Papers

In reinforcement learning with verifiable rewards (RLVR), challenging math questions whose sampled groups remain uniformly incorrect yield no informative update, letting easier questions dominate actor updates. This imbalance can contribute to distribution sharpening: low-$k$ accuracy improves while problem coverage contracts. Hint scaffolding can restore informative updates by revealing a prefix of a verified teacher solution. However, existing hint-based approaches still face two issues: teacher–student distribution mismatch and the need to reduce hint exposure to match no-hint evaluation. We address these issues through two components. Distribution-Aligned Hint Synthesis (DAHS) constructs verified teacher hints conditioned on student-style responses. Backward Hint Annealing (BHA) anneals hint exposure across length buckets as bucket accuracy improves. BHA also uses per-question hint dropout to keep no-hint rollouts throughout RL training. Under the DAPO training framework, our method improves both pass@1 and pass@2048 over DAPO on AIME24/25/26 for Qwen3-1.7B-Base and Qwen3-4B-Base, and improves pass@1 on five secondary math benchmarks (Olympiad, MATH500, Minerva Math, AMC23, and GSM8K). On Llama-3.2-1B-Instruct, the gains concentrate in the high-$k$ regime without recovering the base model's coverage. Overall, when verified teacher prefixes are available, DAHS+BHA mitigates DAPO-induced contraction of high-$k$ problem coverage in the evaluated challenging-math RLVR settings.

Recent progress in AI-enabled constitutive modeling has concentrated on moving from a purely data-driven paradigm to the enforcement of physical constraints and mechanistic principles, a concept referred to as physics augmentation. Classical phenomenological approaches rely on selecting a pre-defined model and calibrating its parameters, while machine learning methods often focus on discovery of the model itself. Sparse regression approaches lie in between, where large libraries of pre-defined models are probed during calibration. Sparsification in the aforementioned paradigm, but also in the context of neural network architecture, has been shown to enable interpretability, uncertainty quantification, but also heterogeneous software integration due to the low-dimensional nature of the resulting models. Most works in AI-enabled constitutive modeling have also focused on data from a single source, but in reality, materials modeling workflows can contain data from many different sources (multi-modal data), and also from testing other materials within the same materials class (multi-fidelity data). In this work, we introduce physics augmented finite element model updating (paFEMU), as a transfer learning approach that combines AI-enabled constitutive modeling, sparsification for interpretable model discovery, and finite element-based adjoint optimization utilizing multi-modal data. This is achieved by combining simple mechanical testing data, potentially from a distinct material, with digital image correlation-type full-field data acquisition to ultimately enable rapid constitutive modeling discovery. The simplicity of the sparse representation enables easy integration of neural constitutive models in existing finite element workflows, and also enables low-dimensional updating during transfer learning.

We develop a geometric framework that links objective accuracy to structural recovery in prototype-based clustering. The analysis is algorithm-agnostic and applies to a broad class of admissible loss functions. We define a clustering condition number that compares within-cluster scale to the minimum loss increase required to move a point across a cluster boundary. When this quantity is small, any solution with a small suboptimality gap must also have a small misclassification error relative to a benchmark partition. The framework also clarifies a fundamental trade-off between robustness and sensitivity to cluster imbalance, leading to sharp phase transitions for exact recovery under different objectives. The guarantees are deterministic and non-asymptotic, and they separate the role of algorithmic accuracy from the intrinsic geometric difficulty of the instance. We further show that errors concentrate near cluster boundaries and that sufficiently deep cluster cores are recovered exactly under strengthened local margins. Together, these results provide a geometric principle for interpreting low objective values as reliable evidence of meaningful clustering structure.

Recommendation systems must continuously adapt to evolving user behavior, yet the volume of data generated in large-scale streaming environments makes frequent full retraining impractical. This work investigates how targeted data selection can mitigate performance degradation caused by temporal distributional drift while maintaining scalability. We evaluate a range of representation choices and sampling strategies for curating small but informative subsets of user interaction data. Our results demonstrate that gradient-based representations, coupled with distribution-matching, improve downstream model performance, achieving training efficiency gains while preserving robustness to drift. These findings highlight data curation as a practical mechanism for scalable monitoring and adaptive model updates in production-scale recommendation systems.

