Recent advances in large language models (LLMs) have enabled weight-space programming (WSP), which directly generates neural network weights without gradient-based optimization. However, extending WSP to physical problems remains challenging, particularly in achieving both solution accuracy and physical consistency. In this paper, we propose Physics2Weight (PtW), a framework that extends WSP to solving partial differential equations by directly generating neural network weights and editing architectures based on governing equations and initial and boundary conditions. PtW introduces physics-constrained prompting and physics-guided architecture editing, combined with iterative weight refinement from WSP, to improve solution accuracy and physical consistency without gradient-based optimization. Experiments on four PDEs, namely the heat, advection, Burgers', and Allen--Cahn equations, demonstrate lower solution errors than conventional physics-informed neural networks, while revealing limitations in satisfying the governing equations for nonlinear PDEs.
Enterprises increasingly use large language model agents to execute standard operating procedures (SOPs): long, branching, tool-intensive workflows. Yet there is little controlled evidence for choosing among competing agentic architectures. Holding the model, tasks, and tools fixed and varying only the architecture, we compare four designs (a function-calling single agent, a skill-augmented agent, centralized multi-agent orchestration, and a triage swarm; eight arms) across 13 enterprise SOPs, and introduce a skill-definition quality rubric to build matched high and low-quality skill corpora. Three takeaways emerge. (1)~\textbf{There is no universal best architecture; the right choice is SOP-dependent.} On a given SOP, task success ranges from under $10\%$ on the hardest to near $100\%$ on the easiest; no architecture is statistically separable in aggregate, and added structure only plausibly pays off on complex, branching SOPs (a directional, not yet significant, trend). (2)~\textbf{Added architecture cost must be justified by a task-success gain.} Every coordinated arm (multi-agent or skill-augmented) spends $1.4$--$4.7\times$ the tokens and $2.6$--$5.8\times$ the latency of the single agent for gains that are usually small and rarely significant, making the single agent the efficient default; high-quality skills do lift success (to $73.1\%$ from $67.4\%$) while degraded skills fall below it. (3)~\textbf{Implementation quality outweighs the single- versus multi-agent choice.} A plausible multi-agent implementation loses to the single agent, yet a targeted redesign of it becomes the only arm to significantly beat the baseline ($75.6\%$ vs.\ $67.4\%$).
Black-box LLM fingerprinting aims to determine whether a target model was derived from a source model when only the output responses to input queries are observable. One way to reveal this relationship is to perturb each query and observe the resulting responses of both models, yielding two corresponding sets of response variations. These sets provide two complementary sources of information: the relations among response changes within each set, and the correspondence of response changes between the two sets. Existing approaches, however, have not fully exploited both. We introduce \textbf{SetPrint}, a training-free method that jointly captures both within-set relations and between-set correspondence. SetPrint expresses each response change relative to the mean change of its own set, and then compares changes from the two models that were induced by the same perturbation. On LeaFBench, SetPrint consistently outperforms existing training-free black-box fingerprinting methods. Further analyses attribute these gains to both within-set relations and between-set correspondence.
AI systems now produce formally verified proofs of celebrated open problems, and the cost of obtaining a correct proof is falling toward zero. If correctness becomes cheap, what is left for a research paper to communicate? I argue that a paper is not exhausted by a verified statement. It is a narrative exposed to the criticism of a community, made of an intension (the author's choice of problem and perspective), a scientific chain of thought (the reasoning steps leading to the result), and the arguments supporting each step. Most published mathematics is judged on these two elements rather than on correctness alone, and machine-generated formal scripts convey neither. Borrowing from logic and type theory, I formalise the first as a Derivation-Level Intension and the second as a Directional Extension: a decomposition of a proof into cut-points whose axiomatic footprints reveal its weak links, where the result could be strengthened, and its strong links, which can be transported to other theories. Moreover, I examine five roles AI can play in the research process, including bibliographical search and critical judgment, and show that none of them removes the need to declare these two elements. I mention ways these two concepts could be useful for Journals and for training of young mathematicians in a AI era. If proofs become abundant, the scarce resource is not correctness but a community's shared understanding of why a result matters and how it could be pushed further.
