Large Reasoning Models (LRMs) improve task performance through extended inference-time reasoning. Although previous studies suggest that longer reasoning should lead to more robust safety behavior, we find evidence to the contrary: over-extended reasoning can instead be exploited to systematically weaken refusal behavior. We propose Chain-of-Thought Hijacking, a simple yet effective black-box jailbreak attack that induces LRMs to engage in prolonged benign puzzle-solving reasoning, often lasting more than five minutes, before eliciting harmful compliance. Across HarmBench, CoT Hijacking achieves attack success rates of 99%, 94%, 100%, and 94% on Gemini 2.5 Pro, ChatGPT o4 Mini, Grok 3 Mini, and Claude 4 Sonnet, respectively. To understand why this attack succeeds, we conduct activation probing, attention-pattern analysis, and causal interventions on open-source reasoning models. Our results indicate that refusal behavior depends on a low-dimensional safety signal whose expression weakens as reasoning traces grow longer. In particular, extended benign reasoning shifts attention away from harmful intentions and attenuates refusal-related activations, producing what we call refusal dilution. These findings demonstrate that excessively prolonged reasoning can introduce a systematic jailbreak attack surface. We release our evaluation materials to support reproducibility and further research.

Recent advances in sparse reward policy gradient methods have enabled effective reinforcement learning (RL)-based language model post-training. However, for reasoning tasks such as mathematical problem solving, binarized outcome rewards provide limited feedback on intermediate reasoning steps. While some studies have attempted to address this issue by estimating overall reasoning quality, it remains unclear whether these rewards are reliable proxies for the quality of stepwise reasoning. In this study, we consider reasoning as a structured process and propose TACReward, the reward model that can be seamlessly integrated into sparse reward policy gradient methods without additional human annotation costs or architectural modifications. TACReward aggregates stepwise structural deviations between teacher and policy reasoning using process mining techniques, producing a scalar output reward range of [0, 1] to indicate reasoning quality. Experiments on multiple mathematical reasoning benchmarks demonstrate that integrating the TACReward into sparse reward frameworks encourages the policy model to improve the structural quality of reasoning. Consequently, this leads to consistent performance improvements over existing sparse reward frameworks. Our code and checkpoints are publicly available at https://github.com/Thrillcrazyer/TACReward and https://huggingface.co/Thrillcrazyer/TACReward7B.

Social robot navigation increasingly relies on large language models for reasoning, path planning, and enabling movement in dynamic human spaces. However, relying solely on LLMs for planning often leads to unpredictable and unsafe behaviors, especially in dynamic human spaces, due to limited physical grounding and weak logical consistency. In this work, we introduce NaviWM, a socially-aware robot Navigation World Model that augments LLM reasoning with a structured world model and a logic-driven chain-of-thought process. NaviWM consists of two main components: (1) a spatial-temporal world model that captures the positions, velocities, and activities of agents in the environment, and (2) a deductive reasoning module that guides LLMs through a multi-step, logic-based inference process. This integration enables the robot to generate navigation decisions that are both socially compliant and physically safe, under well-defined constraints such as personal space, collision avoidance, and timing. Unlike previous methods based on prompting or fine-tuning, NaviWM encodes social norms as first-order logic, enabling interpretable and verifiable reasoning. Experiments show that NaviWM improves success rates and reduces social violations, particularly in crowded environments. These results demonstrate the benefit of combining formal reasoning with LLMs for robust social navigation. Additional experimental details and demo videos for this work can be found at: https://sites.google.com/view/NaviWM.

Ensuring that predicted probabilities align with observed frequencies is critical in high-stakes domains such as clinical decision support, autonomous driving and financial risk assessment. Existing calibration methods typically apply a single global transformation or rely on post-hoc binning over predicted confidences, limiting their ability to exploit heterogeneous reliability across sub-populations. We propose Clustered Calibration, a representation-aware framework that identifies sub-populations via clustering in learned feature spaces (e.g., coverage vectors, SHAP values, CNN activations, Transformer embeddings) and fits a soft mixture of cluster-specific parametric calibrators under hierarchical shrinkage toward a global mapping. This design yields context-specific calibration while maintaining global stability. Across six tabular datasets and additional image and text benchmarks, clustered calibration consistently improves or matches strong global calibrators in terms of negative log-likelihood and Brier score, while preserving AUC and accuracy. We further show, both analytically and empirically, that fixed-bin Expected Calibration Error (ECE) can mis-rank soft, region-aware calibrators even when proper scoring rules improve, and we advocate for log-loss and Brier as more reliable bases for model selection in such settings.

