# Research Papers — 2026-09 · bundle page 133 of 266 > Full-text markdown of papers #6601–6650 of 13298 published in 2026-09, newest first. AI-readable mirrors of the full 13298-paper 2026-09 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2026-09/page/133 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/132.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/134.md) ### Analysing the Linearity of Linguistic Relations in Language Model Embedding Spaces Authors: Vasudevan Nedumpozhimana, Fathima Thekkekara, John Kelleher Published: 2026-09-18 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.21655v1 We propose a framework to analyse how strongly different linguistic relations are linearly encoded in language model embedding spaces. We formalise linear encoding via a constrained linear approximation over related and unrelated word pairs and apply this to an extended BATS dataset covering inflectional, derivational, lexicographic, and encyclopedic relations in GloVe, RoBERTa, and ModernBERT. Our experiments show near-perfect linear encodings for inflectional and derivational relations, but substantially higher errors for lexicographic and encyclopedic relations, especially for one-to-many and many-to-many associations. We also find that RoBERTa and ModernBERT generally encode relations more linearly than GloVe. These results indicate that our framework can reveal which relational structures are most linearly accessible in embeddings, offering a compact tool for probing and comparing relational geometry across models. --- ### Configurable Multi-Stage Vision Pipeline for Crop Disease and Pest Diagnosis Authors: Naga Ganesh, Chandrashekar M S, Lakshmi Pedapudi, Aakash Singh, Vineet Singh Published: 2026-09-18 Categories: cs.CV, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.21651v2 FarmerChat is Digital Green's farm advisory service for smallholder farmers. When something looks wrong with a crop, the farmer takes a photograph and sends it, and that photograph is the whole question: no symptom described, no crop named, often no text at all. The service has to determine whether the picture can be used, what crop it shows, and what is wrong with it, from images taken on cheap phones in a field, in poor light and with a moving camera. The system doing this today cannot be adjusted. It has no adjustable thresholds for photograph rejection, crops and problems cannot be added, and there is no confidence cut-off to set. We study about 1.16 million photographs sent to FarmerChat from Ethiopia, India, Kenya and Nigeria. The production quality gate rejected 46.8% of the images it judged, over a quarter of those reaching diagnosis returned no crop name, and 35.8% of the labelled problems filed under "disease" are pests, identifiable without the crop. We therefore split the work into three stages: a quality gate (M0), a crop detector (M1), and a disease or pest detector (M2). Route A fills all three with one fine-tuned vision-language model (Qwen3-VL-4B) answering in a single call. Route B fills each with a small specialist model (DaViT, YOLO26). We replace our production GPT-4o quality gate with a small MobileNetV3 gate at 86.9% F1 in 12 ms. On one test set scored the same way for every system, a hierarchical DaViT-Base achieves 95.41% crop accuracy against 91.46% for the production baseline. It also leads on diagnosis and never declines to answer, while every language model in the comparison leaves a large share of rows with no diagnosis. The fine-tuned model retains two capabilities the specialists do not have: one call for all three stages, and a request for a better photograph when the image cannot support an answer. --- ### SynthDemo-RL: Breaking the Zero-Reward Barrier in VLA Adaptation with LLM-Guided Synthetic Demonstrations Authors: Hiroaki Kingetsu, Hiroaki Kurihara, Kaoru Yokoo, Kenji Fukumizu, Manohar Kaul Published: 2026-09-18 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.21650v1 Fine-tuning Vision-Language-Action (VLA) models commonly relies on human teleoperation demonstrations, while reinforcement learning (RL) with sparse binary rewards faces an exploration challenge when successful trajectories are rarely sampled. We propose SynthDemo-RL, a teacher-student framework in which an automated teacher converts simulator-privileged state into successful manipulation trajectories, a VLA student is distilled from them by supervised fine-tuning (SFT), and PPO with binary task-success rewards refines the student. We study reward coverage, the fraction of tasks for which at least one success is observed under the fixed evaluation protocol, as a complement to the average success rate. On LIBERO-PRO, a public benchmark of perturbed LIBERO tasks for which no demonstrations exist, 27 of 57 scored tasks are at exactly 0% success for a pi_0.5 policy fine-tuned on the original LIBERO tasks. Direct PPO from this policy, under the same PPO recipe and the same RL compute as SynthDemo-RL's refinement stage, rescues 10 of these 27 tasks and leaves 17 at 0%. SynthDemo-RL, with 50 synthesized trajectories per task and no new human demonstrations, rescues all 27 and reaches average success rates of 97.8% and 97.1% on the Position and Task axes of LIBERO-PRO, respectively. On standard LIBERO, the same pipeline reaches 96.0% with no human demonstrations, within 1.7 points of pi_0.5 trained on 50 human demonstrations per task. We further validate the pipeline on RoboTwin 2.0 and verify that trajectories from a policy trained in a MuJoCo twin execute open-loop on a physical robot. --- ### Riemannian Neural Hamiltonian Flows: Geodesic Symplectic Transport and Interpretability Authors: Vincent Souveton Published: 2026-09-18 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.21647v1 Hamiltonian normalizing flows are attractive generative models because their phase-space maps are invertible and volume preserving, but most neural constructions are formulated in Euclidean space. We introduce Riemannian Neural Hamiltonian Flows, which combine the fixed kinetic energy of a Riemannian manifold, a learned scalar potential, and an explicit geodesic leapfrog integrator. Our analysis explains how the learned Hamiltonian can be made interpretable. Every normalizable potential defines an implicit profile, and the position marginal initially accelerates along the relative score between that profile and the base. The matched potential is the interpretable specialization for which the implicit profile is the target. In the isotropic Gaussian case, the mechanism corresponds to a phase-space rotation. A local harmonic analysis extends this result around each mode of a general target on a manifold. The gap between the learned and the matched potential is the sum of a residual memory of the base and a bias of the model, and the two potentials agree when the position base has been transferred to the momentum. This can be achieved when the former is broader than the target. Numerical experiments on Euclidean, hyperbolic, and spherical spaces show competitive sample quality and numerical cost against a Riemannian continuous normalizing flow, and confirm the interpretability of the learned potential. --- ### Chinese Competitive Debating Dataset and Benchmark Authors: Zongrui Yang, Haoyuan Li, Zhongsheng Wang, Zhirui Zeng, Pengqian Han, Yi Zhou, Yuting Wang, Jiamou Liu Published: 2026-09-18 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.21637v1 Debate adjudication requires tracking how arguments develop through interaction, yet existing datasets rarely combine fine-grained debate transcripts with professional judgments collected during real competitions under a shared rubric. We introduce a dataset and benchmark for evaluating large language models' understanding of competitive Chinese-language debate at the match, stage, and speaker levels. We organized 182 matches and recruited 120 professional judges, with each match independently adjudicated by three judges using a predefined rubric. After excluding matches with incomplete records, the dataset contains 148 matches, 2,698 stages, and 20,542 exchange units, with manually verified transcripts and segmentation. It preserves original stage scores, match votes, best-debater ballots, and adjudication rationales. We define three tasks: winner-tendency prediction, stage-score prediction, and best-debater prediction. Zero-shot evaluation of multiple large language models yields a highest winner-prediction accuracy of 66.2%, a highest Pearson correlation of 0.250 between model stage scores and mean human ratings, and a highest best-debater prediction accuracy of 56.8%. The dataset and benchmark provide a testbed for studying large language models' understanding of interactive argumentation and their agreement with professional judges. --- ### Steering LLMs Responses Towards Moral Foundations on the Norwegian MFQ-30 Authors: Hans Andersen, David Dichas Published: 2026-09-18 Categories: cs.CL, cs.AI, cs.CY arXiv: https://arxiv.org/abs/2609.21636v1 Recent work applies human psychometric questionnaires to large language models to elicit moral and value profiles, but it is not clear whether these instruments measure anything stable in models or whether the resulting profiles can be moved toward a target human population. We administer the Norwegian Moral Foundations Questionnaire (MFQ-30) to six open-weight LLMs and compare their foundation profiles to a sample of N = 1,282 Norwegian respondents. We test two steering interventions, prompt-level persona steering and activation-level ActAdd. Half the models engage with the questionnaire under our attention check. The other half default to flat or central-tendency outputs that look near-human on average without tracking item content. A neutral Nordic-respondent persona, written without any distributional information from the human sample, brings the engaging models 44-77% closer to the Norwegian mean in Mahalanobis $d^2$. One-pair ActAdd at a fixed mid-layer flattens the foundation profile rather than steering individual foundations. For at least one model the same persona that shifts the profile also induces engagement that was absent at baseline, a concrete instance of the cognitive phantoms that Peereboom et al. (2025) warn about. --- ### Extending Decoupled Attention to Dense Prediction and Masked Training for Multi-Channel Images Authors: Umar Marikkar, Sameed Husain, Muhammad Awais, Sara Atito Published: 2026-09-18 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.21629v1 Multi-Channel imaging (MCI) data differs fundamentally from natural images, as each channel records a semantically distinct signal rather than a colour band. To adapt vision encoders to MCI data, Multi-Channel Vision Transformers (MC-ViTs) tokenize each channel independently and concatenate the resulting tokens into one sequence, and the channel count is no longer fixed by the architecture. Self-attention is then computed