Tabular foundation models (TFMs) predict from a labelled context in a single forward pass. A recent benchmark reported that they decisively outperform tree ensembles for customer churn, but it compared them with untuned gradient-boosted trees (GBDTs) and scored only ranking metrics, while a retention campaign is judged in money. We re-evaluate TabPFN-2.5, TabPFN-3.5, TabPFN-3.5-Fast and TabICL v2 against Optuna-tuned GBDTs on five public telecom churn datasets over 10 seeded splits. We separate three questions: whether a model ranks churners well, whether that ranking earns more expected maximum profit (EMP), and whether its probabilities can drive a threshold policy. With a split-aware test and Holm correction over 20 pre-declared comparisons, TabPFN-3.5 earns significantly more EMP than the best tuned GBDT on Cell2Cell (1.9%) and Iranian (7.6%), and these two gains also hold against each of the three tuned GBDTs separately. On Orange it is ahead in all 10 splits (14.3%), but the gain is not significant after correction (Holm p = 0.155). The splits resample the same datasets, so these results do not stand for independent operators. TabICL v2, the earlier benchmark's leader, beats the best tuned GBDT only on Iranian, not all three of them, and underpredicts Orange churn by 2.1 points on a 7.3% base rate. On the datasets and splits where we tested them, a substantially larger tuning budget and a duplicate-aware split leave the main conclusions unchanged. Retuning the baseline cuts the earlier headline gap from 6.1 to 4.0 PR-AUC points without closing it. The licence of the TabPFN-3.5 weights we used (License v1.0), not scoring cost, is what stands between its gains and production use.
The rapid proliferation of autonomously acting AI agents makes the governance of agentic systems urgent. We propose the competence circuit, summarized as IKC + KC + C, that decomposes an agent's real-world impact into three pathways: intelligence-mediated discovery, procedural know-how, and direct control. Using the canonical fast-takeoff debate as a stress test, we formalize the conditions for intelligence-mediated fast takeoff as a conjunction jointly sufficient for fast catastrophic supremacy, showing that most rely on rapid knowledge acquisition or control attainment rather than intelligence alone. We operationalize epistemic intelligence as hypothesis-selection efficiency and validate this framework via stochastic simulation on NK fitness landscapes across six scenarios spanning the conjunctive chain. An analytic bound supported by the simulations shows that intelligence-driven speedup is capped by the ratio of competing intelligence levels regardless of task difficulty. The simulations also expose a fast-takeoff dilemma: a large intelligence-driven speedup multiplier materializes only when the hypothesis space is noisy, but the absolute timescale remains constrained by the time required for capability-critical tests, so a large multiplier need not imply rapid progress. Simulations also indicate that institutional chokepoints disproportionately slow down superintelligent agents when mandatory authorization delays are large relative to the negligible time these agents spend on unproductive search. We argue that the dominant safety levers are restricting control and enforcing knowledge boundaries rather than limiting intelligence alone, and this applies to agentic systems at any capability level. Purely digital domains weaken these bottlenecks, strengthening the case for architectures that intentionally "put physics in the loop."
Reinforcement Learning has achieved tremendous success in recent years. Extending its scope from the traditional complex games such as Atari, Go, Pokeman, and Chess, reinforcement learning has caught researchers' attention on modelling real-life scenarios. Personalized medicine and the dynamic treatment regimen are two such preliminary research fields; researchers are focusing their mindsets currently. Estimating the optimal dynamic treatment regimens from the observational real clinical data is being popularized among the researchers, widening its scope towards more advanced versions of reinforcement learning algorithms. This research is about a research study that focused on reinforcement learning techniques that have been approached in the context of cancer chemotherapy treatments. Primarily this thesis addresses the result of the study by means of the state-of-the-art on reinforcement learning techniques used in generating the optimal policy for sequential treatments, including the existing online reinforcement learning approaches. Thereby a simulation for cancer chemotherapy via traditional Q-learning would be presented. Secondly, the study will extend towards using offline reinforcement learning in the optimal policy generation. Thus, the proposed solution follows a data-driven approach, which works on the real human colon cancer clinical data. The novelty of the proposed solution resides in the implementation of an offline reinforcement learning approach by means of a model-based architecture using the conservative Q-learning and the model-based offline policy optimization algorithm. As per the experimental results and the evaluations, it was revealed that the proposed Chemo RL model caters well the problem of cancer chemotherapy treatment scheduling and serves as a supportive model for the junior oncologists in deciding their treatment decisions.
