Deep Research Agents increasingly automate survey generation, yet whether they match human experts at retrieving essential papers and organizing them into expert-like taxonomies remains unclear. Existing benchmarks emphasize writing quality or citation correctness, while standard clustering metrics ignore hierarchical structure. We introduce TaxoBench, a benchmark of 72 highly cited LLM surveys with expert-authored taxonomy trees and 3,815 papers mapped to paper categories. TaxoBench evaluates (1) retrieval via Recall/Precision/F1, and (2) organization at a leaf level (paper-to-category assignment) and a hierarchy level via two new metrics: Unordered Semantic Tree Edit Distance (US-TED/US-NTED) and Semantic Path Similarity (Sem-Path). Two modes are supported: Deep Research (topic-only, end-to-end) and Bottom-Up (expert paper set provided, organization-only). To distinguish disagreement with a single expert reference from genuine model failure, we explicitly partition findings into capability-based (reference-free) and alignment-based (reference-dependent) groups. Evaluating 7 Deep Research Agents and 12 frontier LLMs reveals a dual bottleneck. On the capability side, the best agent retrieves only 20.92% of expert-cited papers, and 1,000 model taxonomies show 75.9% sibling overlap, 51.2% MECE violations, and 83.4% structural imbalance, all detectable without any reference. On the alignment side, all 12 LLMs converge to Sem-Path 28-29%, well below 47-58% achieved by three independent human-annotator groups on the same paper sets. Our benchmark is publicly available at https://github.com/KongLongGeFDU/TaxoBench.

Large language models often produce human-like moral judgments, but it is unclear whether this reflects an internal conceptual structure or superficial ``moral mimicry.'' Using Moral Foundations Theory (MFT) as an analytic framework, we study how moral foundations are encoded, organized, and expressed across 14 base and instruction-tuned LLMs spanning four model families (Llama, Qwen2.5, Qwen3-MoE, Mistral) and scales from 7B to 70B. We employ a multi-level approach combining (i) layer-wise analysis of MFT concept representations and their alignment with human moral perceptions, (ii) pretrained sparse autoencoders (SAEs) over the residual stream to identify sparse features that support moral concepts, and (iii) causal steering interventions using dense MFT vectors and sparse SAE features. We find that models represent and distinguish moral foundations in a manner that aligns with human judgments, and that this moral geometry naturally emerges from pretraining and is selectively rewired by post-training. At a finer scale, SAE features show clear semantic links to specific foundations, suggesting partially disentangled mechanisms within shared representations. Finally, steering along either dense vectors or sparse features produces predictable shifts in foundation-relevant behavior, demonstrating a causal connection between internal representations and moral outputs. Together, our results provide mechanistic evidence that moral concepts in LLMs are distributed, layered, and partly disentangled, suggesting that pluralistic moral structure can emerge as a latent pattern from the statistical regularities of language alone.

The draft IMO MASS Code requires autonomous and remotely supervised maritime vessels to detect departures from their operational design domain, enter a predefined fallback that notifies the operator, permit immediate human override, and avoid changing the voyage plan without approval. Meeting these obligations in the alert-to-takeover gap calls for a short-horizon, human-overridable fallback maneuver. Classical maritime autonomy stacks struggle when the correct action depends on meaning (e.g., diver-down flag means people in the water, fire close by means hazard). We argue (i) that vision-language models (VLMs) provide semantic awareness for such out-of-distribution situations, and (ii) that a fast-slow anomaly pipeline with a short-horizon, human-overridable fallback maneuver makes this practical in the handover window. We introduce Semantic Lookout, a camera-only, candidate-constrained VLM fallback maneuver selector that selects one cautious action (or station-keeping) from water-valid, world-anchored trajectories under continuous human authority. On 40 harbor scenes we measure per-call scene understanding and latency, alignment with human consensus (model majority-of-three voting), short-horizon risk-relief on fire hazard scenes, and an on-water alert->fallback maneuver->operator handover. Sub-10 s models retain most of the awareness of slower state-of-the-art models. The fallback maneuver selector outperforms geometry-only baselines and increases standoff distance on fire scenes. A field run verifies end-to-end operation. These results support VLMs as semantic fallback maneuver selectors compatible with the draft IMO MASS Code, within practical latency budgets, and motivate future work on domain-adapted, hybrid autonomy that pairs foundation-model semantics with multi-sensor bird's-eye-view perception and short-horizon replanning. Website: kimachristensen.github.io/bridge_policy