Reinforcement Learning from AI Feedback (RLAIF) relies on LLM judges as preference measurement instruments, yet these instruments are fundamentally limited by random measurement errors -- stochastic fluctuations that manifest as preference cycles (e.g., $A \succ B \succ C \succ A$), occurring in 5-9% of evaluations across state-of-the-art models. While repeated sampling mitigates noise by averaging multiple judgments, it treats each comparison in isolation and fails to exploit the structural constraints that distinguish systematic signals from random noise. We introduce Topological Consensus Rewards (TCR), a framework that leverages transitivity as a denoising mechanism via topological majority voting: systematic signals reinforce each other through transitive chains, while random errors cluster into topologically exposed cycles. TCR approximates the Maximum Acyclic Subgraph to filter stochastic noise from preference signals. We also propose Cycle Incidence Rate (CIR) as a diagnostic metric that measures the proportion of samples containing preference cycles. Under our noise model, these cycles arise primarily from stochastic measurement errors rather than genuine intransitivity. Experiments on Arena-Hard, MT-Bench, and WritingBench demonstrate that TCR consistently outperforms pairwise baselines and classical ranking algorithms, while exhibiting robust performance across different judge models.

Ensemble-based data assimilation (DA) methods have become increasingly popular due to their inherent ability to address nonlinear dynamic problems. However, these methods often face a trade-off between analysis accuracy and computational efficiency, as larger ensemble sizes required for higher accuracy also lead to greater computational cost. In this study, we propose a novel machine learning-based data assimilation approach that combines the traditional ensemble Kalman filter (EnKF) with a fully connected neural network (FCNN). Specifically, our method uses a relatively small ensemble size to generate preliminary yet suboptimal analysis states via EnKF. A FCNN is then employed to learn and predict correction terms for these states, thereby mitigating the performance degradation induced by the limited ensemble size. We evaluate the performance of our proposed EnKF-FCNN method through numerical experiments involving Lorenz systems and nonlinear ocean wave field simulations. The results consistently demonstrate that the new method achieves higher accuracy than traditional EnKF with the same ensemble size, while incurring negligible additional computational cost. Moreover, the EnKF-FCNN method is adaptable to diverse applications through coupling with different models and the use of alternative ensemble-based DA methods.

High-confidence errors in large language models are often treated as fragile failures. We study an alternative: some errors may be false fixed points, locally stable, internally coherent, and confidently wrong. This separates robustness from truth-tracking. We develop the separation through a Kantian commitment-gate framing and a minimal linear feedback model in which stability and correctness can diverge. Across three open-weight models, overconfident wrong items are not systematically more locally fragile than confidently correct items under our hidden-state sensitivity probes. Abstention-aware self-critique reduces overconfident wrong commitments by sacrificing coverage, and C3-R, a rule-based explicit feedback gate, sharpens that tradeoff rather than eliminating it. These results motivate, but do not establish, high signal-to-noise (high-SNR) inertia and representational compression as possible mechanisms for stable miscalibration.

2048 is a stochastic single-player game involving 16 cells on a 4 by 4 grid, where a player chooses a direction among up, down, left, and right to obtain a score by merging two tiles with the same number located in neighboring cells along the chosen direction. This paper presents that a variant 2048-4x3 12 cells on a 4 by 3 board, one row smaller than the original, has been strongly solved. In this variant, the expected score achieved by an optimal strategy is about $50724.26$ for the most common initial states: ones with two tiles of number 2. The numbers of reachable states and afterstates are identified to be $1,152,817,492,752$ and $739,648,886,170$, respectively. The key technique is to partition state space by the sum of tile numbers on a board, which we call the age of a state. An age is invariant between a state and its successive afterstate after any valid action and is increased two or four by stochastic response from the environment. Therefore, we can partition state space by ages and enumerate all (after)states of an age depending only on states with the recent ages. Similarly, we can identify (after)state values by going along with ages in decreasing order.

