Learning from noisy labels remains a major challenge in medical image analysis, where annotation demands expert knowledge and substantial inter-observer variability often leads to inconsistent or erroneous labels. Despite extensive research on learning with noisy labels (LNL), the robustness of existing methods in medical imaging has not been systematically assessed. To address this gap, we introduce LNMBench, a comprehensive benchmark for Label Noise in Medical imaging. LNMBench encompasses \textbf{10} representative methods evaluated across 7 datasets, 6 imaging modalities, and 3 noise patterns, establishing a unified and reproducible framework for robustness evaluation under realistic conditions. Comprehensive experiments reveal that the performance of existing LNL methods degrades substantially under high and real-world noise, highlighting the persistent challenges of class imbalance and domain variability in medical data. Motivated by these findings, we further propose a simple yet effective improvement to enhance model robustness under such conditions. The LNMBench codebase is publicly released to facilitate standardized evaluation, promote reproducible research, and provide practical insights for developing noise-resilient algorithms in both research and real-world medical applications.The codebase is publicly available on https://github.com/myyy777/LNMBench.

In our prior work, LayerPipe, we had introduced an approach to accelerate training of convolutional, fully connected, and spiking neural networks by overlapping forward and backward computation. However, despite empirical success, a principled understanding of how much gradient delay needs to be introduced at each layer to achieve desired level of pipelining was not addressed. This paper, LayerPipe2, fills that gap by formally deriving LayerPipe using variable delayed gradient adaptation and retiming. We identify where delays may be legally inserted and show that the required amount of delay follows directly from the network structure where inner layers require fewer delays and outer layers require longer delays. When pipelining is applied at every layer, the amount of delay depends only on the number of remaining downstream stages. When layers are pipelined in groups, all layers in the group share the same assignment of delays. These insights not only explain previously observed scheduling patterns but also expose an often overlooked challenge that pipelining implicitly requires storage of historical weights. We overcome this storage bottleneck by developing a pipeline--aware moving average that reconstructs the required past states rather than storing them explicitly. This reduces memory cost without sacrificing the accuracy guarantees that makes pipelined learning viable. The result is a principled framework that illustrates how to construct LayerPipe architectures, predicts their delay requirements, and mitigates their storage burden, thereby enabling scalable pipelined training with controlled communication computation tradeoffs.

Large reasoning models (LRMs) often incur significant key-value (KV) cache overhead, due to their linear growth with the verbose chain-of-thought (CoT) reasoning. This incurs both memory overhead and throughput bottlenecks, limiting efficient deployment. To reduce KV cache size during inference, we first investigate the effectiveness of existing KV cache eviction methods for CoT reasoning. Interestingly, we find that due to unstable token-wise scoring and reduced effective KV budget caused by padding, state-of-the-art (SoTA) eviction methods fail to maintain accuracy in multi-batch settings. Additionally, these methods often generate longer sequences than the original model without eviction, as semantic-unaware token-wise eviction leads to repeated revalidation during reasoning. To address these issues, we present \textbf{SkipKV}, a \textbf{\textit{training-free}} KV compression method that performs selective \textit{eviction} and \textit{generation}, operating at a coarse-grained, sentence-level sequence removal for efficient CoT reasoning. In specific, it introduces a \textit{sentence-scoring metric} to identify and remove highly similar sentences while maintaining semantic coherence. To suppress redundant generation, SkipKV dynamically adjusts a steering vector to update the hidden activation states during inference, enforcing the LRM to generate concise responses. Extensive evaluations on multiple reasoning benchmarks demonstrate that SkipKV achieves up to $\mathbf{26.7}\%$ higher accuracy compared to baseline methods, at a similar compression budget. Additionally, compared to SoTA, SkipKV yields up to $\mathbf{1.6}\times$ shorter generation length while improving throughput by up to $\mathbf{1.7}\times$. Our code is released at: \href{https://github.com/TTTTTTris/SkipKV}{https://github.com/TTTTTTris/SkipKV}.

