Papers

We study complete-information debate systems in which a probabilistic finite-state verifier reads the alternating messages of a prover and a refuter. Demirci, Say, and Yakary{\i}lmaz showed that every language in $\mathsf{P}$ has such debates checkable with a constant number of random bits and arbitrarily small weak error. Their strong-error construction, which also counts nontermination as failure, did not permit arbitrary error reduction. We close this gap: for every $L\in\mathsf{P}$ and every $\varepsilon>0$, there is a constant-space verifier using a constant number of private coin tosses that has perfect completeness and strong error at most $\varepsilon$. The verifier simulates a polynomial-time alternating multihead finite automaton, privately spot-checking one of its input heads. The key observation is that, on a nonmember, the refuter may concede any round in which the prover first misreports a head reading. This ensures termination against every prover when the refuter follows the specified strategy, and permits strong-error reduction by repetition.

Current embodied systems largely rely on pretrained capabilities that remain fixed after deployment, limiting their ability to learn from physical interaction. We introduce MachEmbodied-Brain (ME-Brain), a self-evolving embodied system organized around a closed loop of action execution, experience acquisition, experience evolution, and improved execution. Evolvable Memory consolidates multimodal trajectories into hierarchical, reusable experience; Cognitive Core transforms physical experience into transferable skills; and the Action Model combines event-driven keyframes, EventCell local-world prediction, and action-conditioned memory modulation to focus computation on decision-critical moments, regions, and historical evidence. Together, these modules shift embodied intelligence from train-and-freeze to deploy-and-evolve without model retraining. Cognitive Core outperforms the strongest comparison models by 8.2 and 9.6 points on embodied and agent benchmarks. The Action Model achieves 47.88% mean success on RoboMME, a 3.26-point improvement over the strongest baseline. On RoboDojo, it reaches a 21.51 mean Score and 16.03% success rate, exceeding $\pi_{0.5}$ by 10.10 and 9.12 points. On the six-task ME-RealBench, ME-Brain achieves a 69.5 mean Score and 66.7% success rate, outperforming DM0.5 by 12.8 and 11.7 points, respectively.

Modern GPUs are no longer physically symmetric. Die scaling leads to both manufacturing-driven floorsweeping and cache and memory partitioning. The former creates chip-specific compute topologies, while the latter causes non-uniform memory access. These asymmetries are substantial. Topology-oblivious compute unit allocation can lead to up to 1.33x performance variation, while remote accesses increase HBM latency by up to 67% and nearly double L2 latency. However, these asymmetries are hidden behind the GPU's logical resource abstractions and can vary across chips. We develop lightweight characterization methods to uncover per-chip compute topology and memory affinity. We then use the discovered information to make existing fine-grained scheduling asymmetry-aware, considering not only how many resources are allocated but also which physical resources are assigned. Across full-GPU kernel execution, intra-application multiplexing, and inter-application co-location, asymmetry-aware scheduling improves mainstream kernels by up to 1.22x, multiplexed LLM inference by up to 14.3%, and avoids up to 1.33x performance variation.

In the context of a new approach towards quantum foundation, the Born rule is proved under weak conditions through several steps. Some of the steps turn out to have connections to statistical theory. One step gives a generalized likelihood principle, a result of independent interest. The sole additional assumption behind the Born rule is a consequence of the likelihood principle of statistics. This principle is discussed from several points of view. The whole approach towards quantum foundation advocated here, leads to a simpler and more intuitive pure state concept than the traditional one in quantum mechanics. It also provides links to statistical theory and to Andrei Khrennikov's macroscopical quantum-like models, in particular, models connected to cognition and decisions. Finally, the approach has connections to relativity theory and to quantum field theory. It is a hope that this approach now can be seen as a good alternative to the traditional formal foundation of quantum theory. The approach is here discussed in the case of finite-values parameters, but continuous parameters and parameters taking a countable set of values can be included in the theory by approximating them with finite-valued ones.