In computational cytology, detecting malignancy on whole-slide images is difficult because malignant cells are morphologically diverse yet vanishingly rare amid a vast background of normal cells. Accurate detection of these extremely rare malignant cells remains challenging due to large class imbalance and limited annotations. Conventional weakly supervised approaches, such as multiple instance learning (MIL), often fail to generalize at the instance level, especially when the fraction of malignant cells (witness rate) is exceedingly low. In this study, we explore the use of one-class representation learning techniques for detecting malignant cells in low-witness-rate scenarios. These methods are trained exclusively on slide-negative patches, without requiring any instance-level supervision. Specifically, we evaluate two OCC approaches, DSVDD and DROC, and compare them with FS-SIL, WS-SIL, and the recent ItS2CLR method. The one-class methods learn compact representations of normality and detect deviations at test time. Experiments on a publicly available bone marrow cytomorphology dataset (TCIA) and an in-house oral cancer cytology dataset show that DSVDD achieves state-of-the-art performance in instance-level abnormality ranking, particularly in ultra-low witness-rate regimes ($\leq 1\%$) and, in some cases, even outperforming fully supervised learning, which is typically not a practical option in whole-slide cytology due to the infeasibility of exhaustive instance-level annotations. DROC is also competitive under extreme rarity, benefiting from distribution-augmented contrastive learning. These findings highlight one-class representation learning as a robust and interpretable superior choice to MIL for malignant cell detection under extreme rarity.

We present FDM (Fan Duality Model), a linear sequence architecture that resolves the fundamental tension between memory efficiency and associative recall in sequence modeling. FDM separates sequence processing into two components: a wave component (recurrent scan via phase-preserving Givens rotations) that compresses long-range patterns into a fixed-size complex hidden state, and a particle component (local-global cache) that retrieves specific tokens via learned associative addressing with W+K=272 slots independent of sequence length N. This yields strictly O(1) decode memory: 867 MB fixed across all prompt lengths 128-8,192 tokens, versus Transformer's 853-4,247 MB (4.9x reduction at N=8,192). Beyond the architecture, we discover that jointly training the wave and particle components leads to suboptimal convergence. We propose Freeze-Scan, a two-phase training strategy that freezes the recurrent scan and optimizes the cache jointly with embeddings, achieving PPL=64.9 on WikiText-103 in 44K steps – a 7.5x improvement over full fine-tuning (PPL=487). On Multi-Query Associative Recall (MQAR), FDM achieves 0.966 accuracy, surpassing Transformer (0.606) by 59.5%, while pure scan without cache scores only 0.011, confirming the necessity of the particle component. Finally, we introduce Holographic Reference Beam Decoding, interpreting the complex hidden state h_t as a holographic plate encoding the entire temporal history. Using the current input x_t as a reference beam to modulate h_t reduces PPL by up to 2.13 points (PPL=62.79) with a 4-head orthogonal reference beam using only 1.3M additional parameters, providing empirical support for the holographic interpretation. Code and pretrained weights: https://github.com/YasongFan/FDM

We analyze a simple one-hidden-layer neural network with ReLU activation functions and fixed biases, with one-dimensional input and output. We study both continuous and discrete versions of the model, and we rigorously prove the convergence of the learning process with the $L^2$ squared loss function and the gradient descent procedure. We also prove the spectral bias property for this learning process. Several conclusions of this analysis are discussed; in particular, regarding the structure and properties that activation functions should possess, as well as the relationships between the spectrum of certain operators and the learning process. Based on this, we also propose an alternative activation function, the full-wave rectified exponential function (FReX), and we discuss the convergence of the gradient descent with this alternative activation function.

In this work, CausalVAE is introduced as a plug-in structural module for latent world models and is attached to diverse encoder-transition backbones. Across the reported benchmarks, competitive factual prediction is preserved and intervention-aware counterfactual retrieval is improved after the plug-in is added, suggesting stronger robustness under distribution shift and interventions. The largest gains are observed on the Physics benchmark: when averaged over 8 paired baselines, CF-H@1 is improved by +102.5%. In a representative GNN-NLL setting on Physics, CF-H@1 is increased from 11.0 to 41.0 (+272.7%). Through causal analysis, learned structural dependencies are shown to recover meaningful first-order physical interaction trends, supporting the interpretability of the learned latent causal structure.