Exponentiated gradient descent (EGD), a biologically motivated optimisation algorithm that respects Dale's law, produces log-normally distributed synaptic weights at convergence, in alignment with experimental observations in neuroscience. Since the marginal distribution of geometric Brownian motion (GBM) at any fixed time is log-normal, this convergence property reveals a natural connection between EGD and GBM-based stochastic processes. We propose a multiplicative score-based generative model with GBM as a forward noising process and derive its corresponding reverse-time SDE in both the ambient space and in the $\log$-transformed space. We derive two multiplicative samplers by discretising the corresponding reverse-time SDEs: a sign-agnostic sampler obtained directly from the ambient-space reverse-time SDE, and a sign-preserving sampler, which we refer to as the Dale-Langevin sampler, obtained via the Lamperti transform. We connect the framework to Mirrored Langevin Dynamics, showing that the convex function driving EGD in optimisation precisely governs the Dale-Langevin sampler. While the standard Stein score, defined as $\nabla \log p_{\boldsymbol{X}}(\boldsymbol{x})$ for a random vector $\boldsymbol{X}$ evaluated at $\boldsymbol{x}$, comes up naturally in the additive noise based diffusion models, in the multiplicative setting, we encounter a modified version of the Stein score for sampling, which we refer to as the {\it Hyvärinen score}: $\boldsymbol{x} \circ \nabla \log p_{\boldsymbol{X}}(\boldsymbol{x})$. To estimate the score, we propose a new multiplicative denoising score-matching objective (M-DSM), prove its equivalence to the multiplicative explicit score-matching loss and show that it subsumes the non-negative score matching loss. Experimental results on MNIST, Fashion-MNIST, Kuzushiji-MNIST, and CIFAR-10 to validate the generative capability of the proposed framework.

Music performance is a distinctly human activity, intrinsically linked to the performer's ability to convey, evoke, or express emotion. Machines cannot perform music in the human sense; they can produce, reproduce, execute, or synthesize music, but they lack the capacity for affective or emotional experience. As such, music performance is an ideal candidate through which to explore aspects of collaboration between humans and machines. In this paper, we introduce the witheFlow system, designed to enhance real-time music performance by automatically modulating audio effects based on features extracted from both biosignals and the audio itself. The system, currently in a proof-of-concept phase, is designed to be lightweight, able to run locally on a laptop, and is open-source given the availability of a compatible Digital Audio Workstation and sensors.

Recent Vision-Language-Action (VLA) models show strong generalization capabilities, yet they lack introspective mechanisms for anticipating failures and requesting help from a human supervisor. We present \textbf{INSIGHT}, a learning framework for leveraging token-level uncertainty signals to predict when a VLA should request help. Using $π_0$-FAST as the underlying model, we extract per-token \emph{entropy}, \emph{log-probability}, and Dirichlet-based estimates of \emph{aleatoric and epistemic uncertainty}, and train compact transformer classifiers to map these sequences to help triggers. We explore supervision regimes for strong or weak supervision, and extensively compare them across in-distribution and out-of-distribution tasks. Our results show a trade-off: strong labels enable models to capture fine-grained uncertainty dynamics for reliable help detection, while weak labels, though noisier, still support competitive introspection when training and evaluation are aligned, offering a scalable path when dense annotation is impractical. Crucially, we find that modeling the temporal evolution of token-level uncertainty signals with transformers provides far greater predictive power than static sequence-level scores. This study provides the first systematic evaluation of uncertainty-based introspection in VLAs, opening future avenues for active learning and for real-time error mitigation through selective human intervention.

Is basic visual understanding really solved in state-of-the-art VLMs? We present VisualOverload, a slightly different visual question answering (VQA) benchmark comprising 2,720 question-answer pairs, with privately held ground-truth responses. Unlike prior VQA datasets that typically focus on near global image understanding, VisualOverload challenges models to perform simple, knowledge-free vision tasks in densely populated (or, overloaded) scenes. Our dataset consists of high-resolution scans of public-domain paintings that are populated with multiple figures, actions, and unfolding subplots set against elaborately detailed backdrops. We manually annotated these images with questions across six task categories to probe for a thorough understanding of the scene. We hypothesize that current benchmarks overestimate the performance of VLMs, and encoding and reasoning over details is still a challenging task for them, especially if they are confronted with densely populated scenes. Indeed, we observe that even the best model (o3) out of 37 tested models only achieves 19.6% accuracy on our hardest test split and overall 69.5% accuracy on all questions. Beyond a thorough evaluation, we complement our benchmark with an error analysis that reveals multiple failure modes, including a lack of counting skills, failure in OCR, and striking logical inconsistencies under complex tasks. Altogether, VisualOverload exposes a critical gap in current vision models and offers a crucial resource for the community to develop better models. Benchmark: http://paulgavrikov.github.io/visualoverload