Language models frequently produce plausible yet incorrect reasoning traces that are difficult to verify. We investigate fine-tuning models to use Prolog as an external symbolic reasoning tool, training Qwen2.5-3B-Instruct with Group Relative Policy Optimization (GRPO) on a cleaned version of GSM8K (which we release as gsm8k-prolog-prover). We systematically vary prompt structure, reward composition (execution, syntax, semantics, structure), and inference protocol (single-try, multiple-try, and two agentic modes). Our reinforcement learning approach outperforms supervised fine-tuning on GSM8K, and the resulting 3B model achieves zero-shot performance on MMLU-STEM and MMLU-Pro competitive with 7B few-shot baselines. Most importantly, we identify an accuracy--auditability trade-off: configurations tuned for correctness alone learn to delegate reasoning to natural language and use Prolog only for the final computation, while configurations rewarded for symbolic structure produce fully auditable programs at a cost in accuracy. We interpret this trade-off as a form of reward hacking and discuss its implications for deploying neurosymbolic systems in safety-critical domains. The source code for our experiments is available under https://github.com/aisilab/Prolog-as-a-Tool

Dexterous manipulation is limited by both control and design, without consensus as to what makes manipulators best for performing dexterous tasks. This raises a fundamental challenge: how should we design and control robot manipulators that are optimized for dexterity? We present a co-design framework that learns task-specific hand morphology and complementary dexterous control policies. The framework supports 1) an expansive morphology search space including joint, finger, and palm generation, 2) scalable evaluation across the wide design space via morphology-conditioned cross-embodied control, and 3) real-world fabrication with accessible components. We evaluate the approach across multiple dexterous tasks, including in-hand rotation with simulation and real deployment. Our framework enables an end-to-end pipeline that can design, train, fabricate, and deploy a new robotic hand in under 24 hours. The full framework will be open-sourced and available on our website: https://an-axolotl.github.io/HouseofDextra/ .

In this study, we evaluate the efficacy of the Mamba architecture bioacoustics by introducing BioMamba, a Mamba-based audio representation model for wildlife sounds. We pre-train a BioMamba using self-supervised learning on a large audio corpus and evaluate it on the BEANS benchmark across diverse classification and detection tasks. Compared to the state-of-the-art Transformer-based model (AVES), BioMamba achieves comparable performance while significantly reducing VRAM consumption. Our results demonstrate Mamba's potential as a computationally efficient alternative for real-world environmental monitoring.

Neural oscillators, originating from second-order ordinary differential equations (ODEs), have demonstrated strong performance in stably learning causal mappings between long-term sequences or continuous temporal functions, as well as in accurately approximating physical systems. However, theoretically quantifying the capacities of their neural network architectures remains a significant challenge. In this study, the neural oscillator consisting of a second-order ODE followed by a multilayer perceptron (MLP) is considered. Its upper approximation bound for approximating causal and uniformly continuous operators between continuous temporal function spaces and that for approximating uniformly asymptotically incrementally stable second-order dynamical systems are derived. The established proof method of the approximation bound for approximating the causal continuous operators can also be directly applied to state-space models consisting of a linear time-continuous complex recurrent neural network followed by an MLP. Theoretical results reveal that the approximation error of the neural oscillator for approximating the second-order dynamical systems scales polynomially with the reciprocals of the widths of two utilized MLPs, thus overcoming the curse of parametric complexity. The convergence rates of two established approximation error bounds are validated through four numerical cases. These results provide a robust theoretical foundation for the effective application of the neural oscillator in science and engineering.

Autonomous navigation in partially observable environments requires agents to reason beyond immediate sensor input, exploit occlusion, and ensure safety while progressing toward a goal. These challenges arise in many robotics domains, from urban driving and warehouse automation to defense and surveillance. Classical path planning approaches and memoryless reinforcement learning often fail under limited fields of view (FoVs) and occlusions, committing to unsafe or inefficient maneuvers. We propose a hierarchical navigation framework that integrates a Deep Transformer Q-Network (DTQN) as a high-level subgoal selector with a modular low-level controller for waypoint execution. The DTQN consumes short histories of task-aware features, encoding odometry, goal direction, obstacle proximity, and visibility cues, and outputs Q-values to rank candidate subgoals. Visibility-aware candidate generation introduces masking and exposure penalties, rewarding the use of cover and anticipatory safety. A low-level potential field controller then tracks the selected subgoal, ensuring smooth short-horizon obstacle avoidance. We validate our approach in 2D simulation and extend it directly to a 3D Unity-ROS environment by projecting point-cloud perception into the same feature schema, enabling transfer without architectural changes. Results show consistent improvements over classical planners and RL baselines in success rate, safety margins, and time to goal, with ablations confirming the value of temporal memory and visibility-aware candidate design. These findings highlight a generalizable framework for safe navigation under uncertainty, with broad relevance across robotic platforms.