We introduce IatroBench, a benchmark with two axes of harm (commission and omission), comprising 60 pre-registered clinical scenarios, tested on 6 models. Matched scenarios are framed as a patient query and a doctor consultation, differing in register and request (with the implication of supervision by a treating physician in the latter). We analyse the responses of five different models and find that all share more information in the doctor framing than the patient framing (which we call "framing-contingent withholding"). For example, a model with strong safety training provides a benzodiazepine tapering schedule to a doctor, but does not provide this schedule to a patient who requests it. We use Claude Opus 4.6 for structured evaluation, and Gemini 3 Flash as our primary judge, to score model responses against a physician's rubrics. Our primary judge agrees with physicians' omission scores about as well as physicians agree with each other. We find a decoupling gap of +0.38 (p = 0.003) on average across models. With our primary judge (checked by physicians) the decoupling gap is +0.22 (95% CI 0.10-0.36, p = 0.0014). We find three distinct patterns underlying this gap, exemplified by each of the models below. In the doctor framing, Claude Opus demonstrates that it has the information, and withholds it in the patient framing. Llama 4 performs poorly in both framings, meaning the decoupling gap cannot distinguish between withholding and incompetence. Finally, GPT-5.2 (excluded from this analysis) failed to return text for 33.2% of doctor responses, compared to 0% of layperson responses. In 86.6% of cases that we score (through our structured evaluation) as having omission harms, our primary judge (Gemini 3 Flash) scores zero omission harm. Because our scenarios are designed to pit safety against helpfulness, these statistics hold only for this distribution.

Large Multimodal Models (LMMs) achieve state-of-the-art performance in high-stakes domains like healthcare, yet their reasoning remains opaque. Current interpretability methods, such as attention mechanisms or post-hoc saliency, often fail to faithfully represent the model's decision-making process, particularly when integrating heterogeneous modalities like time-series and text. We introduce Tree-of-Evidence (ToE), an inference-time search algorithm that frames interpretability as a discrete optimization problem. Rather than relying on soft attention weights, ToE employs lightweight Evidence Bottlenecks that score coarse groups or units of data (e.g., vital-sign windows, report sentences) and performs a beam search to identify the compact evidence set required to reproduce the model's prediction. We evaluate ToE across six tasks spanning three datasets and two domains: four clinical prediction tasks on MIMIC-IV, cross-center validation on eICU, and non-clinical fault detection on LEMMA-RCA. ToE produces auditable evidence traces while maintaining predictive performance, retaining over 0.98 of full-model AUROC with as few as five evidence units across all settings. Under sparse evidence budgets, ToE achieves higher decision agreement and lower probability fidelity error than other approaches. Qualitative analyses show that ToE adapts its search strategy: it often resolves straightforward cases using only vitals, while selectively incorporating text when physiological signals are ambiguous. ToE therefore provides a practical mechanism for auditing multimodal models by revealing which discrete evidence units support each prediction.

To implement the intelligent transportation digital twin (ITDT), unmanned aerial vehicles (UAVs) are scheduled to process the sensing data from the roadside sensors. At this time, generative artificial intelligence (GAI) technologies such as diffusion models are deployed on the UAVs to transform the raw sensing data into the high-quality and valuable. Therefore, we propose the GAI-empowered ITDT. The dynamic processing of a set of diffusion model inference (DMI) tasks on the UAVs with dynamic mobility simultaneously influences the DT updating fidelity and delay. In this paper, we investigate a joint optimization problem of DMI task offloading, inference optimization and UAV trajectory planning as the system utility maximization (SUM) problem to address the fidelity-delay tradeoff for the GAI-empowered ITDT. To seek a solution to the problem under the network dynamics, we model the SUM problem as the heterogeneous-agent Markov decision process, and propose the sequential update-based heterogeneous-agent twin delayed deep deterministic policy gradient (SU-HATD3) algorithm, which can quickly learn a near-optimal solution. Numerical results demonstrate that compared with several baseline algorithms, the proposed algorithm has great advantages in improving the system utility and convergence rate.