Device-cloud collaboration holds promise for deploying large language models (LLMs), leveraging lightweight on-device models for efficiency while relying on powerful cloud models for superior reasoning. A central challenge in this setting is determining, for each incoming query, whether it should be processed locally or offloaded to the cloud. Existing approaches typically rely on external routers, which often struggle to determine difficulty from the prompt itself, especially for tasks involving complex reasoning. Motivated by this limitation, we propose enabling on-device LLMs to decide internally whether to invoke cloud assistance at inference time, with this capability instilled through reinforcement learning based post-training. Casting on-device LLM post-training as a reward maximization problem, we design hierarchical rewards to encourage local problem solving and judicious cloud offloading. To solve the resulting problem, we develop an algorithm featuring a group-level policy gradient that stabilizes optimization, together with adaptive prompt filtering that provides complementary learning signals to mitigate policy collapse (i.e., exclusive local execution or exclusive cloud offloading). Extensive experiments on on-device-scale LLaMA and Qwen models across multiple reasoning benchmarks show that our method consistently outperforms baselines and significantly narrows the gap to full cloud LLMs.

Standard attention mechanisms in transformers employ static token representations that remain unchanged across all pair-wise computations in each layer. This limits their representational alignment with the potentially diverse relational dynamics of each token-pair interaction. While they excel in domains with relatively homogeneous relationships, standard attention's static relational learning struggles to capture the diverse, heterogeneous inter-channel dependencies of multivariate time series (MTS) data--where different channel-pair interactions within a single system may be governed by entirely different physical laws or temporal dynamics. To better align the attention mechanism for such domain phenomena, we propose attention with dynamic relational priming (prime attention). Unlike standard attention where each token presents an identical representation across all of its pair-wise interactions, prime attention tailors each token dynamically (or per interaction) through learnable modulations to best capture the unique relational dynamics of each token pair, optimizing each pair-wise interaction for that specific relationship. This representational plasticity of prime attention enables effective extraction of relationship-specific information in MTS while maintaining the same asymptotic computational complexity as standard attention. Our results demonstrate that prime attention consistently outperforms standard attention across benchmarks, achieving up to 6.5\% improvement in forecasting accuracy. In addition, we find that prime attention achieves comparable or superior performance using up to 40\% less sequence length compared to standard attention, further demonstrating its superior relational modeling capabilities.

Differential diagnosis is an iterative process that integrates patient information with broader medical knowledge. Clinical case series such as the NEJM Clinicopathologic Conferences (CPCs), published continuously since 1923, feature expert physicians who demonstrate diagnostic reasoning to peers, and have been used for decades to evaluate AI. However, prior AI evaluations have largely focused on final diagnostic accuracy rather than nuanced clinical reasoning. Here, we introduce Dr. CaBot, an agentic AI system that emulates an expert diagnostician by generating written and narrated slide-based presentations from an initial case description alone. CaBot recently generated the first AI diagnosis published in the 100+ year history of the NEJM CPCs. In blinded evaluations, physicians misclassified the source of the differential (CaBot vs. physician-written) in 46/62 (74%) of trials and rated them favorably across quality dimensions. When tasked with solving cases for 72 patients with undiagnosed disease from the NIH Undiagnosed Diseases Network, CaBot identified the working diagnosis in 50/72 (69%) of cases from referral notes alone. To promote transparency and research, we also developed CPC-Bench, a physician-validated benchmark based on 7,102 CPCs and 47,648 questions across 10 tasks. We show that CaBot outperforms frontier models on CPC-Bench, and release both CaBot and CPC-Bench publicly to foster progress in clinical AI.