across all channel-patch tokens with no restriction on which channels attend to which, which dilutes the features of individual channels. The Decoupled Vision Transformer (DC-ViT) regulates this by separating updates computed within a channel from updates computed across channels, and by forming a representation per channel before the channels are combined. Its formulation, however, pairs tokens by spatial position, and thus requires the same visible tokens in every channel. Correspondence under independent per-channel masking is recovered by solving a linear assignment between the retained patches of each channel, which allows decoupled attention to be combined with current masked multi-channel training in its standard configuration rather than a restricted one. Across three classification and three segmentation benchmarks spanning fluorescence microscopy, imaging mass cytometry and satellite imaging, including dense prediction at high channel counts, the resulting formulation outperforms the strongest MC-ViT baseline. --- ### Detection is solved, delineation is not: what governs tooth segmentation on panoramic radiographs Authors: Muhammad Rehan, Moaz Amjad, Syed Danial Ahmed, Mariam Adnan, Haider Ali Published: 2026-09-18 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.21628v1 Automatic tooth segmentation and FDI numbering on panoramic radiographs underpins computer-assisted dental diagnosis, yet which factors govern performance remains unclear. We assemble a corpus of 1,422 panoramic radiographs containing 42,142 expert-delineated tooth polygons across the 32-class FDI taxonomy, annotated by 30 dental practitioners and independently reviewed by two others, and use it to isolate input resolution, architecture and anatomical priors under a single evaluation protocol. First, resolution dominates: across a controlled 640/1024/1280 ablation, mask mAP50-95 rises 0.656 -> 0.710 -> 0.717 while mAP50 stays flat at ~0.982. Both gains are significant under a paired bootstrap over images (p < 0.001, p = 0.024); neither mAP50 change is distinguishable from zero. Added resolution buys boundary precision, not detection. Second, architecture is nearly irrelevant in-domain: a query-based transformer with 2.1x the parameters is statistically equivalent to a one-stage detector (95% CI (-0.0064, +0.0064)), only marginally better under domain shift, 5.5x slower on CPU and not executable under standard ONNX runtimes. Third, three targeted interventions fail: a LoRA-adapted self-supervised encoder underperforms, a promptable foundation segmenter degrades masks by 39%, and globally optimal anatomical label assignment yields +0.0007 despite correcting a constraint violated in 40% of out-of-domain predictions. Zero-shot transfer to an independent multi-centre cohort, verified overlap-free, costs 62% of mask mAP50-95 but only 18% of mAP50, reproducing the dissociation. Decomposing masks along the tooth axis localises the residual error to the apical third. Boundary precision is therefore the binding constraint, and effort is better directed at resolution and acquisition diversity than at architectural novelty. --- ### One Prompt Does Not Fit All: Self-Meta-Evolve for Personalized Information Extraction Authors: Hongliang Li, Lu Wang, Yong Xu, Hanyang Chen, Zhitao Hou, Xiaoting Qin, Song Ge, Qingwei Lin, Dongmei Zhang Published: 2026-09-18 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.21626v1 Large language models (LLMs) are increasingly deployed for enterprise information extraction (IE), where the same document must be reorganized differently for each user. Existing prompt optimization methods, however, rely on a single prompt optimized against a global objective, which is misaligned with the inherent user heterogeneity of real workplaces. We formulate enterprise IE as per-user prompt adaptation under interaction feedback and propose Self-Meta-Evolve, a hierarchical framework that maintains a dedicated prompt for each user and continuously refines it through a dual-loop process: an inner loop that edits structured prompts based on persona-conditioned feedback, and an outer loop that evolves the meta-prompt itself by distilling successful editing patterns. To enable scalable training and evaluation, we release a persona-driven IE benchmark of 292 simulated enterprise users, paired with a reproducible persona-generation pipeline grounded in O*NET occupational taxonomies. On this benchmark, Self-Meta-Evolve achieves a 74.58% success rate, outperforming the strongest prompt-optimization baseline by 13.56 absolute points, and reaches 52.54\% within only two iterations. A double-blind human study with twenty real professionals further confirms that prompts adapted by our framework win against static baselines in 71% of pairwise comparisons. --- ### Learned Parametric Emotion Editing: Real-Time Affective Filtering for On-Device Social Media Video Authors: Musa Rochi, Marcel Schubert, Christoph Gebhardt Published: 2026-09-18 Categories: cs.CV, cs.HC arXiv: https://arxiv.org/abs/2609.21624v1 Problematic internet use affects a growing share of the population, yet common interventions, e.g., time limits, blocking, forced breaks, are coercive and easily circumvented. We explore a less restrictive alternative: adapting the emotional intensity of visual content. Prior work has shown that optimization can steer an image's affective content, but its per-image optimization cost makes it impractical for real-time deployment. We instead learn a model that predicts this transformation in a single forward pass: a MobileNetV4 backbone with FiLM-based emotion conditioning outputs parameters for differentiable global transformations. This replaces prior iterative optimization (80 s per image) with a single 3.7 ms forward pass. In a user study (N = 54), the model reduced viewer-reported arousal relative to unedited images, comparably to the grayscale well-being filter, while being rated higher in perceived quality. We integrate the model into an Android app that adapts Instagram video in real time, sustaining 60 fps on a Samsung Galaxy S23. --- ### Calibrating Teacher--Student Discrepancy for On-Policy Distillation Authors: Qiangqiang He, Jin Li, MingCai Chen Published: 2026-09-18 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.21619v1 On-policy distillation (OPD) improves reasoning models by learning the token-level discrepancy between a stronger teacher and an on-policy student. However, this discrepancy does not purely reflect the capability gap between the teacher and the student: it also contains deviations arising from the teacher itself, which are consequently mixed into the observed teacher--student discrepancy and indiscriminately learned by standard OPD during training. This issue is further exacerbated by privileged OPD, where privileged information induces larger teacher-side likelihood shifts, thereby encouraging the student to learn more of the teacher's own deviation. We introduce \textbf{Calibrated On-Policy Distillation (Cal-OPD)}, which estimates the teacher's self-deviation region through positive and negative privileged interventions and calibrates the original teacher--student discrepancy by retaining only the component that lies beyond this region. Experiments on mathematical reasoning benchmarks show that, while retaining only about 52--65\% of the original teacher--student discrepancy as the optimization signal, Cal-OPD consistently outperforms standard OPD and its variants across model scales. --- ### Potential-Field Action Representation for Reinforcement Learning in Contact-Rich Manipulation Authors: Xinyu Liu, Gökhan Solak, Arash Ajoudani Published: 2026-09-18 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.21609v1 Model-free reinforcement learning can acquire contact-rich robotic manipulation skills through trial-and-error interaction, but it often requires the policy to learn both task strategy and low-level motion generation. In this setting, the action representation is critical because it determines how policy outputs are converted into robot motion, shaping both exploration and physical execution. Direct Cartesian command interfaces require the policy to generate motion at every decision step, coupling task-level adaptation with continuous low-level control and increasing the learning burden. We propose PA-RL, a reinforcement-learning framework that uses artificial potential fields as the action representation. Instead of commanding motion directly, the policy adapts the parameters of an energy-like potential field, which generates a state-dependent guidance direction executed through a Cartesian impedance controller. We evaluate PA-RL on peg-in-hole insertion, a representative contact-rich task with nonlinear dynamics and discontinuous contact transitions. In simulation, PA-RL is compared with Cartesian velocity, Cartesian pose, and variable-impedance action spaces using the same RL algorithm. PA-RL is the only method to reach a 100% evaluation success rate within the allotted training time, while the best baseline reaches 92.6%. It also reduces joint-torque variation by 55.4% and Cartesian acceleration variation by 70.8% relative to the best baseline, without explicit motion-quality penalties in the reward. The simulation-trained policy further completes 9/9 real-robot insertions without fine-tuning, demonstrating the deployment feasibility of the learned potential-field interface. --- ### Trading Depth for Time in Recurrent Transformers Authors: Zeyi Huang, Xuehai He, Yong Jae Lee, Yelong Shen Published: 2026-09-18 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.21605v1 Recurrent Transformers increase computational depth through temporal recurrence, feeding each token's high-level hidden state into the computation of the next. This raises a natural question: is additional computation better spent on more temporal steps or greater physical depth? We investigate this question using Latent Recurrent Transformers (LRTs), which retain one backbone forward pass per vocabulary token during decoding and provide a controlled setting for comparing these two ways of adding computation. Specifically, we insert a latent thought token between consecutive vocabulary tokens. Each thought token passes through the same $L$ layers as a vocabulary token, sharing the backbone parameters and providing an additional stage of hidden-state refinement before predicting the next token. We compare this $L$-layer LRT against a $2L$-layer LRT without thought tokens. Both execute $2L$ Transformer blocks per vocabulary token during decoding, but the thought-token model uses fewer parameters. On 16- and 20-layer mixture-of-experts NanoChat backbones, one thought token brings the shallower model within 0.006 and 0.004 bits per byte of its double-depth counterpart, recovering 67% and 81% of the improvement with approximately 48% fewer total parameters. These results suggest that