Identifying governing equations of nonlinear dynamical systems from data is challenging. While sparse identification of nonlinear dynamics (SINDy) and its extensions are widely used for system identification, operator-logarithm approaches use the logarithm to avoid time differentiation, enabling larger sampling intervals. However, they still suffer from the curse of dimensionality. Then, we propose a data-driven method to compute the Koopman generator in a low-rank tensor train (TT) format by taking logarithms of Koopman eigenvalues while preserving the TT format. Experiments on 4-dimensional Lotka-Volterra and 10-dimensional Lorenz-96 systems show accurate recovery of vector field coefficients and scalability to higher-dimensional systems.

Verification and validation of cyber-physical systems (CPS) via large-scale simulation often surface failures that are hard to interpret, especially when triggered by interactions between continuous and discrete behaviors at specific events or times. Existing debugging techniques can localize anomalies to specific model components, but they provide little insight into the input-signal values and timing conditions that trigger violations, or the minimal, precisely timed changes that could have prevented the failure. In this article, we introduce DeCaF, a counterfactual-guided explanation and assertion-based characterization framework for CPS debugging. Given a failing test input, DeCaF generates counterfactual changes to the input signals that transform the test from failing to passing. These changes are designed to be minimal, necessary, and sufficient to precisely restore correctness. Then, it infers assertions as logical predicates over inputs that generalize recovery conditions in an interpretable form engineers can reason about, without requiring access to internal model details. Our approach combines three counterfactual generators with two causal models, and infers success assertions. Across three CPS case studies, DeCaF achieves its best success rate with KD-Tree Nearest Neighbors combined with M5 model tree, while Genetic Algorithm combined with Random Forest provides the strongest balance between success and causal precision.

We establish guarantees for the unique recovery of vector fields and transport maps from finite measure-valued data, yielding new insights into generative models, data-driven dynamical systems, and PDE inverse problems. In particular, we provide general conditions under which a diffeomorphism can be uniquely identified from its pushforward action on finitely many densities, i.e., when the data $\{(\rho_j,f_\#\rho_j)\}_{j=1}^m$ uniquely determines $f$. As a corollary, we introduce a new metric which compares diffeomorphisms by measuring the discrepancy between finitely many pushforward densities in the space of probability measures. We also prove analogous results in an infinitesimal setting, where derivatives of the densities along a smooth vector field are observed, i.e., when $\{(\rho_j,\text{div} (\rho_j v))\}_{j=1}^m$ uniquely determines $v$. Our analysis makes use of the Whitney and Takens embedding theorems, which provide estimates on the required number of densities $m$, depending only on the intrinsic dimension of the problem. We additionally interpret our results through the lens of Perron–Frobenius and Koopman operators and demonstrate how our techniques lead to new guarantees for the well-posedness of certain PDE inverse problems related to continuity, advection, Fokker–Planck, and advection-diffusion-reaction equations. Finally, we present illustrative numerical experiments demonstrating the unique identification of transport maps from finitely many pushforward densities, and of vector fields from finitely many weighted divergence observations.

Synthesizable molecular optimization seeks to improve target properties while ensuring that molecular modifications follow feasible synthetic pathways. Existing synthesis-aware methods typically rely on exploring a large space of candidate transformations defined by reaction templates and purchasable building blocks. This search becomes even more challenging when property improvement requires multiple reaction steps, as the space expands further along the pathway. To address this challenge, we introduce MolReAct, which reformulates molecular optimization as search over compact reaction spaces proposed by a tool-augmented large language model (LLM). At each step, the LLM combines its prior chemical knowledge with cheminformatics tools to identify a molecule-specific set of compatible reactions, preserving synthesizability while making multi-step optimization feasible. Given this compact action space, we further leverage Group Relative Policy Optimization (GRPO) with the terminal oracle reward to improve long-term decision-making over multiple reaction steps. Across diverse molecular optimization tasks, MolReAct achieves the highest Top-10 score on 11 of 14 tasks and the best sample efficiency on 12 of 14 tasks, outperforming existing baselines under limited oracle budgets. Beyond these gains, MolReAct also provides each optimized molecule with a template-grounded synthetic pathway.