Reward models (RMs) are essential in reinforcement learning from human feedback (RLHF) to align large language models (LLMs) with human values. However, RM training data is commonly recognized as low-quality, containing inductive biases that can easily lead to overfitting and reward hacking. For example, more detailed and comprehensive responses are usually human-preferred but with more words, leading response length to become one of the inevitable inductive biases. A limited number of prior RM debiasing approaches either target a single specific type of bias or model the problem with only simple linear correlations, \textit{e.g.}, Pearson coefficients. To mitigate more complex and diverse inductive biases in reward modeling, we introduce a novel information-theoretic debiasing method called \textbf{D}ebiasing via \textbf{I}nformation optimization for \textbf{R}M (DIR). Inspired by the information bottleneck (IB), we maximize the mutual information (MI) between RM scores and human preference pairs, while minimizing the MI between RM outputs and biased attributes of preference inputs. With theoretical justification from information theory, DIR can handle more sophisticated types of biases with non-linear correlations, broadly extending the real-world application scenarios for RM debiasing methods. In experiments, we verify the effectiveness of DIR with three types of inductive biases: \textit{response length}, \textit{sycophancy}, and \textit{format}. We discover that DIR not only effectively mitigates target inductive biases but also enhances RLHF performance across diverse benchmarks, yielding better generalization abilities. The code and training recipes are available at https://github.com/Qwen-Applications/DIR.

AI safety and alignment research has predominantly been focused on methods for safeguarding individual AI systems, resting on the assumption of an eventual emergence of a monolithic Artificial General Intelligence (AGI). The alternative AGI emergence hypothesis, where general capability levels are first manifested through coordination in groups of sub-AGI individual agents with complementary skills and affordances, has received far less attention. Here we argue that this patchwork AGI hypothesis needs to be given serious consideration, and should inform the development of corresponding safeguards and mitigations. The rapid deployment of advanced AI agents with tool-use capabilities and the ability to communicate and coordinate makes this an urgent safety consideration. We therefore propose a framework for distributional AGI safety that moves beyond evaluating and aligning individual agents. This framework centres on the design and implementation of virtual agentic sandbox economies (impermeable or semi-permeable), where agent-to-agent transactions are governed by robust market mechanisms, coupled with appropriate auditability, reputation management, and oversight to mitigate collective risks.

Generating controllable indoor scenes is fundamental to applications in game development, architectural visualization, and embodied AI. However, existing approaches either support a limited input modalities or rely on implicit generation processes that hinder precise control over scene structure and semantics. To address these limitations, we introduce RoomPilot, a unified framework for controllable indoor scene synthesis from multi-modal inputs, including textual descriptions and CAD floor plans. RoomPilot maps heterogeneous inputs into an Indoor Domain-Specific Language (IDSL), which serves as a structured and interpretable semantic representation for describing indoor scenes. Built upon IDSL, RoomPilot presents a hierarchical synthesis pipeline that progressively organizes scenes at the building, room, and object levels, promoting structural coherence and functional consistency across multi-room layouts. Moreover, RoomPilot constructs a curated asset dataset with rich semantic annotations to support high-quality scene synthesis, improving visual realism and appearance consistency. Extensive experiments demonstrate effective multi-modal understanding, fine-grained controllability in scene generation, and improved physical consistency and visual fidelity, marking a significant step toward controllable 3D indoor scene synthesis. Code and model will be available.

Collecting robotic manipulation data is expensive, making it impractical to acquire demonstrations for the combinatorially large space of tasks that arise in multi-object, multi-robot, and multi-environment settings. While recent generative models can synthesize useful data for individual tasks, they do not exploit the compositional structure of robotic domains and struggle to generalize to unseen task combinations. We propose a semantic compositional diffusion transformer that factorizes transitions into robot-, object-, obstacle-, and objective-specific components and learns their interactions through attention. Once trained on a limited subset of tasks, we show that our model can zero-shot generate high-quality transitions from which we can learn control policies for unseen task combinations. Then, we introduce an iterative self-improvement procedure in which synthetic data is validated via offline reinforcement learning and incorporated into subsequent training rounds. Our approach substantially improves zero-shot performance over monolithic and hard-coded compositional baselines, ultimately solving nearly all held-out tasks and demonstrating the emergence of meaningful compositional structure in the learned representations.

