Computer Science (arXiv)

Bitcoin's block reward is scheduled to decline to zero, raising concerns about whether the network can remain secure once miners rely solely on transaction fees. This paper seeks to identify the conditions under which large-scale and persistent deviation from honest mining can arise. We analyze and compare the payoffs of honest and deviating miners in a sequential decision model, and identify a deviation threshold $G_t$ at which honest mining ceases to be privately optimal. Around the 2024 Bitcoin halving, we show that current mining behavior does not exhibit large-scale or structural deviation. However, when the block reward is removed, the $G_t$ criterion implies that deviation can arise even with a very small fraction of transaction fees. Finally, we evaluate three protocol-level mechanisms: Base Fee, Fee Floor, and an adaptive maximum block size rule, and show that their combination raises the deviation threshold and mitigates incentive breakdown in a fee-only regime. These results provide a practical benchmark for assessing Bitcoin's security as block rewards disappear.

The estimation of odometry in legged robots depends on the assumption that the velocity of the foot with respect to the world remains zero during the stance phase. Feedback for the main body velocity is derived from the kinematic serial chain of the feet making accurate leg phase detection is a critical subproblem. A considerable number of studies employ ground reaction force sensors mounted at the tip of the foot to classify, yet these sensors may not be universally available for all legged robots. Additionally, these sensors are often unresponsive to unaccounted disturbances, such as slippage, while the foot remains in contact with the ground. In this study, we propose a self-supervised representation learning framework for contact detection that utilizes the standard sensor set of joint encoders without reliance on force sensor augmentations. We employ learned representations to model the stance and swing phases probabilistically. The experimental results obtained confirm the efficacy of the proposed self-supervised contact detector. Our framework exhibited superior performance in comparison to supervised methods which necessitate sensor set augmentation and labeling, as well as baseline probabilistic approaches. Additionally, we make our code available to the public.

Mutation testing is widely used to evaluate test-suite effectiveness, yet IEC 61131-3 Structured Text (ST) programs still lack a publicly available benchmark that supports reproducible mutation-based research. This gap is especially important because ST is extensively used in Programmable Logic Controllers (PLCs) that operate in real-time, safety-critical industrial environments, where software faults may cause equipment damage, production loss, or unsafe system behavior. To address this need, we present STMutants, a curated mutation testing dataset for industrial automation software. STMutants contains 110 generated first-order mutants derived from 11 ST programs collected from the OSCAT basic library and industrially relevant sources, of which 108 are retained after observability and equivalence screening. The dataset covers seven mutation operator categories adapted from classical taxonomies for the PLC domain, including value, relational, arithmetic, logical, negation, operation insertion/omission, and initialization faults. Each mutant is constructed through a four-phase methodology: fault-type profiling and operator selection, syntactic transformation, compilability verification, and manual equivalence screening with strong inter-rater agreement (kappa = 0.87). To demonstrate the usefulness of the dataset, we evaluate three large language models (LLMs) in a two-phase setting: test-suite generation followed by mutation kill/survive prediction. Across 108 retained mutants, the models achieve mutation detection accuracies of 86.1%, 94.4%, and 86.1%, respectively, with statistical analysis confirming significant performance differences. By providing the first publicly available mutation benchmark for ST programs, STMutants enables reproducible research on automated test generation, mutation analysis, fault localization, and AI-assisted quality assurance for PLC software.

Given the inherently multimodal nature of human experience, vision-language models (VLMs) hold substantial promise for modeling human cognition as it grows and develops with experience. Realizing their potential requires tools for comparing VLMs with human cognitive development across tasks, ages, and populations. We present LEVANTE-bench, a benchmark based on tasks and data from the Learning Variability Network (LEVANTE), which distributes open-source tasks and data measuring children's cognition across languages and cultures. In LEVANTE-bench, we systematically assess VLMs on six tasks, comparing their alignment with children aged 5-12 ($N$ = 1547) across three countries. We compare models at multiple scales, assessing their overall accuracy, their task- and item-level alignment with children, and how well they match children's trial-level error distributions. Alignment was heterogeneous across scales: at the level of tasks and items, more capable models aligned better with humans. However, match to human error distributions varied widely across tasks, and for several tasks, smaller models matched younger children's errors better. In addition, even the best-performing VLMs struggled on matrix reasoning and mental rotation tasks. Thus, current VLM architectures align only partially with the cognitive abilities of children.