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used for post-training Large Language Models, but practical verifiers can make errors. We study how the rate and structure of reward noise affect RLVR in code generation, with a preliminary scientific-reasoning check. In multi-seed Qwen3 8B experiments on MBPP, mean validation reward over two fixed late evaluations is within 1 percentage point of the clean baseline at the tested resampled group-rollout noise rates through 20%, and within about 2 points at 30%. Confidence intervals allow larger losses; these point estimates do not establish a general tolerance threshold. We also examine full-program pass@k, four controlled noise structures, two model-based verifiers, and policy models from three families spanning 4B-9B parameters. We derive conditional advantage distributions for symmetric and asymmetric group noise, including retained format penalties, and show why clipping limits simple gradient-scaling arguments. The analysis identifies information preserved by whole-group corruption and limits on interpreting our asymmetric sweep as a precision-recall comparison. Overall, the results indicate that imperfect verification can support effective RLVR in the tested settings, while aggregate error rates alone do not characterize the learning signal.

The AdamW optimizer, while standard for LLM pretraining, is a critical memory bottleneck, consuming optimizer states equivalent to twice the model's size. Although light-state optimizers like SinkGD attempt to address this issue, we identify the embedding layer dilemma: these methods fail to handle the sparse, high-variance gradients inherent to embeddings, forcing a hybrid design that reverts to AdamW and partially negates the memory gains. We propose SAGE (Sign Adaptive GradiEnt), a novel optimizer that resolves this dilemma by replacing AdamW in this hybrid structure. SAGE combines a Lion-style update direction with a new, memory-efficient $O(d)$ adaptive scale. This scale acts as a "safe damper," provably bounded by 1.0, which tames high-variance dimensions more effectively than existing methods. This superior stability allows SAGE to achieve better convergence. On Llama models up to 1.3B parameters, our SAGE-based hybrid achieves new state-of-the-art perplexity, outperforming all baselines, including SinkGD hybrid, while significantly reducing optimizer state memory.

Monotone variational inequalities (VIs) provide a unifying framework for convex minimization, equilibrium computation, and convex-concave saddle-point problems. Extragradient-type methods are among the most effective first-order algorithms for such problems, but their performance hinges critically on stepsize selection. While most existing theory focuses on ergodic averages of the iterates, practical performance is often driven by the significantly stronger behavior of the last iterate. Moreover, available last-iterate guarantees typically rely on fixed stepsizes chosen using problem-specific global smoothness information, which is often difficult to estimate accurately and may not even be applicable. In this paper, we develop parameter-free extragradient methods with non-asymptotic last-iterate guarantees for constrained monotone VIs. For globally Lipschitz operators, our algorithm achieves an $o(1/\sqrt{T})$ last-iterate rate. We then extend the framework to locally Lipschitz operators via backtracking line search and obtain the same rate while preserving parameter-freeness, thereby making parameter-free last-iterate methods applicable to important problem classes for which global smoothness is unrealistic. Our numerical experiments on bilinear matrix games, LASSO, minimax group fairness, and state-of-the-art maximum entropy sampling relaxations demonstrate wide applicability of our results as well as strong last-iterate performance and significant improvements over existing methods.

Web agents act through long interaction sequences, yet existing benchmarks evaluate only terminal success, discarding all process information and offering little guidance on improvement. In this work, we conduct a process-level analysis of web agents. We introduce WebStep, a benchmark of 1,800 task instances with controlled difficulty and automatic semantic state tracking. Each website exposes a deterministic semantic MDP alongside the GUI: the agent operates on the interface, while the environment records high-level states and transitions in the background, enabling fine-grained analysis without manual annotation. Based on the semantic trajectory, we first show that process metrics reveal differences invisible to outcome evaluation: three agents whose success rates cluster within 34-37% diverge in exploration reach versus execution accuracy. Then, decomposing by skill characterizes the nature of these differences, exposing opposite per-skill rankings hidden within the same website: e.g., on Q&A, Claude CUA outperforms OpenAI CUA by 30% on navigation actions yet underperforms it by 6.7% on inspection, pinpointing a concrete skill to improve even within a domain. Bifurcation analysis further localizes the decisive error that loses the task and shows that this error is agent-specific rather than shared. Finally, these differences widen as tasks grow harder: success rate is similar on easy tasks but separates sharply as exploration becomes more demanding. Our process-level analysis opens a new avenue in web agent evaluation, providing fine-grained and actionable insight into where and how each agent should be improved. Project page: https://jiwanchung.github.io/webstep