The advancement of vision-only Bird's-Eye-View (BEV) perception, a core paradigm for cost-effective autonomous driving, is hindered by the long-standing fundamental trade-off between perception accuracy and on-device deployment efficiency. In this work, we introduce Fast-BEV++, a BEV perception framework that resolves this tension through two fundamental design principles: Fast by Algorithm and Deployable by Design. By decomposing the core view transformation module into a hardware-oriented standard Index-Gather-Reshape pipeline, Fast-BEV++ eliminates dependencies on custom kernels while achieving no less than 3 times speedup over the Fast-BEV baseline across mainstream edge platforms. Empirically, Fast-BEV++ establishes a new state-of-the-art result of 0.488 NDS on the nuScenes 3D object detection benchmark, simultaneously delivering real-time inference at more than 134 FPS via our acceleration design. In particular, our integrated, learnable depth module yields consistent performance gains, maintaining the highest accuracy among comparable methods. Overall, this inherently decomposed architecture enables seamless real-time deployment across diverse production-grade automotive platforms, alleviating hardware limitations without compromising perception accuracy or inference efficiency.

Uniform Meaning Representation (UMR) is a novel graph-based semantic representation which captures the core meaning of a text, with flexibility incorporated into the annotation schema such that the breadth of the world's languages can be annotated (including low-resource languages). While UMR shows promise in enabling language documentation, improving low-resource language technologies, and adding interpretability, the downstream applications of UMR can only be fully explored when text-to-UMR parsers enable the automatic large-scale production of accurate UMR graphs at test time. Prior work on text-to-UMR parsing is limited to date. In this paper, we introduce two methods for English text-to-UMR parsing, one of which fine-tunes existing parsers for Abstract Meaning Representation and the other, which leverages a converter from Universal Dependencies, using prior work as a baseline. Our best-performing model, which we call SETUP, achieves an AnCast score of 84 and a SMATCH++ score of 91, indicating substantial gains towards automatic UMR parsing.

Generative search engines based on large language models (LLMs) are replacing traditional search, fundamentally changing how information providers are compensated. To sustain this ecosystem, we need fair mechanisms to attribute and compensate content providers based on their contributions to generated answers. We introduce MaxShapley, an efficient algorithm for fair credit attribution in generative search pipelines that retrieve external sources before generation. MaxShapley is a special case of the celebrated Shapley value; it leverages a de-composable max-sum utility function to compute attributions with polynomial-time computation in the number of documents, as opposed to the exponential cost of Shapley values. We evaluate MaxShapley on three multi-hop QA datasets (HotPotQA, MuSiQUE, MS MARCO); MaxShapley achieves comparable attribution quality to exact Shapley computation, while consuming a fraction of its tokens--for instance, it gives up to a 9x reduction in resource consumption over prior state-of-the-art methods at the same attribution accuracy. We release open-source code and re-calibrated datasets. An educational demo is available at https://fair-search.com.

Traditional cellular traffic forecasting models are optimized for minimizing symmetric errors, leaving them indifferent to shifting operational priorities. To bridge this gap, we introduce BERTO, a BERT-based framework for traffic prediction and energy optimization in cellular networks. Built on transformer architectures, BERTO achieves high prediction accuracy while enabling a single fine-tuned model to operate across multiple forecasting regimes via natural-language operator prompts. By combining a Balancing Loss Function (BLF) with prompt-based conditioning, BERTO adaptively shifts its forecasting bias toward underprediction or overprediction depending on the operator's desired trade-off between power savings and service quality. This allows the same model to dynamically generate different decision-aware forecasts without retraining or modifying model parameters. Experiments on real-world datasets demonstrate that BERTO can operate across a flexible range of approximately 1.4 kW in power consumption while balancing 9x variation in service level agreement (SLA) violations, making it well suited for intelligent RAN deployments.