Fine-grained, subcell-level dissipation control is essential for achieving robust high-order discontinuous Galerkin (DG) simulations of nonlinear hyperbolic systems in under-resolved regimes while preserving accuracy. This paper proposes a subcell-refined entropy-residual-driven limiting strategy for DG on Legendre-Gauss-Lobatto nodes. The limiter introduces only nearest-neighbor pairwise dissipation within each element, with closed-form coefficients that supply the minimal dissipation required to restore the element entropy inequality. The strategy is a diagonal, locally stable approximation of classical entropy-stable methods, and a generalized subcell framework reveals split-form DG and residual-distribution-based entropy correction schemes as particular choices of the limiting coefficients. For the Euler equations, a physically consistent jump operator separately models thermal and shear entropy production while preserving velocity and pressure equilibrium; a subcell refinement of the Zhang-Shu positivity limiter ensures pointwise positivity. Extensive numerical tests confirm that the scheme maintains optimal high-order accuracy, strictly enforces entropy dissipation, and significantly reduces the difficulty of a posteriori positivity-preserving procedures.

Text-to-Speech (TTS) is an essential module that provides spoken responses in a spoken dialogue system (SDS) centered on a large language model (LLM). To ensure accurate TTS synthesis, responses generated by an LLM must be converted into TTS-readable formats via a Text Normalization (TN) module, imposing strict low-latency requirements in real-time SDS scenarios. Existing TN solutions are largely rule-based, rely on manual engineering, and generalize poorly to unseen patterns. Although an LLM itself can perform TN through prompt engineering, it faces key limitations: high first-token latency due to non-streaming processing, hallucination risks, and degraded intelligence or reasoning when the core LLM module is fine-tuned solely for TN. To address these challenges, we propose StreamTN, a lightweight LLM-based Chinese streaming TN model. Built on Qwen3-0.6B, StreamTN employs a dual-track streaming framework in which input tokens and output tokens are processed on two parallel tracks, enabling low-latency real-time inference without complex prompting. Moreover, task-specific fine-tuning yields superior TN performance and fewer hallucinations than rule-based systems and general-purpose LLMs. We also introduce a TN benchmark that spans diverse text scenarios, providing a comprehensive evaluation standard for speech generation in spoken dialogue systems. Experiments demonstrate the effectiveness of StreamTN in accuracy and inference latency.

Many computational fluid dynamics (CFD) algorithms solve partial differential equations by discretization, resulting in large and sparse systems of linear equations. While iterative methods are widely used to solve these systems classically, most existing quantum linear system solvers target the solution through matrix inversion rather than approximating it using an iterative procedure. In this work, we develop a quantum implementation of the Jacobi method based on quantum singular value transformation (QSVT). By reformulating the Jacobi iteration as a polynomial transformation of a block-encoded operator, the algorithm requires only a constant ancilla overhead with respect to the number of iterations while maintaining a circuit depth that scales linearly with the iteration number. We demonstrate the algorithm for one- and two-dimensional Poisson problems, including the pressure Poisson equation arising in Chorin's projection method for the lid-driven cavity flow. The proposed algorithm provides a promising building block for future quantum implementations of multigrid methods and preconditioning techniques, bringing quantum algorithms closer to established CFD solution strategies.

Unsupervised anomaly detection (UAD) aims to localize abnormal regions in medical scans without pixel-level annotations. A typical strategy seeks to reconstruct a pseudo-healthy image that preserves subject-specific anatomy. Recently, diffusion models have been proposed to perform UAD. However, these methods rely on heuristic noise schedules or synthetic corruptions to balance subject-specificity and anomaly removal. In this work, we propose an alternative formulation of UAD as a Bayesian inverse problem under a diffusion prior. First, we introduce a latent spatial anomaly mask that models pixel-wise consistency between a test image and its latent corresponding pseudo-healthy image. Then, we propose an approximation of the unknown generation process that links healthy anatomy, anomalies, and the observed image, enabling a well-defined likelihood within the Bayesian framework. Building on recent advances in diffusion-based inverse problem methods, we jointly infer the pseudo-healthy image and the anomaly mask via annealed posterior sampling. We evaluate our approach on FDG PET (ADNI) and FLAIR MRI (BraTS 2021), demonstrating improved anomaly localization performance compared to other diffusion-based approaches and validating the contribution of our introduced model. Our code is available at https://github.com/HuguesRoy/UAD_DAPS.