Achieving peak GPU performance remains a significant challenge as the system throughput is constrained by host-device synchronization delays and kernel scheduling overheads, even with aggressive kernel optimizations and batch processing. Furthermore, existing approaches often underutilize hardware resources such as compute cores and copy engines due to scheduling overheads. To address these problems, we propose a CUDA runtime framework for task-parallel pipelines to minimize the synchronization overheads and the gap between kernel executions. The proposed solution combines two innovations: (1) a multi-stream task-parallel pipeline programming model that leverages event-chaining and work-stealing mechanisms to fully utilize available hardware resources; (2) a graph-based execution flow with per-stream buffers to ensure memory safety for multiple in-flight jobs running concurrently. Extensive evaluations on representative real-world workloads show 1.15--1.44X speedup and reduce scheduling overheads by 18--54% compared to state-of-the-art CUDA graph baselines.

Automatic text summarization has become increasingly important due to the rapid growth of digital textual information. This paper presents a Multi-Model Adaptive Summarization Framework designed to improve the robustness and quality of abstractive text summarization. Relying on a single model often leads to inconsistent summarization quality across articles with varying structures and topics. To address this limitation, the proposed framework integrates multiple fine-tuned transformer-based summarization models and introduces an adaptive selection mechanism. In this framework, each model independently generates a candidate summary for the same input article. The generated summaries are then evaluated using automatic evaluation metrics that capture both lexical similarity and semantic relevance. Based on these scores, the framework selects the highest-quality summary as the final output. The models are fine-tuned and evaluated on the widely used CNN/DailyMail news summarization dataset. Experimental results demonstrate that the proposed framework achieves the highest BERTScore among all compared methods with a score of 88.63%. It also outperforms several LLMs such as GPT3-D2, Falcon-7b, and Mpt-7b, highlighting its effectiveness and robustness. These findings highlight the effectiveness of leveraging multiple transformer-based models within an adaptive selection strategy to improve the quality and robustness of automatic text summarization systems.

Reverse engineering (RE) is a critical activity in software engineering and cybersecurity, supporting tasks such as malware analysis, vulnerability discovery, legacy system maintenance, and firmware inspection. Despite its importance, there is limited empirical understanding of the challenges, topics, and knowledge gaps faced by RE practitioners in real-world settings, and no publicly available dataset has systematically captured RE discussions from developer Q&A forums. In this paper, we present REStack, a large-scale dataset of RE discussions collected from Stack Overflow and the dedicated Reverse Engineering Stack Exchange site. The dataset comprises over 12,000 RE-related posts spanning more than 15 years. Using Latent Dirichlet Allocation (LDA) with Genetic Algorithm (GA)-based hyperparameter optimization, followed by manual topic labeling, we identify 23 semantically coherent RE topics grouped into six high-level thematic categories. The dataset is further enriched with metadata and difficulty indicators derived from community interaction signals, such as unanswered rates and response times. Our analysis reveals that RE discussions are predominantly practical and task-oriented, with strong emphasis on debugging, decompilation, and system-level analysis, while topics related to memory, firmware, and file format analysis exhibit high difficulty and unresolved rates. Beyond empirical characterization, REStack provides a reusable resource for empirical studies, educational research, and the development and evaluation of AI- and LLM-based developer assistance tools for RE. By releasing the dataset and accompanying scripts, this work aims to facilitate reproducible research and advance data-driven support for RE practice.