Incorporating metadata in Large Language Models (LLMs) pretraining has recently emerged as a promising approach to accelerate training. However prior work highlighted only one useful signal-URLs, leaving open the question of whether other forms of metadata could yield greater benefits. In this study, we investigate a wider range of metadata types and find other types of metadata, such as fine-grained indicators of document quality that can also accelerate pretraining when prepended. We identify a common feature among effective metadata: they encode information at a finer granularity. We further introduce metadata appending as a means of improving training efficiency, where predicting an appropriate metadata as auxiliary task can help speed up pretraining. In addition, learnable meta-tokens trained with masked loss can recover part of the speedup by inducing quality-aware latent structure. Using probing, we analyze latent representations to understand how metadata shapes learning. Together, these results yield practical guidelines for integrating metadata to improve both the efficiency and effectiveness of LLM pretraining.

Reliable depth estimation under real optical conditions remains a core challenge for camera vision in systems such as autonomous robotics and augmented reality. Despite recent progress in depth estimation and depth-of-field rendering, research remains constrained by the lack of large-scale, high-fidelity, real stereo DSLR datasets, limiting real-world generalization and evaluation of models trained on synthetic data as shown extensively in literature. We present the first high-resolution (5472$\times$3648px) stereo DSLR dataset with 18000 images, systematically varying focal length and aperture across complex real scenes and capturing the optical realism and complexity of professional camera systems. For 9 scenes with varying scene complexity, lighting and background, images are captured with two identical camera assemblies at 10 focal lengths (28-70mm) and 5 apertures (f/2.8-f/22), spanning 50 optical configurations in 2000 images per scene. This full-range optics coverage enables controlled analysis of geometric and optical effects for monocular and stereo depth estimation, shallow depth-of-field rendering, deblurring, 3D scene reconstruction and novel view synthesis. Each focal configuration has a dedicated calibration image set, supporting evaluation of classical and learning based methods for intrinsic and extrinsic calibration. The dataset features challenging visual elements such as multi-scale optical illusions, reflective surfaces, mirrors, transparent glass walls, fine-grained details, and natural / artificial ambient light variations. This work attempts to bridge the realism gap between synthetic training data and real camera optics, and demonstrates challenges with the current state-of-the-art monocular, stereo depth and depth-of-field methods. We release the dataset, calibration files, and evaluation code to support reproducible research on real-world optical generalization.

Fabrication uncertainty arising from tolerance accumulation, material imperfection, and positioning errors remains a critical barrier to automated robotic assembly in construction, particularly for contact-rich manipulation tasks governed by friction and geometric constraints. This paper investigates the deployment of diffusion policy learning on construction-scale industrial robots to enable robust, high-precision assembly under such uncertainty, using tight-fitting mortise and tenon timber joinery as a representative case study. Sensory-motor diffusion policies are trained using teleoperated demonstrations collected from an industrial robotic workcell equipped with force/torque sensing. A two-phase experimental study evaluates baseline performance and robustness under randomized positional perturbations up to 10 mm, far exceeding the sub-millimeter joint clearance. The best-performing policy achieved 100% success under nominal conditions and 75% average success under uncertainty. These results provide initial evidence that diffusion policies compensate for misalignments through contact-aware control, representing a step toward robust robotic assembly in construction under tight tolerances.

Probing has emerged as a promising method for monitoring large language models (LLMs), enabling cheap inference-time detection of concerning behaviours. However, natural examples of many behaviours are rare, forcing researchers to rely on synthetic or off-policy LLM responses for training probes. We systematically evaluate how off-policy data influences probe generalisation across eight distinct LLM behaviours. Testing linear and attention probes across multiple LLMs, we find that training data generation strategy can significantly affect probe performance, though the magnitude varies greatly by behaviour. The largest generalisation failures arise for behaviours defined by response ``intent'' (e.g., strategic deception) rather than text-level content (e.g., usage of lists). We then propose a useful test for predicting generalisation failures in cases where on-policy test data is unavailable: successful generalisation to incentivised data (where the model was coerced) strongly correlates with high performance against on-policy examples. Based on these results, we predict that current deception probes may fail to generalise to real monitoring scenarios. We find that off-policy data can yield more reliable probes than on-policy data from a sufficiently different setting. This underscores the need for better monitoring methods that handle all types of distribution shift.

