance admits a necessary lower bound on its decay rate, implying fundamental constraints on how quickly uncertainty dissipates, regardless of the predictor used. Based on these dynamics, we derive a budget-optimal probing principle and introduce a predictability phase diagram that organizes tasks into three distinct regimes: Static-Sufficient, Dynamic-Critical, and Noise-Dominant. Extensive experiments on synthetic and real-world benchmarks validate these theoretical regimes and demonstrate the efficiency of our probing strategy.

Standard accuracy metrics cannot explain why LLMs handle variable tracking but fail on semantically equivalent loops. We study an internal lifecycle of code reasoning in which models first brew the answer, making it linearly recoverable many layers before it becomes self-decodable, and then diverge into one of four resolution outcomes: Resolved, Overprocessed, Misresolved, or Unresolved. Understanding this lifecycle matters because similar task accuracies can mask fundamentally different failure modes that surface-level evaluation cannot detect. We introduce a dual diagnostic framework pairing layer-wise linear probing with Context-Stripped Decoding (CSD) and apply it to six code-reasoning task families across 16 models spanning Qwen, Llama, and DeepSeek architectures. All four outcomes carry substantial mass in every task family: overall Resolved is only 41.5%, with multiple tasks below 30%. Controlled sweeps over structure, depth, and operators expose task-specific failure bottlenecks: Function Call Resolved plunges from 61.1% to 2.5% as call depth increases from one to three. Across architectures and scales, the brewing scaffold remains stable, with normalized brewing duration 24-42% across all 16 models, while resolution success varies with capability. This indicates that the scaffold is a stable empirical regularity across the tested decoder-only Transformer families, whereas resolution success covaries with capability, scale, and training. Code: https://github.com/euyis1019/llm-brewing

Large language model (LLM) web agents are usually deployed as tool callers: each turn, the model reads a fresh page observation and emits one structured tool action. When every action is a low-level primitive, horizons grow quickly and so do policy-facing LLM completions, dominating latency and cost on benchmarks such as Mind2Web and WebArena. Recent systems therefore wrap repeated interaction fragments as web skills: callable tools built from successful trajectories or induced programs, so one call can replace several primitives. However, prior skill libraries are still triggered mainly by instruction similarity or coarse site metadata, which yields low skill reuse on held-out sites and leaves much of the potential step and token reduction on the table. We present SkillMigrator, an agent that learns reusable web skills and transfers them across sites by matching layout structure rather than specific element references. Each induced skill is stored as a transferable interaction pattern (TIP): the skill paired with a structural sketch of the snapshot at induction time. At test time, SkillMigrator retrieves TIPs by layout similarity and grounds their references on the live page. The rest of the stack is standard: accessibility-snapshot observations with stable references, and fixed tool calling over primitives plus skill invocations. Compared with the state-of-the-art approaches, SkillMigrator reduces the average LLM-action count on successful trajectories by 8-10% across both WebArena and Mind2Web at matched success rate.

Datasets in practical document processing scenarios typically grow over time, and their class annotations undergo continuous refinement. This creates significant re-annotation efforts, which are time-consuming and costly. A promising remedy is to re-annotate only a small subset of available documents manually and apply semi-supervised learning techniques that leverage both labelled and unlabelled data. Although there are numerous approaches to tackle this problem for classification, there exists no adaptation for the problem of re-classifying object detection instances, e.g. for document layout analysis. To this end, we propose Bounding Box Label Propagation (BBLP), a pseudo-labelling framework for object detection. An object encoder integrates visual, textual, and positional embeddings from object detection samples to come up with a joint embedding that can be used for Label Propagation on partially annotated datasets in a plug-and-play fashion. Evaluation results indicate that the proposed approach produces high-quality class annotations of bounding boxes. In the D4LA layout analysis dataset, it achieves a mAP of 54.0%, corresponding to 81.6% of fully supervised performance, while using only 10% labelled data. Our work demonstrates the potential of Label Propagation for object detection and lays the groundwork for reducing manual annotation efforts in real-world document processing applications.

