# Research Papers — 2026-09 · bundle page 231 of 266 > Full-text markdown of papers #11501–11550 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/231 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/230.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/232.md) ### Deep belief networks are exact Authors: Gleb Smirnov Published: 2026-09-04 Categories: cs.AI, cs.LG, math.PR arXiv: https://arxiv.org/abs/2609.05572v1 We prove that every strictly positive probability distribution on \(\{-1,1\}^n\) is represented exactly by a sigmoid belief network with finite parameters. This answers a question of Sutskever and Hinton. The proof upgrades their probability-sharing approximation to exact representation using Brouwer's fixed-point theorem. --- ### Grounded Skill Synthesis from Code at Scale for Agentic Intelligence Authors: Yongqi Tong, Pan Wang, Hang Wang, Jianshe Li, Xin Zhang, Jiang-Ming Yang, Wei Wu Published: 2026-09-04 Categories: cs.SE, cs.CL arXiv: https://arxiv.org/abs/2609.05571v1 Reusable skills give agents transferable procedural knowledge, making scalable acquisition essential for extending agents beyond prior experience. Existing methods face two limitations: trajectory-based synthesis requires interactions with specific environments, while document-derived skills may lack executable evidence and verification. Source code offers a complementary path: it requires no prior agent experience yet provides executable evidence for grounding abstractions. We present Code2Skill, a fully automated pipeline that transforms selected code units into implementation-anchored records of atomic operations, composite workflows, and recurring patterns, then verifies each record through source-body-blind reconstruction and source-aware comparison. Applied to 19,769 popular, actively maintained GitHub repositories, Code2Skill produces CodeSkillBank, a grounded bank of 1,006,822 accepted records with workflow, boundary, provenance, and source-evidence metadata. Across 72 protocol-matched evaluations covering nine model settings and eight benchmarks, models augmented with retrieved CodeSkillBank skills improve by 11.7% on average over matched baselines and outperform them in 57 cases. Under a unified downstream interface, Code2Skill also outperforms trajectory-derived skill banks on all seven shared benchmarks, showing that repository-derived skills can provide useful procedural knowledge before agents accumulate sufficient interaction experience. Skills synthesized from tested AI-generated code achieve a 93.50% pass rate, compared with 93.00% for human-written code, providing initial evidence that the pipeline can expand with the growing volume of AI-generated software. Overall, Code2Skill transforms procedural knowledge embedded in repositories into grounded, verifiable, and transferable agent skills. --- ### WorldSculpt: Generating Compositional Worlds from Grounded Videos Authors: Muyao Niu, Jixuan He, Ruihan Yu, Lian Fu, Yonghao Yu, Zheng-Hui Huang, Yifan Zhan, Fengbo Lan, Yongtao Ge, Yinqiang Zheng, Kaipeng Zhang, Zhixiang Wang Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05416v2 We study the problem of generating a compositional 3D representation of a cluttered scene containing hundreds of objects. The goal is to represent the scene as a collection of individual object meshes placed in a shared world frame, as required by downstream applications such as gaming, AR/VR, simulation, and robotics. This task is challenging in densely cluttered scenes, where objects heavily occlude one another and each view reveals only a fraction of their geometry. Geometry-based approaches typically reconstruct the scene as a single representation and leave incomplete geometry in occluded regions, while existing compositional methods with generative priors are largely limited to relatively simple scenes. We show that complex scenes with hundreds of objects can instead be generated compositionally by adapting a strong single-object 3D generative prior to multi-view observations. We instantiate this paradigm with Pixal3D, extending it with a multi-view conditioning pathway that grounds object generation in multiple posed observations. Although the model is finetuned entirely on single objects in canonical space, it generalizes to large scenes with severe occlusion without any scene-level training, demonstrating the feasibility and scalability of this paradigm. We further introduce UE-MeshyScene, a photorealistic benchmark of densely cluttered scenes with hundreds of objects, per-object annotations, and ground-truth meshes. Across single-object, controlled multi-object, and UE-MeshyScene evaluations, our method consistently outperforms prior approaches, with larger gains as scene complexity and occlusion increase. Finally, we demonstrate broader applicability by converting generated 3DGS worlds, such as Marble and HY-World 2.0, into compositional mesh scenes. --- ### WorldSculpt: Generating Compositional Worlds from Grounded Videos Authors: Muyao Niu, Jixuan He, Ruihan Yu, Lian Fu, Yonghao Yu, Zheng-Hui Huang, Yifan Zhan, Fengbo Lan, Yongtao Ge, Yinqiang Zheng, Kaipeng Zhang, Zhixiang Wang Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05416v1 We study the problem of generating a compositional 3D representation of a cluttered scene containing hundreds of objects. The goal is to represent the scene as a collection of individual object meshes placed in a shared world frame, as required by downstream applications such as gaming, AR/VR, simulation, and robotics. This task is challenging in densely cluttered scenes, where objects heavily occlude one another and each view reveals only a fraction of their geometry. Geometry-based approaches typically reconstruct the scene as a single representation and leave incomplete geometry in occluded regions, while existing compositional methods with generative priors are largely limited to relatively simple scenes. We show that complex scenes with hundreds of objects can instead be generated compositionally by adapting a strong single-object 3D generative prior to multi-view observations. We instantiate this paradigm with Pixal3D, extending it with a multi-view conditioning pathway that grounds object generation in multiple posed observations. Although the model is finetuned entirely on single objects in canonical space, it generalizes to large scenes with severe occlusion without any scene-level training, demonstrating the feasibility and scalability of this paradigm. We further introduce UE-MeshyScene, a photorealistic benchmark of densely cluttered scenes with hundreds of objects, per-object annotations, and ground-truth meshes. Across single-object, controlled multi-object, and UE-MeshyScene evaluations, our method consistently outperforms prior approaches, with larger gains as scene complexity and occlusion increase. Finally, we demonstrate broader applicability by converting generated 3DGS worlds, such as Marble and HY-World 2.0, into compositional mesh scenes. --- ### UniMate: One Unified Model to Animate Diverse Skeletons Authors: Linzhan Mou, Jiahui Lei, Zhiyang Dou, Chenyue Cai, Chaoyue Song, Adam Finkelstein, Szymon Rusinkiewicz Published: 2026-09-04 Categories: cs.CV, cs.GR, cs.LG arXiv: https://arxiv.org/abs/2609.05415v1 Recent advances in automatic rigging now deliver animation-ready 3D assets at scale, yet generating the motion to drive them remains a bottleneck. Existing learned animators are topology-constrained: they rely on category-specific templates or require per-skeleton fine-tuning and reference motions at inference. We present UniMate, a unified foundation model that synthesizes articulated motion for arbitrary skeletons from a rigged 3D asset and a text prompt, with no test-time optimization or per-skeleton retraining. UniMate introduces a topology-aware diffusion transformer, which integrates skeletal topology into attention via three mechanisms: (1) a graph-aware attention bias from pairwise joint relations and geodesic distances; (2) a spectral rotary position embedding generalizing RoPE to arbitrary kinematic trees via the graph Laplacian; and (3) a global topological conditioner attention-pooled from the rest-pose skeleton. We also curate UniML3D, 13,006 motion sequences spanning bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid objects with unified canonicalization and text pairing. Trained on this dataset, UniMate outperforms state-of-the-art baselines in quality, generalization, and efficiency, and supports zero-shot cross-topology transfer, in-betweening, expansion, and text-guided editing. Our project page is available at https://linzhanmou.com/unimate/. --- ### WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data Authors: Ji Soo Lee, Xilun Chen, Pierce Chuang, Ashish Shenoy, Jason Wei, Dohwan Ko, Hyunwoo J. Kim, Benoit Corda Published: 2026-09-04 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.05405v1 