We study when and how multimodal input improves knowledge graph question answering (KGQA) in large vision--language models. Using a controlled evaluation over 1,500 WebQSP and CWQ questions and six VLMs within a fixed KGQA pipeline, we compare text-only prompting against text-plus-graph visualization at the per-question level and analyze performance changes as a function of reasoning complexity, model size, and image layout. Adding a graph image yields an average gain of 1.85 F1 points, driven by strong precision improvements on low-complexity questions (+4.65 F1). Mixed-effects and regression analyses show that these multimodal gains do not depend strongly on model scale, but are highly sensitive to graph layout: image-only settings sharply degrade performance, hierarchical Graphviz layouts outperform less organized renderings, and decomposing large subgraphs into separated paths yields the largest improvements. Taken together, the results indicate that visual prompting primarily aids KGQA by enhancing prediction precision in low-complexity contexts, provided that visual design respects cognitive and perceptual constraints. To optimize both accuracy and inference efficiency, we argue for adaptive visual prompting policies that selectively append diagrams to small, visually sparse subgraphs with well-structured layouts, avoiding unnecessary visual clutter and token overhead on complex queries. We conclude by relating these findings to a visual notation design theory, outlining how principled diagram encoding can guide future multimodal VLM prompting.
Do large language models represent psychological constructs internally, or do they only generate outputs that resemble psychometric structure? We address this question using a cross-persona paradigm across 14 model variants from two families (Llama 3 and Qwen 2.5, 0.5B-14B, base and instruct). Given responses on one psychological scale from real individuals (N = 272 Taiwanese university students), models predict responses on six other scales. At the behavioral level, larger variants reproduce and amplify human cross-scale correlations; amplification persists after correcting the seven-scale human correlations for measurement attenuation. We then examine whether this structure is reflected internally. Contrastive direction analysis finds geometry aligned with the human matrix in larger instruct variants, but little alignment in base variants under this assay. Across models, geometry strength is associated with behavioral amplification (r = 0.68, p = 0.008; partial r = 0.69 controlling for log size). To relate internal structure to output behavior, a matched activation-probing paradigm shows that probe slopes vary less than behavioral slopes (1.38-1.77 vs 0.52-1.53). A synthetic control shows that ridge probing can inflate these slopes; after subtracting a baseline based on a reliability proxy from another analysis, adjusted slopes remain associated with behavior (r = 0.88). We term the measured probe-behavior gap readout attenuation. These results establish decodability and geometric alignment under the tested assays, without showing that the measured directions mediate the cross-persona answers.
Intelligent landslide detection is essential for disaster prevention. However, existing models primarily focus on isolated visual tasks, require extensive labeled data, and lack deep analytical and scene-understanding capabilities. To address these limitations, we propose LandslideFM, the first unified foundation model that connects visual perception tasks with contextual interpretation of landslides. Specifically, LandslideFM adapts to multiple visual tasks via a Source-Agreement Graduated Trust semi-supervised learning framework that leverages limited labeled data alongside abundant unlabeled data. Furthermore, predicted landslide regions provide explicit spatial guidance to a Vision-Language Model adapted through parameter-efficient fine-tuning, enabling comprehensive landslide reasoning. To support training and evaluation in this domain, we also present LandslideVL, the first benchmark for intelligent landslide analysis, combining visual imagery and newly introduced textual prompts. Comprehensive experiments across public datasets and LandslideVL demonstrate that LandslideFM achieves superior quantitative and qualitative performance. Most importantly, LandslideFM surpasses general-purpose VLMs in specialized landslide reasoning and scene analysis. Our code is released at https://github.com/wangshanwen001/LandslideFM.