Financial multimodal reasoning requires agents to coordinate numerical computation, retrieval, visual interpretation, and temporal grounding across heterogeneous evidence sources. Existing tool-augmented agents improve execution fidelity, yet remain largely stateless across episodes, repeatedly rediscovering reasoning strategies and failure patterns. In high-stakes financial settings, this leads to unreliable tool routing, noisy retrieval, and hallucination-prone reasoning. We present FinAcumen, a financial reasoning agent framework centered on selective experience memory for tool-augmented multimodal reasoning. FinAcumen accumulates financially grounded reasoning experience from prior trajectories, distilling successful strategies and failure-derived cautionary rules into a persistent memory bank. During inference, retrieved experiences condition reasoning only when semantic relevance exceeds a calibrated threshold, while irrelevant memory is explicitly suppressed through a fallback mechanism. A deterministic financial tool environment further grounds numerical computation, retrieval, visual decoding, and answer verification.Across four financial multimodal reasoning benchmarks, FinAcumen consistently improves a frozen 8B vision-language model over finance-specialized models and approaches leading proprietary general-purpose models. Further analysis shows that selective experience activation improves reasoning reliability under retrieval uncertainty. Our code is anonymously available at https://anonymous.4open.science/r/FinAcumen

Generalist embodied agents require more than object recognition: they must reason about spatial relations, actions, procedures, human intentions, environmental constraints, and commonsense consequences from situated visual observations. Yet existing visual and embodied question answering benchmarks often provide limited control over the reasoning dependencies being tested, making it difficult to distinguish grounded embodied reasoning from shortcut-driven visual or linguistic pattern matching. We present ERQA-Plus, a diagnostic benchmark for reasoning in embodied AI. ERQA-Plus contains 1,766 question-answer instances grounded in 711 robot-centric images and organized according to a structured taxonomy spanning perceptual, action-centric, social-interaction, navigation-environmental, and contextual commonsense reasoning. The dataset is constructed using a multi-stage generation and validation pipeline that combines taxonomy-guided question generation, automatic quality judging, iterative revision, and human assessment to improve visual grounding, answer validity, and reasoning quality. We benchmark representative general-purpose vision-language models and embodied models, including LLaVA-NeXT-8B, Prismatic-7B, MiniCPM-V-4.5-8B, Qwen3-VL, RoboRefer-8B, and RoboBrain2.5-8B. Although the strongest model, Qwen3-VL-32B, achieves 83.4% overall accuracy and 61.4 SBERT score, category-level results reveal persistent weaknesses in spatial reasoning, procedural reasoning, event prediction, and intention inference. ERQA-Plus therefore provides a fine-grained evaluation framework for measuring not only whether embodied agents answer correctly, but also which forms of embodied reasoning they can and cannot perform reliably. The dataset is available https://huggingface.co/datasets/huggingdas/erqa-plus and the project page at https://github.com/LUNAProject22/erqa-plus.

Building Management Systems (BMS) are essential for optimizing energy efficiency and operational performance in modern buildings. However, the lack of standardization across BMS points from different manufacturers creates significant barriers to integration and data utilization. While the Brick schema offers a standardized ontology for building systems, mapping BMS points to appropriate Brick classes presents three critical challenges: (i) the extensive number of Brick classes (936 in the latest version), (ii) limited domain-specific knowledge in large language models (LLMs), and (iii) substantial manual effort required for verification. To address these challenges, we propose Brick-DICL, a two-stage dynamic in-context learning framework for automated Brick schema classification. Brick-DICL consists of two primary components: metadata-RAG, which retrieves relevant examples to enhance LLMs' domain knowledge, and class-RAG, which narrows down potential Brick classes to address the large classification space. Additionally, we implement a multi-LLM filtering mechanism that compares predictions across multiple models, flagging low-confidence classifications for human review. As a result: (i) General: Brick-DICL is applicable to any building management system regardless of manufacturer or metadata format; (ii) Novel and Powerful: as the first dynamic in-context learning approach for Brick schema classification, Brick-DICL achieves significant classification accuracy improvements on building datasets, outperforming existing methods; (iii) Efficient: our multi-LLM filtering strategy reduces manual verification effort, enabling rapid digital building onboarding. Extensive experiments demonstrate Brick-DICL's effectiveness across diverse building datasets, accelerating the path toward standardized, interoperable building management systems.