Recent advances in wearable sensing enable continuous monitoring of physiological and behavioral signals, yet existing benchmarks rarely evaluate whether AI systems can reason over a real user's longitudinal wearable record. We introduce WearableQA, a benchmark comprising 4,084 10-option multiple-choice questions constructed from the wearable time series, blood biomarkers, and demographics of 200 real users, each with up to 500 days of daily measurements. WearableQA preserves authentic wearable distributions that include device noise and inter-individual variability. To evaluate distinct reasoning capabilities, we introduce 16 question types organized along two complementary axes: data versus health reasoning, which distinguishes computation over longitudinal measurements from physiological interpretation; and single- versus cross-signal reasoning, which separates reasoning about individual signals from the integration of multiple signals. To construct reliable questions at scale, we adopt a dual-grounding framework that combines literature-grounded physiological findings with statistically validated population-grounded physiological patterns. This enables the capture of meaningful relationships observed in real-world wearable data. Evaluation of 14 proprietary and open-source LLMs demonstrates that WearableQA effectively differentiates model capabilities, with performance ranging from 19.6% to 72.9% against a 10% chance baseline. Moreover, WearableQA remains far from solved: most models achieve accuracies below 60%. Overall, WearableQA provides a realistic and diagnostic benchmark for evaluating LLM reasoning over real-world wearable data. --- ### Diffusion TV: Experiencing Diffusion Models through Tangible, Embodied Interaction Authors: Sihwa Park Published: 2026-09-04 Categories: cs.HC, cs.AI arXiv: https://arxiv.org/abs/2609.05404v1 Diffusion TV is an interactive AI art installation that offers a tangible and embodied experience of diffusion models through a modified CRT TV. By physically manipulating the TV's antenna, audiences control the clarity of AI-generated images and sounds, metaphorically enacting the denoising process that underlies diffusion-based generation. Using the tuning knob, participants switch between three channels featuring AI-generated animals from the Past (extinct species), Present (endangered species), and Future (speculative creatures), situating the interaction within a temporal and ecological narrative. Through continuous audiovisual feedback and physical interaction, Diffusion TV foregrounds the generative process over final outputs, allowing audiences to explore intermediate states as experiential material. Rather than providing explicit technical explanation, the work presents an alternative, embodied mode of explainable AI that invites exploratory engagement with and reflection on generative technologies. --- ### RegionFed: Federated Learning for Personalized Query Understanding in Heterogeneous Retail Environments Authors: Quoc H. Nguyen, Ali Lafzi, Abhijeet Phatak, Siddharth Pratap Singh, Rohit Upadhyay, Yogananda Domlur Seetharama, Chittaranjan Tripathy Published: 2026-09-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.05403v1 Retail search systems serve diverse geographic regions with distinct query patterns, vocabularies, and product preferences, creating significant data heterogeneity that challenges both privacy-preserving training and model personalization. Federated learning offers a natural solution for privacy, but standard FL methods produce global models that sacrifice regional performance, while existing personalized FL approaches operate at the parameter level and catastrophically collapse on modern transformers (below 10\% accuracy on T5) due to tied embeddings and LayerNorm interactions. We introduce RegionFed, an \textit{architecture-robust} federated learning framework that sidesteps this failure by operating entirely at the gradient level. RegionFed uses the $\ell_2$ conflict between regional and global gradients as a unified signal that (i) diagnoses heterogeneity, (ii) routes each region to the cheapest sufficient personalization strategy, and (iii) adaptively controls personalization strength. Because it treats models as differentiable black boxes, RegionFed deploys on T5-Small, T5-3B, RoBERTa, and CNN with zero code changes, providing large gains on transformers (where parameter-level methods collapse) and consistent improvements on CNNs. Across three public datasets (Amazon ESCI, Amazon Reviews, LEAF-FEMNIST) and four architectures, RegionFed-Meta achieves 92.27\%, closing the gap to the privacy-violating centralized upper bound (Centralized + Regional Weighting: 92.04\%, $Δ$=0.23pp, within 1$σ$) while providing $(ε{\approx}0.60)$-differential privacy and $\mathcal{O}(1/\sqrt{T})$ convergence. --- ### Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models Authors: Wonje Jeung, Sangyeon Yoon, Hyesoo Hong, Yoonjun Cho, Dongjae Jeon, Bumjun Kim, Jean Oh, Youngjae Yu, Albert No Published: 2026-09-04 Categories: cs.RO, cs.CL arXiv: https://arxiv.org/abs/2609.05401v1 Vision-language models are increasingly used as reward functions for robotic learning, but this role requires paraphrase invariance: the same trajectory should receive the same reward under semantically equivalent goal descriptions. We show that current VLM reward models often violate this property. Paraphrasing the instruction alone can substantially change predicted progress scores, and can even flip identical robot behavior between failure and success. To measure this failure mode, we introduce ROBORMBENCH, a benchmark with 2,390 real-robot trajectories, ground-truth progress labels, and 21,673 verified paraphrases spanning lexical, syntactic, and action-goal rewrites. Across proprietary and open-source VLMs, paraphrase-induced instability is widespread and severe, grows under more divergent rewrites, and is not reliably reduced by scale or explicit reasoning. Dedicated reward models trained with trajectory-grounded supervision are substantially more stable. These results show that paraphrase robustness is a core requirement for reliable VLM-based reward modeling in robotics. --- ### A Generalizable Feature Extractor for Alzheimer's-Related Brain MRI Tasks Authors: Reza Rajabli, D. Louis Collins Published: 2026-09-04 Categories: cs.CV, q-bio.QM arXiv: https://arxiv.org/abs/2609.05400v1 When there is not enough labeled data to properly train deep learning models, transfer learning can help. We still do not fully understand how effective it is in neuroimaging, especially for Alzheimer's disease research. It is also not clear if these transferred models can work on new datasets without being retrained for each specific task. We evaluate whether a compact, supervised pretrained model can serve as a reusable foundation model for downstream neuroimaging tasks. We freeze the 7.18 million weights of a 3D CNN previously trained for brain-age prediction, and adapt it to each task using Low-Rank Adaptation (LoRA), requiring only ~1% additional trainable parameters. We evaluate generalizability in six experiments. Adapting the model to classify cognitively normal versus Dementia on ADNI gave an AUC of 0.964 on held-out folds (Experiment #1). Applying that adapted model unchanged to OASIS-3, with no retraining, gave an AUC of 0.871 (Experiment #2). Reusing its output logit together with age and a cognitive score distinguished stable from progressing MCI with an AUC of 0.828 (Experiment #3). Adapting the same backbone to predict amyloid positivity from structural MRI gave an AUC of 0.804 (Experiment #4). Finally, the same approach estimated ICV-normalized hippocampal and white matter hypointensity volumes directly from the T1w image, with R^2 of 0.80 and 0.91 respectively, tasks normally addressed with much larger U-Net networks (Experiments #5 and #6). A compact model supervised on brain age can therefore serve as a reusable backbone, adapting to each task with ~1% additional parameters and transferring to an unseen cohort without any training. Our findings suggest that a carefully trained brain age model can serve as an effective foundation model for Alzheimer's related tasks, even under strict data constraints. --- ### From Interpretability Methods to Interpretable Models Authors: Julien Colin, Nuria Oliver, Thomas Serre Published: 2026-09-04 Categories: cs.CV, cs.HC arXiv: https://arxiv.org/abs/2609.05399v1 More than a decade in, explainable AI (XAI) for computer vision has assembled a mature toolbox: attribution, feature visualization, concept-based, and circuit-based methods. Yet almost all of the field's effort has gone into building and comparing these methods, and little into the question they were meant to answer---how interpretable are our models, and are we making progress as they evolve? We argue for shifting the