Multi-modal novel view synthesis (NVS) combining RGB and thermal imagery enables precise 3D scene reconstruction with visual and thermal information. However, existing methods typically rely on precisely calibrated RGB-thermal image pairs or stereo setups, limiting scalability and practical deployment. To address this, we introduce a framework for unpaired RGB-thermal NVS that leverages VGGT, a 3D feed-forward transformer architecture, to independently estimate camera poses for each modality. The pose sets are then aligned using the Procrustes algorithm with a cross-modal feature matcher, enabling joint registration without paired calibration. Building on this alignment, we further propose a multi-modal 3D Gaussian Splatting approach that learns directly from unpaired RGB and thermal images. Experiments on diverse scenes demonstrate that our method achieves competitive performance in thermal view synthesis while maintaining RGB fidelity. Moreover, we show that existing reconstruction approaches can produce modality-specific reconstructions that lack cross-modal consistency. We thus introduce a benchmarking framework to rigorously evaluate both per-modality image synthesis and the multi-modal coherence of reconstructed scenes.

Retrieval systems underpin modern AI applications -- spanning visual search, recommendation engines, and multi-modal question answering. Modern multi-stage retrieval systems require the joint optimization of highly coupled parameters, yet traditional hyperparameter optimization (HPO) methods -- including Tree-structured Parzen Estimators (TPE) and Gaussian Process Bayesian Optimization -- rely on an independence assumption that fundamentally prevents them from navigating these coupled configuration spaces. We address this limitation with a phase-aware large language model (LLM) agent that conditions each proposal on its full optimization history, navigating the coupled parameter space across phase-partitioned exploration, exploitation, and fine-tuning stages. Evaluated on the HICO-DET human-object interaction retrieval benchmark using Intel VDMS (Visual Data Management System), our agent outperforms Optuna TPE by +33.3% and VDTuner by +34.2% under SIEVE (Safeguarded Index Evaluation of Vector-search Efficiency, a quality-constrained throughput metric), delivering a 15.3x throughput gain over UniIR. Validation across three benchmarks confirms that the agent's advantage grows with the degree of parameter coupling: +33.3% on HICO-DET (high coupling), methods converge within 1% on GLDv2 (moderate coupling) and within 3.6% on SIFT1M (near-independent control). Cross-system validation on Milvus confirms the optimizer ranks first on all three datasets without modification, demonstrating transferability across vector database management system (VDBMS) platforms.

Identifying subtypes of complex conditions, such as Inflammatory Bowel Disease (IBD), often requires capturing latent patterns in longitudinal omics data. However, these data are typically high-dimensional, sparsely sampled, and irregularly observed over time, posing substantial challenges for conventional (bi)clustering and functional data analysis methods. We propose Tri-SfSVD, a unified sparse functional Singular Value Decomposition framework for discovering biclusters and triclusters in longitudinal data. Unlike existing functional biclustering methods that rely on ad hoc imputation or enforce restrictive shape-homogeneity assumptions, Tri-SfSVD integrates continuous trajectory estimation with simultaneous subject, feature, and temporal selection within a single optimization framework. By imposing sparse penalties across subjects, variables, and temporal subregions, the proposed method works directly on observed data to uncover localized structures at the subject, subject-feature, and subject-feature-time levels. Extensive simulations demonstrate that Tri-SfSVD outperforms existing approaches in high-dimensional settings. Applied to IBD multi-omics data, the method identified three biclusters linking sample clusters with distinct IBD-related clinical characteristics to microbial pathway groups associated with specific bacterial taxa, providing interpretable subject-pathway associations for characterizing disease heterogeneity. Applied to multi-channel EEG data, the method identified three triclusters linking sample clusters with distinct alcohol-related phenotypes to localized brain activity patterns, including subgroup differences separated by temporal subregions within the same spatial region.