temporal thinking offers a parameter-efficient alternative to increasing physical depth in recurrent Transformers. --- ### Reducing Barriers to Academic Support: Evaluating a Course-Specific RAG System for Addressing Help-Seeking Disparities in Higher Education Authors: Andy Gray, Jake Hobbs Published: 2026-09-18 Categories: cs.AI, cs.CY, cs.HC arXiv: https://arxiv.org/abs/2609.21600v1 Access to academic support is a key determinant of student success, yet students experience it unequally: some readily seek help from lecturers or tutors, while others hesitate due to anxiety, fear of judgement, uncertainty about expectations, or low confidence in their understanding. This may be especially evident in computing education, where programming tasks are cumulative and cognitively demanding. Although students increasingly turn to general-purpose generative AI tools, these can produce responses that are inaccurate, insufficiently contextualised, or misaligned with module expectations. This study presents and evaluates Beacon, a course-specific Retrieval-Augmented Generation (RAG) system providing private, immediate, module-aligned academic support. Grounding responses in approved teaching materials, Beacon was designed to lower barriers to help-seeking while encouraging independent learning. Using a design-based research approach, Beacon was developed iteratively and evaluated via mixed methods, combining questionnaires and semi-structured interviews with students and staff at a Higher Education institution. Students described Beacon's responses as closely aligned with module content and more trustworthy than unrestricted generative AI tools, valuing its use of pseudocode and scaffolded explanations over direct solutions. Although participants remained cautious about trusting AI-generated responses without verification, they viewed the system as a valuable first point of support before consulting lecturers or official resources. The findings suggest that carefully designed course-specific AI systems may reduce barriers to academic support by occupying an intermediary space between independent study and formal support. Rather than replacing educators, educational AI may be most valuable when it broadens access to guidance while preserving the pedagogical role of lecturers. --- ### Beyond Accuracy: Centroid-Guided Contrastive Loss for Structured Fraudulent Job Posting Detection Authors: Syed Ali Ahmed, Malaika Raza, Muhammad Shoaib Siddiqui, Muhammad Rafi Published: 2026-09-18 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.21599v1 Fraudulent job posting detection aims to identify job advertisements that are corrupted either through fake content, misleading information, or negative intent, disrupting the online eco-system of job-seekers and employers. Existing studies in this domain lack effective methods to simultaneously achieve high accuracy and meaningful structure of latent-space representations that capture subtleties among fake posts. To this end, we propose Centroid-Guided Contrastive Loss (CGCL), a loss function which unifies classification with densely formulated clustering to consistently reshape latent-space through a centroid-driven top-$k$ push-and-pull mechanism. The complementary nature of CGCL enables the model to enforce accurate decision boundaries and maintain high clustering compactness, effectively capturing both class separability and latent structure. Extensive experiments demonstrate the state-of-the-art (SOTA) performance of our method on EMSCAD, a public benchmark dataset. The code associated with this work is available at: https://github.com/ali-ahmed925/CGCL_code --- ### HAT: Hypothesis-Anchored Tracking for Video Monocular Spacecraft Pose Estimation Authors: André Lopo, Atabak Dehban, Rodrigo Ventura Published: 2026-09-18 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.21597v1 Monocular 6-DoF pose estimation of non-cooperative targets is important for on-orbit servicing and debris removal. A single-image estimator can confuse near-symmetric spacecraft orientations, and tracking can preserve an incorrect pose. We present Hypothesis-Anchored Tracking (HAT), a causal framework that uses inter-frame motion to select among competing CAD-based pose hypotheses before alignment and fusion. Rather than independently choosing the highest-scoring hypothesis in each image, HAT retains competing orientation histories and selects a pose to anchor the relative trajectory estimated by monocular SLAM. Sparse anchors and pose fusion provide per-frame estimates after initialization without revising past outputs. The method requires only a calibrated RGB sequence, a metric CAD model, and target image regions, which can be supplied by detection or segmentation. The pretrained pose and SLAM networks require no target-specific training or fine-tuning. We evaluate two versions, Mega-HAT and Pico-HAT, using MegaPose and PicoPose, on SPARK-2024, SwissCube and SHIRT, with YCB-Video assessing performance outside the space domain. Using one temporal configuration per method, the arithmetic means of the four dataset-wise comparisons show 9.4% lower mean pose error and 3.76 times the sustained input FPS for Mega-HAT relative to independent MegaPose, and 23.9% lower mean pose error and 2.42 times the FPS for Pico-HAT relative to independent PicoPose. Mega-HAT ablations on SPARK and an offline reference examine component contributions and the effect of revising past estimates. --- ### Evaluating In-Context Learning and Retrieval Strategies for Devanagari Post-OCR Correction Authors: Abhishek Bhandari, Gaurav Harit Published: 2026-09-18 Categories: cs.CL, cs.CV arXiv: https://arxiv.org/abs/2609.21595v1 In-context learning using Large Language Models (LLMs) offers a compelling path to training-free post-OCR correction, yet its effectiveness for Devanagari script remains entirely unexplored. We present the first systematic evaluation of LLMs (3B-32B) for post-OCR correction in Hindi and Marathi, comparing three in-context example retrieval strategies: domain-random selection, dense semantic retrieval, and our proposed CharBM25, which retrieves examples by character n-gram BM25 similarity over OCR inputs to target shared error patterns with the test sentence. Across a 20,000-sentence benchmark spanning five news domains, retrieval strategy is the decisive factor in correction quality: CharBM25 outperforms domain-random selection by 2.8-4.0pp absolute WER on Hindi and 2.9-3.8pp on Marathi, using character trigrams, which consistently outperform bigrams and unigrams. Scale dominates performance: Gemma-3-27B achieves WER reductions of 55.0% for Hindi and 33.3% for Marathi under CharBM25-5. Few-shot gains are capacity-gated: models below 8B do not reliably improve over the OCR baseline, and on Marathi the smallest models (3B) degrade more sentences than they improve. Marathi is persistently harder to correct than Hindi across all scales, reflecting its greater morphological complexity. These findings establish CharBM25 as an effective, GPU-free retrieval strategy that matches or exceeds dense retrieval at negligible computational cost, and show that combining it with a general-purpose LLM of 12B+ parameters delivers reliable, training-free Devanagari post-OCR correction without task-specific fine-tuning. Dataset: https://huggingface.co/datasets/AbhishekBhandari/Devanagari-OCR-ICL-Benchmark --- ### A benchmark dataset and baseline methods for four-dimensional STEM diffraction patterns Authors: Yuyan Guan, Haoran Zhang, Zian Mao, Antong Yang, Caifei Li, Jialong Wang, Chuying Ouyang, Hong Wang, Xiaoqin Zeng, Yujun Xie Published: 2026-09-18 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.21593v1 Four-dimensional scanning transmission electron microscopy (4D-STEM) records a two-dimensional diffraction pattern at each electron-probe position, yielding spatially resolved reciprocal-space information but large, heterogeneous data volumes. Here we describe 4D-ImageNet, a collection of 174,000 diffraction patterns comprising 145,000 experimental patterns selected from 29 acquisitions and 29,000 multislice simulations. The experimental data cover acquisition-level labels for Ag, Au, mixed Au-Ag, CoO, Pd and ZnO specimens across multiple fields of view, scan dimensions, camera lengths and exposure times. Each acquisition contributes 5,000 quality-ranked patterns with source scan coordinates and acquisition metadata. A set-prediction detector provides model-derived pseudo-labels for the direct-beam position and Bragg-disk centres, with a confidence score for each disk. The simulation data cover 13 crystal structures and include Euler rotations, reciprocal-space sampling and approximate low-index beam directions. A grouped mixed-domain masked-reconstruction benchmark is provided to assess leakage-resistant loading and evaluation across experimental and simulated data. The dataset is intended for representation learning, disk detection, diffraction-pattern retrieval, orientation analysis and simulation-to-experiment studies. --- ### Periodic Neural Mapping for Unsteady Rotor-Blade Pressure and Aeroelastic Load Prediction Authors: Lionel Salesses, Joachim Dominique, Tariq Benamara, Théo Flament, Franck Mastrippolito Published: 2026-09-18 Categories: physics.flu-dyn, cs.LG arXiv: https://arxiv.org/abs/2609.21590v1 Accurate prediction of unsteady aerodynamic loads remains a major challenge in turbomachinery design. High-fidelity Computational Fluid Dynamics (CFD) simulations are expensive, while aeroelastic Quantities of Interest (QoI) depend sensitively on the temporal evolution of the pressure field. This work introduces periodic Fourier Neural Mapping (p-FNM), a neural-operator framework for predicting unsteady pressure distributions on turbine rotor blades simulated using the chorochronic numerical hypothesis. The architecture embeds temporal periodicity into the model and learns a continuous mapping from operating conditions and time to pressure fields. Unlike sequential latent-space approaches, p-FNM predicts pressure fields independently at any time, avoiding error accumulation while preserving temporal continuity. The model is evaluated on a database of unsteady rotor-blade simulations and compared with a reduced-order baseline based on a variational autoencoder and recurrent neural network, refered as the Temporal Prediction Model (TPM). Performance is assessed for pressure fields and Generalized Aerodynamic Forces (GAFs), the primary aeroelastic QoI. Across all training datasets, p-FNM consistently outperforms TPM. On the largest dataset, p-FNM achieves a pressure-field mean absolute percentage error of 0.46% and a GAF-magnitude prediction error of 4.42%, corresponding to improvements of 60.7% and 77.6%, respectively. The minimum weighted phase error reaches 0.060 rad, demonstrating accurate preservation of the temporal characteristics of the aerodynamic response. The results show