As we deploy machine learning systems in the real world, a core challenge is to maintain a model that is performant even as the data shifts. Such shifts can take many forms: new classes may emerge that were absent during training, a problem known as open-set recognition, and the distribution of known categories may change. Guarantees on open-set recognition are mostly derived under the assumption that the distribution of known classes, which we call the background distribution, is fixed. In this paper we develop CoLOR, a method that is guaranteed to solve open-set recognition even in the challenging case where the background distribution shifts. We prove that the method works under benign assumptions that the novel class is separable from the non-novel classes, and provide theoretical guarantees that it outperforms a representative baseline in a simplified overparameterized setting. We develop techniques to make CoLOR scalable and robust, and perform comprehensive empirical evaluations on image and text data. The results show that CoLOR significantly outperforms existing open-set recognition methods under background shift. Moreover, we provide new insights into how factors such as the size of the novel class influences performance, an aspect that has not been extensively explored in prior work.

Early detection of malignant lung nodules remains constrained by size and growth based screening criteria, often delaying diagnosis. We present an integrated AI system that jointly performs nodule detection and malignancy assessment directly at the nodule level from low dose CT scans, within a unified CADe/CADx framework. Unlike conventional pipelines separating detection and diagnosis, our approach targets malignant nodules directly, redefining evaluation at the point where clinical decisions are made. To address limitations in dataset scale and explainability, the system consists of a Large Ensemble Model (LEM) combining ensembles of shallow deep learning and feature based models. It was trained and evaluated on 25,709 scans with 69,449 annotated nodules, with external validation on an independent cohort. It achieved an AUC of 0.98 internally and 0.945 externally, outperforming all growth based metrics, Lung RADS size based triage, European volume and VDT based screening criteria, radiologists, and leading AI models. The model maintains high sensitivity at low false positive rates, excels for small and early stage cancers, and enables malignancy assessment up to one year earlier than radiologists for indeterminate and slow growing nodules. This approach has the potential to streamline lung cancer screening workflows and support earlier, more actionable clinical decision making.

Reflections of active markers in the environment are a common source of ambiguity in onboard visual relative localization. This work presents a novel approach that exploits these typically unwanted reflections for onboard relative localization in heterogeneous multi-UAV teams. The method operates without prior knowledge of robot size or predefined marker configurations, remains independent of surface properties, and explicitly accounts for uncertainties caused by surface irregularities, including dynamic water surfaces relevant for marine deployments. We validated the approach in both indoor and outdoor experiments, demonstrating reliable operation across varying lighting conditions and achieving greater effective range (above 30 m) and accuracy than state-of-the-art methods. The video is available under the following link: https://youtu.be/y0zp8cIwkig.

Scalable Vector Graphics are a standard representation for editable visual design, yet they are usually authored as single view two dimensional illustrations. This limits their use in applications that require object level assets to remain coherent when observed, edited, or animated from different viewpoints. We present SVG360, a framework that converts a single input SVG into geometrically and visually consistent multiview SVG assets. The key challenge is that direct per view generation or vectorization produces view dependent regions, fragmented paths, and unstable colors, making the resulting SVGs difficult to edit as a coherent object. SVG360 addresses this problem through a view consistent vectorization pipeline. It first lifts the rasterized input into a view conditioned object representation and renders target views under prescribed cameras. It then propagates part identity across neighboring views using a spatial memory mechanism adapted from video segmentation, establishing consistent region decomposition, path correspondence, and color assignment without task specific retraining. Finally, each view is reconstructed as an editable SVG through structure aware vectorization, where redundant paths are consolidated and local geometry is optimized while preserving boundaries and semantic parts. Experiments on object level SVG assets show that SVG360 improves multiview consistency, reduces path redundancy, and better preserves fine structures compared with direct per view vectorization. By turning a single view SVG into a coherent 360 degree vector asset, SVG360 expands vector graphics from static illustration toward editable multiview content for design, animation, and structured visual editing.