Computed Tomography Angiography (CTA) is widely used to reconstruct vascular geometry from projection measurements, with conventional approaches such as Filtered Back-Projection (FBP) and Iterative Reconstruction (IR) forming the clinical standard. Blood flow is subsequently estimated through Computational Fluid Dynamics (CFD) simulations, which require vascular geometry and boundary conditions to be specified a priori. Since the geometry is fixed prior to flow estimation, the recovery of unknown anatomical features (e.g., missing branches or stenoses) is precluded. In this work, we present a fluid-physics-constrained reconstruction framework that leverages topology optimization (TO) to jointly recover vascular geometry and blood velocity directly from time-resolved CTA sinograms. The formulation couples a steady incompressible flow model with a transient advection-diffusion contrast transport model, mapped to sinogram space through a differentiable projection operator. The recovered velocity fields provide hemodynamic information and can support downstream estimation of wall shear stress and flow distribution, without requiring a separate CFD pipeline. The proposed method is demonstrated on synthetic phantoms under varying sparsity and noise levels, and on representative projection data.

Prompt ambiguity is a common source of failure in large language models, but is difficult to localize because it is a latent property of the prompt, while existing attribution methods are designed to explain observable outputs such as logits or generated tokens. We introduce PRIG, a gradient attribution method that uses a probe logit to attribute latent ambiguity to token positions. Specifically, PRIG trains a linear probe to distinguish clear prompts from ambiguous prompts and attributes the probe score to earlier token representations in the residual stream. To enable token-level evaluation, we construct synthetic ambiguity datasets across coding, math, and writing by rewriting one task-critical sentence per prompt, and complement them with a human-written gold benchmark. In this setting, PRIG localizes ambiguous spans substantially better than gradient attribution baselines, achieving 0.840 AUROC on the combined synthetic benchmark and 0.891 AUROC on the gold set. It also outperforms GPT-5.4 on sentence-level ambiguity identification and retains useful signal out-of-domain. These results establish PRIG as a practical tool for identifying which parts of a prompt are ambiguous. More broadly, they suggest that latent prompt properties can be localized through intermediate representations, rather than through output-level attribution.

Pipeline parallelism enables training of large language models that exceed single-device memory, yet inter-stage activation communication becomes the dominant bottleneck when trained on low-bandwidth networks. Recent work in this area has proposed using fixed orthogonal projections to compress activations. However, this still results in a significant performance degradation and requires a number of non-standard adaptations to constrain the optimization. A natural alternative is to learn a low rank projection for each pipeline stage, however maintaining the necessary orthogonality of these projectors during training remains a challenge. We present Manifold Aware Projection Learning (MAPL), a method that treats inter-stage compression as a learnable orthogonal projection under explicit Stiefel manifold (orthogonal matrices) constraints. Rather than prescribing a fixed global subspace, MAPL lets each pipeline stage discover and continuously adapt its own task-optimal compression subspace via manifold-constrained steepest descent. To recover token-specific signals at stage boundaries, we introduce per-stage factorized anchor embeddings that allow for full-rank activation reconstruction with negligible communication overhead. We further show that we can incorporate residual vector quantization after projection with a streaming codebook synchronization protocol that amortizes dictionary communication. Across LLaMA models from 150M to 1B parameters we show that MAPL can be easily applied to the existing pipeline and can achieve high compression with neglibile performance degradation with a drastically improved tradeoffs in performance vs. compression compared to Subspace Networks.