Offline preference optimization methods are efficient for large language models (LLMs) alignment. Direct Preference optimization (DPO)-like learning, one of the most popular approaches, stands out for its efficiency in reward modeling. However, these methods typically follow the convention to use Bradley-Terry (BT) reward modeling that faces several critical assumptions, including the requirement for pairwise training data, model distribution shifting, human rationality assumption, etc. To address these limitations, we propose a general framework for offline preference optimization methods, Adaptive Preference Optimization with Utility Anchor (UAPO), which introduces an anchoring function to estimate the uncertainties brought from preference data annotation. Our method enables training even in scenarios where the data is unpaired, significantly enhancing data utilization efficiency. Moreover, the anchor design makes UAPO more robust in the training process. Experimental results demonstrate that UAPO achieves competitive outcomes without the strict dependency on data pairing, paving the way for more flexible and effective preference optimization methods.

Pruning is a typical acceleration technique for compute-bound models by removing computation on unimportant values. Recently, it has been applied to accelerate Vision-Language-Action (VLA) model inference. However, existing acceleration methods focus on local information from the current action step and ignore the global context, leading to >20% success rate drop and limited speedup in some scenarios. In this paper, we point out spatial-temporal consistency in VLA tasks: input images in consecutive steps exhibit high similarity, and propose the key insight that token selection should combine local information with global context of the model. Based on this, we propose SpecPrune-VLA, a training-free, two-level pruning method with heuristic control. (1) Action-level static pruning. We leverage global history and local attention to statically reduce visual tokens per action. (2) Layer-level dynamic pruning. We prune tokens adaptively per layer based on layer-wise importance. (3) Lightweight action-aware controller: We classify actions as coarse- or fine-grained by the speed of the end effector and adjust pruning aggressiveness accordingly. Extensive experiments show that SpecPrune-VLA achieves up to 1.57$\times$ speedup in LIBERO simulation and 1.70$\times$ on real-world tasks, with negligible success rate degradation.

High-Intensity Focused Ultrasound (HIFU) is a non-invasive therapy, yet its safety is often degraded by severe acoustic interference during continuous ultrasound guidance. Conventional HIFU interference suppression methods heavily rely on proprietary raw Radio-Frequency (RF) data or complex hardware synchronization, limiting their clinical utility and preventing real-time implementation. To address this limitation, we propose Manifold-Constrained Hyper-Connections Diffusion (mHC-Diff), an image-domain diffusion framework for real-time interference suppression without specialized hardware synchronization, disentangling complex interference from anatomical structures while ensuring high reconstruction fidelity. To achieve clinical real-time application, our approach employs a two-stage strategy: (i) anatomy-aware prior acquisition, where a diffusion model is trained with multi-step UNet as a highfidelity Teacher; and (ii) efficiency distillation, where this prior is distilled into a one-step Student via knowledge distillation to achieve real-time throughput. Extensive validation on a clinically representative dataset across diverse therapeutic scenarios shows that mHC-Diff achieves superior restoration (26.65 dB PSNR), while enabling real-time inference (~20 FPS) on a single NVIDIA RTX 4090, providing a ~6.8x speedup over iterative diffusion baselines (e.g., HIFU-Diff). By eliminating the requirement for specialized hardware synchronization and proprietary RF access, this image-domain framework ensures compatibility and facilitates real-time interference suppression during ultrasound-guided HIFU interventions.

Decision theories offer principled methods for making choices under various types of uncertainty. Algorithms that implement these theories have been successfully applied to a wide range of real-world problems, including materials and drug discovery. Indeed, they are desirable since they can adaptively gather information to make better decisions in the future, resulting in data-efficient workflows. In scientific discovery, where experiments are costly, these algorithms can thus significantly reduce the cost of experimentation. Theoretical analyses of these algorithms are crucial for understanding their behavior and providing valuable insights for developing next-generation algorithms. However, theoretical analyses in the literature are often inaccessible to non-experts. This monograph aims to provide an accessible, self-contained introduction to the theoretical analysis of commonly used probabilistic decision-making algorithms, including bandit algorithms, Bayesian optimization, and tree search algorithms. Only basic knowledge of probability theory and statistics, along with some elementary knowledge about Gaussian processes, is assumed.