Topic monitoring is the task of labelling the topic of each token in a model's response during generation. Activation-based monitors that run at generation time have so far been trained to detect harmful content and are limited to two or three classes. We introduce TopicScope, a framework for training causal topic classifiers on SAE features that assign a topic to every token, so that the label can change partway through a response. We train three detector architectures on synthetic multi-topic conversations covering seven topics. Across Gemma-3 1B and 4B, monitors trained on features extracted from prefill hidden states were able to generalise to decoding. We also found that the classifiers are robust to perturbations, losing at most 5 macro-F1 points when off-topic asides are inserted, topics are removed or sections are reordered. Ablations conducted to understand the contribution of the SAE revealed that classifiers reaching baseline performance without it were several times larger. Our findings show that compact topic monitors can be trained inexpensively in prefill and deployed during generation. More broadly, our work shows where SAEs are useful beyond safety classification. Our code, dataset and checkpoints are available at https://github.com/Mercity-AI/SAE-Guardrails and https://huggingface.co/collections/Mercity/online-guardrails-using-sparse-autoencoders-6ac602940468c62fefb2b9e8
While multimodal agents now carry out complex data analysis with tools, their ability to turn decades of satellite imagery into reliable land-change analysis, crucial for Earth observation science, remains unmeasured. Existing benchmarks fall short: they pose questions on curated images or give agents preprocessed inputs, leaving out the raw multi-sensor archives where such analysis begins. We introduce EarthRecord, a benchmark built on the raw record of 56 sites: four decades of Landsat, MODIS, VIIRS, and Sentinel-2 data, with cloud masking and cross-sensor harmonization left to the agent. It poses 238 land-change questions across ten analysis tasks, scored against withheld references from global products, agreement among methods, and visual interpretation. Two controls test what answers rest on: the closed book removes the record, and instrument pairs alter one satellite's data while nothing changes on the ground. Evaluations of twelve multimodal agents reveal that leading agents complete most analyses end to end but falter when an answer requires mapping from imagery. Further analysis shows that the record moves answers from prior knowledge to the pixels but not yet to the real change on the ground: when one satellite's data are altered, half of the right answers follow the alteration. EarthRecord thus provides a principled platform for measuring and advancing agents from pixels to land change.
Closed-loop autonomous experimentation can accelerate scientific discovery, but implementation effort limits accessibility. Agentic systems may lower this barrier, although sustained operation on physical instruments remains largely unexplored. Here, using coding-agent infrastructure grounded in validated scientific methods, we demonstrate autonomous synchrotron X-ray experiments on nanoparticle superlattices under evolving, human-defined objectives. Complete human-agent and agent-agent transcripts enable an audit that reconstructs decision provenance and characterizes the division of cognitive labor. Across three experiments, 61-67% of messages involved interpretation or judgment, while the share of methodological and scientific contributions increased from 46% in the setup-focused first experiment to 56-61% in later experiments. Provenance showed that 59-68% of error corrections involved multiple participants and that about half of all decisions incorporated the input of another participant. These results may guide design principles for multi-agent scientific systems and inform future agent harnesses, model training, and accessible autonomous experimentation.