Tool-use agent traces identify messages and API calls, but procedural analyses also need explicit units of action and inspectable links to their evidence. We present Canonical Procedural Actions (CPAs), an annotation protocol that records a procedural function, its first agent-event anchor, the agent events that realize it, and separate contextual evidence. Multiple actions may share a message anchor without an inferred within-message order. A retail case study produces a versioned 24-entry codebook through open induction, recorded consolidation, and successive application audits. Two isolated LLM contexts annotate 32 trajectories disjoint from development at the trajectory level, producing 499 and 491 occurrences with anchor-label overlap A=0.982. Requiring identical context-event references reduces overlap to 0.798. These are structural repeatability measures, not semantic accuracy: 16 of 26 task IDs also occur in development, and historical tool payloads were truncated to 110 characters. Retrospective controls show that collapsing all labels raises overlap to 0.986, while simple endpoint rules reproduce the tool-anchored portion with 0.997 overlap. Assistant-message actions have 0.971 overlap, with a per-label minimum of 0.816. Applying the frozen codebook to 244 further trajectories yields 4,058 records, including eight diagnostic outcomes. The contribution is an explicit, auditable annotation instrument and a case study of its construction and measurement limits; human-reference validity and downstream utility remain to be established.

This paper investigates energy-efficient power control in 6G in-X subnetworks. We consider a graph neural network (GNN) framework that captures inter-subnetwork interference and the underlying wireless topology to optimize transmit powers. Three energy efficiency (EE) formulations are studied: (i) network-centric, which maximizes total network energy efficiency; (ii) subnetwork-centric, which maximizes the average energy efficiency per subnetwork; and (iii) a multi-objective approach, which balances energy efficiency and sum-rate performance. Extensive simulations in industrial factory settings with 3GPP channel models demonstrate that the GNN effectively learns interference-aware power allocation policies, significantly outperforming maximum power transmission and existing GNN based power control solution. Results showed network EE gains of up to 1341%, average per-device EE improvements of up to 1302%, and sum-rate enhancements up to 24.7% relative to a maximum transmit power policy, depending on the chosen optimization formulation and trade-off settings.

Operating conditions and fault severity change both the initial states and control modes of active distribution networks. Direct capacity aggregation can therefore lose the power responses required for transmission-system dynamic analysis. This paper develops a procedure for generating low-order physical electromagnetic transient equivalents from a small set of representative operating conditions. Device power, ride-through states, load mechanical states, and local sequence voltages are combined with electrical distance to form response families. Response reconstruction and terminal simulation errors determine the retained architecture, after which bounded sensitivity calculations determine effective dynamic coefficients. The architecture and coefficients are then fixed. For a new operating condition, capacity and power balances, voltage-dependent load relations, and an initialization correction generate the model parameters from static inputs and power-flow results. A representative severe fault establishes a template for each fault type, which is reused across fault depths. Aggregation and switching-time bounds explain the roles of voltage coherence and state separation. Comparisons with direct aggregation show improved terminal responses across operating conditions and several fault types, with the largest gains under severe unbalanced faults. The method provides a reusable physical equivalent for operating-condition studies and fault-depth screening.

Landau solution is a special family of solutions to the stationary Navier Stokes equations, which is important for the study of stationary problems such as asymptotic behavior and regularity. In this paper we prove that the Landau solution is rigid in rotated self-similar class when rotation parameter $\alpha$ sufficiently small or large. The arguments in the two parts rely on different approaches. For small rotation, we use the compactness argument and establish a classification lemma of the self-similar kernel of the linearization around any Landau solution, this method can also be extended to DSS and RDSS cases. For large rotation, the strong angular dissipation produces additional coercivity on the non-axisymmetric component, which forces the solution to reduce into Landau solution.