that GAF prediction is more challenging than pressure-field prediction and that temporal coherence is critical for accurately predicting spectral aerodynamic quantities. These findings demonstrate the potential of periodic neural operators for reduced-order modeling and aeroelastic analysis in turbomachinery. --- ### Predictive Suppression Layers for Communication-Efficient Spiking Neural Networks Authors: Aidin Attar, Michele Rossi Published: 2026-09-18 Categories: cs.NE, cs.LG, eess.SP arXiv: https://arxiv.org/abs/2609.21583v1 Feedforward Spiking Neural Networks (SNNs) typically propagate every generated spike indiscriminately, disregarding whether the information is redundant from an information-theoretic perspective. This lack of selectivity induces high redundancy in inter-layer communication, creating an expensive overhead, e.g., in scenarios involving many-core neuromorphic hardware or communication-dominated Internet-of-Things (IoT) where features are transmitted wirelessly. To address this challenge, we trade localized processing for leaner network channels by introducing a minimal predictive coding framework for SNNs. We propose two layer variants sharing a predictor block: error units, which transmit signed spiking residuals, and predictive suppression, which uses residual magnitude to dynamically gate and forward only unpredictable, "surprising" activity. Evaluated on the N-MNIST and Spiking Heidelberg Digits (SHD) datasets using diagnostic metrics that decouple local processing from cross-layer communication, our new predictive coding layers achieve significant communication savings. Numerical results reveal a three-fold reduction in communicated activity, while increasing the task accuracy for both datasets. The latter finding is notable, and suggests that predictive coding layers not only minimize communication overhead, but also produce output feature vectors with a higher representation power. --- ### GestureFAR: Streaming Co-Speech Gesture Generation with Flow Autoregression Authors: Pinxin Liu, Haiyang Liu, Jiahao Luo, Junhua Huang, Chunhao Zou, Luchuan Song Published: 2026-09-18 Categories: cs.CV, cs.GR, cs.HC arXiv: https://arxiv.org/abs/2609.21576v1 Generating natural co-speech gestures from streaming speech is essential for embodied conversational agents, where motion must be produced while a user is still speaking. Recent streaming gesture systems make online generation possible by autoregressing over discrete motion tokens, but this design compresses high-dimensional continuous motion into finite codebooks and can limit the realism and diversity of generated gestures. To preserve both causality and continuous expressiveness, we propose \textbf{GestureFAR}, a flow-autoregressive framework for streaming co-speech gesture generation. First, GestureFAR autoregresses over causal continuous motion latents, using a transformer to model streaming audio-motion context and a per-token flow-matching head to sample the next latent from a continuous distribution. Second, we introduce a head-only flow distillation strategy that freezes the causal backbone and distills the multi-step per-token flow head into a single network evaluation using consistency and distribution-matching objectives. This keeps the model token-causal while removing the main latency bottleneck for live interaction. Experiments on BEAT2 show that GestureFAR significantly improves the quality--latency trade-off among streaming-capable methods, preserving strong gesture quality while enabling real-time token-causal generation. Project Page: https://andypinxinliu.github.io/GestureFAR --- ### Micro-Collaborative Poisoning: A Distributed Attack on RAG Systems Authors: Pedro Pereira, Eva Maia, Isabel Praça Published: 2026-09-18 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.21573v1 Retrieval-Augmented Generation (RAG) improves large language models by grounding outputs in external knowledge sources, but this dependency also creates a surface for poisoning attacks. This paper introduces Micro-Collaborative Poisoning, a distributed attack in which a false target claim is divided across multiple locally plausible documents instead of being concentrated in a single malicious passage. We evaluate the attack across 108 RAG configurations by varying dataset, retriever architecture, retrieval depth, database composition, number of poisoned databases, and generator model. The results indicate that Micro-Collaborative Poisoning is not driven by a single dominant poisoned passage, but by the accumulation of weak adversarial signals across retrieved sources. Increasing top-$k$ and poisoning multiple databases make it more likely that these signals will appear together in the retrieved context, while clean database diversity and stronger retrievers can reduce their influence. The document-level poisoning visibility analysis further shows that this threat is difficult to expose through isolated document inspection, since Micro-Collaborative Poisoning achieves downstream influence while leaving a weaker explicit poisoning signature than direct poisoning. --- ### CityLearn v3: A Configurable Simulation and Evaluation Framework for Realistic Control Studies of Renewable Energy Communities Authors: Tiago Fonseca, Luis Lino Ferreira, Armando Sousa, Ava Mohammadi, Zoltan Nagy Published: 2026-09-18 Categories: cs.MA, cs.AI arXiv: https://arxiv.org/abs/2609.21570v1 Renewable energy communities (RECs) coordinate buildings, photovoltaic generation, batteries, electric vehicles and flexible loads. Controller studies often simplify changing participation, equipment availability, service deadlines and data quality, so lower cost or peak demand can conceal missed services or infeasible power requests. This paper presents CityLearn v3, a configurable simulation and evaluation framework for REC control studies under these conditions. It represents changing members and assets, flexible-load deadlines, demand-response requests, local energy sharing, and data or equipment failures within one simulation environment. Building and phase power limits constrain controllable requests, while a declared timestep preserves consistent power-to-energy accounting. The framework records controller inputs and distinguishes requested actions from those applied to the simulated equipment. Reference controllers, service- and constraint-aware performance indicators, and trajectory exports support comparisons within and across communities. Software checks and application examples examine service delivery, electrical constraints, settlement and changing scenarios; a synthetic high-frequency trace replay illustrates how aggregation can conceal short peaks without changing annual energy. Together, these records allow aggregate performance to be interpreted alongside service failures, action reductions and participant-level outcomes. --- ### Weighted Quantum Signal Processing: Low-Depth Polynomial Approximation with Applications to Kolmogorov-Arnold Networks Authors: Rohit Sarma Sarkar, Rupayan Bhattacharjee, Elias F. Combarro, Michele Grossi, Lirandë Pira, Carmen G. Almudéver, Sergi Abadal, Eduard Alarcon Published: 2026-09-18 Categories: quant-ph, cs.CC, cs.LG arXiv: https://arxiv.org/abs/2609.21567v2 Quantum Signal Processing is a powerful quantum framework for generating and approximating univariate polynomials. However, QSP is often limited by circuit-depth bottlenecks and parity constraints on the class of realizable polynomials. In this work, we introduce Weighted Quantum Signal Processing, an extension of QSP in which a weight function is assigned to the central rotation operator. This formulation provides a deeper understanding of QSP, which emerges as the special case of WQSP with unit weights. The choice of weights determines the structure and expressive capabilities of WQSP circuits. When the weights are natural numbers greater than one, WQSP reduces to a pruned version of QSP, revealing parameter redundancies in the standard framework. Through appropriate selection of integer weights, WQSP achieves linear-to-exponential reductions in the number of parameters required to realize arbitrary bounded univariate polynomials while preserving approximation quality. For generic weights, we establish corresponding approximation error bounds and show that, in many cases, the approximation is exact. We analyze WQSP from both a deterministic perspective, where polynomial generation is formulated as the solution of a linear system, and a quantum machine learning perspective, where WQSP serves as a structured and expressive quantum learning model. We further employ this learning framework to parameterize learnable activation functions in Kolmogorov--Arnold Networks for multivariate function approximation. Our results show that WQSP provides a compact, flexible, and theoretically grounded framework for realizing arbitrary univariate polynomials while requiring significantly fewer trainable parameters than conventional QSP. This yields expressive and parameter-efficient neural architectures, highlighting the potential of WQSP as a scalable primitive for quantum-enhanced machine learning. --- ### GameLogicBench: Evaluating Coding Agents on Runtime Game Logic with Tick-Level State Assertions Authors: Xinyu Che, Yunfei Ge, Shihao Li, Yanchen Liu, Hang Yan, Xinping Lei, Yanghai Wang, Zixuan Dong, Yifan Yao, Qianqian Xie, Letian Zhu, Jiaheng Liu Published: 2026-09-18 Categories: cs.SE, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.21562v2 Coding agents can modify and test code across large software projects. Game development is a domain where agents must implement gameplay rules. A game can end in a valid state even after violating its rules during the run. Current game-development benchmarks replay fixed examples, score videos, or ask another model to judge the result. However, no existing benchmark checks game rules throughout execution across varied evaluator-selected scenarios while ensuring exactly reproducible verdicts. We introduce GameLogicBench, a benchmark of 72 gameplay-logic tasks in Godot projects. An automated evaluator checks each game's rules at every simulation tick. Across 403 hand-designed scenarios, seeded parameter variations produce 1,451 test cases. To ensure that the evaluator measures behavior rather than implementation choice, it must accept different correct implementations for each task while rejecting mutants, implementations with one required capability removed. The tasks span isolated mechanics, multi-system interactions, and repository-scale features. Across 20 combinations of language models and scaffolds, the best observed run solves 52.78% of tasks. Under Claude Code, all twelve models solve fewer tasks as task scope expands from isolated mechanics, through interacting systems, to repository-scale features. Agents inspect code more often and make more tool calls on repository-scale tasks than on isolated-mechanic