Evaluating large language models (LLMs) is important for understanding their capabilities, comparing competing systems, and supporting the deployment of reliable models in practice. For open-ended tasks, pairwise evaluation has become a popular paradigm, in which two responses to the same prompt are compared and the resulting judgments are aggregated into an overall ranking. A central challenge of this paradigm is intransitivity: the induced comparison outcomes may fail to support any coherent global ranking. For example, one may observe cyclic preferences such as $A \succ B \succ C \succ A$, or inconsistencies involving ties such as $A \equiv B\equiv C\neq A$. Such contradictions make the resulting leaderboard unstable and challenging to interpret. In this paper, we propose a prompt perturbation framework for improving the consistency of pairwise LLM evaluation. Our approach generates perturbed variants of each prompt, uses the resulting comparison graphs to identify and filter out structurally inconsistent comparison patterns, and then applies standard ranking methods to the filtered comparisons. A key feature of the proposed framework is that graph-level structural consistency is incorporated explicitly into the evaluation pipeline before ranking aggregation. This provides a simple and principled way to reduce cyclic inconsistencies and improve the reliability of LLM rankings.

Safe and efficient trajectory planning in unknown, cluttered 3D environments constitutes a critical bottleneck for deploying Unmanned Aerial Vehicles (UAVs) in real-world applications. This challenge is further exacerbated by the limited field-of-view (FOV) and sensing range of onboard sensors. Many existing methods either make simplistic assumptions about unexplored space or rely on conservative heuristics such as speed limits or fixed perception patterns, reducing efficiency and generalizing poorly across different sensor types. In this work, we propose a novel planning framework that directly integrates active perception into trajectory optimization, thereby improving safety while preserving efficiency. The perception constraints are derived from the UAV's dynamic model and formulated in the sensor coordinate frame, which enables precise handling of FOV geometry. The velocity-triggered activation mechanism enables the planner to balance perception and motion efficiency. We introduce an active perception sub-trajectory segment with parametric start-time optimization, mitigating collision risks from late obstacle detection. Our formulation enables active perception during arbitrary 3D maneuvers, extending beyond prior methods designed mainly for horizontal motion. All constraints and penalties are incorporated into a differentiable optimization problem, so the planner requires only a simple front-end global path for guidance, rather than a computationally expensive perception-aware path generator. Extensive simulations and real-world experiments demonstrate robust performance across diverse unknown environments with varying sensor configurations.

Memory has become a standard substrate for self-evolving agents, yet retaining experience is not the same as learning how to evolve through it. Existing memory agents can store trajectories, retrieve reflections, or accumulate skills, but often lack the holistic competence to select useful experience, act on it, write reusable knowledge, and maintain a growing repository. We introduce OPD-Evolver, a slow-fast co-evolution framework that cultivates such an agent evolver through on-policy self-distillation. In the fast loop, OPD-Evolver interacts with a four-level memory hierarchy to read, use, write, and maintain experience for rapid test-time evolution. In the slow loop, outcome-calibrated memory attribution and privileged hindsight distill these four abilities into the deployable policy. Across multi-domain benchmarks, OPD-Evolver surpasses memory systems such as ReasoningBank by up to 11.5%, and training-based methods such as Skill0 by ~5.8%. Further analysis shows that OPD-Evolver internalizes high-value experience and memory management, enabling OPD-Evolver-9B to challenge giant counterparts such as Qwen3.5-397B-A17B and Step-3.5-Flash, pointing beyond memory-augmented agents toward genuinely qualified agent evolvers.

Fine-grained action recognition in egocentric video is challenging for Vision-Language Models (VLMs): actions often differ only in small visual cues, and a single model tends to be biased toward a subset of these cues. We propose Divide, Deliberate, Decide, a fully-local, zero-shot multi-agent framework in which (i) a VLM orchestrator chunks the video and proposes a top-k candidate label list per segment, (ii) an ensemble of heterogeneous VLM specialists, drawn from different open model families, engages in a structured deliberation that includes a peer-consultation round of questions, and (iii) agent rankings are aggregated with a Borda count and the orchestrator re-ranks its own prediction in light of the specialists' evidence. The entire pipeline runs locally with no fine-tuning. Experiments show that our method positively improves zero-shot action recognition performance over the baseline, highlighting the influence of a heterogeneous deliberation step, showing that the gain stems from decorrelated model priors rather than from additional compute.