We study the problem of adversarially robust PAC learning. In this framework, the learner observes independent samples from an unknown distribution over $\mathcal{X} \times \{0,1\}$, as in classical PAC learning. However, given a perturbation map $\mathcal{U} : \mathcal{X} \to 2^{\mathcal{X}}$ known to the learner, the goal is to output, with high probability, a predictor that correctly classifies every perturbation $z \in \mathcal{U}(x)$ of most future examples $(x,y)$ drawn from the same underlying distribution. We determine the optimal $\mathcal{U}$-independent sample complexity of this problem in both the realizable and agnostic settings. More specifically, for every concept class $\mathcal{H}$ of $\operatorname{VC}$ dimension $d$, we prove upper bounds of $\mathcal{O} \big( d/\epsilon + \log(1/\delta)/\epsilon \big)$ in the realizable setting and $\mathcal{O} \big( d/\epsilon^2 + \log(1/\delta)/\epsilon^2 \big)$ in the agnostic setting, together with an optimal first-order refinement of the latter. These bounds match the corresponding lower bounds for classical PAC learning. Consequently, and perhaps surprisingly, adversarial robustness incurs no additional distribution-free statistical cost, uniformly over all perturbation maps. Our bounds improve exponentially on those of [Montasser, Hanneke, and Srebro; COLT '19]. On the technical side, we present short and elementary proofs based on a new algorithmic principle that we call binomial-bagging. We believe that binomial-bagging and its analysis may be of independent interest.

The effectiveness of agent memory ultimately depends on whether the underlying LLM gives each memory in context an appropriate degree of influence over its response. Yet this capability has remained largely overlooked. To assess this capability, we introduce MemCalib, a benchmark grounded in realistic memory-system scenarios for evaluating memory use and advancing optimization algorithms. Results on the MemCalib test set reveal that frontier open- and closed-source models struggle to use memory appropriately. They frequently over-use or under-use memory rather than matching each proposition's actual use to its target level, leading to biased, low-quality responses. Experiments with common post-training algorithms, including group relative policy optimization and on-policy self-distillation, further reveal a clear directional skew: trained models improve in one direction while deteriorating in the other. We therefore propose MemCalib-RL, an ordered bidirectional counterfactual credit-assignment algorithm that separates over- and under-use signals and localizes their credit to response tokens through exact atom ablation. Results across model families and scales (Qwen3-8B, Ministral-3-8B-Instruct, and Qwen3.5-35B-A3B) show that MemCalib-RL achieves the best overall performance while better balancing over-use and under-use, with gains generalizing beyond MemCalib in external benchmark evaluation. Further experiments support its design choices and robustness and provide insight into its training dynamics.

We study the parameter dependence of weighted Bergman kernels on fixed bounded domains in $\mathbb C^n$. Our main result establishes real-analytic dependence on $t\in(-1,\infty)$ for the kernels associated with $\delta^t\,dV$ on bounded pseudoconvex domains with $C^2$ boundary, where $\delta$ is the Euclidean distance to the boundary. The parameter derivatives satisfy factorial estimates in $C^\ell(S\times S)$ for every $S\Subset\Omega$ and $\ell\ge0$, uniformly on compact parameter intervals. The proof combines weighted $L^2$ estimates for $\bar\partial$ with a holomorphic family of bounded operators and gives a local holomorphic extension with values in a fixed weighted Bergman space. For weight families smooth jointly in space and parameter up to the boundary, we also express all parameter derivatives in terms of iterated weighted Bergman projections and complete exponential Bell polynomials. This formula implies preservation of the Gevrey class $G^s$, $s\ge1$, under uniform parameter estimates up to the boundary. We show that the same derivative formula holds for the weights $-t\log\delta$.

This paper proposes a novel discretely entropy conserving/stable flux reconstruction (FR) scheme for nonlinear conservation laws. The scheme combines a hybridized split-form differential operator with a generalized entropy projection, enabling discrete entropy conservation for arbitrary nodal distributions and arbitrary correction functions without modifying the original quadrature weights. The formulation is built directly upon the classical FR framework and is shown to preserve conservation properties while achieving entropy conservation in the semi-discrete sense. Numerical experiments using the isentropic vortex problem confirm the theoretically predicted accuracy and demonstrate robust entropy behavior with entropy-dissipative interface fluxes.

This paper is devoted to the derivation of the dispersion relation of the Geodesic Acoustic Mode in pair plasmas, i.e. assuming that ions and electrons have the same mass. Geodesic Acoustic Modes are plasma perturbations playing a crucial role in turbulence regulation, and therefore in the determination of the plasma confinement in tokamaks. Experiments with pair plasmas, like electron-positron plasmas, have been proposed with different kinds of confinements, and aim to study fundamental processes in plasma physics and understanding the formation of the early universe.