Absolute Pose Regression (APR) has emerged as a compelling paradigm for visual localization. However, APR models typically operate as black boxes, directly regressing a 6-DoF pose from a query image, which can lead to memorizing training views rather than understanding 3D scene geometry. In this work, we propose a geometrically-grounded alternative. Inspired by novel view synthesis, which renders images from intermediate geometric representations, we reformulate APR as its inverse that regresses the underlying 3D representations directly from the image, and we name this paradigm Geometric Representation Regression (GRR). Our model explicitly predicts two disentangled geometric representations in the world coordinate system: (1) a raymap's directions to estimate camera rotation, and (2) a corresponding pointmap to estimate camera translation. The final camera pose is then recovered from these geometric components using a differentiable deterministic solver. This disentangled approach, which separates the learned visual-to-geometry mapping from the final pose calculation, introduces a strong geometric prior into the network. We find that the explicit decoupling of rotation and translation predictions measurably boosts performance. We demonstrate state-of-the-art performance on 7-Scenes and Cambridge Landmarks datasets, validating that modeling the inverse rendering process is a more robust path toward generalizable absolute pose estimation.

Research in bioacoustics, neuroscience, and linguistics often uses birdsong as a proxy to acquire knowledge across diverse areas. This requires audio models to annotate and parse the birdsong. Developing such models requires precise, syllable-level annotated training data. Therefore, automated methods that reduce annotation costs are in demand. This work presents a data-efficient birdsong annotator called Residual Multi-Layer Perceptron Recurrent Neural Network. It then presents a three-stage training pipeline for developing reliable birdsong syllable detectors with minimal annotation. The first stage is self-supervised learning from unlabeled data. Two of the most successful pretraining paradigms are explored, namely, masked prediction and online clustering. The second stage is supervised training with effective data augmentation to produce a robust frame-level syllable detector for each individual. The third stage is a semi-supervised post-training step that refines each individual's model using unlabeled data. The effectiveness of this approach is demonstrated for the Canary song in extreme label-scarcity scenarios. From a signal-processing perspective, the Canary song exhibits one of the most challenging spectro-temporal patterns for algorithmic time-series annotation: rapid vocalizations, brief inter-syllabic intervals, fast and broadband frequency sweeps, and spectrally similar syllables that require fine-grained features to distinguish. Hence, a successful syllable detection algorithm for Canary also establishes a robust baseline for other birds. This methodological generalization is validated in a case study of Bengalese Finch song annotation. Finally, the potential of self-supervised embeddings is assessed for linear probing and unsupervised birdsong analysis.

Diffusion probabilistic models have set a new standard for generative fidelity but are hindered by a slow iterative sampling process. A powerful training-free strategy to accelerate this process is Schedule Optimization, which aims to find an optimal distribution of timesteps for a fixed and small Number of Function Evaluations (NFE) to maximize sample quality. To this end, a successful schedule optimization method must adhere to four core principles: effectiveness, adaptivity, practical robustness, and computational efficiency. However, existing paradigms struggle to satisfy these principles simultaneously, motivating the need for a more advanced solution. To overcome these limitations, we propose the Hierarchical-Schedule-Optimizer (HSO), a novel and efficient bi-level optimization framework. HSO reframes the search for a globally optimal schedule into a more tractable problem by iteratively alternating between two synergistic levels: an upper-level global search for an optimal initialization strategy and a lower-level local optimization for schedule refinement. This process is guided by two key innovations: the Midpoint Error Proxy (MEP), a solver-agnostic and numerically stable objective for effective local optimization, and the Spacing-Penalized Fitness (SPF) function, which ensures practical robustness by penalizing pathologically close timesteps. Extensive experiments show that HSO sets a new state-of-the-art for training-free sampling in the extremely low-NFE regime. For instance, with an NFE of just 5, HSO achieves a remarkable FID of 11.94 on LAION-Aesthetics with Stable Diffusion v2.1. Crucially, this level of performance is attained not through costly retraining, but with a one-time optimization cost of less than 8 seconds, presenting a highly practical and efficient paradigm for diffusion model acceleration.