Data-driven Prognostics and Health Management (PHM) uses time-varying condition-monitoring data to diagnose system states and estimate remaining useful life in engineered assets. These tasks are central to maintenance planning, but industrial PHM data are often fragmented, partially observed, and poorly labeled, which hinders supervised learning. Foundation models offer a route toward reusable predictive systems, yet most time-series foundation models are designed for forecasting and assume long, coherent, regularly sampled sequences. To address this gap, we propose a framework for applying Tabular Foundation Models to industrial time series using in-context learning, and we evaluate them on a variety of PHM tasks. By converting raw unit-level signals into tabular rows, we show that these models perform well across multiple tasks - including prognostics, and diagnostics - and are highly data efficient. We compare them directly with sequence models, transformer baselines, and gradient-boosted trees under a common evaluation protocol. The results indicate that tabular foundation models achieve the best average ranks across prognostic and diagnostic tasks. Our findings further show that PFN-based models are competitive in low-data regimes, that temporal context can be preserved in the tabular representation, and that performance depends on representative context construction under subsampling. These results demonstrate that tabular foundation models provide a practical and general interface for heterogeneous PHM problems.

Diffusion Models (DM) have revolutionized text-driven generation by enabling the synthesis of high-quality, photorealistic visual content from user prompts. Whereas prior advances in visual generation such as VAEs and GANs were primarily evaluated on perceptual or visual similarity metrics such as FID PSNR, DM advances have fostered the development of more advanced Human Preference Metrics (HPM) that model and quantify human judgment as scalar values. However, DMs synthesize content using an inherently stochastic process where random noise seeds generation. The initial random noise directly affects the quality of generated outputs, both qualitatively and quantitatively. This influence is pronounced in smaller models for local deployment scenarios. Given this phenomenon, we first investigate to what extent we can predict scalar HPM scores prior to committing compute resources for generation. Further, we then investigate to what extent we can leverage such prediction to improve the quality of generated images, and also study which HPMs are best suited for this task. Our investigation reveals that not only is this possible, but that it is feasible to achieve negligible hardware overhead.

Security misconfigurations remain a leading cause of OS-level compromise, and manually keeping systems compliant with standards like Defense Information Systems Agency (DISA) Security Technical Implementation Guides (STIGs) is a tedious and expensive process. Existing compliance automation tools can reduce some of this burden, but they depend on static, pre-written corrective actions. In this paper, we introduce SHIELDS, a multi-agent system that uses large language models (LLMs) to approach OS hardening as an iterative, feedback-driven process. Instead of applying fixed remediations, SHIELDS continuously proposes fixes and refines them based on feedback from target system execution and validation scans. We evaluate the system across multiple virtual machine configurations using six contemporary LLMs ranging from 20B to 400B parameters, and find that SHIELDS successfully remediates up to 73% of scan findings. Our results also suggest that success in this setting depends less on model size (parameter count) than on effective tool use and information gathering, paving a practical path toward reducing the burden of security compliance in environments where compute is limited or security and privacy needs drive local model use.

Protein binder design has largely optimized for affinity alone, leaving conformational selectivity unaddressed: for allosteric targets such as kinases, nuclear receptors, and GPCRs, a binder that engages both active and inactive states provides no functional specificity regardless of how tightly it binds. We introduce AlloGen, a modular framework that decouples backbone generation from a learned state-selectivity scorer $Q_θ$, an SE(3)-invariant interface graph transformer trained via a two-phase curriculum that first learns interface geometry before imposing conformational discrimination. Because $Q_θ$ is fully differentiable and generator-agnostic, it integrates with any backbone generator as a passive reranker or an active gradient-based guide without retraining. Across a diverse benchmark of proteins spanning multiple families and conformational mechanisms, AlloGen consistently identifies binders that preferentially recognize desired structural states while rejecting alternative conformations. Experimental validation on calmodulin further demonstrates that these computational selectivity signals translate to physical molecules, yielding de novo peptides that bind the desired holo conformation while exhibiting no detectable binding to the apo state. Together, these results establish conformational selectivity as a learnable property and provide a general framework for state-selective protein binder design.