Vision Language Models (VLMs) have achieved impressive performance on spatial reasoning benchmarks, yet these evaluations mask critical weaknesses in understanding object interactions. Current benchmarks test high level relationships ('left of,' 'behind', etc.) but ignore fine-grained spatial understanding needed for real world applications: precise 3D localization, physical compatibility between objects, object affordances and multi step spatial planning. In this work, we present BOP-ASK, a novel large scale dataset for object interaction reasoning for both training and benchmarking. Our data generation pipeline leverages 6D object poses from the Benchmark for Object Pose Estimation (BOP) datasets from which we derive fine grained annotations such as grasp poses, referred object poses, path planning trajectories, relative spatial and depth relationships, and object-to-object relationships. BOP-ASK comprises over 150k images and 33M question answer pairs spanning six tasks (four novel), providing a rich resource for training and evaluating VLMs. We evaluate proprietary and open sourced VLMs, and conduct human evaluations on BOP-ASK-core, a contributed test benchmark. We also release BOP-ASK-lab, an out-of-distribution benchmark with images not sourced from BOP, enabling testing of generalization. Our experiments demonstrate that models trained on BOP-ASK outperform baselines and exhibit emergent capabilities such as precise object and grasp pose estimation, trajectory planning, and fine-grained object-centric spatial reasoning in cluttered environments.

Phase-time arrays, which integrate phase shifters (PSs) and true-time delays (TTDs), have emerged as a cost-effective architecture for generating frequency-dependent rainbow beams in wideband sensing and localization. This paper proposes an end-to-end deep learning-based scheme that simultaneously designs the rainbow beams and estimates user positions. Treating the PS and TTD coefficients as trainable variables allows the network to synthesize task-oriented beams that maximize localization accuracy. A lightweight fully connected module then recovers the user's angle-range coordinates from its feedback of the maximum quantized received power and its corresponding subcarrier index after a single downlink transmission. Compared with existing analytical and learning-based schemes, the proposed method reduces overhead by an order of magnitude and delivers consistently lower two-dimensional positioning error.

Multimodal Large Language Models advance multimodal representation learning by acquiring transferable semantic embeddings, thereby substantially enhancing performance across a range of vision-language tasks, including cross-modal retrieval, clustering, and classification. An effective embedding is expected to comprehensively preserve the semantic content of the input while simultaneously emphasizing features that are discriminative for downstream tasks. Recent approaches demonstrate that MLLMs can be adapted into competitive embedding models via large-scale contrastive learning, enabling the simultaneous optimization of two complementary objectives. We argue that the two aforementioned objectives can be decoupled: a comprehensive understanding of the input enables the embedding model to achieve superior performance on downstream tasks via contrastive learning. In this paper, we propose CoMa, a compressed pre-training phase, which serves as a warm-up stage for contrastive learning. Experiments demonstrate that with only a small amount of pre-training data, we can transform an MLLM into a competitive embedding model. CoMa achieves new state-of-the-art results among MLLMs of comparable size on the MMEB, realizing optimization in both efficiency and effectiveness. Our project is available at https://github.com/Trustworthy-Information-Access/CoMa.

Evaluating log summarization systems is challenging due to the lack of high-quality reference summaries and the limitations of existing metrics like ROUGE and BLEU, which depend on surface-level lexical overlap. We introduce REFLEX, a reference-free evaluation metric for log summarization based on large language model (LLM) judgment. REFLEX uses LLMs as zero-shot evaluators to assess summary quality along dimensions such as relevance, informativeness, and coherence, without requiring gold-standard references or human annotations. We show that REFLEX produces stable, interpretable, and fine-grained evaluations across multiple log summarization dataset, and more effectively distinguishes model outputs than traditional metrics. REFLEX provides a scalable alternative for evaluating log summaries in real-world settings where reference data is scarce or unavailable.