No feature ranking can be simultaneously faithful, stable, and complete when features are collinear. For collinear pairs, ranking reduces to a coin flip. We prove this impossibility, quantify it for four model classes, resolve it via ensemble averaging (DASH), and machine-verify it with 305 Lean 4 theorems. We characterize the complete attribution design space: exactly two families of methods exist – faithful-complete methods (unstable, with rankings that flip up to 50% of the time) and ensemble methods like DASH (stable, reporting ties for symmetric features) – and no method lies outside this dichotomy. The impossibility is quantitative: the attribution ratio diverges as 1/(1-rho^2) for gradient boosting, is infinite for Lasso, and converges for random forests. DASH (Diversified Aggregation of SHAP) is provably Pareto-optimal among unbiased aggregations, achieving the Cramer-Rao variance bound with a tight ensemble size formula. In a survey of 77 public datasets, 68% exhibit attribution instability. Switching to conditional SHAP does not escape the impossibility when features have equal causal effects. The framework includes practical diagnostics – a Z-test workflow and single-model screening tool – and has direct consequences for fairness auditing: SHAP-based proxy discrimination audits are provably unreliable under collinearity. The design space theorem, diagnostics, and impossibility are mechanically verified in Lean 4 (305 theorems from 16 axioms, 0 sorry) – to our knowledge, the first formally verified impossibility in explainable AI.

The creation of cinematic-quality animal effects necessitates the precise modeling of muscle and fur dynamics, a process that remains both labor-intensive and computationally expensive within traditional production workflows. While generative diffusion models have shown promise in diverse artistic workflows, their capacity for high-fidelity animal simulation remains largely unexploited. We present MoZoo, a generative dynamics solver that bypasses conventional refinement to synthesize high-fidelity animal videos from coarse meshes under multimodal guidance. We propose Role-Aware RoPE (RAR-RoPE) which employs role-based index remapping to synchronize motion alignment while decoupling reference information via fixed temporal offsets. Complementing this, Asymmetric Decoupled Attention partitions the latent sequence to enforce a unidirectional information flow, effectively preventing feature interference and improving computational efficiency. To address the scarcity of high-quality training data, we introduce MoZoo-Data, a synthetic-to-real pipeline that leverages a rendering engine and an inverse mapping approach to construct a large-scale dataset of paired sequences. Furthermore, we establish MoZooBench, a comprehensive benchmark with 120 mesh-video pairs. Experimental results demonstrate that MoZoo achieves high-fidelity fur simulation across diverse animal skeletons and layouts, preserving superior temporal and structural consistency.

Same-model self-verification, prompting a model to audit its own predicted answer, is a plausible confidence signal for selective prediction, but its practical value remains unclear once strong likelihood-based baselines are taken seriously. We evaluate self-verification against two such baselines, LL-AVG and LL-SUM, on ARC-Challenge and TruthfulQA-MC across multiple model families, scales, and prompt variants. We measure not only correctness ranking, but also abstention quality through AURC and operating-point analyses. The results are sharply task- and model-dependent. On ARC-Challenge, self-verification substantially improves over LL-AVG for Phi-2 and the Qwen models, with the largest gains appearing in Qwen-7B. On TruthfulQA-MC, however, the signal is less reliable: smaller models can become prompt-sensitive, DeepSeek-R1-Distill-8B degrades relative to LL-AVG, and LL-SUM often remains the stronger practical baseline. We therefore do not treat self-verification as a general-purpose uncertainty estimator. In this setting, it is better understood as a conditional confidence signal whose value depends on task type, model family, prompt formulation, and, crucially, the baseline it must beat.