Learning semantics is essential for deep learning models to be interpretable and better aligned with human reasoning. Concept-based models approach this by representing classes through meaningful semantic abstractions, but typically treat all concepts as a flat, unstructured set learned at a single neural network layer. This overlooks a fundamental property of human semantic understanding: concepts being organized hierarchically, from general to specific. While deep networks do learn a hierarchy of visual features, this structure is rarely aligned with explicit semantic hierarchies. Drawing on Formal Concept Analysis, we demonstrate that formal concept lattices provide principled semantic scaffolds to guide neural network learning. These lattices naturally identify where in the network concepts should be learned based on their level of generality. This allows the model to develop staged, semantically grounded representations throughout its depth. Empirical results on real-world datasets show that our models produce more interpretable embeddings, support more effective interventions, and learn concept representations that are both meaningful and hierarchically structured.

A grand Motzkin path with air pockets is a non-empty lattice path in the first and fourth quadrant of $\mathbb{Z}^2$, starting at the origin $(0,0)$, ending on the $x$-axis, and consisting of up-steps $(1, 1)$, horizontal steps $(1, 0)$, down-steps $(1, -k)$ where $k \geq 1$, and with no consecutive down-steps. A {grand Dyck path with air pockets} is a grand Motzkin path with air pockets that uses no horizontal steps. We present the first known 2-Gray codes for grand Motzkin paths with air pockets. Setting the number of horizontal steps to zero in our algorithm yields the first known 2-Gray codes for grand Dyck paths with air pockets. Our three-stage algorithm generates each path in constant amortized time per string, using $O(n^2)$ memory. We also provide enumeration formulae for grand Motzkin paths and grand Dyck paths with air pockets.

Post-training Vision-Language-Action (VLA) models into policies that can be reliably deployed on real robots remains a major bottleneck. SFT and DAgger exploit failure signals only indirectly, and reward-based RL is bottlenecked by the difficulty of real-world reward design and of training reliable critics. We present FlowPRO, a reward-free offline reinforced fine-tuning framework for flow-matching VLAs. Algorithmically, we propose RPRO (Robotic Flow-matching Proximalized Preference Optimization), a preference-optimization objective tailored to the flow-matching action head of VLA models. RPRO pairs a contrastive optimizer with an explicit proximal regularizer that anchors the absolute magnitude of the implicit reward, thereby eliminating the reward-hacking failure mode of plain Flow-DPO. On the data side, a teleoperated intervention-and-rollback paradigm produces naturally paired positive and negative trajectories $(τ^w, τ^l)$ on a real robot from a single operator action; a Smooth Interpolation procedure, combined with batch mixing, then converts these sparse corrections into dense per-state supervision while preserving the base policy's capabilities. On four long-horizon bimanual tasks, FlowPRO attains the highest success rate, outperforming four representative baselines, and ablations confirm the contribution of each loss component.

A long-running append-mostly sequence, such as an edit log, event store, or versioned working set, is usually tiered into a bounded hot stratum and colder folded summaries. This saves memory but breaks stable references: a handle minted while a record is hot may later be resolved after the record has moved into a digest, after it has been superseded, or while a fold is in flight. We define the resulting cross-tier anomalies--dangling, stale, corrupt, and snapshot-skewed resolution--and present the Cascade Log, a reference-stable tiered append structure. The structure keeps a single persistent coalescing interval map over handles as the sole authority on each live version; folding a contiguous run replaces many singleton entries by one digest-backed interval node, and immutable roots provide snapshot tokens. Its cost is characterized by the fragmentation $A$, the number of index pieces, namely live handles plus maximal same-digest runs. The index uses $Θ(A)$ space, resolves a point in $O(\log A)$, reports a $k$-handle range in $O(\log A+k)$, and performs $a$ appends and $s$ supersedes in $O((a/B+s)\log A)$ update work for fold block size $B$. Matching lower bounds show that $Ω(A)$ space and $Ω(\log A+k)$ ordered range cost are unavoidable, and an adversary can force $A=Θ(s)$. Thus the index is sublinear on append-dominated histories and grows linearly only under fragmenting edits. A reference implementation and reproducible experiments to $10^6$ records validate the anomaly-freedom and the fragmentation bounds.