Reinforcement learning with verifiable rewards (RLVR) has emerged as a prominent paradigm for enhancing the reasoning capabilities of large language models (LLMs). However, the entropy of LLMs usually collapses during RLVR training, leading to premature convergence to suboptimal local minima and hindering further performance improvement. Although various approaches have been proposed to mitigate entropy collapse, a comprehensive study of entropy in RLVR remains lacking. To bridge this gap, we conduct extensive experiments to investigate the entropy dynamics of LLMs trained with RLVR and analyze how model entropy correlates with response diversity, calibration, and performance across various benchmarks. Our results identify three key factors that influence entropy: the clipping thresholds in the optimization objective, the number of off-policy updates, and the diversity of the training data. Furthermore, through both theoretical analysis and empirical validation, we demonstrate that tokens with positive advantages are the primary drivers of entropy collapse. Motivated by this insight, we propose Positive-Advantage Reweighting, a simple yet effective approach that regulates model entropy by adjusting the loss weights assigned to tokens with positive advantages during RLVR training, while maintaining competitive performance.

Deep learned (DL) models for image recognition have been shown to fail to generalize to data from different devices, populations, etc. COVID-19 detection from Chest X-rays (CXRs), in particular, has been shown to fail to generalize to out-of-distribution (OOD) data from new clinical sources not covered in the training set. This occurs because models learn to exploit shortcuts - source-specific artifacts that do not translate to new distributions - rather than reasonable biomarkers to maximize performance on in-distribution (ID) data. Rendering the models more robust to distribution shifts, our study investigates the use of fundamental noise injection techniques (Gaussian, Speckle, Poisson, and Salt and Pepper) during training. Our empirical results demonstrate that this technique can significantly reduce the performance gap between ID and OOD evaluation from 0.10-0.20 to 0.01-0.06, based on results averaged over ten random seeds across key metrics such as AUC, F1, accuracy, recall and specificity. Our source code is publicly available at https://github.com/Duongmai127/Noisy-ood

Reasoning over temporal and numerical data, such as time series, is a crucial aspect of fact-checking. While many systems have recently been developed to handle this form of evidence, their evaluation remains limited by existing datasets, which often lack structured evidence, provide insufficient justifications for verdicts, or rely on synthetic claims. In this paper, we introduce TSVer, a new benchmark dataset for fact verification focusing on temporal and numerical reasoning with time-series evidence. TSVer contains 304 real-world claims sourced from 41 fact-checking organizations and a curated database of 400 time series covering diverse domains. Each claim is annotated with time frames across all pertinent time series, along with a verdict and justifications reflecting how the evidence is used to reach the verdict. Using an LLM-assisted multi-step annotation process, we improve the quality of our annotations and achieve an inter-annotator agreement of kappa=0.77 on verdicts. We also develop a baseline for verifying claims against time-series evidence and show that even the state-of-the-art reasoning models like Gemini-2.5-Pro are challenged by time series, achieving a 63.57 accuracy score on verdicts and an Ev2R score of 47.36 on verdict justifications.

We present an ongoing initiative to provide open, very large, high-quality, and richly annotated textual datasets for almost 200 languages. At 30 trillion tokens, this is likely the largest generally available multilingual collection of LLM pre-training data. These datasets are derived from web crawls from different sources and accompanied with a complete, open-source pipeline for document selection from web archives, text extraction from HTML, language identification for noisy texts, exact and near-deduplication, annotation with, among others, register labels, text quality estimates, and personally identifiable information; and final selection and filtering. We report on data quality probes through contrastive and analytical statistics, through manual inspection of samples for 24 languages, and through end-to-end evaluation of various language model architectures trained on this data. For multilingual LLM evaluation, we provide a comprehensive collection of benchmarks for nine European languages, with special emphasis on natively created tasks, mechanisms to mitigate prompt sensitivity, and refined normalization and aggregation of scores. Additionally, we train and evaluate a family of 57 monolingual encoder-decoder models, as well as a handful of monolingual GPT-like reference models. Besides the monolingual data and models, we also present a very large collection of parallel texts automatically mined from this data, together with a novel parallel corpus synthesized via machine translation.

Despite their impressive performance, vision-language models (VLMs) still struggle on culturally situated inputs. To understand how VLMs process culturally grounded information, we study the presence of culture-sensitive neurons, i.e., neurons whose activations show preferential sensitivity to inputs associated with particular cultural contexts. We examine whether such neurons are important for culturally diverse visual question answering and where they are located. Using the CVQA benchmark, we identify neurons of culture selectivity and perform diagnostic tests by deactivating the neurons flagged by various identification methods. Experiments on three VLMs across 25 cultural groups demonstrate the existence of neurons whose ablation disproportionately harms performance on questions about the corresponding cultures, while having limited effects on others. Moreover, we introduce a new margin-based selector Contrastive Activation Margin (ConAct) and show that it outperforms probability- and entropy-based methods in identifying neurons associated with cultural selectivity. Finally, our layer-wise analyses reveal that such neurons are not uniformly distributed: they cluster in specific decoder layers in a model-dependent way.