field's focus from methods to models, along two complementary lines. One is already within reach: existing tools let us characterize and compare what different models represent and compute. The other is harder, and largely neglected: whether a model can actually be understood by the humans who rely on it---the independent evaluators on whom trust and certification depend, not the experts confirming what they already expect. It can only be measured, not inferred. We review why the toolbox is mature enough to support both, survey the thin body of work comparing models, draw a parallel to systems neuroscience, and close with a model-centric XAI agenda. --- ### CrossDepth: Geometry-Constrained Attention for Generalizable Multi-View Surround Depth Estimation Authors: Samer Abualhanud, Max Mehltretter Published: 2026-09-04 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.05397v1 Reliable 3D understanding of the surrounding environment is a core requirement for autonomous driving. Multi-view surround camera rigs provide broad scene coverage, but the spatially adjacent images typically overlap only minimally. Consequently, the depth of most pixels must be inferred from monocular appearance cues. These cues can appear differently across images and may therefore be interpreted differently by the depth estimation model. We target two main sources of cross-image inconsistency: differences in camera intrinsics and the limited receptive field of each image. We address the former by conditioning the features on per-pixel camera-aware ray embeddings, enabling the network to account for camera-dependent variations in monocular cues. We address the latter by extending each pixel's context beyond its own image through cross-image attention constrained to geometrically plausible regions, derived from the calibrated rig setup. The model is trained in a fully self-supervised manner based on photometric consistency. Evaluations on DDAD and nuScenes show improved overall depth accuracy and cross-image depth consistency over state-of-the-art self-supervised methods under in-domain and cross-domain evaluation. Code is available at https://abualhanud.github.io/CrossDepthPage/. --- ### A Deep Generative Model for Synthesizing Labeled Wireless Signals Authors: Yuxiao Li, Keke Hu, Santiago Mazuelas, Yuan Shen Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05396v1 Wireless signals with position-related labels are pivotal for both performance evaluation and model training in the realm of wireless sensing. However, acquiring real-world datasets is often challenged by significant measurement and labeling costs. Traditional methods for synthesizing labeled wireless signals typically rely on environmental models, leading to extensive hyper-parameter tuning and inadequate realism for comprehensive model training purposes. To address these limitations, we introduce a novel deep learning (DL)-based method, namely Inter-Instance Generative Adversarial Networks (IIns-GAN), to generate realistic labeled wireless signals. The generated signals are particularly adaptive to different environment scenarios and well-suited for various model training tasks, including distance estimation and environment identification. We have conducted extensive experiments on public Ultra-Wideband (UWB) datasets to evaluate the realism and utility of the generated signals. The results demonstrate that the signals generated by IIns-GAN mirror the physical characteristics of real-world measurements, and significantly contribute to the improvement of model training in diverse wireless sensing tasks. --- ### Multi-Step Tool-Calling over Korean Open Public APIs: A Benchmark and a Data-Synthesis Recipe Authors: Dain Kim, Eungi Cho, Kyumin Kim, Shinyeong Noh, Kyuseong Lim Published: 2026-09-04 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.05395v1 Data-sovereignty regulations increasingly require public institutions to deploy open-source, on-premise LLM agents that chain multiple tool-calls across live government APIs. However, open-source models consistently underperform in this multi-step setting, and no existing benchmark measures the gap. We introduce the Korean Open Public API Benchmark (KOPA-Bench), comprising 145 real-world tasks. To close this gap, we present EDGE, an Execution-grounded Dynamic Graph for tool-calling data synthEsis driven by live execution. EDGE builds a graph of how each tool's output can feed another's input, keeps only the links that succeed when actually called against the live APIs, and traverses these verified links to synthesize executable multi-step trajectories. Fine-tuned via GRPO on the resulting dataset, our 9B model nearly matches the untuned 27B model from the same family, improving substantially not only on KOPA-Bench but also on the BFCL benchmark. --- ### Think-Verify-Revise: Neuro-Symbolic Visual Reasoning with Vision-Language Models and Dynamic Logic Tensor Networks Authors: Homayoun Afshari, Pietro Basci, Alessandro Russo, Lia Morra Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05388v1 Visual reasoning tasks require a system to jointly perceive visual content and apply formal relational constraints---a combination that neither pure neural nor purely symbolic approaches handle well in isolation. This paper proposes a Neuro-Symbolic (NeSy) framework that closes this gap by tightly coupling a Vision-Language Model (VLM) for automatic First-Order Logic (FOL) rule induction with a Dynamic Logic Tensor Network (D-LTN) for differentiable rule verification, in a closed iterative feedback loop. The VLM receives a small set of labelled visual examples and proposes candidate FOL rules conforming to a strict grammar (Think); the D-LTN is automatically assembled from these rules at runtime and evaluates them grounding on CNN-produced visual embeddings (Verify); and verification failures are fed back to guide the VLM's next hypothesis (Revise). Evaluated on the ViSudo-PC benchmark across four visual domains (MNIST, EMNIST, KMNIST, FMNIST), the system induces valid Sudoku constraint rules using only three training examples as visual context. The proposed method achieves AUC scores matching or outperforming previous methods (NeuPSL, LTN), showing the potential for automatic rule discovery through VLM. Code is available at https://github.com/homayoun-afshari/nesy. --- ### Necessary or Sufficient? Evaluating LLM Explanations With Behavioural Evidence Authors: Urja Pawar, Rajitha Ramanayake, Nabeel Kemal, Ashwin Kandath, Owen O'Neill, Guillaume Bourgeon, Houssem Chatbri, Christopher Martin, Vadim Pertsovskiy Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05385v2 LLM decision components that can operate within agent workflows often produce action-relevant recommendations or judgements together with explanations. Operators may use the named factors to monitor a system, diagnose errors, or decide when to escalate an output. Such use assumes that the explanations agree with the component's observable decision behaviour. We test two interpretations of the named factors: necessity, meaning that changing a factor would change the output, and sufficiency, meaning that retaining it while removing other changeable information would preserve the output. We evaluate these interpretations in two synthetic use cases: recommending advisors to clients and judging prompts for harmfulness or risk. Models return an output and the top three factors that most influenced it. Controlled black-box interventions estimate a necessity score for each factor by measuring how often changing it changes the output, and a sufficiency score by measuring how often retaining it preserves the output. Across eight models from the Claude, GPT, and Gemini families, the mean Spearman correlations between the cited ranking and the necessity and sufficiency scores are 0.349 and 0.354 for advisor recommendation, and 0.431 and 0.580 for prompt monitoring. Furthermore, an uncited factor scores above the lowest-scoring cited factor in 57.6% of advisor responses under necessity and 58.1% under sufficiency; the corresponding prompt-monitoring rates are 25.8% and 8.9%. The cited top three contain useful information but do not reliably identify the three factors with the strongest measured influence under necessity or sufficiency. The framework provides a black-box reliability check for explanations used in agent oversight while remaining scoped to individual LLM decisions. --- ### Necessary or Sufficient? Evaluating LLM Explanations With Behavioural Evidence Authors: Urja Pawar, Rajitha Ramanayake, Nabeel Kemal, Ashwin Kandath, Owen O'Neill, Guillaume Bourgeon, Houssem Chatbri Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05385v1 LLM decision components that can operate within agent workflows often produce action-relevant recommendations or judgements together with explanations. Operators may use the named factors to