Let \(\Kl(1;q)\) denote the normalized Kloosterman sum modulo \(q\). We prove that, for each sign, there are \(\gg X/\log X\) square-free moduli \(q\in(X,2X]\) with at most four prime factors for which \(\Kl(1;q)\) has that sign. The proof combines the Selberg sieve with equidistribution and mean-square estimates for Kloosterman sums; the uniform estimate for shifted sieve weights established in Part~I controls the contribution of moduli having small prime factors.

Evidence about sexual violence in Bangladesh is fragmented across individual reports, while official and civil-society statistics rarely provide reusable case-level detail. We examine how structured analysis can make one part of this fragmented record legible without treating it as prevalence data. Our corpus contains 2,811 articles timestamped 2013-2023 from the Prothom Alo publishing ecosystem; 2,794 include parseable metadata about reported victims, alleged perpetrators, incidents, legal responses, and locations. We combine descriptive and spatial analysis of these records with thematic analysis of 115 convenience-sample survey responses collected in late 2020. The corpus documents many young, female, and student victims, frequent acquaintance and neighbor relationships, uneven geographic documentation, and substantial missingness in legal outcomes. Respondents most often describe weak enforcement, insecurity, patriarchal socialization, education gaps, and community inaction as conditions enabling persistence. These findings characterize news documentation and public perceptions; they do not estimate incidence, geographic risk, or causality. We contribute an uncertainty-aware framing for HCI research using sensitive, low-resource news data and identify design requirements for provenance, privacy, validation, and responsible communication.

3D Gaussian Splatting (3DGS) provides high-fidelity scenes for large-scale embodied simulation, but constructing large-scale urban assets remains constrained by expensive equipment and delayed quality feedback. Preset surveys can leave complex surfaces insufficiently observed, with defects discovered only after reconstruction, requiring return visits and repeated processing. We present OpenFlyScan, a quality-guided aerial reconstruction system for consumer drones that integrates a GS quality model, a reacquisition planner, and a custom-designed mobile app. The model learns from GS rendering errors to predict regional reconstruction quality. Based on these predictions, the planner then generates complementary reacquisition strips to be executed through the app, which also supports automated oblique surveys and data transfer without additional hardware on board. Across real aerial scenes, the model effectively identifies regions that are likely to be poorly reconstructed. In the Expo West field experiment, targeted reacquisition improves PSNR at additional views by 10.95 dB. With consumer drones, OpenFlyScan integrates capture, targeted reacquisition, and reconstruction to support rapid, low-cost urban asset creation. Code and models will be made publicly available at https://openflyscan.github.io/.

The Iris framework for concurrent separation logic has been widely used for program verification research. An important factor contributing to the framework's adoption is its high-quality mechanization in Rocq. This mechanization uses a number of Rocq features in sophisticated ways, including a carefully constructed algebraic hierarchy for modeling separation logic resources, and a proof mode for embedded separation logic proofs, which combines custom Ltac with extensible typeclasses. The library has been developed for over a decade with dozens of contributors, with an emphasis on modularity and maintainability. We explore how Lean features like flexible metaprogramming and quotient types can simplify the design and usage of Iris. Using these features, we provide a novel implementation of the Iris proof mode with improved performance, simplify the handling of equivalences through quotient types, build a variant of Diaframe proof automation, and provide convenience features like automatic construction of Iris fixed points. Building on Lean lets us integrate with the extensive Mathlib library, allowing us to re-use results from this library for program verification tasks that have heavy mathematical dependencies, as we demonstrate with an application to probabilistic program verification.

We prove that a complete, two-sided, stable capillary minimal hypersurface in $\R^4_+$ is flat provided its contact angle $\theta$ satisfies $|\cot\theta|<\kappa_*$, where $\kappa_*=4\Gamma(3/4)^2/\Gamma(1/4)^2\approx 0.45$.