Reconstructing high-fidelity LiDAR simulation scenes from real-world driving logs is essential for cost-effective safety validation of autonomous driving systems. While simulators based on Neural Radiance Fields (NeRF) achieve realistic scene reconstruction, their large-scale adoption is limited by slow rendering speeds. Recent 3D Gaussian splatting (3DGS) approaches have improved rendering efficiency through explicit primitive representation and differentiable rasterization. However, existing 3DGS-based methods remain largely restricted to rigid vehicles and produce redundant primitives through unconstrained densification. To address these limitations, we propose URL-Sim, a unified and regulated dynamic LiDAR simulation framework. URL-Sim constructs a dynamic Gaussian scene graph that decouples the modeling of traffic participants, including vehicles, pedestrians, and cyclists, enabling accurate reconstruction of heterogeneous dynamic objects within a single framework. It further introduces a regulated Gaussian densification strategy that ranks primitives with multi-dimensional importance scores and allocates Gaussians to the most critical regions while keeping the total primitive count bounded. Experiments on the Waymo and KITTI-360 datasets demonstrate that URL-Sim achieves higher quantitative accuracy and visual quality than existing baselines under comparable primitive budgets, and accurately recovers the geometry and poses of diverse dynamic traffic participants.
Automatic hexahedral meshing for complex CAD models has long been regarded as the "holy grail" of mesh generation. The fundamental challenge arises from feature misalignment, where orthogonal or interacting features coexist, making direct hexahedral meshing infeasible and requiring decomposition into simpler, meshable blocks. While existing decomposition methods, including frame-field pipelines and learning-based techniques, have reduced manual intervention, they can remain sensitive to geometric and topological complexity and may fail to produce valid decompositions for challenging feature configurations. Specifically, their action spaces are typically restricted to surface-based splitting, which often yields invalid or unmeshable decompositions. To address these limitations, we propose DecompRL, a deep reinforcement learning framework for CAD decomposition toward hexahedral meshing. We design a set of topology-aware internal-splitting actions that leverage both global topological structure and local geometric property to effectively separate misaligned features. Furthermore, we introduce geometric rewards together with meshability feedback to guide the learning process toward valid decompositions and high-quality meshes. Extensive experiments on ABC, Fusion360, NIST, and DeepCAD show that DecompRL achieves the highest success rate among the evaluated learning-based methods on three of the four benchmarks, while maintaining competitive mesh quality. The measured LMFF-to-DecompRL runtime ratio reaches up to two orders of magnitude; this compares the full LMFF procedure with DecompRL decomposition alone, excluding subsequent mesh generation. The largest SR improvement is observed on Fusion360, where DecompRL reaches a success rate of 59.6%.
Video applications are increasingly prevalent, yet their high data volume and temporal dependency make robust wireless transmission challenging under limited communication resources. To address this issue, we propose a token importance-aware semantic communication (Ti-SemCom) framework for bandwidth-efficient video transmission. Specifically, object and motion information is jointly extracted to identify communication-relevant regions, which are hierarchically associated with semantic tokens according to their importance. The transmitted tokens are then dynamically selected based on both token importance and the available bandwidth, prioritizing more informative video content under resource constraints. At the receiver, untransmitted regions reuse recently reconstructed references to further reduce redundant transmission. Experiments on the CAVIAR dataset show that Ti-SemCom can achieve an 88.3\% reduction in communication cost while maintaining competitive reconstruction quality.
Assisted history matching remains non-unique because many geological realisations can reproduce equivalent production responses with substantial differences between individual well matches. This ambiguity increases when a field comprises several depositional settings and modelling strategies. The paper applied a graph autoencoder approach to history matching of a benchmark Brugge field that comprises different depositional zones. Each stratigraphic zone is modelled with its dedicated generative model that comprises a joint latent space for optimisation history matching loops. Graph-based generators operate directly on the active cells of the corner-point grid, so its irregular geometry is handled without conversion to a regular lattice. The optimisation search in the 160-dimensional joint latent space is done with the Covariance Matrix Adaptation Evolution Strategy. The multi-component objective function includes static well data mismatch, dynamic production mismatch, and a geological realism penalty based on the Decoder induced latent metric. Latent space diagnostics with topological data analysis are applied prior to history matching to test whether one generator is sufficient or whether a multiple generator formulation is required. The results show that the zone-based formulation gives coherent latent spaces, preserves the main porosity and permeability relations that vary from one zone to another, and supports history matching while keeping the solutions within the learned geological manifold. The matched models reproduce field oil and water production within the prior uncertainty range and provide a comparable well-level match to previously published Brugge studies based on Ensemble Kalman Filter, Particle Swarm Optimisation with principal component analysis of parameterisation, and Multiple Kernel Learning with Particle Swarm Optimisation.