monitor a system, diagnose errors, or decide when to escalate an output. Such use assumes that the explanations agree with the component's observable decision behaviour. We test two interpretations of the named factors: necessity, meaning that changing a factor would change the output, and sufficiency, meaning that retaining it while removing other changeable information would preserve the output. We evaluate these interpretations in two synthetic use cases: recommending advisors to clients and judging prompts for harmfulness or risk. Models return an output and the top three factors that most influenced it. Controlled black-box interventions estimate a necessity score for each factor by measuring how often changing it changes the output, and a sufficiency score by measuring how often retaining it preserves the output. Across eight models from the Claude, GPT, and Gemini families, the mean Spearman correlations between the cited ranking and the necessity and sufficiency scores are 0.349 and 0.354 for advisor recommendation, and 0.431 and 0.580 for prompt monitoring. Furthermore, an uncited factor scores above the lowest-scoring cited factor in 57.6% of advisor responses under necessity and 58.1% under sufficiency; the corresponding prompt-monitoring rates are 25.8% and 8.9%. The cited top three contain useful information but do not reliably identify the three factors with the strongest measured influence under necessity or sufficiency. The framework provides a black-box reliability check for explanations used in agent oversight while remaining scoped to individual LLM decisions. --- ### Reflection-aware Generative Novel View Synthesis Authors: GeonU Kim, Shin Dong-Yeon, Tae-Hyun Oh Published: 2026-09-04 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.05382v1 We propose Ref-GeNVS, a training-free, reflection-aware method for generative novel view synthesis (NVS) in mirror scenes. Existing multi-view diffusion models often fail to recognize the mirror in the scene and cannot exploit reflected content for scene generation. To fix this issue without additional training, our key idea is to treat a mirror image as two complementary views. From input images, we estimate the mirror plane and reflect camera poses to form virtual views. Based on this virtual view setup, we propose a two-stage generation method consisting of Mirror-gated attention and Reflection injection, which enables reflection-consistent NVS by explicitly leveraging reflection relationships in a multi-view diffusion model. Ref-GeNVS inherits the strong generalizability of the multi-view diffusion backbone, while it does not require finetuning. On synthetic and real scenes including mirrors, Ref-GeNVS outperforms recent generative NVS methods by generating reflection-consistent and contextually coherent novel views, revealing scene structure visible only through mirrors. Project page: https://kim-geonu.github.io/Ref-GeNVS/ --- ### Molecular Déjà Vu: Digit-Level Retrieval of Published Values in Frontier Language Models Authors: Matthias Busch, Marius Tacke, Sviatlana V. Lamaka, Mikhail L. Zheludkevich, Christian J. Cyron, Roland C. Aydin, Christian Feiler Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05381v1 Large language models (LLMs) are increasingly evaluated on molecular property benchmarks, but accuracy cannot distinguish a model that predicts a property from one that retrieves a published number. We audit 22 frontier models on 12 regression benchmarks for verbatim retrieval and find that it is widespread but relatively benchmark-specific: on five datasets more than $50\%$ of the LLMs show verbatim retrieval, while on the remaining datasets it appears only in isolated cells. We run our experiments at two reasoning levels and find that reasoning changes retrieval. The same experiments, on the same molecules and with the same prompt, are flagged $89\%$ more often at the higher reasoning level than at the lowest one. Finally, we test a way to interrupt retrieval in our most contaminated cases, and find that the strongest models in some cases still recognise a combination of transformed SMILES strings and original labels. Furthermore, suppressing retrieval moves the prediction errors of the different models closer together in relative terms, while their differing use of verbatim retrieval spreads them apart. This indicates that the general predictive capability of an LLM is not determined solely by the amount of memorised values. This work provides an overview of the amount and depth of verbatim retrieval in molecular regression benchmarks using LLMs. --- ### What Matters, When? Diagnosing and Improving Conditional Visual Grounding in Visuomotor Imitation Policies Authors: Vivek Chavan, Pengtao Xie, Yahuan Shi, Oliver Heimann, Kevin Haninger, Jörg Krüger Published: 2026-09-04 Categories: cs.RO, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.05376v1 Visuomotor imitation policies can achieve high performance under in-distribution visual conditions yet fail when visually similar objects or receptacles are introduced. We study this behavior as a problem of conditional visual grounding: the visual target required for successful control changes with the manipulation phase and, in more complex tasks, with the observed task state. Using Action Chunking with Transformers (ACT), we systematically introduce distractor objects and receptacles with controlled color and shape similarity and localize failures to picking and placement. We find that distractor sensitivity is specific to both the type of visual similarity and the manipulation stage. Guided by this diagnosis, we evaluate distractor augmentation, phase-dependent attention regularization, and appearance-based visual prompting as complementary interventions for improving target selection while preserving spatial information required for control. These interventions substantially improve robustness in simulation and on a physical UR3e. We further examine the same failure pattern in a pretrained vision-language-action policy on a state-conditioned instrument-handling task, where the observed state of a medical instrument determines the correct destination. Together, the results show that visual distractors can cause incorrect object or destination selection even when the underlying manipulation skill remains intact, and that explicitly improving target selection can substantially recover performance across distinct visuomotor policy-learning regimes. --- ### CUA-Universe: A Scalable and Dynamic Environment for Hybrid GUI+CLI Agents Authors: Haoting Shi, Wenhao Wang, Weicheng Fang, Yaozhong Liang, Tian Jin, Pengxiang Zhao, Guangyi Liu, Siheng Chen, Yanfeng Wang Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05374v1 Computer-use agents have advanced on benchmarks like OSWorld and AndroidWorld, but still act mostly through the GUI, often producing inefficient trajectories. Real-world computer work is hybrid, combining visual-state inspection with precise, high-throughput command-line operations, so capable agents must coordinate both modalities over shared application state. Yet scalable hybrid environments remain scarce because supporting both GUI and CLI over real applications typically requires substantial manual engineering for each application. Existing agents also struggle to use the two interfaces complementarily: CLI-native agents lack visual perception for tasks involving interface state or layout, while GUI-native agents are inefficient for operations better executed through commands. We introduce CUA-Universe, a scalable environment-to-data pipeline that turns real desktop software into hybrid GUI+CLI environments. App-Forge adapts applications into reproducible VMs and command-line surfaces it discovers, wraps, or generates, scaling to 16 applications; Task-Weave synthesizes diverse hybrid tasks of controllable difficulty from reusable operations over seed files; and Path-Steer steers rollouts along efficient hybrid paths and harvests verified trajectories for post-training. Training on this data shifts behavior from inefficient GUI interaction and brittle CLI scripting toward effective GUI+CLI orchestration. Our 9B model improves both success and efficiency on CUA-Verse (Score +39.3 pts; -37% steps, -60% tokens), OSWorld (SR +16.8 pts; -57% steps, -44% tokens), and OSWorld-MCP (Score +7.84 pts; -27% steps, -30% tokens). CUA-Universe provides a scalable path toward more capable and efficient computer-use agents. --- ### When LLM Decompilers Recompile More and Preserve Less Authors: Chang Liu, Edward Raff, Kristopher Micinski Published: 2026-09-04 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.05370v1 Decompilation recovers high-level source from compiled machine code and serves as a foundation for security tasks such as vulnerability detection and malware analysis. Traditional decompilers like Ghidra and Hex-Rays expose