Accurate evaluation of procedural communication compliance is essential in simulation-based training, particularly in safety-critical domains where adherence to compliance checklists reflects operational competence. This paper explores a lightweight, deployable approach using prompt-based inference with open-source large language models (LLMs) that can run efficiently on consumer-grade GPUs. We present Prompt-and-Check, a method that uses context-rich prompts to evaluate whether each checklist item in a protocol has been fulfilled, solely based on transcribed verbal exchanges. We perform a case study in the maritime domain with participants performing an identical simulation task, and experiment with models such as LLama 2 7B, LLaMA 3 8B and Mistral 7B, running locally on an RTX 4070 GPU. For each checklist item, a prompt incorporating relevant transcript excerpts is fed into the model, which outputs a compliance judgment. We assess model outputs against expert-annotated ground truth using classification accuracy and agreement scores. Our findings demonstrate that prompting enables effective context-aware reasoning without task-specific training. This study highlights the practical utility of LLMs in augmenting debriefing, performance feedback, and automated assessment in training environments.

In many real-world applications involving static environments, the spatial layout of objects remains consistent across instances. However, state-of-the-art object detection models often fail to leverage this spatial prior, resulting in inconsistent predictions, missed detections, or misclassifications, particularly in cluttered or occluded scenes. In this work, we propose a graph-based post-processing pipeline that explicitly models the spatial relationships between objects to correct detection anomalies in egocentric frames. Using a graph neural network (GNN) trained on manually annotated data, our model identifies invalid object class labels and predicts corrected class labels based on their neighbourhood context. We evaluate our approach both as a standalone anomaly detection and correction framework and as a post-processing module for standard object detectors such as YOLOv7 and RT-DETR. Experiments demonstrate that incorporating this spatial reasoning significantly improves detection performance, with mAP@50 gains of up to 4%. This method highlights the potential of leveraging the environment's spatial structure to improve reliability in object detection systems.

One central goal of robotics is to enable robots to interact with the physical world. Traditional manipulation studies primarily focus on single robots and relatively small objects. However, factory and domestic environments often require large-object manipulation, such as moving tables, where multiple robots must work collaboratively. Existing studies still lack a generalizable framework that can handle diverse objects, tasks, and robot team sizes. In this work, we propose CollaBot, a generalist framework for simultaneous collaborative manipulation. First, we use SEEM for scene segmentation and target-object extraction. Then, we propose a collaborative grasping framework that decomposes the task into local grasp pose generation and global coordination. Finally, we design a two-stage planning module to generate collision-free trajectories for task execution. Experimental results across different settings with varying objects, tasks, and numbers of robots indicate that our framework achieves a 72% success rate. This marks a substantial improvement over behavior cloning-based methods, validating the advantages of the proposed framework in complex multi-robot cooperative tasks. Real-world experiments further demonstrate the feasibility of our method in practical applications.

The cognitive reality of irregular morphological patterns has been debated for decades: do speakers extend them to novel forms, or are they lexical artifacts? A neural network trained on distributional input offers a learnability test: if it recovers the pattern, the pattern is learnable from input statistics alone. We apply this test to the Spanish L-shaped morphome, where the first-person singular indicative stem appears in every present subjunctive cell despite lacking apparent phonological or semantic motivation. We further ask whether the frequency of irregular verbs in the input modulates generalization, evaluating transformers under three frequency conditions (10%, 50%, 90% irregular) and comparing them to human behavioral data. On full-form production from pseudoword inputs all models performed poorly, but all three conditions produced the correct stem more often than humans (43--49% vs. 33%). Response preferences revealed a clear divergence: humans consistently favored regular inflections, whereas models preferred irregular forms more as their proportion in training grew. Models in the naturalistic and balanced conditions were also sensitive to phonological similarity between pseudowords and real Spanish irregular verbs, an effect absent in humans. The L-shaped morphome is thus learnable from distributional input alone, but models generalize it qualitatively differently from humans.

Overestimation in evaluating large language models (LLMs) has become an increasing concern. Due to the contamination of public benchmarks or imbalanced model training, LLMs may achieve unreal evaluation results on public benchmarks, either intentionally or unintentionally, which leads to unfair comparisons among LLMs and undermines their realistic capability assessments. Existing benchmarks attempt to address these issues by keeping test cases permanently secret, mitigating contamination through human evaluation, or repeatedly collecting and constructing new samples. However, these approaches fail to ensure reproducibility, transparency, and high efficiency simultaneously. Moreover, the extent of overestimation in current LLMs remains unquantified. To address these issues, we propose ArxivRoll, a dynamic evaluation framework inspired by one-time pad encryption in cryptography. ArxivRoll comprises two key components: \emph{i) SCP (Sequencing, Cloze, and Prediction)}, an automated generator for private test cases, and \emph{ii) Rugged Scores (RS)}, metrics that measure the proportion of public benchmark contamination and training bias. Leveraging SCP, ArxivRoll constructs a new benchmark every six months using recent articles from ArXiv and employs them for one-time evaluations of LLM performance. Extensive experiments demonstrate the high quality of our benchmark, and we provide a systematic evaluation of current LLMs. The source code is available at https://github.com/liangzid/ArxivRoll/.