Large Vision-Language Models (LVLMs) typically process visual inputs as a prefix to the language decoder. As the model autoregressively generates text, this initial visual information inevitably undergoes "dilution" leading the model to over-rely on language priors and hallucinate objects. Existing interventions attempt to correct this by contrasting logits or iteratively refining outputs, but they incur prohibitive latency costs. We propose Residual-Update Directed DEcoding Regulation (RUDDER), a framework that counters visual dilution by creating a persistent visual anchor. We extract a robust evidence direction (CARD) directly from the model's prefill residual updates, and inject it into the decoding process. This injection is modulated by an adaptive gate, the Beta Gate, which acts as a trust mechanism and ensures the visual reminder is applied only when necessary. Experiments on LLaVA-1.5 (7B/13B), Idefics2, InstructBLIP, and Qwen2.5-VL demonstrate that RUDDER consistently mitigates hallucination (with greedy decoding, RUDDER reduces CHAIR_S by an average of 24.4% and CHAIR_i by 23.6% relative) and scales effectively across architectures, all while maintaining >96.0% throughput.

Global plant maps of plant traits, such as leaf nitrogen or plant height, are essential for understanding ecosystem processes, including the carbon and energy cycles of the Earth system. However, existing trait maps remain limited by the high cost and sparse geographic coverage of field-based measurements. Citizen science initiatives offer a largely untapped resource to overcome these limitations, with over 50 million geotagged plant photographs worldwide capturing valuable visual information on plant morphology and physiology. In this study, we introduce PlantTraitNet, a multi-modal, multi-task uncertainty-aware deep learning framework that predictsfour key plant traits (plant height, leaf area, specific leaf area, and nitrogen content) from citizen science photos using weak supervision. By aggregating individual trait predictions across space, we generate global maps of trait distributions. We validate these maps against independent vegetation survey data (sPlotOpen) and benchmark them against leading global trait products. Our results show that PlantTraitNet consistently outperforms existing trait maps across all evaluated traits, demonstrating that citizen science imagery, when integrated with computer vision and geospatial AI, enables not only scalable but also more accurate global trait mapping. This approach offers a powerful new pathway for ecological research and Earth system modeling.

Short-video recommenders such as Douyin must exploit extremely long user behavior histories without breaking latency or cost budgets. We present an end-to-end industrial recommender system that scales long-sequence recommendation modeling to 10K-length histories in production. First, we introduce Stacked Target-to-History Cross Attention (STCA), which replaces history self-attention with stacked cross-attention from the target to the history, reducing complexity from quadratic to linear in sequence length and enabling efficient end-to-end training over long user behavior sequences. Second, we propose Request Level Batching (RLB), a user-centric batching scheme that aggregates multiple targets for the same user/request to share the user-side encoding, substantially lowering sequence-related storage, communication, and compute without changing the learning objective. Third, we design a length-extrapolative training strategy -- train on shorter windows, infer on much longer ones -- so the model generalizes to 10K-scale histories without additional training cost. Across offline and online experiments, we observe predictable, monotonic gains as we scale history length and model capacity, mirroring the scaling law behavior observed in large language models. Deployed at full traffic on Douyin, our system delivers significant improvements on key engagement metrics while meeting production latency, demonstrating a practical path to scaling end-to-end ultra-long sequence recommendation to the 10K regime.

Multiple-choice cloze questions are commonly used to assess linguistic proficiency and comprehension. However, generating high-quality distractors remains challenging, as existing methods often lack adaptability and control over difficulty levels, and the absence of difficulty-annotated datasets further hinders progress. To address these issues, we propose a novel framework for generating distractors with controllable difficulty by leveraging both data augmentation and a multitask learning strategy. First, to create a high-quality, difficulty-annotated dataset, we introduce a two-way distractor generation process to produce diverse and plausible distractors. These candidates are filtered and then categorized by difficulty using an ensemble QA system. Second, this newly created dataset is used to train a difficulty-controllable generation model via multitask learning. Experimental results demonstrate that our method generates high-quality distractors across difficulty levels and substantially outperforms GPT-4o in aligning distractor difficulty with human perception.