whatever they cannot resolve as visible placeholders and often emit pseudocode that will not compile or execute; LLM-based decompilers produce clean, idiomatic C and are now judged almost entirely by recompilability and re-executability: whether the output builds and passes its shipped input/output tests. We show that these metrics can reward the wrong path: a function may recompile and pass every shipped test yet diverge on other legitimate inputs, and a disclosed vulnerability may disappear from the recompiled code with no visible trace of the crash. Neither failure is caught by existing suites. To address this gap, we propose Decompile-Diverge, a behavioral comparison oracle not relying on fixed or hand-crafted tests: for each function it synthesizes a driver, grows a fuzzing corpus from the reference, and reruns the decompiled code on the same inputs to detect changes in the function's behavior. Across eight systems in nine configurations on established LLM decompilation corpora, candidates that pass every shipped test still diverge from the original on our input corpus: 4.9% overall, and as many as 13% for a single system. On 300 real GitHub library functions and 287 CVE-grounded functions, recompilability and behavioral agreement can come apart: the strongest refinement LLM lifts Ghidra's build rate from 75% to 90%, while its Matched rate falls from 74% to 62%; on disclosed vulnerabilities, up to one tenth exhibit Crash Absence in its output. Source-level analysis traces this divergence to introduced fields, types, callees, and guards that replace the visible unknowns traditional tools leave behind. --- ### Towards Neuro-Symbolic Procedural Reasoning for Long-Horizon Vision-Language-Action Manipulation Authors: Vivek Chavan, Yahuan Shi, Oliver Heimann, Kevin Haninger, Jörg Krüger Published: 2026-09-04 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2609.05369v1 Vision-language-action (VLA) models can execute short manipulation skills, but remain brittle in long-horizon procedures requiring persistent task state, dependency-aware reasoning, conditional decisions, and reliable grounding. We investigate a neuro-symbolic framework that combines learned VLA control with explicit task graphs and multimodal procedural memory. Task graphs encode action dependencies, valid transitions, and branch conditions, while memory maintains the active step, completed actions, textual context, and task-relevant visual evidence. Together, these structures guide object selection, destination grounding, subgoal dispatch, and verification of expected state transitions. Human demonstrations provide additional spatial and temporal guidance through gaze or saliency cues. To isolate their effect on policy learning, our initial study bypasses cross-view gaze transfer and directly annotates pseudo-gaze in robot-view teleoperation videos. The resulting guidance is used during VLA fine-tuning and inference. We study two long-horizon manipulation domains, workspace clearing and surgical-instrument handling, which require ordered execution, visually grounded decisions, and conditional branching. We evaluate correct-object and destination selection, subtask completion, task progress, step-order consistency, complete-task success, and procedural or execution mistakes. This work positions structured symbolic reasoning and demonstration-derived visual guidance as complementary mechanisms for reliable long-horizon VLA manipulation. --- ### Design Docs Are All You Need: An AI-native Machine-Learning Performance Tool Authors: Samuel Kushnir, Kimia Noorbakhsh, Kavya Sreedhar, Liqun Cheng, Ming Liu, Parthasarathy Ranganathan, Mohammad Alizadeh, Fred Kjolstad, Suvinay Subramanian Published: 2026-09-04 Categories: cs.PL, cs.AI arXiv: https://arxiv.org/abs/2609.05364v1 Machine-learning performance modeling is a uniquely hostile terrain for long-lived software: the assumptions baked into today's abstractions are invalidated by tomorrow's models and systems, forcing perpetual refactoring of performance-modeling frameworks. Meanwhile, AI coding agents have become fast and capable enough that regenerating an entire library is cheaper than paying down the tech debt of incrementally patching it. We describe SMART, a rigorous symbolic performance-modeling library for ML systems whose main branch contains almost no code: the repository is a DAG of self-contained natural-language design docs, coding sub-agents regenerate the implementation from only the docs on new version updates, and every human change is a natural-language edit to a doc--self-documenting by construction. Two ingredients make regeneration reliable: (i) a design-doc style built around step-by-step worked examples that act as in-context demonstrations for the generating agents, and (ii) a minimal, recursively defined operator IR with symbolic (SymPy) cost expressions, a fast analytical roll-up mode for large sweeps, and a slow modulo-scheduling mode for fine-grained schedule studies. Regenerated implementations reproduce hand-audited reference models--including DeepSeek-V3 serving on a TPU pod slice--to round-off precision, suggesting that design docs--not code--can be the durable artifact for ML-systems co-design tools. --- ### Distill Globally, Adapt Locally: Reasoning Distillation and Product-Type Test-Time Training for Scalable Trade-Up Recommendation Authors: Siliang Liu, Mohammad Ghasemi, Sapan Patel, Amin Banitalebi-Dehkordi Published: 2026-09-04 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.05363v1 Trade-up recommendation identifies higher-quality alternatives that preserve a customer's purchase intent while offering upgraded benefits. Large language models (LLMs) can reason about such distinctions, but applying them directly to hundreds of millions of product pairs is operationally impractical. We introduce a two-level framework that distills LLM reasoning into an efficient non-generative student and adapts its decision boundary to product-type-specific trade-up criteria. At Level 1, a retrieval-augmented few-shot LLM teacher generates structured relation labels and natural-language rationales. These rationales supervise a compact embedding-pair classifier through alignment and contrastive objectives; at inference, the student uses only two precomputed 768-dimensional product embeddings, with no LLM calls or text generation. On a fixed human-annotated benchmark of 8,352 pairs, a 15.5M-parameter four-class reasoning-distilled student achieves AUC 0.924 (95% CI (0.918, 0.929)), compared with 0.912 for the four-class label-only student. At Level 2, product-type test-time training (PT-TTT) uses few-shot demonstrations to optimize lightweight category-specific adapters over the frozen student. PT-TTT improves AUC from 0.924 to 0.941 and average precision from 0.920 to 0.940. On a 100K-pair proxy catalog, the distilled student on a single eight-GPU machine is approximately 5,000x faster and 10,000x lower in estimated cost than direct LLM inference. --- ### MEOX: Compact Multimodal Mixture-of-Experts for Earth Observation Authors: Mohanad Albughdadi Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05351v2 Recent advances in Earth Observation representation learning accommodate heterogeneous sensors and missing observations, often through larger architectures. We present MEOX (Multimodal Earth Observation with eXperts), a multimodal masked autoencoder with a 2.939 million-parameter encoder and 3.115 million parameters in total. Sensor-specific adapters, explicit validity signals, and a shared sparse-expert block preserve modality-dependent processing before a learned patch-wise fusion. Four metadata tokens then accompany a single spatial sequence through fourteen further encoder blocks. Shared expert projections with private low-rank residuals constrain parameter growth, while rotary attention supports downstream spatial grids different from pretraining. The model is pretrained on 1.228 million MMEarth64 samples using modality-balanced masked reconstruction and structured sensor dropout. Frozen transfer is evaluated on six GEO-Bench tasks at both 64 and 224 pixels. The model reaches 64.42% mean intersection-over-union on cashew segmentation at 64 pixels and 90.56% average accuracy on EuroSAT at 224 pixels, exceeding the corresponding reported CSMoE results. BigEarthNet finetuning reaches 72.95% micro-average precision. Routing diagnostics distinguish expert participation, spatial dependence, modality association, and functional contribution. A held-out WorldCover probe measures a 0.64-percentage-point benefit from metadata, while retrieval separates same-sensor semantics from cross-sensor alignment. These results demonstrate sensor-flexible representation learning and strong task transfer using a compact parameter budget. --- ### MEOX: Compact Multimodal Mixture-of-Experts for Earth Observation Authors: Mohanad