Large Language Models have achieved impressive performance on reasoning-intensive tasks, yet optimizing their reasoning efficiency remains an open challenge. While Test-Time Scaling (TTS) improves reasoning quality, it often leads to overthinking, wasting tokens on redundant computations. This work investigates how to efficiently and adaptively guide current model' test-time scaling without additional training. Inspired by the concept of momentum in physics, we propose Momentum Uncertainty-guided Reasoning (MUR), which dynamically allocates thinking budgets to critical reasoning steps by tracking and aggregating stepwise uncertainty over time. To support flexible inference-time control, we introduce gamma-control, a simple mechanism that tunes the reasoning budget via a single hyperparameter. We provide in-depth theoretical proof to support the superiority of MUR in terms of stability and biases. MUR is comprehensively evaluated against various TTS methods across four challenging benchmarks (MATH-500, AIME24, AIME25, and GPQA-diamond) using different sizes of recent Qwen3 models (1.7B, 4B, and 8B). Results demonstrate that MUR reduces computation by by over 45% on average while improving accuracy from 0.33 to 3.46%.

The Web of Agents (WoA) transforms the document-centric Web into an environment of autonomous agents acting on users' behalf, a vision newly tractable as large language models (LLMs) mature. We argue that across three decades the WoA has undergone a \emph{semantic-effort migration} in chronological order: from platform-side coordination (Multi-Agent Systems, Generation~I), through data-side annotation (Semantic Web, Generation~II), to model-side interpretation (LLM-era, Generation~III). The central Gen~II~$\rightarrow$~Gen~III transition within this trajectory, which we call the \emph{semantics-in-data $\rightarrow$ semantics-in-models} shift, is predictive: each generation's failure modes and current open problems follow from where that generation located its semantic effort. The survey makes five contributions: (i)~a unified evolutionary narrative spanning 1990--2026; (ii)~a four-dimensional comparative framework (semantic foundation, communication paradigm, locus of intelligence, discovery mechanism) applied uniformly across all three generations; (iii)~classification of sixteen representative systems on these dimensions, including hybrid LLM--knowledge-graph and computer-use agents; (iv)~coverage of the November~2024--August~2026 institutional convergence (Linux Foundation's Agentic AI Foundation, A2A v1.0, MCP November~2024 launch and November~2025 specification, Visa/Mastercard/Stripe payment-network protocols, EU AI Act phased enforcement, the NIST AI Agent Standards Initiative, International AI Safety Report 2026); and (v)~seven named lessons grounded in cross-generational evidence paired with seven generation-invariant challenges that persist regardless of which protocol prevails. Further progress depends less on protocol design than on the socio-technical infrastructure now being assembled by standards bodies, regulators, and commercial payment networks.

Decentralized finance (DeFi) has introduced a new era of permissionless financial innovation but also led to unprecedented market manipulation. Without centralized oversight, malicious actors coordinate shilling campaigns and pump-and-dump schemes across various platforms. We propose a Multi-Agent Reinforcement Learning (MARL) framework for decentralized manipulation detection, modeling the interaction between manipulators and detectors as a dynamic adversarial game. This framework identifies suspicious patterns using delayed token price reactions as financial indicators.Our method introduces three innovations: (1) Group Relative Policy Optimization (GRPO) to enhance learning stability in sparse-reward and partially observable settings; (2) a theory-based reward function inspired by rational expectations and information asymmetry, differentiating price discovery from manipulation noise; and (3) a multi-modal agent pipeline that integrates LLM-based semantic features, social graph signals, and on-chain market data for informed decision-making.The framework is integrated within the Symphony system, a decentralized multi-agent architecture enabling peer-to-peer agent execution and trust-aware learning through distributed logs, supporting chain-verifiable evaluation. Symphony promotes adversarial co-evolution among strategic actors and maintains robust manipulation detection without centralized oracles, enabling real-time surveillance across global DeFi ecosystems.Trained on 100,000 real-world discourse episodes and validated in adversarial simulations, Hide-and-Shill achieves top performance in detection accuracy and causal attribution. This work bridges multi-agent systems with financial surveillance, advancing a new paradigm for decentralized market intelligence. All resources are available at the Hide-and-Shill GitHub repository to promote open research and reproducibility.