The shipping industry has a significant impact on the global economy, emphasizing the need for operational availability and safety through the use of effective maintenance techniques. During the last decades, predictive maintenance (PdM) has emerged as a promising solution compared to the existing conventional maintenance systems. This is because it offers several advantageous functions, such as damage predictions for vessel components, reduced downtime, improved and extended life of machinery, as well as higher safety during voyages. However, existing methodologies developed for performing PdM do not provide explanations of their results to users, so that they can understand the failures that may occur. To address this limitation, this paper proposes a novel framework based on a fuzzy decision tree and a deep residual neural network, aiming to perform explainable PdM on naval vessels. The proposed framework is able to generate fuzzy local rules based on the dataset used, and can provide explanations of its outcomes, using cause-and-effect relationships, in a way that are understandable to users, thereby gaining their trust. Experiments using a publicly available dataset demonstrate the effectiveness of the proposed framework, as it achieves an accuracy of 99.24%.

Neural networks, both convolution or transformer based, are essential for modern computer vision systems. However, they are vulnerable to small perturbations, almost imperceptible to humans, which significantly alter the model's prediction. These adversarial attacks are often considered to be a significant threat to the implementation of neural networks in safety-critical applications. Most attacks utilize the white-box threat model and therefore require full access to the target model, making them unrealistic to use in practice. We propose a novel approach under the more realistic black-box threat model that utilizes concepts from reinforcement learning to optimize perturbations with a non-differentiable target model. Reinforcement learning algorithms have already been optimized to be query efficient, making them an ideal starting point when designing black-box adversarial attacks. We show the success of our reinforcement learning inspired black-box adversarial attack (RIBA) in generating adversarial perturbations using only a small number of queries to the target model, by comparing it to state of the art attacks on different models on the Cifar10 and ImageNet data sets. RIBA takes $25.4\%$ fewer median queries to generate attacked images against a ResNet-18 on Cifar10 and $22.5\%$ fewer median queries to fool a Vit-B/16 model on ImageNet. Additionally, we demonstrate that RIBA can match the performance of white-box attacks on an adversarially trained model.

We prove an asymptotic formula for correlations of divisor sums with quadratic cutoff functions shifted independently by \(s,u\in[0,1/2]\). The error is \(O_g(Y/(\log Y)^2)\), uniformly for \(Y^{1/5}\leq R\leq Y^{1/3}\), including coincident shifts. The main term is an explicit bilinear form in the first and second derivatives of the cutoff functions. We obtain it by taking the residue in the outer Mellin variable and evaluating the remaining double integral by Laplace inversion. The present paper establishes the divisor-sum estimate. Its application to the sign-change problem, including the choice of the prime cutoff and the remaining parameters, is treated in Part II.

In this work, we characterize the glassy character of the Bose-glass phase in a disordered Bose-Hubbard model using three complementary diagnostics: the finite-temperature spectral function, spatial inhomogeneity, and the inverse participation ratio. Spectral analysis, obtained from finite-temperature Green's function and random phase approximation, shows that disorder introduces localized low-energy states within the Mott gap, eventually closing the gap at sufficiently strong disorder. Spatial inhomogeneity, calculated using the Gutzwiller ansatz, increases sharply with disorder and then saturates at moderate disorder strengths. The inverse participation ratio has been calculated from the exact diagonalization of a small system, and it reveals enhanced localization with stronger disorder. Increasing temperature suppresses these disorder-induced features: the low-energy spectral weight diminishes, the spatial inhomogeneity varies more smoothly, and the IPR decreases. Taken together, these diagnostics show that disorder drives the development of glassy character, while thermal fluctuations gradually wash out its signatures. The early saturation of spatial inhomogeneity compared with the continued evolution of low-energy excitations and the inverse participation ratio demonstrates that no single diagnostic fully captures the evolution of the Bose-glass state, highlighting the need for a combined characterization.

Large language models (LLMs) have shown remarkable progress in natural language understanding, yet their effectiveness in specialized fields like astronomy and astrodynamics remains limited due to challenges in multi-step reasoning, symbolic manipulation, and domain-specific terminology. To address this, we present Taramandal-GPT (Constellation-GPT), a domain-adapted framework built on the Qwen3-8b backbone, enhanced with a Retrieval-Augmented Generation (RAG) pipeline and a fallback mechanism for improved contextual precision. We evaluate it on the Astrodynamics Problems Benchmark (APBench), a dataset of 299 questions covering foundational to advanced levels of space science. Using a dual evaluation method - numeric margin-based scoring and semantic similarity assessment - Taramandal-GPT achieves competitive performance against state-of-the-art open- and closed-source models, with notable strength in thinking-intensive tasks. These results highlight the value of specialized LLMs for domains demanding accuracy and interpretability, positioning Taramandal-GPT as a step toward reliable Artificial Intelligence (AI) assistants for astrophysics, spacecraft engineering, and space exploration.