Online bilevel optimization (OBO) is a powerful framework for machine learning problems where both outer and inner objectives evolve over time, requiring dynamic updates. Current OBO approaches rely on deterministic \textit{window-smoothed} regret minimization, which may not accurately reflect system performance when functions change rapidly. In this work, we introduce a novel search direction and show that both first- and zeroth-order (ZO) stochastic OBO algorithms leveraging this direction achieve sublinear {stochastic bilevel regret without window smoothing}. Beyond these guarantees, our framework enhances efficiency by: (i) reducing oracle dependence in hypergradient estimation, (ii) updating inner and outer variables alongside the linear system solution, and (iii) employing ZO-based estimation of Hessians, Jacobians, and gradients. Experiments on online parametric loss tuning and black-box adversarial attacks validate our approach.

Estimating the cognitive complexity of reading comprehension (RC) items is crucial for assessing item difficulty before it is administered to learners. Unlike syntactic and semantic features, such as passage length or semantic similarity between options, cognitive features that arise during answer reasoning are not readily extractable using existing NLP tools and have traditionally relied on human annotation. In this study, we examine whether large language models (LLMs) can estimate the cognitive complexity of RC items by focusing on two dimensions-Evidence Scope and Transformation Level-that indicate the degree of cognitive burden involved in reasoning about the answer. Our experimental results demonstrate that LLMs can approximate the cognitive complexity of items, indicating their potential as tools for prior difficulty analysis. Further analysis reveals a gap between LLMs' reasoning ability and their metacognitive awareness: even when they produce correct answers, they sometimes fail to correctly identify the features underlying their own reasoning process.

Fair graph clustering seeks partitions that respect network structure while maintaining proportional representation across sensitive groups, with applications spanning community detection, team formation, resource allocation, and social network analysis. Many existing approaches enforce rigid constraints or rely on multi-stage pipelines (e.g., spectral embedding followed by $k$-means), limiting trade-off control, interpretability, and scalability. We introduce \emph{DFNMF}, an end-to-end deep nonnegative tri-factorization tailored to graphs that directly optimizes cluster assignments with a soft statistical-parity regularizer. A single parameter $λ$ tunes the fairness--utility balance, while nonnegativity yields parts-based factors and transparent soft memberships. The optimization uses sparse-friendly alternating updates and scales near-linearly with the number of edges. Across synthetic and real networks, DFNMF achieves substantially higher group balance at comparable modularity, often dominating state-of-the-art baselines on the Pareto front. The code is available at https://github.com/SiamakGhodsi/DFNMF.git.

Deep learning models have achieved remarkable accuracy in chest X-ray diagnosis, yet their widespread clinical adoption remains limited by the black-box nature of their predictions. Clinicians require transparent, verifiable explanations to trust automated diagnoses and identify potential failure modes. We introduce CXR-LanIC (Language-Grounded Interpretable Classifier for Chest X-rays), a novel framework that addresses this interpretability challenge through task-aligned pattern discovery. Our approach trains transcoder-based sparse autoencoders on a BiomedCLIP diagnostic classifier to decompose medical image representations into interpretable visual patterns. By training an ensemble of 100 transcoders on multimodal embeddings from the MIMIC-CXR dataset, we discover approximately 5,000 monosemantic patterns spanning cardiac, pulmonary, pleural, structural, device, and artifact categories. Each pattern exhibits consistent activation behavior across images sharing specific radiological features, enabling transparent attribution where predictions decompose into 20-50 interpretable patterns with verifiable activation galleries. CXR-LanIC achieves competitive diagnostic accuracy on five key findings while providing the foundation for natural language explanations through planned large multimodal model annotation. Our key innovation lies in extracting interpretable features from a classifier trained on specific diagnostic objectives rather than general-purpose embeddings, ensuring discovered patterns are directly relevant to clinical decision-making, demonstrating that medical AI systems can be both accurate and interpretable, supporting safer clinical deployment through transparent, clinically grounded explanations.