tasks. Most unsuccessful submissions are runnable, but implement some required game behavior incorrectly. We compared versions of our benchmark evaluator built with and without validation using mutants. Without this validation, incorrect agent submissions passed. A separate analysis finds agents copying code from public repositories when network access is open. Reliable evaluation thus depends both on what the tests reject and on what external code agents can access. --- ### On Repulsive and Attractive Teachers: Separating Correctness from Behavior in Self-Distillation Authors: Anton Baumann, Akmal Ashirmatov, Leo Schmidt-Traub, Frederike Lübeck, Jonas Hübotter, Thomas Kleine Buening, Andreas Krause Published: 2026-09-18 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.21561v1 On-policy self-distillation provides dense, token-level supervision by conditioning a model on privileged information and distilling the resulting teacher distribution back into the model. However, privileged information can change not only what the teacher knows, but also how it behaves, entangling correctness-relevant learning signals with unintended behavioral shifts. We study this effect in reasoning tasks by contrasting attractive self-distillation, which moves the model toward a privileged teacher, with repulsive self-distillation, which moves it away from a privileged teacher. We find that both objectives can induce strong and opposing behavioral shifts: attraction suppresses exploratory reasoning and promotes shorter, more confident responses, whereas repulsion increases response length, can trigger unintended switches into a model's latent thinking mode, and ultimately becomes unstable. Motivated by these observations, we study contrastive self-distillation, which combines attraction toward a correct-solution-conditioned teacher with repulsion from an incorrect-solution-conditioned teacher. In contrast to prior work that combines such distillation signals with a GRPO objective, we isolate the self-distillation objective and study its behavior on its own. We find that the shared behavioral shifts of the two teachers largely cancel, leaving a token-level signal that more directly reflects correctness. Across non-thinking, instruct-only, and already-thinking models, this contrastive objective improves reasoning performance while maintaining stable response lengths. --- ### MIRAGE: Multi-Perspective Creative Language Model Reasoning with Reinforcement Learning Guidance Authors: Arash Lagzian, Srinivas Anumasa, Dianbo Liu Published: 2026-09-18 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.21554v1 Recent advances in Large Language Models (LLMs) have revolutionized artificial intelligence and how human interact with AIs. Despite impressive advancements, LLMs struggle with complex mathematical, scientific, and logical tasks. Inspired by human cognitive flexibility - our ability to dynamically switch mental perspectives - we propose MIRAGE (Multi-perspective Inference-time Reasoning via Agent-Guided Exploration), a novel inference-time creative thinking framework. MIRAGE includes a Selector that prioritizes effective conceptual perspectives (e.g., algebraic, probabilistic) and a Reasoner that sequentially solves tasks until a confident solution emerges, otherwise aggregating multiple perspectives. Tested on GSM8K, MATH500, MMLU-Pro, and Game-of-24 benchmarks, MIRAGE consistently outperforms methods like Chain-of-Thought and diverse prompting ensembles, significantly boosting accuracy with minimal inference overhead, providing a scalable solution for practical applications. --- ### OneBid: A Unified Auto-Bidding Foundation Model for Diverse oCPX Advertising Scenarios Authors: Yewen Li, Peng Jiang, Yitian Li, Pengfei Lv, Xialong Liu, Peng Jiang, Qingpeng Cai Published: 2026-09-18 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.21550v1 Auto-bidding is central to computational advertising, where strategies must maximize advertisers' conversion value under economic constraints. It has evolved from rule-based controllers to reinforcement learning and generative methods such as Decision Transformer (DT). Yet these methods increasingly mismatch the prevailing optimized cost-per-X (oCPX) paradigm, which spans heterogeneous scenarios (e.g., registration, purchase), each served by a separate model, leading to fragmented pipelines and underexploring cross-scenario modeling. Inspired by foundation models like LLMs, unifying these oCPX scenarios into one model raises three challenges: multi-objective control, scalable capacity under strict latency, and safe offline policy improvement. We present OneBid, a unified auto-bidding foundation model that learns a reusable backbone from heterogeneous oCPX logs and adapts it to scenario-specific deployments via offline post-training. Building on DT, OneBid extends single Return-to-Go conditioning to two atomic signals, Return-to-Go for conversion value and Cost-to-Go for cost ratio, plus value-aware regularization on next-action prediction. To absorb distributional heterogeneity, we design a sequence-level Mixture-of-Experts architecture, where shared experts encode cross-scenario knowledge and sparsely-routed experts capture scenario-specific patterns at low latency, yielding consistent scaling with model size and data. During post-training, we align the backbone with scenario preferences via Critic-guided Relative Offline Policy optimization (CROP): a learned critic scores candidate actions group-relatively, avoiding the unsafe online exploration of GRPO-style fine-tuning while constraining policy shift to reduce OOD risk. Validated via online A/B tests and fully deployed at Kuaishou, OneBid delivers an overall +2.2% ADVV gain on oCPX Ads, peaking at +13.1% in the ROAS scenario. --- ### Dual-Interest Sequential Product Recommendation With Multi-Granular SSM Authors: Shuiying Liao, P. Y. Mok Published: 2026-09-18 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.21548v1 Sequential recommendation aims to predict the next item a user will interact with based on their historical behavior. Advances in Transformers have significantly improved sequential recommendation but are still limited by cost efficiency. Although State Space Models (SSMs) have recently enabled efficient long-range modeling, most existing methods encode each item with a single static contextual role, overlooking the phenomenon of item polysemy. In fact, the same item often plays different semantic roles depending on user context, and existing methods are limited in capturing dynamic behavior across different temporal granularities. In this work, we propose DSRec, a novel dual-interest cross-SSM model that explicitly disentangles item roles across long-term and short-term semantic context. Sequential items are encoded into long-term interest embeddings that capture stable preferences via historical aggregation, and a short-term interest branch that emphasizes local session intent modulated by inter-click time intervals. These interest embeddings are processed through distinct SSM encoders: a full-sequence Mamba for long-term modeling, and a time-modulated SSM that dynamically adjusts state evolution based on temporal gaps. To enable effective cross-granularity alignment, we adopt a residual cross-fusion mechanism that exchanges contextual information between the two branches while preserving semantic independence. Experiments on public benchmarks demonstrate that DSRec outperforms other state-of-the-art methods. --- ### From Retrieval to Recognition:How Vision--Language Models Become OCR Specialists Authors: Yuanxiang Huangfu, Hanmeng Zhong, Linqing Chen, Jeffrey Tiong Jee Hui Published: 2026-09-18 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.21543v1 Does a general vision--language model acquire specialized OCR ability by developing a new reading circuit or by reusing an existing mechanism? We address this question in the setting of full-sequence OCR, rather than local-answer retrieval. Using an evidence-grounded protocol with held-out causal interventions, we identify sparse and stable OCR-head sets in GLM-OCR, MinerU2.5, and PaddleOCR-VL-1.6. We then investigate the mechanistic origin of these OCR heads by comparing them with independently identified textual retrieval/copy heads in general VLMs. Across two general VLMs, visual OCR heads strongly overlap independently identified textual retrieval/copy heads, yielding untuned top-20 intersections of 73.3% and all-head Spearman correlations of 0.677-0.886. The overlap and causal interventions suggest that full-sequence OCR operates as dense sequential multimodal copy-and-paste, repeatedly retrieving visual evidence and routing it to the current output position. Finally, we examine how this shared circuit changes as a general VLM becomes an OCR specialist. Matched base-to-specialized comparisons show that OCR specialization largely preserves head identity, retaining 17-20 of the top 20 heads per task with all-head rank correlations of 0.874-0.942, while redistributing their functional and causal strengths. --- ### Purification and Regulation: Comorbidity-Aware Multi-Label Few-Shot Learning for Medical Image Classification Authors: Ying-Chih Lin, Po-Chih Kuo, Yong-Sheng Chen Published: 2026-09-18 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.21541v1 Multi-label few-shot learning (MLFSL) remains a significant challenge in medical image analysis (MIA). Current metric-based meta-learning methods face two critical limitations in MIA. First, conventional prototype generation often entangles irrelevant disease information, leading to contaminated prototypes and degraded performance. Second, prior studies typically enforce inter-class separability in embedding space, largely neglecting the inherent correlations among diseases. To overcome these challenges, we propose Prototype Purification and Regulation (PPR), a novel MLFSL framework for MIA. PPR first performs prototype purification by leveraging sample-level comorbidity scores to emphasize disease-specific features, producing purified prototypes that better characterize each disease. Building upon these purified prototypes, PPR further addresses the underexplored problem of inter-class prototype distance in MIA by incorporating disease-level comorbidity statistics to adaptively regulate inter-class similarity, forming a comorbidity-aware embedding space. Overall, PPR sequentially enables the model to capture pure disease features and inter-class relationships for reliable MLFSL in MIA. Extensive experiments across four chest X-ray benchmark datasets, including cross-domain evaluation, show that PPR consistently outperforms state-of-the-art methods, significantly improving disease detection while demonstrating robust generalization and clinical applicability. --- ### MACE: Memory-Agent Co-Evolution with Adaptive Memory Graphs for Multi-Agent Systems Authors: Kairui