Existing approaches to 3D semantic urban scene generation predominantly rely on voxel-based representations, which are bound by fixed resolution, challenging to edit, and memory-intensive in their dense form. In contrast, we advocate for a primitive-based paradigm where urban scenes are represented using compact, semantically meaningful 3D elements that are easy to manipulate and compose. To this end, we introduce PrITTI, a latent diffusion model that leverages vectorized object primitives and rasterized ground surfaces for generating diverse, controllable, and editable 3D semantic urban scenes. This hybrid representation yields a structured latent space that facilitates object- and ground-level manipulation. Experiments on KITTI-360 show that primitive-based representations unlock the full capabilities of diffusion transformers, achieving state-of-the-art 3D scene generation quality with lower memory requirements, faster inference, and greater editability than voxel-based methods. Beyond generation, PrITTI supports a range of downstream applications, including scene editing, inpainting, outpainting, and photo-realistic street-view synthesis. The source code and more results can be found at https://raniatze.github.io/pritti/.

Carbon dioxide (CO$_2$) emissions are critical indicators of both environmental impact and various industrial processes, including livestock management. We introduce CarboFormer, a lightweight semantic segmentation framework for Optical Gas Imaging (OGI), designed to detect and quantify CO$_2$ emissions across diverse applications. Our approach integrates an optimized encoder-decoder architecture with specialized multi-scale feature fusion and auxiliary supervision strategies to effectively model both local details and global relationships in gas plume imagery while achieving competitive accuracy with minimal computational overhead for resource-constrained environments. We contribute two novel datasets: (1) the Controlled Carbon Dioxide Release (CCR) dataset, which simulates gas leaks with systematically varied flow rates (10-100 SCCM), and (2) the Real Time Ankom (RTA) dataset, focusing on emissions from dairy cow rumen fluid in vitro experiments. Extensive evaluations demonstrate that CarboFormer achieves competitive performance with 84.88\% mIoU on CCR and 92.98\% mIoU on RTA, while maintaining computational efficiency with only 5.07M parameters and operating at 84.68 FPS. The model shows particular effectiveness in challenging low-flow scenarios and significantly outperforms other lightweight methods like SegFormer-B0 (83.36\% mIoU on CCR) and SegNeXt (82.55\% mIoU on CCR), making it suitable for real-time monitoring on resource-constrained platforms such as programmable drones. Our work advances both environmental sensing and precision livestock management by providing robust and efficient tools for CO$_2$ emission analysis.

Deep reasoning is fundamental for solving complex tasks, especially in vision-centric scenarios that demand sequential, multimodal understanding. However, existing benchmarks typically evaluate agents with fully synthetic, single-turn queries, limited visual modalities, and lack a framework to assess reasoning quality over multiple steps as required in real-world settings. To address this, we introduce Agent-X, a large-scale benchmark for evaluating vision-centric agents multi-step and deep reasoning capabilities in real-world, multimodal settings. Agent- X features 828 agentic tasks with authentic visual contexts, including images, multi-image comparisons, videos, and instructional text. These tasks span six major agentic environments: general visual reasoning, web browsing, security and surveillance, autonomous driving, sports, and math reasoning. Our benchmark requires agents to integrate tool use with explicit, stepwise decision-making in these diverse settings. In addition, we propose a fine-grained, step-level evaluation framework that assesses the correctness and logical coherence of each reasoning step and the effectiveness of tool usage throughout the task. Our results reveal that even the best-performing models, including GPT, Gemini, and Qwen families, struggle to solve multi-step vision tasks, achieving less than 50% full-chain success. These findings highlight key bottlenecks in current LMM reasoning and tool-use capabilities and identify future research directions in vision-centric agentic reasoning models. Our data and code are publicly available at https://github.com/mbzuai-oryx/Agent-X