Reinforcement learning with verifiable rewards (RLVR) has become the mainstream technique for training LLM agents. However, RLVR highly depends on well-crafted task queries and corresponding ground-truth answers to provide accurate rewards, which requires significant human effort and hinders the scaling of RL processes, especially in agentic scenarios. Although a few recent works explore task synthesis methods, the difficulty of generated agentic tasks can hardly be controlled to provide effective RL training advantages. To achieve agentic RLVR with higher scalability, we explore self-play training for deep search agents, in which the learning LLM utilizes multi-turn search engine calling and acts simultaneously as both a task proposer and a problem solver. The task proposer aims to generate deep search queries with well-defined ground-truth answers and increasing task difficulty. The problem solver tries to handle the generated search queries and output the correct answer predictions. To ensure that each generated search query has accurate ground truth, we collect all the searching results from the proposer's trajectory as external knowledge, then conduct retrieval-augmentation generation (RAG) to test whether the proposed query can be correctly answered with all necessary search documents provided. In this search self-play (SSP) game, the proposer and the solver co-evolve their agent capabilities through both competition and cooperation. With substantial experimental results, we find that SSP can significantly improve search agents' performance uniformly on various benchmarks without any supervision under both from-scratch and continuous RL training setups. The code is at https://github.com/Qwen-Applications/SSP.

Archaeological predictive modelling estimates where undiscovered sites are likely to occur by combining known locations with environmental and geospatial variables, presenting a positive-unlabeled (PU) learning challenge where confirmed sites are rare and most locations are unlabeled rather than truly negative. To overcome this, we propose asymmetric dual pseudolabeling (DPL), an end-to-end deep learning method that learns from sparse positives directly from multi-band geospatial imagery without hand-crafted feature engineering or assumptions about site absence, and evaluate on two prominent archaeological datasets. On the Sagalassos dataset, evaluated against an independent, held-out field survey, DPL outperforms the LAMAP baseline by 12% in F1 and 29% in Recall, while LAMAP maintains advantages in probability ranking. Standard supervised baselines fail catastrophically when negatives are uncertain; positive-only training collapses to predicting everywhere, es- tablishing empirical bounds. On the Cyprus dataset, a pure PU setting without confirmed negatives, SL inverts probability rankings while DPL recovers discrimination. DPL ensembles produce interpretable probability surfaces supporting survey planning, enabling effective site discovery from minimal labeled data.

We present an experimental recipe for studying the relationship between training data and language model (LM) behavior. We outline steps for intervening on data batches -- i.e., ``rewriting history'' -- and then retraining model checkpoints over that data to test hypotheses relating data to behavior. Our recipe breaks down such an intervention into stages that include selecting evaluation items from a benchmark that measures model behavior, matching relevant documents to those items, and modifying those documents before retraining and measuring the effects. We demonstrate the utility of our recipe through case studies on factual knowledge acquisition in LMs, using both cooccurrence statistics and information retrieval methods to identify documents that might contribute to knowledge learning. Our results supplement past observational analyses that link cooccurrence to model behavior, while demonstrating that extant methods for identifying relevant training documents do not fully explain an LM's ability to correctly answer knowledge questions. Overall, we outline a recipe that researchers can follow to test further hypotheses about how training data affects model behavior. Our code is made publicly available to promote future work.

Generative AI systems are increasingly assisting and acting on behalf of end users in practical settings, from digital shopping assistants to next-generation autonomous cars. In this context, safety is no longer about blocking harmful content, but about preempting downstream hazards like financial or physical harm. Yet, most AI guardrails continue to rely on output classification based on labeled datasets and human-specified criteria,making them brittle to new hazardous situations. Even when unsafe conditions are flagged, this detection offers no path to recovery: typically, the AI system simply refuses to act--which is not always a safe choice. In this work, we argue that agentic AI safety is fundamentally a sequential decision problem: harmful outcomes arise from the AI system's continually evolving interactions and their downstream consequences on the world. We formalize this through the lens of safety-critical control theory, but within the AI model's latent representation of the world. This enables us to build predictive guardrails that (i) monitor an AI system's outputs (actions) in real time and (ii) proactively correct risky outputs to safe ones, all in a model-agnostic manner so the same guardrail can be wrapped around any AI model. We also offer a practical training recipe for computing such guardrails at scale via safety-critical reinforcement learning. Our experiments in simulated driving and e-commerce settings demonstrate that control-theoretic guardrails can reliably steer LLM agents clear of catastrophic outcomes (from collisions to bankruptcy) while preserving task performance, offering a principled dynamic alternative to today's flag-and-block guardrails.