Social Stories are widely used to support autistic children in understanding and preparing for everyday situations, but creating stories that are appropriately tailored to each child's needs remains labor-intensive for practitioners. Existing digital tools support story assembly and delivery, but much of the work of writing, visual preparation, and personalization remains manual. Recent AI-based approaches have enabled automated story generation, yet offer limited support for practitioner oversight, context-sensitive personalization, and the use of supportive visuals grounded in individual learner profiles. We present AdaptED Stories, a practitioner-guided system for authoring, personalizing, and delivering Social Stories in special-education contexts. The system uses student profiles to draft story text and visuals, supports review and refinement by practitioners, and includes reading-session support with comprehension activities and session records. We report findings from a practitioner-informed design process, assessments of generated stories and visuals, and a usability study with seven special-education practitioners. Our findings suggest that AI assistance can reduce story-preparation burden and support more individualized story creation, alongside the importance of practitioner oversight, cultural and contextual specificity, and designing for varied learner needs. These findings contribute design implications for AI-assisted accessibility tools in special-education.

Multimodal large language models (MLLMs) fail at fine-grained visual questions less because they cannot reason than because they never see the evidence: high-resolution images are downsampled before encoding, so the model answers from linguistic priors. The standard remedies are expensive: annotated answers (SFT), hand-engineered verifiers (RLVR), or a large external teacher (on-policy distillation). We ask whether the visual evidence itself can supply the signal for free. We formalize the contrastive evidence gap, the per-token log-likelihood ratio that a model assigns to its own output when conditioned on a question-relevant region versus an irrelevant one, and study it across Qwen2.5-VL-7B, Qwen3-VL-8B, and Qwen3-VL-30B-A3B on V*Bench. Our main positive result is training-free: selecting the candidate crop under which the model's answer distribution is most peaked, using a single-view, label-free criterion, discovers the answer-bearing region with no bounding boxes, training, or labels. It localizes the target 4.4 to 5.1 times better than chance and raises fine-grained accuracy from 70 percent to 85 percent at inference. We further show that the gap is complementary to the model's own confidence. Combining them predicts correctness better than either alone, with AUC up to 0.99, and flags confidently wrong answers, with AUC ranging from 0.97 to 1.00 within the high-confidence subset. All effects concentrate on perception-bottleneck questions and vanish on a global-context control. Finally, we report an honest negative result: converting the same signal into a training method, gated self-distillation (SEG-Distill), does not outperform the base model at pilot scale across three gate designs, while more aggressive gating degrades accuracy. The signal is real, but converting it into training gains remains an open problem.

Reinforcement learning (RL) has substantially improved the reasoning capabilities of vision-language models (VLMs), yet it can induce chain-of-thought (CoT) obfuscation, an operational, non-intentional regime where models achieve high accuracy while producing reasoning traces that are less grounded and harder to monitor. While prior work documents this decay at the behavioral level, the underlying representational correlates of such drift remain poorly understood. In this work, we find that obfuscation is accompanied by a representational shift, where RL-induced optimization makes task-agnostic template features less separable from visually grounded content in the model's activation space. Based on this insight, we propose TAME, an activation-level intervention framework that leverages Sparse Autoencoders (SAEs) to regularize VLM internal features during RL training. Specifically, TAME utilizes an anti-obfuscation feedback loop to reduce behavioral obfuscation and applies an asymmetric SAE constraint to specific internal directions, penalizing template activations while preserving visual evidence. By directly intervening on the features associated with obfuscation, our method improves the transparency and visual grounding of reasoning traces. Experiments on the VIRL-39k and SPA-VL benchmarks across two model families show that TAME improves CoT monitorability (by up to +30.9 points on VIRL and +16.7 points on SPA-VL over GRPO), while largely preserving task accuracy and general-capability performance across four standard VLM benchmarks, with task-specific trade-offs. Our code is available at https://github.com/henrymao2004/TAME.