We present a neuromuscular speech interface that translates electromyographic (EMG) signals recorded from orofacial muscles during speech articulation directly into audio. We find that self-supervised speech (S3) representations are strongly linearly related to the electrical power of muscle activity: a simple linear mapping predicts EMG power from S3 representations with a correlation of r = 0.85. In addition, EMG power vectors associated with distinct articulatory gestures form structured, separable clusters. Together, these observations suggest that S3 models implicitly encode articulatory mechanisms, as reflected in EMG activity. Leveraging this structure, we map EMG signals into the S3 representation space and synthesize speech, enabling end-to-end EMG-to-speech generation without explicit articulatory modeling or vocoder training. We demonstrate this system with a participant with amyotrophic lateral sclerosis (ALS), converting orofacial EMG recorded while she silently articulated speech into audio.

Despite their successes in vision and language, foundation models have stumbled in pathology, revealing low accuracy, instability, and heavy computational demands. These shortcomings stem not from tuning problems but from deeper conceptual mismatches: dense embeddings cannot represent the combinatorial richness of tissue, and current architectures inherit flaws in self-supervision, patch design, and noise-fragile pretraining. Biological complexity and limited domain innovation further widen the gap. The evidence is clear-pathology requires models explicitly designed for biological images rather than adaptations of large-scale natural-image methods whose assumptions do not hold for tissue.

Autoregressive neural surrogates offer computational acceleration for fluid dynamics but inherently suffer from error accumulation and non-physical drift during long-term rollouts. Although hybrid strategies combining surrogate models and physics-based solvers have been proposed, they are limited to manual implementations for low-dimensional benchmarks. In this study, we propose an OpenFOAM-based hybrid framework, XRePIT (eXtensible Residual-based Physics-nformed Transfer learning), characterized by its fastness, robustness, and scalability. Unlike prior manual implementations (e.g., RePIT), XRePIT integrates a fully automated open-source workflow that manages the state transition between a neural surrogate and a traditional numerical solver (OpenFOAM) based on a monitored residual threshold. Using 3D buoyancy-driven flow as a testbed, we demonstrate that this residual-guided coupling enables stable long-term simulation-ell beyond the stability horizon of standalone surrogates. Our results indicate that the hybrid loop achieves up to 2.91x wall-clock acceleration while maintaining relative L2 errors within O(1E-03) Furthermore, we benchmark the framework's extensibility by introducing a finite-volume-based Fourier neural operator (FVFNO), confirming that the stabilizing effect of the residual guardrail is agnostic to the underlying neural architecture. This study provides a deployable methodology for fast, robust, and automated hybrid simulation in 3D unsteady flow.

As LLMs are increasingly deployed in real-world interactions, their social reasoning in interpersonal communication becomes critical. To explore their capabilities, we introduce SCRIPTS, a 1.1k-dialogue dataset in English and Korean, sourced from movie scripts and propose a social reasoning task based on SCRIPTS that evaluates the capacity of LLMs to infer the social relationships (e.g., friends, lovers) between speakers in each dialogue. Evaluating nine models on our task, current LLMs achieve around 75--80% on the English dataset and 58--69% in Korean, and models predict an Unlikely relationship in 10--25% of responses in both languages. Furthermore, we find that thinking models and chain-of-thought prompting provide minimal benefits for social reasoning and occasionally amplify social biases. In sum, there are significant limitations in current LLMs' social reasoning capabilities, especially for Korean, highlighting the need for efforts to develop socially-aware LLMs across languages.