Albughdadi Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05351v1 Recent advances in Earth Observation representation learning accommodate heterogeneous sensors and missing observations, often through larger architectures. We present MEOX (Multimodal Earth Observation with eXperts), a multimodal masked autoencoder with a 2.939 million-parameter encoder and 3.115 million parameters in total. Sensor-specific adapters, explicit validity signals, and a shared sparse-expert block preserve modality-dependent processing before a learned patch-wise fusion. Four metadata tokens then accompany a single spatial sequence through fourteen further encoder blocks. Shared expert projections with private low-rank residuals constrain parameter growth, while rotary attention supports downstream spatial grids different from pretraining. The model is pretrained on 1.228 million MMEarth64 samples using modality-balanced masked reconstruction and structured sensor dropout. Frozen transfer is evaluated on six GEO-Bench tasks at both 64 and 224 pixels. The model reaches 64.42% mean intersection-over-union on cashew segmentation at 64 pixels and 90.56% average accuracy on EuroSAT at 224 pixels, exceeding the corresponding reported CSMoE results. BigEarthNet finetuning reaches 72.95% micro-average precision. Routing diagnostics distinguish expert participation, spatial dependence, modality association, and functional contribution. A held-out WorldCover probe measures a 0.64-percentage-point benefit from metadata, while retrieval separates same-sensor semantics from cross-sensor alignment. These results demonstrate sensor-flexible representation learning and strong task transfer using a compact parameter budget. --- ### Who Should Grade My Work? Student Perspectives on Transparent AI-Assisted Writing Assessment in Higher Education Authors: Rayed AlGhamdi Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05346v1 The integration of GenAI tools into higher education assessment raises important questions about how students understand, interpret, and respond to AI-mediated evaluation. As instructors increasingly explore AI tools for providing feedback, prior research has examined whether GenAI-generated feedback improves writing performance and how students perceive its usefulness; comparatively little is known, however, about how students interpret such evaluation when they are explicitly informed that an AI system, rather than a human instructor, produced the feedback and the score. This study reports findings from a qualitative pedagogical inquiry conducted in an undergraduate technical communication course for computing students at a Saudi public university. Thirteen male undergraduate computing students completed an in-class handwritten writing task; the scanned submissions were evaluated by ChatGPT using a rubric-based prompt aligned with the task objectives. Students were then explicitly informed that ChatGPT had generated the score and feedback and were invited to reflect on the evaluation in writing. Inductive thematic analysis of these reflections identified four themes: perceived usefulness of feedback; awareness of AI's contextual and pedagogical limitations; conditional trust, distinguishing feedback utility from evaluative authority; and reflection on the institutional and pedagogical role of the human instructor. Participants accepted GenAI feedback as useful for surface-level revision but consistently positioned the human instructor as the appropriate authority over grading decisions. The study identifies this as a distinction between feedback utility and evaluative authority, two judgments that students treat as analytically separate rather than as opposite ends of a single approval scale... --- ### Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory Portability Authors: Ankit Goyal, Jaideep Ray Published: 2026-09-04 Categories: cs.AI, cs.CL, cs.IR arXiv: https://arxiv.org/abs/2609.05339v1 Model upgrades are routine; memory migrations are not. An agent can keep the same memory store and still forget: a new model may interpret old notes differently, mixed embedding versions may break retrieval, and repair may fail without the original evidence. We compare memory as the same history is preserved verbatim for long-context reading (LC-RAW), divided into chunks for retrieval-augmented generation (RAG), compressed by a model into natural-language notes (NOTES), or normalized into a fixed-schema knowledge graph (KG-fixed). The study uses 48 synthetic histories with randomized answer codes, exact scoring, and two open-weight models with sub 10 billion parameters. Our measurements show that fixed-schema structures transfer reliably, with KG-fixed accuracy changing by only $+0.0004 \pm 0.0020$ following a writer swap. Conversely, compressed NOTES exhibit high model coupling, with accuracy shifting asymmetrically by $+9.91$ or $-13.28$ percentage points depending on the specific migration direction. In RAG systems, partial embedding migrations using a 50/50 mixed index capture only a 4.96-point accuracy improvement, forfeiting the majority of the 11.90-point gain achieved through full re-embedding. Diagnostic decomposition attributes 80% ($0.467 \pm 0.014$) of the NOTES accuracy deficit to information lost during initial construction, whereas retrieval failures drive 81% ($0.364 \pm 0.012$) of the RAG deficit. Finally, store-only repair of NOTES fails to reach a 90% performance recovery target in all 48 test cases, whereas retaining the raw source history enables successful recovery in 34 of 48 cases for one tested direction. These findings highlight the necessity of direction-specific migration testing, strict embedding space isolation, and the retention of source histories for memory repair. --- ### Variational Continuation for Double Pendulum Periodic Orbits Authors: Leo Yao, Ziming Liu, Max Tegmark Published: 2026-09-04 Categories: cs.LG, math-ph, nlin.CD, physics.class-ph arXiv: https://arxiv.org/abs/2609.05337v1 We present a Hessian-based approach to numerically continue periodic orbits in dynamical systems. A loop (periodic orbit candidate) is parametrized as a Fourier series; a loss function is defined based on the deviation of the loop from the physical differential equations. Unlike previous work relying on hand-derived Jacobians, our method automates the process by leveraging automatic differentiation, a common machine learning technique. The continuation direction can be determined by the flat directions of the loss landscapes (directions with zero eigenvalues), making the search of periodic orbits efficient and guided. Our method is integrator-free, precisely initializes oscillations around unstable fixed points, and efficiently detects orbit family intersections and subharmonic bifurcations. As a demonstration, we present full continuations of periodic double pendulum oscillations from fixed points, showing bifurcations along orbit families and categorizing branches of periodic orbits. In particular, we find periodic orbits where both pendulum masses are never simultaneously at rest, which to our knowledge has been missing in the literature. --- ### The History Is the Detector: Executing CVE Patch History, End-to-End Authors: Qiushi Wu, Kevin Eykholt, Youngja Park, Xiaokui Shu, Dhilung Kirat, Douglas Lee Schales, Ian Molloy Published: 2026-09-04 Categories: cs.CR, cs.AI, cs.SE arXiv: https://arxiv.org/abs/2609.05335v1 Public vulnerability databases collect rich information about known software flaws, including their weakness types, affected components, and related patches. Fixing commits provide the exact code changes that removed these flaws. While these records capture why the original code was unsafe, they are documented mainly for human inspection rather than automated reuse. Consequently, the same unsafe conditions may still exist elsewhere in code without a known advisory, leaving much of this detection knowledge unused. We present BUGSTONE-E2E, a framework that transforms vulnerability history into executable detection rules and validates their findings. First, BUGSTONE-E2E mines reusable rules from verified fixing commits, capturing scan anchors, fix semantics, and CVE provenance and organizing them by CWE and language. Second, detection follows a funnel-shaped pipeline: early stages process a large pool of candidates using lightweight analysis, while later stages apply increasingly capable and expensive models to a shrinking set of targets. Specifically, BUGSTONE-E2E first enumerates call sites matching rule anchors using Tree-sitter, then removes benign sites using lightweight heuristics without LLM calls. Next, LLM-based agents inspect the remaining candidates guided by the rule. Following this inspection, the system re-triages surviving candidates and builds runtime verifications, then generates scope-checked patches validated via two-sided differential tests. Using 