Social media platforms are a primary source of information for many users. Although research explores various interventions to promote prosocial content or demote harmful content on platforms, there are few systematic methods for evaluating competing intervention designs before conducting costly human-subject experiments. We develop a simulation-based evaluation pipeline built on the OASIS framework. We generate 13,835 LLM-based agents, each with a unique profile. Over 20 time steps, we expose them to real-world Twitter/X posts and let them comment, like, repost, or follow. We use those agents to systematically evaluate three interventions that increase exposure to verified news: adding content, upranking content, and celebrity source cues. All interventions significantly increase news engagement, with upranking news to prominent feed positions producing the largest effects. These effects hold across agents' gender, age, and education levels. We validate our injection treatment results against a comparable human-subject study and find strong directional alignment and similar dose-normalized efficiency. Our results suggest that LLM-based agent simulations can support early-stage evaluation of platform interventions before human-subject deployment, and upranking may be more effective than news volume or source.
Offline web-behavior benchmarks frequently insert the demonstrated action into a finite candidate set and then measure reranking. We show that this protocol can create a candidate-support confound: when the target is outside training support while all distractors are sampled from that support, support membership identifies the gold action without modeling behavior. We formalize this effect with support-identifiability rate (SIR), construct support-balanced candidate sets, and evaluate a target-free dual-path proposer that combines support retrieval with open-vocabulary DOM/LM proposals. On filtered OPeRA (4,282/469/992 train/dev/test steps), the original protocol has SIR 1.0000 on target-OOV rows and a support-only rule reaches 1.0000 OOV Exact@1. Support balancing reduces SIR to 0.0000; the strongest conditional reranker obtains 0.7214 overall Exact@1, 0.7988 on IV rows, and 0.5668 on OOV rows. The target-free dual-path system reaches Proposal Recall@16 0.6915 and end-to-end Exact@1 0.5212, compared with oracle conditional Exact@1 0.7450. On Mind2Web, the same audit yields SIR 0.2380 and an oracle-to-proposal gap of 0.2125. A 200-row semantic audit labels 0.7150 of exact-key OOV targets as semantically novel, 0.2450 as locator/string variation, and 0.0400 as ambiguous (Cohen's $\kappa=0.8173$). These measurements show why support composition must be controlled alongside candidate accuracy when offline results are used to reason about deployment.
Multimodal humor detection requires interpreting language, vocal delivery, and visual behavior jointly. Expressive fusion must also regulate how interactions of different orders enter the final judgment. We propose M\"obius Interaction Filtering (MoIF), which learns a modality-subset prediction function and imposes order-selective spectral regularization through its aggregation structure. Private temporal encoders and balanced reciprocal interaction construct supervised unimodal, bimodal, and trimodal predictors. Their uniform aggregate is exactly a M\"obius-domain filter with a geometrically decreasing interaction-order response. This formulation retains higher-order representations while regulating their contribution to prediction. Controlled UR-FUNNY perturbations reveal a pronounced decline in label-aligned signal-to-noise ratio across interaction orders. The induced filter improves the final ratio by 1.62-2.79 times relative to the co-trained trimodal predictor while preserving its own clean class-aligned signal under perturbation. Five-seed experiments achieve 70.6% accuracy and weighted F1 on UR-FUNNY, exceeding thirteen reproduced methods on these detection metrics. CMU-MOSI and CMU-MOSEI evaluations establish applicability to sentiment prediction, and cross-seed analyses show more stable estimates on all three datasets. MoIF connects the organization of multimodal interactions to their statistical reliability through an executable spectral constraint. Code is available at https://github.com/D-ST-Sword/MoIF.