A guard model fine-tuned on entirely benign data can lose all safety alignment – not through adversarial manipulation, but through standard domain specialization. We demonstrate this failure across three purpose-built safety classifiers – LlamaGuard, WildGuard, and Granite Guardian – deployed as protection layers in agentic AI pipelines, and show that it originates in the destruction of latent safety geometry: the structured harmful – benign representational boundary that guides classification. We extract per-layer safety subspaces via SVD on class-conditional activation differences and track how this boundary evolves under benign fine-tuning. Granite Guardian undergoes complete collapse – refusal rate drops from 85% to 0%, CKA falls to zero, and 100% of outputs become ambiguous – a severity exceeding prior findings on general-purpose LLMs, explained by the specialization hypothesis: concentrated safety representations are efficient but catastrophically brittle. To mitigate this, we propose Fisher-Weighted Safety Subspace Regularization (FW-SSR), a training-time penalty combining (i) curvature-aware direction weights derived from diagonal Fisher information and (ii) an adaptive $\lambda_t$ that scales with task-safety gradient conflict. FW-SSR recovers 75% refusal on Granite Guardian (CKA = 0.983) and reduces WildGuard's Attack Success Rate to 3.6% – below the unmodified baseline – by actively sharpening the safety subspace rather than merely anchoring it. Across all three models, structural representational geometry (CKA, Fisher score) predicts safety behavior more reliably than absolute displacement metrics, establishing geometry-based monitoring as a necessary component of guard model evaluation in agentic deployments.

Reinforcement learning (RL) has become a central post-training tool for improving the reasoning abilities of large language models (LLMs). In these systems, the rollout, the trajectory sampled from a prompt to termination, including intermediate reasoning steps and optional tool or environment interactions, determines the data the optimizer learns from, yet rollout design is often underreported. This survey provides an optimizer-agnostic view of rollout strategies for RL-based post-training of reasoning LLMs. We formalize rollout pipelines with unified notation and introduce Generate-Filter-Control-Replay (GFCR), a lifecycle taxonomy that decomposes rollout pipelines into four modular stages: Generate proposes candidate trajectories and topologies; Filter constructs intermediate signals via verifiers, judges, critics; Control allocates compute and makes continuation/branching/stopping decisions under budgets; and Replay retains and reuses artifacts across rollouts without weight updates, including self-evolving curricula that autonomously generate new training tasks. We complement GFCR with a criterion taxonomy of reliability, coverage, and cost sensitivity that characterizes rollout trade-offs. Using this framework, we synthesize methods spanning RL with verifiable rewards, process supervision, judge-based gating, guided and tree/segment rollouts, adaptive compute allocation, early-exit and partial rollouts, throughput optimization, and replay/recomposition for self-improvement. We ground the framework with case studies in math, code/SQL, multimodal reasoning, tool-using agents, and agentic skill benchmarks that evaluate skill induction, reuse, and cross-task transfer. Finally, we provide a diagnostic index that maps common rollout pathologies to GFCR modules and mitigation levers, alongside open challenges for building reproducible, compute-efficient, and trustworthy rollout pipelines.

Passenger assistance services are essential for accessible rail travel, yet demand varies substantially across stations and over time, creating challenges for workforce planning and staff rostering. This paper presents a data-driven decision support framework for forecasting station-level passenger assistance demand and translating forecasts into workforce plans. The forecasting component applies a horizon-aware Prophet modelling approach using multi-source operational data, while the planning component maps demand forecasts to staffing requirements under service and operational constraints through an interpretable red-amber-green risk framework. The approach has been implemented within a production-grade system to support routine planning and staffing decisions across LNER-managed stations. Results demonstrate improved forecast accuracy relative to year-on-year baseline methods, with absolute error reduced by up to 76.9%, and show that forecast-informed staffing is associated with an approximate 50% reduction in failed passenger assistance deliveries attributable to staff availability. These findings highlight the value of integrating interpretable forecasting with operational work.

Physics-guided sampling with diffusion priors has recently shown strong performance in solving complex systems of partial differential equations (PDEs) from sparse observations. However, these methods are typically evaluated on benchmark problems that do not fully demonstrate their ability to generate temporally consistent solutions of time-dependent PDEs, often focusing instead on reconstructing a single snapshot. In this work, we apply these methods to gas-phase reaction kinetics problems governed by the advection-reaction-diffusion (ARD) equation, providing a setting that more closely reflects realistic laboratory experiments. We demonstrate that guided sampling can be used to reconstruct full spatiotemporal trajectories, rather than isolated states. Furthermore, we show that these methods generalise to previously unseen parameter regimes, highlighting their potential for real-world applications.