Yang, Minghao An, Xunkai Li, Ziheng Yi, Zekai Chen, Guangyuan He, Rong-Hua Li Published: 2026-09-18 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.21533v1 LLM-based multi-agent systems generate collaboration traces that record how agents plan tasks, verify intermediate results, and repair failures. Reusing these procedures requires preserving an action's prerequisites and the outputs needed by subsequent agents. Our empirical studies show that grouping these dependencies into functional memory units improves their retention, while connecting units increases retrieval of the units and links jointly required by a task. The preferred combination of units also changes between instructions and checklists, even when each combination's content is fixed across formats. Updating choices from the outcomes of each combination and format pairing outperforms scoring combinations and formats separately. These findings motivate MACE, a memory-agent co-evolution framework that adapts memory organization and agent memory use through execution feedback. Its MemGoG structure represents functional units as subgraphs of related conditions, actions, and outputs, connecting them through support, conflict, and repair relations. MACE Loop selects task-relevant units and relations within a memory budget and provides each agent with instructions or checklists for its current operation. It records the selected units, presentation formats, agent outputs, and task outcomes to update unit scores and relations for retrieval and inform subsequent presentation choices. Across eight benchmarks, MACE outperforms ten baselines with an average score of 81.11%, compared with 78.97% for the strongest baseline, SAGE. --- ### OpenMAS-GCom. A Diagnostic Benchmark for Graph-enhanced Multi-Agent Systems Authors: Kairui Yang, Xunkai Li, Kaixiang Zhang, Minghao An, Zekai Chen, Yuxuan Ba, Rong-Hua Li Published: 2026-09-18 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.21527v1 Graph-enhanced multi-agent systems (G-MAS) coordinate large language model agents through communication graphs and role assignments, which determine how agents exchange information and divide responsibilities. However, final-score comparisons across systems combine differences in models, communication patterns, roles, and computation costs, making performance differences difficult to attribute to specific communication structures, role assignments, and information flows. To address this evaluation attribution problem, we introduce OpenMAS-GCom, a benchmark for diagnosing how these components affect G-MAS performance through controlled interventions. We represent systems through collaboration units, communication links, shared intermediate information, and execution rules. OpenMAS-GCom compares original systems with versions modified by changing one component while keeping tasks, models, prompts, and budget limits fixed. We rewire communication edges, remove specialist or critic agents, replace intermediate messages with incorrect content, and disable workers during execution. The benchmark evaluates 17 single-agent, ordinary multi-agent, and graph-enhanced configurations on 29 datasets across six domains. We add 400 G-MAS-Complex tasks requiring agents to combine information from multiple documents, resolve conflicting records, and return specified values with source identifiers. Experiments show larger mean losses after specialist removal than after critic removal, different performance degradation under incorrect messages and worker failures despite similar original scores, and different configurations achieving the highest accuracy and accuracy per token on G-MAS-Complex. --- ### IncentRL: The Trade-Off Between Preference Guidance and Task Performance Authors: Xuening Wu, Yanlan Kang, Shenqin Yin Published: 2026-09-18 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.21525v1 Preference-based reward shaping can guide reinforcement learning, but adding preference signals to the reward may unintentionally change the task being optimized. We address this problem with IncentRL, a framework that introduces preference guidance while explicitly characterizing its effect on external-task performance. IncentRL adds a Kullback--Leibler (KL) penalty between a specified outcome distribution and a preferred distribution. For finite discounted Markov decision processes with bounded shaping costs, we derive an external-value perturbation bound, establish a sufficient strict-action-gap condition for preserving the original optimal policy, and characterize the large-weight regime through discounted cumulative preference cost. Exact examples clarify the limits of these guarantees, including tied optima and support mismatch. We study a practical implementation using a hand-designed, distance-based outcome proxy, a fixed preference distribution, and score-weighted coefficient search. On MiniGrid DoorKey-8x8, the reported three-seed mean success rate after two million training steps reaches 98\% with coefficient 0.01, compared with 90.5\% for the reported zero-coefficient baseline, while the search progressively shifts toward smaller coefficients. Together, these results provide a principled view of the central trade-off in preference-based RL: using additional guidance to improve learning without excessively distorting the original task objective. The current experiments remain descriptive and do not yet isolate KL shaping from simpler alternatives. --- ### What Must Survive? Exact Task-Information--State Frontiers Under Partial Task Revelation Authors: Ronald Katende Published: 2026-09-18 Categories: cs.LG, cs.IT arXiv: https://arxiv.org/abs/2609.21523v2 A state may need to be compressed before its exact downstream task is known. For a finite family of linear task operators $\{T_u\}_{u\in\U}$, we characterize exactly how much state must survive when only one of $K$ coarse task messages is available at compression time: $$p^*(K)= \min_{\substack{\Pcal\text{ partition of }\U\\|\Pcal|\le K}} \max_{C\in\Pcal}\rank(T_C),$$ where $T_C$ stacks the tasks that remain unresolved within cell $C$. Thus task information reduces retained state precisely by separating tasks whose joint task-visible subspace is expensive. We derive the dual bit frontier, an irreducible common-core floor, and a sharp singular-value characterization for nonzero tolerance. For overlapping task families, a cumulative overlap deficit places the exact frontier within an additive $Δ$ of weighted load balancing. Optimal partitioning is strongly NP-hard, yet direct-sum families admit classical approximation guarantees and a PTAS. We stress-test the theory on learned task operators from CIFAR-10, Fashion-MNIST, SVHN, and Burgers observables. For Burgers with $K=2$ and $\eps=0.35$, a constructive partition attains $p=4$, while balanced random partitions require median $p=6$, with held-out worst standardized RMSE $0.0368$. The results distinguish uniform task sufficiency from distributional and decision-level performance. --- ### Refine Then Fusion: Training-Free 3D Point Cloud Adaptation with Priority Refinement and Multi-Modal Knowledge Fusion Authors: Hang Cheng, Yan Chen, Mingyu Fan, Long Zeng Published: 2026-09-18 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.21522v1 Recent pre-trained foundation models provide rich multi-modal priors for downstream 3D vision tasks. However, the effectiveness of these representations in few-shot scenarios is limited by two fundamental challenges: High-dimensional features often contain substantial channel redundancy and task-irrelevant noise, while the reliability of different modalities varies across samples. Consequently, direct aggregation of heterogeneous representations overlooks sample-dependent modality reliability and may obscure the discriminative cues essential. To address these limitations, we propose Refine Then Fusion(RTF), a training-free framework for few-shot 3D recognition. RTF first identifies discriminative feature channels by jointly modeling inter-class similarity and intra-class stability, thereby decoupling domain-specific knowledge refinement from the cached representations of pre-trained models. It then introduces a reliability-aware fusion mechanism that estimates sample-wise modality reliability from the distribution shifts induced by feature refinement, enabling adaptive aggregation of multi-modal representations. Furthermore, RTF constructs a memory cache that integrates instance-level support features with class-level prototypes to infer query labels. Extensive experiments on five benchmarks demonstrate that RTF consistently outperforms single-modal baselines, partial-fusion variants, and existing lightweight adaptation methods, achieving state-of-the-art few-shot 3D recognition performance without gradient optimization, additional training data, auxiliary training, or parameter updates. --- ### VidOmni-Bench: A Benchmark for Fine-Grained Video Understanding via Spatio-Temporal Event Verification across Complexity and Duration Authors: Changbeen Kim, Junwon Chang, Kipyo Kim, Risa Shinoda, Kuniaki Saito, Donghyun Kim Published: 2026-09-18 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.21521v1 While Video Large Language Models (Video-LLMs) have recently demonstrated strong performance, reliably evaluating their fine-grained video understanding remains challenging. Existing benchmarks often rely on question answering or ground-truth caption matching, where models may succeed through superficial cues and incomplete annotations. To this end, we introduce VidOmni-Bench, a benchmark that requires models to verify whether each event in dense video captions is supported by the video. VidOmni-Bench consists of 500 videos spanning five complexity types and diverse durations from 4 seconds to 90 minutes. After collecting videos along these axes, we use diverse Video-LLMs to generate dense captions and obtain human-verified sentence-level labels, where sentences containing incorrect events serve as hard negatives for evaluation. Our experiments on VidOmni-Bench reveal three key findings: (i) Video-LLMs frequently generate hallucinated descriptions in dense video captioning; (ii) they also struggle as verifiers, failing to reliably detect plausible but incorrect event descriptions; and (iii) model weaknesses vary across video complexity and duration, revealing diverse, model-specific bottlenecks in current Video-LLMs. --- ### Learning-to-Optimize as the Missing Architectural Layer of AI-Native Networks Authors: Giambattista Amati, Federica Mangiatordi, Pierpaolo Salvo, Emiliano Pallotti, Simone Angelini Published: 2026-09-18 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.21519v1 Artificial Intelligence (AI) is becoming a fundamental design principle of future AI-native communication networks, enabling autonomous resource management, adaptive control, and zero-touch network operation. While current AI-native architectures