19,325 high-severity CVEs from 2022 to 2026, BUGSTONE-E2E identifies 2,710 fixing commits and constructs 1,033 detection rules across 56 CWE families, packaged into 172 skills. When applied across 14 programs, it produced runtime evidence for 644 findings. These results demonstrate that CVE history can be turned into an executable workflow, transforming past vulnerabilities into reproducible detection and repair. --- ### Lightweight Vision Transformer Compression for On-Device Plant Disease Detection in Resource-Constrained Agricultural Field Conditions Authors: Mahadev Sunil Kumar, Bhavika Gondi, Desaisetty Venkata Satya Sai Swapnith, Gangireddy Rahul Jogi, Sudheesh Manalil, Arnab Raha, Amitava Mukherjee, Parthasarathy Seethapathy, G. Gopakumar Published: 2026-09-04 Categories: cs.CV, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.05334v1 Chilli (Capsicum annuum) is one of India's most economically significant crops, yet its productivity is persistently threatened by diseases that are difficult to identify without expert intervention. While Vision Transformers (ViTs) have achieved high classification accuracy, their large computational footprint makes deployment on resource constrained devices challenging. Existing compression approaches typically address pruning, quantization, and knowledge distillation in isolation, leaving the potential benefits and interactions of their combined application insufficiently explored. We propose a unified Vision Transformer compression framework that combines Hessian-Balanced Adaptive Block Pruning (H-BAC), guided by second-order sensitivity estimation, with quantization and attention-based knowledge distillation. To systematically identify the most effective configuration within each compression family, each technique is first evaluated independently through controlled ablation studies, after which the best-performing components are integrated into a sequential deployment pipeline tailored to real-world agricultural constraints. On a chilli 3-class village-split dataset with a genuine cross-village, cross-device out-of-distribution test split, the resulting compressed models match or exceed the 95.13% FP32 baseline's accuracy, alongside 74-98% model size reduction, and the fully integrated compression pipeline achieves a 54.5x size reduction (327.42 MB to 6.01 MB) at 95.13 +/- 2.32% accuracy across four tested configurations. A direct comparison further reveals that, on this dataset, a directly-trained student of the same final size, without pruning or distillation, reaches comparable accuracy of 94.87%, at the same 6.01 MB INT8 size, indicating where H-BAC and knowledge distillation are, and are not yet shown to be, worth their computational cost. --- ### Technical Manual for a Toolkit for Measuring Contextual Individuation in Transformer Language Models Authors: José Luciano Verçosa Marques, Frederico Jorge Heitmann, Daniel Omar Perez, Marcelo Vinicius de Paula, Tárcio André dos Santos Barros Published: 2026-09-04 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.05333v1 A transformer language model assigns a single, context-independent vector to a word type at its embedding layer, yet is widely believed to individuate that word's occurrences by context in its later layers. Testing this belief cleanly requires a construct that holds the word form fixed while its context and intended sense vary in a controlled, labeled way. This manual documents an open toolkit built around such a construct, which we call a bridge form: a single written word that recurs, unchanged, across two or more subject domains with a different sense in each. We describe, and justify, every stage of the pipeline: the declarative specification of bridge forms and their source domains, corpus acquisition from Wikipedia, occurrence localization, layer-wise representation extraction, a domain-pairwise silhouette measurement of separation in the model's representation space, and a paired visualization protocol. Each design choice is presented together with the methodological failure mode it is meant to avoid (sense contamination from overly broad category labels, the multi-group bias of the silhouette coefficient, subword-tokenization misalignment, and axis-comparability artifacts in dimensionality-reduced plots, among others). This manuscript is a methodological and implementation reference: it does not report or interpret empirical outcomes of running the toolkit on any particular model or bridge-form set. The toolkit, its full source, and the corpora used to exercise it are archived separately (Section 9) under a persistent identifier, and are intended to be cited as an instrument by studies that use it to produce and interpret empirical results. --- ### Embedded Graph Flows for Categorical Graph Generation Authors: Ethan Ma, Zihan Wang, Chris Siu Yeung Chow, Xinguo Feng, Qingqing Li, Rui Jiang, Naipeng Dong, Guangdong Bai Published: 2026-09-04 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.05328v1 Generating categorical graphs requires choosing node and edge types that form a coherent structure without depending on node order. Many graph generators encode categories as fixed one-hot vectors, which can impose an artificial geometry in which categories are equidistant. We propose Embedded Graph Flows (EGF), a generative model that learns continuous embeddings for node and unordered-edge categories and transports Gaussian noise towards these learnt endpoints using a permutation-equivariant graph transformer. A terminal readout maps the embeddings back to discrete graph categories. Across molecular benchmarks, EGF achieved competitive performance. On QM9, EGF gives the best result on all four reported metrics among the three methods, including a Fréchet ChemNet Distance (FCD) of 0.150, compared with 0.717 for the categorical-diffusion baseline DiGress and 0.812 for the bridge-based baseline GruM. When applied to larger molecules in ZINC250k, EGF retains the lowest maximum mean discrepancy (MMD) using the neighbourhood subgraph pairwise distance kernel (NSPDK), indicating close agreement with the local substructures of the reference molecules. Our code is available at https://github.com/Trusted-System-Lab/EGF. --- ### LLM-Driven Algorithm Design for Quantum Circuit Synthesis based on Binary Decision Diagrams Authors: Yoonju Sim, Federico Berto, Chuanbo Hua, Jinkyoo Park, Changhyun Kwon Published: 2026-09-04 Categories: cs.AI, cs.AR arXiv: https://arxiv.org/abs/2609.05327v1 Quantum circuits are central to implementing quantum algorithms on quantum devices, where quantum gates must be reversible. Many quantum algorithms rely on Boolean functions, which must therefore be implemented reversibly within quantum circuits. Reversible circuit synthesis provides a way to translate such Boolean functions into reversible circuits. Binary decision diagrams (BDDs) offer a scalable approach to this task, but the resulting BDDs and circuits depend heavily on variable ordering. Existing ordering heuristics commonly minimize BDD size because it is closely tied to the circuit size. However, BDD size is an imperfect proxy for the quantum cost of the synthesized circuit (QCC). We propose \texttt{QuantumEvo}, an evolutionary framework that uses an LLM as a heuristic generator for QCC-aware BDD variable ordering. Instead of predicting orderings directly, \texttt{QuantumEvo} searches over ordering heuristics initialized from multiple heuristic families. Candidate heuristics directly manipulate variable orderings using standard BDD operations and are selected by downstream QCC. The discovered heuristic, HGA-QE, modifies the sifting step inside a genetic algorithm so that the procedure is better aligned with QCC. Across the benchmark set, HGA-QE achieves a 70.9\% tie-or-win rate against the per-function best baseline and is strictly best on 13.5\% of the functions. The results demonstrate broadly competitive QCC performance, with HGA-QE showing a clearer relative advantage in strict wins on the two benchmark suites drawn from sources different from the data used for heuristic discovery. --- ### RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks? Authors: Zhenxuan Fan, Bo Zhang, Yutong Lin, Yuqian Yuan, Juekai Lin, Liang Liang, Zhuoyi Huang, Wenqiao Zhang, Juncheng Li, Siliang Tang, Jun Xiao, Yueting Zhuang Published: 2026-09-04 Categories: cs.RO, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.05324v1 Vision-Language-Action (VLA) models have shown promising progress in language-conditioned robotic manipulation. However, existing datasets and benchmarks mainly evaluate task completion under predefined settings, offering limited insight into model reasoning under increasing spatial and procedural complexity. We introduce \textbf{RoboSPA} (\textbf{Robo}t \textbf{S}patial-\textbf{P}rocedural \textbf{A}ssessment), a large-scale robotic manipulation dataset and