Fault intensity diagnosis (FID) plays a pivotal role in intelligent manufacturing while neglecting dependencies among target classes hinders its practical deployment. This paper introduces a novel and general framework with deep hierarchical knowledge loss (DHK) to achieve hierarchical consistent representation and prediction. We develop a novel hierarchical tree loss to enable a holistic mapping of same-attribute classes, leveraging tree-based positive and negative hierarchical knowledge constraints. We further design a focal hierarchical tree loss to enhance its extensibility and devise two adaptive weighting schemes based on tree height. In addition, we propose a group tree triplet loss with hierarchical dynamic margin by incorporating hierarchical group concepts and tree distance to model boundary structural knowledge across classes. The joint two losses significantly improve the recognition of subtle faults. Extensive experiments are performed on four real-world datasets from various industrial domains (three cavitation datasets from SAMSON AG and one publicly available dataset) for FID, all showing superior results and outperforming recent state-of-the-art FID methods.

Real-time voice assistants must revise task state when users interrupt mid-response, but existing spoken-dialog benchmarks largely evaluate turn-based interaction and miss this failure mode. We introduce EchoChain, a controlled benchmark for evaluating full-duplex state-update reasoning under mid-speech interruptions. EchoChain identifies three recurring failure patterns in post-interruption continuations: contextual inertia, interruption amnesia, and objective displacement. The benchmark generates scenario-driven conversations and injects interruptions at a standardized point relative to assistant speech onset, enabling controlled cross-model comparison. In a paired half-duplex control, total failures drop by 40.2% relative to interrupted runs, indicating that many errors are driven by state-update reasoning under interruption rather than task difficulty alone. Across evaluated real-time voice models, no system exceeds a 50% pass rate, showing substantial room for improvement in mid-generation state revision. EchoChain provides a reproducible benchmark for diagnosing state-update reasoning failures in full-duplex voice interaction.

Automated diagnosis from chest CT has improved considerably with deep learning, but models trained on skewed datasets tend to perform unevenly across patient demographics. However, the situation is worse than simple demographic bias. In clinical data, class imbalance and group underrepresentation often coincide, creating compound failure modes that neither standard rebalancing nor fairness corrections can fix alone. We introduce a two-level objective that targets both axes of this problem. Logit-adjusted cross-entropy loss operates at the sample level, shifting decision margins by class frequency with provable consistency guarantees. Conditional Value at Risk aggregation operates at the group level, directing optimization pressure toward whichever demographic group currently has the higher loss. We evaluate on the Fair Disease Diagnosis benchmark using a 3D ResNet-18 pretrained on Kinetics-400, classifying CT volumes into Adenocarcinoma, Squamous Cell Carcinoma, COVID-19, and Normal groups with patient sex annotations. The training set illustrates the compound problem concretely: squamous cell carcinoma has 84 samples total, 5 of them female. The combined loss reaches a gender-averaged macro F1 of 0.8403 with a fairness gap of 0.0239, a 13.3% improvement in score and 78% reduction in demographic disparity over the baseline. Ablations show that each component alone falls short. The code is publicly available at https://github.com/Purdue-M2/Fair-Disease-Diagnosis.

Digital forensic investigations increasingly rely on heterogeneous evidence such as images, scanned documents, and contextual reports. These artifacts may contain explicit or implicit expressions of harm, hate, threat, violence, or intimidation, yet existing automated approaches often assume clean text input or apply vision models without forensic justification. This paper presents a case-driven multimodal approach for hate and threat detection in forensic analysis. The proposed framework explicitly determines the presence and source of textual evidence, distinguishing between embedded text, associated contextual text, and image-only evidence. Based on the identified evidence configuration, the framework selectively applies text analysis, multimodal fusion, or image-only semantic reasoning using vision language models with vision transformer backbones (ViT). By conditioning inference on evidence availability, the approach mirrors forensic decision-making, improves evidentiary traceability, and avoids unjustified modality assumptions. Experimental evaluation on forensic-style image evidence demonstrates consistent and interpretable behavior across heterogeneous evidence scenarios.