increasingly embed intelligence across network functions, they provide little guidance on how optimisation knowledge should be systematically generated, transferred, and exploited by AI models. This paper argues that the Learning-to-Optimize (L2O) represents the missing architectural layer between optimisation and AI-native intelligence. Rather than viewing optimisation merely as an online decision engine, the proposed paradigm redefines optimisation algorithms as offline knowledge generators that produce high-quality supervisory information for neural surrogate models. The resulting models inherit optimisation expertise while enabling low-latency runtime inference suitable for dynamic network environments. A generic four-stage L2O workflow is introduced, comprising optimisation, knowledge generation, surrogate learning, and runtime inference. Unlike existing Learning-to-Optimize approaches, which primarily focus on algorithm acceleration, the proposed framework establishes L2O as an architectural abstraction applicable across heterogeneous communication and computing systems. The proposed paradigm is illustrated by an NR-V2X relay-selection problem, in which optimisation-generated solutions from a Mixed-Integer Linear Programming (MILP) solver are used to train a Graph Neural Network that can reproduce near-optimal decisions in real time. The presented perspective positions Learning-to-Optimize as a key architectural enabler for future AI-native networks. --- ### 2D GauSS-MI: Efficient Active Scene Reconstruction with Balanced Visual and Geometric Quality Authors: Yuhan Xie, Jia Pan Published: 2026-09-18 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.21516v1 Active reconstruction requires efficient active view selection to achieve high-quality reconstruction within limited onboard computational resources. Existing methods face challenges in adequately balancing visual and geometric quality with the computational efficiency required for real-time operation. In this work, we present an active reconstruction framework based on 2D Gaussian Splatting (2DGS). We develop an efficient online 2DGS mapping pipeline for incremental RGB-D observations and introduce a probabilistic reliability model that characterizes the view-dependent reconstruction quality of individual 2D Gaussian splats. Building on this model, we formulate 2D Gaussian Splatting Shannon Mutual Information (2D GauSS-MI), a mutual-information-based metric that exploits the explicit surface orientation of 2DGS to evaluate the expected information gain of candidate views. The proposed metric enables active view selection to account for both visual and geometric reconstruction quality. We evaluate the proposed system against three state-of-the-art baselines on eight Replica scenes. Experimental results demonstrate that our method achieves a favorable balance between visual and geometric reconstruction quality with substantially lower computational cost and competitive model storage. --- ### ServeGuard: Verifiable, Bounded-Residual Confinement of Operator-Invisible Channels Without Revealing the Certified Read Factor Authors: Dominik Dahlem, Rui Vieira Published: 2026-09-18 Categories: cs.CR, cs.LG arXiv: https://arxiv.org/abs/2609.21515v1 Third-party adapters for open-weight language models ship as opaque weight matrices; a recipient cannot check whether an adapter hides a backdoor without trusting the publisher or inspecting the weights, the publisher's core asset. For one important class (payloads placed where a safety monitor is structurally blind), detection is unsound as a defense: every detector that factors through the declared monitor is invariant on its blind subspace, and honest and backdoored adapters overlap on every blind-subspace statistic we evaluate, because benign adaptation uses that subspace too. Rather than detect this channel, we make it structurally \emph{absent} and prove that we did. The publisher builds the adapter to read the input only through directions the monitor covers and proves this in zero knowledge, revealing nothing about the read factor it certifies. The certificate is cheap because the expensive part, identifying the monitor's blind spot, is a deterministic function of the \emph{public} base model, so only one linear identity is proved; the served residual is the base model's own public floor, not a prover-chosen tolerance. The result is \emph{ServeGuard}, a supply-chain primitive: the publisher ships a \emph{proof-carrying adapter} whose proof lets a consumer or regulator verify, without the certified read factor and without trusting the publisher, that the adapter carries no hidden channel of this class relative to the declared monitor; an admission-time typing guard binds the guarantee to the adapter bytes admitted at serving time. Across eight checkpoints up to 7B from four families, the monitoring budget is architectural: the measured frontier saturates at the value-path rank on grouped-query checkpoints but not on multi-head ones. On a 0.5B model confinement is nearly free for benign adaptation, making monitor quality the security lever. --- ### 2nd Place Solution to the HANDS 2026 Workshop Challenge-Dexterous Grasp Motion Track: Single-Shot Trajectory Warping for Grasp Motion Generation Authors: Muneeb A. Khan, Woojin Kim, Shinwoo Kim, Muhammad Munsif, Binod Bhattarai, Seungryul Baek Published: 2026-09-18 Categories: cs.RO, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.21511v1 This report describes our 2nd place solution to the HANDS 2026 workshop challenge (Dexterous Grasp Motion track) in conjunction with ECCV 2026. In this challenge, we address grasp motion generation for the 12-DoF LinkerHand O6, aiming to produce physically plausible reach-and-lift trajectories for unseen objects from randomized initial hand poses in simulation. This task is particularly challenging because each grasp requires a per-step policy to make approximately $70$ twelve-dimensional decisions, with errors accumulating over time, while test objects and physical dynamics may differ from those encountered during training. To address these challenges, we propose editing a single successful GraspM3 demonstration instead of generating the motion step by step: a policy observes the object once and outputs a 12-D warp of the demonstration, which is then replayed open-loop. Moreover, we train the warp policy with one-step PPO over all $4{,}824$ training objects in parallel. As a result, our method achieved success rates of $94.61\%$ on the easy track, the highest of all submissions, and $57.18\%$ on the hard track of the private test set. --- ### The Communication Bottleneck: A Round-Trip Study of Tree-Structured Expression Serialization in Language Models Authors: Xavier Suau, Alex Ferrando de las Morenas, Luca Zappella, Samy Bengio Published: 2026-09-18 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.21509v2 When language models reason in chain-of-thought or exchange free-text intermediates, they serialize structured information into natural language. How much tree-structured compositional content survives this bottleneck? We propose a round-trip protocol that answers this question empirically for tree-structured expressions. A generator converts a procedurally generated arithmetic expression into a word problem, a separate extractor recovers the expression from the word problem alone, and symbolic equivalence provides an exact oracle. Evaluating all pairwise combinations of sixteen models yields a communication matrix whose marginals separate generation quality from extraction quality. Three main findings emerge. First, the channel is lossy and asymmetric: swapping which model generates and which extracts shifts accuracy by up to 60.4 points, and the best pair reaches 92.9% by combining different models on each end rather than the same model on both. Second, at least 73.6% of round-trip failures originate at generation, and difficulty is driven by tree structure (operator count, depth, right-branching) rather than model family. Third, the channel is trainable: ~3600 fine-tuning examples that share the evaluation's operators and tree shapes lift every open-weight model above untrained Gemini-3.1-Pro, an upper bound under matched semantics. A disjoint-domain regime with new operators and vocabulary also raises every open-weight model, confirming the gain is not an artifact of matched semantics, though a gap to the frontier remains. Together these results identify tree-structured expression serialization as a primary limiting factor when models communicate hierarchical structure through natural language. --- ### Adaptive World Memory 3D Foundation Model for Scalable 3D Mapping, Localization, and Rendering Authors: Tianchen Deng, Guole Shen, Yilin Shen, Wenhua Wu, Yilin Fang, Ziqi Ma, Tianjun Zhang, Shenghai Yuan, Wolfram Burgard, Hesheng Wang Published: 2026-09-18 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.21502v1 Recent 3D foundation models enable generalizable geometric reasoning from RGB images but remain limited in persistent memory, scalability, and renderable scene modeling. We present a memory-centric 3D foundation model for scalable robotic localization, reconstruction, and Gaussian rendering. Its core is an adaptive world memory mechanism that combines transformer-based gated updates with test-time temporal-spatial regulation. Learned gates control recurrent memory propagation, while temporal state evolution and spatial observation-state consistency regulate token-wise updates and forgetting over long image sequences. To support large-scale mapping, we organize memory into local submaps and integrate progressive mapping and tracking, loop closure, and SL(4)-based global refinement to maintain local accuracy and global consistency. A Gaussian reconstruction head decodes memory-enhanced features into renderable primitives, unifying camera pose estimation, dense point-cloud reconstruction, and photorealistic rendering within a single model. Experiments on public benchmarks and self-collected datasets from diverse robotic platforms demonstrate improved trajectory accuracy, reconstruction completeness, and rendering quality over existing 3D foundation reconstruction and SLAM baselines. These results support adaptive memory as a foundation for persistent robotic world modeling. The dataset and code will be made publicly available at \href{https://github.com/dtc111111/AWM-3DFM}{https://github.com/dtc111111/AWM-3DFM}. --- ### VoxelTTO: Voxel-Aligned Feed-Forward 3D Gaussian Splatting with Test-Time Optimization Authors: Yibin Zhao, Yihan Pan, Yangwen Li, Jun Nan, Jianjun Yi Published: 2026-09-18 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.21498v1 Recent feed-forward 3D Gaussian Splatting (3DGS) methods typically regress pixel-aligned Gaussian primitives, often causing excessive overlap and artifacts, while inaccuracies in predicted camera poses can lead