Replicating AI research is a crucial yet challenging task for large language model (LLM) agents. Existing approaches often struggle to generate executable code, primarily due to insufficient background knowledge and the limitations of retrieval-augmented generation (RAG) methods, which fail to capture latent technical details hidden in referenced papers. Furthermore, previous approaches tend to overlook valuable implementation-level code signals and lack structured knowledge representations that support multi-granular retrieval and reuse. To overcome these challenges, we propose Executable Knowledge Graphs (xKG), a pluggable, paper-centric knowledge base that automatically integrates code snippets and technical insights extracted from scientific literature. When integrated into three agent frameworks with two different LLMs, xKG shows substantial performance gains (10.9% with o3-mini) on PaperBench, demonstrating its effectiveness as a general and extensible solution for automated AI research replication. Code is available at https://github.com/zjunlp/xKG.

Multimodal large language models (MLLMs) have rapidly advanced from perception tasks to complex multi-step reasoning, yet reinforcement learning with verifiable rewards (RLVR) often leads to spurious reasoning since only the final-answer correctness is rewarded. To address this limitation, we propose AutoRubric, a framework that integrates RLVR with process-level supervision through automatically collected rubric-based generative rewards. Our key innovation lies in a scalable self-aggregation method that distills consistent reasoning checkpoints from successful trajectories, enabling problem-specific rubric construction without human annotation or stronger teacher models. By jointly leveraging rubric-based and outcome rewards, AutoRubric achieves state-of-the-art performance on six multimodal reasoning benchmarks and substantially improves reasoning faithfulness in dedicated evaluations.

Recent efforts to accelerate LLM pretraining have focused on computationally-efficient approximations that exploit second-order structure. This raises a key question for large-scale training: how much performance is forfeited by these approximations? To probe this question, we establish a practical upper bound on iteration complexity by applying full Gauss-Newton (GN) preconditioning to transformer models of up to 150M parameters. Our experiments show that full GN updates yield substantial gains over existing optimizers, achieving a 5.4x reduction in training iterations compared to strong baselines like SOAP and Muon. Furthermore, we find that a precise layerwise GN preconditioner, which ignores cross-layer information, nearly matches the performance of the full GN method. Collectively, our results suggest: (1) the GN approximation is highly effective for preconditioning, implying higher-order loss terms may not be critical for convergence speed; (2) the layerwise Hessian structure contains sufficient information to achieve most of these potential gains; and (3) a significant performance gap exists between current approximate methods and an idealized layerwise oracle.

Large reasoning models exhibit long chain-of-thought reasoning with complex strategies such as backtracking and self-verification. Yet, these capabilities typically require resource-intensive post-training. We investigate whether such behaviors can be elicited in large models without any gradient updates. To this end, we propose a decoding-time approach, ThinkLogit, which utilizes logit arithmetic to transfer these capabilities from a substantially smaller reasoning guider to a large non-reasoning target. We further show that we can boost performance by training the guider to correct the target's errors using preference optimization over mixed model outputs, a setup we refer to as ThinkLogit-DPO. We evaluate these methods across six reasoning benchmarks spanning math, science, and coding domains using the Qwen2.5-32B guided by R1-Distill-Qwen-1.5B, a model 21x smaller. Our experiments demonstrate that ThinkLogit and ThinkLogit-DPO achieve a relative improvement of 21.5% and 24.2%, respectively, over the target model. Moreover, ThinkLogit remains effective even when the guider and target come from different model families. Crucially, our method requires zero training for the large model and would incur minimal inference overhead when logits are computed in parallel, presenting a practical solution for enabling long reasoning at scale.

Parallel test-time scaling (TTS) is a pivotal approach for enhancing large language models (LLMs), typically by sampling multiple token-based chains-of-thought in parallel and aggregating outcomes through voting or search. Recent advances in latent reasoning, where intermediate reasoning unfolds in continuous vector spaces, offer a more efficient alternative to explicit Chain-of-Thought, yet whether such latent models can similarly benefit from parallel TTS remains open, mainly due to the absence of sampling mechanisms in continuous space, and the lack of probabilistic signals for advanced trajectory aggregation. This work enables parallel TTS for latent reasoning models by addressing the above issues. For sampling, we introduce two uncertainty-inspired stochastic strategies: Monte Carlo Dropout and Additive Gaussian Noise. For aggregation, we design a Latent Reward Model (LatentRM) trained with step-wise contrastive objective to score and guide latent reasoning. Extensive experiments and visualization analyses show that both sampling strategies scale effectively with compute and exhibit distinct exploration dynamics, while LatentRM enables effective trajectory selection. Together, our explorations open a new direction for scalable inference in continuous spaces. Code and checkpoints released at https://github.com/ModalityDance/LatentTTS