Reference-driven image generation has made rapid progress on identity preservation, but reliable viewpoint control across different subjects remains poorly understood. The difficulty is not merely generating a new image of the target subject: the model must infer the implicit viewpoint of one subject and transfer it to another subject using only image-level evidence, without camera poses, depth, or ray-based conditions. In this setting, existing generators conditioned on multiple image references often rely on spurious semantic correlations, which lead to viewpoint drift, part-level structural mismatches, and missing or unsupported target-specific content. We formulate this challenge as cross-subject viewpoint alignment and propose RAVA, a retrieval-augmented framework that supplies explicit geometric evidence before generation. RAVA first learns a cross-instance viewpoint embedding that retrieves target-subject images aligned with the anchor viewpoint, then applies a LogDet-based subset selection strategy to retain a compact reference set that is both view-consistent and structurally complementary. The selected references are finally consumed by a fine-tuned multi-reference image generator. Experiments show that generic semantic embeddings are nearly random for this task, while the proposed retriever substantially improves viewpoint retrieval quality. On cross-subject generation, RAVA consistently outperforms zero-shot baselines and stronger retrieval alternatives under the same generation backbone. These results indicate that cross-subject viewpoint alignment benefits from retrieval-augmented geometric grounding rather than relying on end-to-end generation alone.

Estimating human proficiency from video is a key challenge for automated skill assessment, with applications in sports coaching, music pedagogy, surgical training, and workplace learning. Existing approaches often focus on individual scenarios or rely on shared multi-view aggregation, limiting their ability to adapt to heterogeneous camera viewpoints and activity domains. We introduce SkillMoV, a unified, parameter-efficient framework for multi-scenario proficiency estimation from synchronized multi-view video. At its core, SkillMoV introduces a Mixture-of-View Projector (MoVP), which adapts the mixture-of-experts paradigm to camera-specific view features. MoVP is composed of four stages: (i) a Mixture-of-View soft router with twelve expert MLPs that learns view-dependent expert preferences without camera-identity supervision; (ii) cross-view attention to align synchronized cameras; (iii) learnable prototype anchoring to condition the representation on class-level reference vectors; and (iv) a prototype-conditioned gated projection that produces the final skill embedding. We evaluate SkillMoV on EgoExo4D across six skill domains and three separately trained view configurations: Ego, Exos, and Ego+Exos. SkillMoV reaches 50.17% overall accuracy in the Exos setting with a single model trained jointly across all scenarios, surpassing the strongest reported Exos result among the compared methods by 3.57 percentage points. In Ego+Exos, SkillMoV remains close to the best reported result in that setting (47.63% versus 48.20%). Ablations on the selected Exos configuration validate each component: MoV routing contributes +6.61 pp over attentive aggregation, cross-view attention +4.92 pp, prototype anchoring +4.07 pp, and stochastic view dropout +3.90 pp. Through LoRA adaptation, SkillMoV trains only 23.32% of its parameters and adds limited measured overhead relative to a LoRA-only baseline.

Compressing large language models reduces memory use and inference cost, but it can also create failures that standard benchmarks miss. A pruned model may still perform well on multiple-choice evaluations, yet fail to answer the same question in open generation. We ask what pruning changes: does it erase the correct answer, or does it make the answer harder to produce as the top output? We study this question with multilingual question answering, tracking the same questions before and after pruning. We find a benchmark illusion. Under high-sparsity pruning, especially Wanda, models often fail in greedy open generation while still selecting the correct answer under multiple-choice scoring. In these recognition-only errors, the answer is usually not gone, but demoted: it often reappears with beam search, sampling, or one in-context example. Overall, multiple-choice benchmarks can overstate the usability of compressed LLMs, creating an evaluation blind spot. Compressed models should be tested on what they can produce, not only on what they can recognize.

The widespread deployment of face recognition (FR) systems exposes personal images shared on social media and public platforms to identity linkage and privacy risks. Existing adversarial privacy protection methods can degrade unauthorized FR performance but are not compatible with generative face editing. Artificial intelligence-driven face editing tools are gaining popularity, which has significantly increased user demand for personalized portrait generation and social sharing. However, current editing methods often preserve identity features, making the edited images still susceptible to tracking by malicious FR systems. Thus, this paper proposes Flux-Guard, a privacy-preserving face editing framework based on adversarial attacks, which integrates face editing and privacy protection within a unified generative process. Specifically, we design a flow trajectory control method to align semantic manipulations with the generative process and introduce latent-space adversarial optimization with an adaptive perceptual-loss-driven weighting strategy, dynamically adjusting adversarial strength to maximize attack effectiveness while preserving visual quality. Extensive experiments demonstrate that Flux-Guard supports face editing while significantly improving attack success rates against cross-domain face recognition models on the CelebA-HQ and LADN datasets. Furthermore, evaluation results for commercial APIs have confirmed its effectiveness in real-world applications. The code is released at https://github.com/JLMWang/Flux-Guard.