to misalignment in novel-view synthesis (NVS). We present VoxelTTO, a feed-forward framework for reconstructing geometrically accurate 3DGS scenes from an arbitrary number of images and optional camera parameters. VoxelTTO aggregates dense image features into a global voxel representation and decodes Gaussians from voxel features, breaking the pixel-to-Gaussian correspondence. To exploit known camera parameters while keeping the pretrained visual foundation model (VFM) parameters frozen, we introduce test-time optimization (TTO) that adapts lightweight LoRA modules using pose supervision. We further replace vanilla 3DGS rasterization with stochastic solid volume rendering during training and inference, improving geometric fidelity. Training updates only the voxel-aligned Gaussian reconstruction modules, requiring 80 GPU hours. Experiments on Replica, Tanks and Temples, and DTU demonstrate improved RGB-D NVS and camera-pose estimation relative to prior methods. --- ### PolyBridgeBench: Benchmarking Multimodal LLMs for Physics-Grounded Bridge Design Authors: Zicheng Zhao, Dongyin Chen, Rui Xu, Yinghui Xu Published: 2026-09-18 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.21493v1 Multimodal large language models, or MLLMs, perform well at visual understanding and structured generation, yet these capabilities do not establish whether an engineering design will work when executed. Existing benchmarks assess spatial reasoning, structural validity, or physics-grounded construction, but they do not determine whether MLLMs can synthesize complete load-bearing structures and repair them after simulator execution exposes a failure. We introduce PolyBridgeBench, an executable benchmark for multimodal bridge design. A model receives a visual scene and structured engineering constraints and generates a complete node--member--material topology. Deterministic legality checks gate execution in a native dynamic physics simulation. Following an execution failure, the benchmark returns temporal visual evidence from the failed rollout and evaluates repair under a fixed interaction budget. Separate measurements of deterministic validity, dynamic functional success, and post-failure recovery identify the stage at which design fails. Experiments with six representative MLLMs across 189 levels expose a substantial gap between deterministic validity and dynamic success, pronounced sensitivity to material budgets, and limited post-failure recovery under the primary strict-budget setting. --- ### LogicTrack: Auditing Reasoning Trajectories of Large Language Models with Formal Logic Solvers Authors: Jingyu Hu, Shu Yang, Weiru Liu, Di Wang Published: 2026-09-18 Categories: cs.AI, cs.LO, cs.SC arXiv: https://arxiv.org/abs/2609.21492v1 Chain-of-Thought (CoT) reasoning has been shown to improve the performance of large language models (LLMs), yet existing optimization methods largely rely on outcome-based feedback, leaving the logical validity of intermediate reasoning steps largely unverified. To address the gap whereby LLMs arrive at correct final answers through logically flawed intermediate reasoning chains, we propose LogicTrack, a neuro-symbolic framework that audits reasoning trajectories by auto-formalizing each reasoning step into symbolic representations and verifying it with automated theorem provers. LogicTrack introduces Solver-Based Backtracking Reward (SBR), a step-wise scoring mechanism that quantifies logical soundness and guides backtracking tree search at inference time. We further extend LogicTrack to construct supervised fine-tuning (SFT) data with backtracking traces from its trajectories, enabling fine-tuned models to internalize step-wise auditing as an intrinsic capability. Extensive experiments across 8 reasoning benchmarks and 7 LLMs demonstrate that LogicTrack effectively improves both the verifiability of reasoning chains and final answer pass rate, thereby enhancing overall CoT quality and trustworthiness in high-stakes domains. --- ### Benchmarking Gender Bias in Machine Translation Evaluation Metrics across Occupations Authors: Orfeas Menis Mastromichalakis, Giorgos Filandrianos, Wafaa Mohammed, Giuseppe Attanasio, Chrysoula Zerva Published: 2026-09-18 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.21490v1 Gender bias remains a persistent concern in machine translation (MT), affecting both generated translations and their automatic evaluation. When a source text leaves a person's gender unspecified, translations may realize that person using masculine or feminine forms, and both MT systems and evaluation metrics may exhibit systematic preferences between these alternatives despite the source providing no basis for such a distinction. We study this behavior in the WMT 2026 Automated Translation Quality Evaluation Systems Shared Task using an occupation-balanced subset of GAMBIT+. We consider seven English-source language pairs, six from the original dataset, targeting Arabic, Czech, Greek, Icelandic, Russian, and Ukrainian, and extend the original resource with German. The subset contains 1,308 masculine/feminine translation pairs per target language, with three examples for each of the 436 ISCO-08 occupational groups. We evaluate shared-task submissions and baselines for score prediction and error annotation, examining the direction, magnitude, and frequency of gender-related differences. We find an overall tendency for masculine translations to receive higher scores, as well as differences per occupation following stereotypical gender representations, although the strength and consistency of this preference vary considerably across evaluators and languages. Our results show that gender bias remains present in MT evaluation, but that capturing its extent requires looking beyond a single aggregate measure to complementary dimensions of evaluator behavior. --- ### Driving on Registers, Reasoning on Risk: Risk-Aware Occupancy for Register-Based End-to-End Autonomous Driving Authors: Jiaxing Chen, Hengduo Zou, YuKai Qin, Yiren Zhao, Lidong Yu, Bolin Gao Published: 2026-09-18 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.21486v1 Multimodal trajectory prediction improves behavioral coverage in end-to-end autonomous driving, but existing methods remain limited by sparse scene representations. Incomplete evidence leads to low-quality candidate generation and unreliable ranking among geometrically similar trajectories. On a register-based baseline, bad and poor candidates constitute 19.74% of the candidate set, while the oracle-best candidate ranks only 33.9th on average. We propose RRDrive, which introduces risk-aware occupancy as a dense, temporally aligned, and trajectory-queryable representation. Its global structure guides high-quality multimodal generation, while candidate-conditioned risk queries support fine-grained selection. We further construct RiskOcc4D-NAVSIM with automatic risk annotations. RRDrive achieves a selected-trajectory PDMS of 0.951, representing a 1.5% relative improvement over the baseline (0.937), and improves the average candidate PDMS by 7.7%. In challenging scenes, it improves candidate PDMS by 30.2% and increases the Spearman correlation among good candidates by 0.41, from 0.26 to 0.67. To move beyond this oracle setting, we further develop an external RiskOcc predictor, a perception module that estimates risk-aware occupancy directly from sensor inputs. The competitive performance validates the representation's feasibility. --- ### HE-Guardrail: A Homomorphic Guardrail Against Jailbreak Attacks for Encrypted Large Language Model Inference Authors: Byeongseo Min, Yongwoo Lee, Young-Sik Kim, Yongjune Kim Published: 2026-09-18 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.21484v1 Homomorphic encryption (HE) has emerged as a promising approach to privacy-preserving machine learning (PPML), enabling computation directly over encrypted data. In HE-based PPML, a client submits an encrypted input to the server, which evaluates models such as large language models (LLMs) without access to the underlying plaintext. However, we identify a critical security vulnerability in this setting: HE-LLM inference is vulnerable to malicious clients that submit adversarial prompts, such as jailbreak attacks. The same confidentiality that protects benign clients also prevents the server from inspecting incoming prompts or generated responses, making adversarial attempts difficult to detect or block and potentially allowing successful attacks to remain entirely invisible to the server. To address this vulnerability, we propose HE-Guardrail, a framework that evaluates guardrail mechanisms entirely over encrypted data and homomorphically controls whether the target-model response is returned to the client. We instantiate HE-Guardrail with three representative guardrails - Llama Guard, JBShield, and GradSafe. Our results show that HE-Guardrail closely reproduces the decisions of the corresponding plaintext guardrails in the encrypted domain, with distinct security-efficiency-utility trade-offs. --- ### Adaptive Rollout Truncation Based on Epistemic Uncertainty for Efficient Offline World Model Training Authors: Nikodem Sebastian Zymla, Laurin Thiele, Johannes Pitz Published: 2026-09-18 Categories: cs.RO, cs.LG arXiv: https://arxiv.org/abs/2609.21482v1 Accurate neural world models are central to model-based robotics, where they enable robots to predict future states from previously observed trajectories. Multi-step autoregressive training improves long-horizon prediction, but fixed rollout horizons also increase computational cost and can amplify early training errors when the model is still inaccurate. Existing training schemes typically use the same rollout length throughout optimization, independent of the model's current predictive reliability. We propose an epistemic uncertainty-driven adaptive rollout strategy for offline world model training following an auto-curriculum training scheme. Instead of always unrolling to a fixed horizon, the model terminates autoregressive rollouts once epistemic uncertainty exceeds a threshold calibrated from a warm-up phase. We study two uncertainty estimators: a five-head ensemble with a shared recurrent backbone and Monte Carlo Dropout. A two-stage warm-up procedure stabilizes uncertainty estimates before we enable adaptive truncation. Experiments on ANYmal-D and ANT show that ensemble-based adaptive truncation matches or improves the prediction accuracy of fixed-horizon training and the RWM-U baseline while requiring substantially fewer cumulative rollout steps. Training a world model on ANYmal-D following the presented approach reaches comparable final performance with the baselines with roughly 72% less rollout computation. These results indicate that epistemic uncertainty is useful not only for downstream policy regularization, but also for making world model training itself more compute-efficient. ---