Tagging of generic dynamic values is important in symbolic-computation and dynamic-language systems, but the trade-offs change as machine architectures and workloads evolve. In particular, old folklore about boxed values, immediate values, and type tags must be recalibrated from time to time. We revisit the performance of badged object headers, low-bit tagging, and two NaN-boxing layouts on a range of platforms in use today, including AArch64 and x86-64 architectures from different manufacturers. The experiments isolate two distinct effects: the cost avoided by not heap-allocating common scalar values, and the cost avoided by obtaining tag information from the value word rather than by performing a heap read. The results show that several local bit operations are often cheaper than opening a heap object to obtain a tag or small value. Low-bit tagging remains the simplest and usually fastest choice for mostly symbolic workloads, while NaN-boxing is close in access cost and avoids the time and space of heap allocation for ordinary floating-point values.

Calibrating thermomechanical material models from experiments is challenging because deformation, temperature, and force responses are strongly coupled, while measurements are usually restricted to specimen surfaces. We present a full-field calibration framework for coupled finite-strain thermomechanical material models using boundary displacement, reaction-force data, and temperature. The forward model is formulated as a near-incompressible thermo-hyperelastic problem with thermomechanical coupling derived from a Helmholtz free energy, and the inverse problem is posed as a PDE-constrained optimization problem with weighted observation terms for the available data streams. Reduced gradients are computed with adjoint sensitivities that are obtained by automatic differentiation, enabling gradient-based calibration of nonlinear transient thermomechanical systems. The formulation is first verified on synthetic examples involving uniform thermal preconditioning and localized transient rod contact, where the ground-truth parameters are recovered from full-field measurements and force observations. The same workflow is then applied to experimental thermomechanical data by first calibrating a hyperelastic mechanical baseline from cyclic equibiaxial loading and subsequently identifying thermal expansion and directional shrinkage parameters from surface-temperature and boundary-force histories. The results demonstrate that coupled thermomechanical parameters can be inferred from experimentally accessible surface data without requiring volumetric observations.

Verifiable reward training has improved mathematical and coding reasoning, but these domains capture only part of step-by-step decision making. Many real-world tasks require finding a high-value feasible plan among many valid alternatives. We introduce OPT*, a scalable family of optimization-style tasks for training and evaluating LLM step-by-step optimization-like reasoning along a complexity axis: each task provides a feasibility checker and evaluator, while a complexity parameter expands the search space without requiring new human labels. This motivates studying these tasks in two regimes: (i) solver-guided online policy optimization, which uses a solver as a value oracle for partial states and applies rank-based reward shaping to reinforce better next steps, and (ii) search-based offline RL when such solvers are unavailable. Theoretically, we relate success in large search spaces to the information a reasoner extracts per unit of search budget. Empirically, we ablate the ingredients that make search efficient on OPT* and show that training on OPT* improves step-by-step optimization-like reasoning.

Patient safety event triage, determining whether a clinical event is reportable under jurisdiction-specific policy, is a high-stakes task typically performed manually by patient safety experts. Although LLMs may support this workflow, reliable evaluation is limited by the lack of benchmarks to capture evidence-grounded policy reasoning, proactive information seeking for incomplete reports, and principled abstention in irreducibly ambiguous cases. We address this gap with a policy-grounded construction methodology centered on the clause card, a structured representation that factorizes regulatory text into auditable decision specifications. Combining clause cards with anchor-driven instantiation and closed-loop verification, our scalable pipeline produces narratives with by-construction ground truth and naturally supports generating missing information and uncertain variants. We instantiate this method on Minnesota's 29 Reportable Adverse Health Events, producing PSEBench, a 5,074-case benchmark with an agentic evaluation environment. Evaluation on 15 representative LLMs reveals consistent capability trends, demonstrates the benchmark's utility, and identifies actionable gaps toward reliable LLM-based patient safety event triage.