benchmark for diagnosing embodied reasoning in VLA models. \texttt{RoboSPA} focuses on two core dimensions, Fine-Grained Spatial Reasoning and Long-Horizon Procedural Planning, covering 10 task categories and 56 base tasks. Each task is instantiated across five difficulty levels, yielding 280 variants with increasing spatial ambiguity and procedural complexity. We collect 527K trajectories across multiple embodiments and diverse scenes. Beyond binary success rate, \texttt{RoboSPA} introduces diagnostic metrics for more detailed evaluation. Experiments on representative VLA models show that current systems still struggle with complex spatial relations, precise low-level execution, and memory-intensive planning. These results establish \texttt{RoboSPA} as a challenging diagnostic benchmark for developing more capable, reliable, and generalizable embodied agents. Our data and code are available at https://github.com/fanzhenxuan/RoboSPA. --- ### Scalable Detection of Fossil Palynomorphs in Multifocal Digital Microscopy Images Authors: Abbas Shaikh, Praise Mayor, Patrick Ainlay-Vazquez, Aditya Viswanathan, Teon Golden, Eric Zhang, Ingrid C. Romero, Alexander E. White, Scott Wing, Arko Barman Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05323v1 Palynomorphs (microscopic, organic-walled fossils such as pollen, spores, and dinoflagellates) are important high-resolution records of past climates and are critical to the study of ancient ecosystems. Existing methods rely on manual analysis of high-resolution, multifocal digital microscopy images, which is slow and time-consuming and requires researchers to compromise on the scale of their investigations. To the best of our knowledge, our work proposes the first ever scalable end-to-end pipeline for automated palynomorph detection in whole slide images that addresses this bottleneck through: (1) efficient methods for decomposing and compressing digitized multifocal microscope slide images into tractable 2-dimensional tiles for analysis; (2) benchmarking modern object detection models, including RF-DETR, for the detection of palynomorphs, achieving an AP@50 of 0.879; (3) an efficient algorithm for the synthesis of detection outputs across large-scale, high-resolution images; and (4) an I/O optimization resulting in faster inference time. Our methods drastically reduce the time required for palynomorph detection in a single slide from often days of manual inspection to under one hour of automated analysis, enabling palynological research at a substantially greater scale. --- ### Adaptive Gated Deepfake Detection for Low-Resolution and Resource-Constrained Environments Authors: Vaishnavi Sen, Cody Laurie, Rashida Hasan Published: 2026-09-04 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.05320v1 Deepfake detection models often rely on high-quality inputs, fixed inference paths, and computationally expensive architectures, limiting their use in low-resolution and resource-constrained settings. This paper proposes AdaGate-DF, an adaptive gated deepfake detection framework that uses image-quality cues to route samples through a dual multi-exit system so high-quality images can exit earlier and save compute. We evaluated AdaGate-DF against MaD-CoRN, DefakeHop++, and ShuffleNetV2 on two benchmark datasets (Celeb-DF and FaceForensics++) under multiple configurations to test image resolution dependence and training and inference efficiency. On Celeb-DF, AdaGate-DF achieves an AUC of 0.9370, outperforming MaD-CoRN and DefakeHop++ while maintaining a low inference latency. Resolution-based testing shows consistent improvement as input resolution increases, reaching an AUC of 0.9708 at 384 by 384. The FaceForensics++ results highlight that AdaGate-DF remains effective under class imbalance, following competitive results with evaluated models. Overall, AdaGate-DF demonstrated a practical balance between detection performance, uncertainty-aware prediction, and computational efficiency for variable-quality deepfake detection. --- ### Optimal Rates for Agentic Networked Information Aggregation Authors: MohammadHossein Bateni, Zahra Hadizadeh, MohammadTaghi Hajiaghayi, Mahdi JafariRaviz, Shayan Taherijam Published: 2026-09-04 Categories: cs.LG, cs.GT, econ.TH arXiv: https://arxiv.org/abs/2609.05318v1 Building on the pioneering paper of Kearns, Roth, and Ryu (SODA'26), we study information aggregation in a networked learning model. The model captures a central pattern in agentic AI: each agent sees only part of the data and passes on only its own conclusion. Their model considers a linear regression problem with the mean squared error (MSE) loss. Agents sit in a DAG and each sees only a subset of the features and its parents' predictions, fits a linear predictor, and passes only its prediction forward. The benchmark is the full-feature learner that sees all raw features. A path of depth $D$ is $M$-covered if every block of $M$ consecutive agents collectively sees all raw features. Kearns, Roth, and Ryu proved that the excess mean squared error of the last agent on such a path is $O(M/\sqrt D)$, and gave a cyclic instance with excess error $Ω(M/D)$ for $D<M^2$. We close this gap: the correct rate is constant up to depth $M^2$, and $Θ(M^2/D)$ beyond it. We first give a sharper analysis of the cyclic instance and improve its lower bound to $Ω(\sqrt{M/D})$ for $D<M^2$. We then construct, for every depth $D\ge M^2$, an $M$-covered path of depth $D$ with excess error $Ω(M^2/D)$. The same instance gives the constant lower bound for all $D < M^2$. We also show that for any fixed distribution the excess error contracts geometrically along the path, ruling out any single instance that witnesses any polynomial lower bound at every depth. Finally, we prove the same optimal rate for logistic classification in the logit-passing model of Bateni et al., which considers the binary cross-entropy (BCE) loss. The same improved upper bound of $O(M^2/D)$ holds, and we transfer all the regression lower bounds by showing that on those examples the logistic path follows the least-squares path up to rescaling. --- ### Large Language Models for HVAC Operations in Building Energy Systems: A Critical Review of Methods, Applications, and Deployment Readiness Authors: Alexander Neubauer, Tianzhen Hong, Han Li, Mengbo Yu, Amin Darbandi, Yannick Fürst, Martin Kriegel Published: 2026-09-04 Categories: cs.AI, cs.CL, eess.SY arXiv: https://arxiv.org/abs/2609.05314v1 Building automation systems generate rich sensor data yet remain insight-poor because heterogeneous point naming, missing metadata, and fragmented documentation obstruct their operational use. This systematic review analyses and codes 66 peer-reviewed studies on large language models (LLMs) for HVAC operations published between 2023 and March 2026. Each study is classified across five application families and three LLM method families and assessed for evidence realism, deployment readiness, and the responsibility boundary between the LLM and physical HVAC decisions. The corpus is concentrated in building energy modelling (BEM, 32 of 66 papers), while load forecasting remains too sparse for subfield-level conclusions. Only four studies reach pilot-level evidence, and none reports sustained operational deployment. No study was classified as ready-now for industry adoption; three were near-term and 63 research-only. Nevertheless, several bounded, human-in-the-loop uses merit near-term trials, including point-name normalisation, document-grounded operator support, BEM workflow assistance, and advisory interfaces around physics-based controllers. Conventional machine learning (ML), model predictive control (MPC), reinforcement learning (RL) and ontology-based tools remain more adopted for high-frequency control, short-horizon numerical forecasting, and well-posed ontology mapping, while autonomous agentic operation and unvalidated occupant proxies remain research-stage. Current evidence therefore supports LLMs primarily as semantic and workflow layers rather than autonomous HVAC controllers. Future work should prioritise field-validated benchmarks, orchestration evaluation under operational constraints, and LLM-MPC/RL architectures with bounded latency and verifiable safety properties. --- ### How Does mHC Use Its Residual Streams? Selective Routing and Near-Identity Mixing Authors: Pengxiang Zhao, Xing Li, Xianzhi Yu, Wei Guo, Zhenhua Dong Published: 2026-09-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.05309v1 Hyper-Connections and their manifold-constrained variant mHC widen a residual pathway from one stream to n, yet how trained models use this capacity remains unclear: how broadly blocks read and write, how strongly the residual pathway mixes streams, and whether the streams carry distinct representations. We examine these properties in the four-stream residual pathway of DeepSeek-V4-Flash using effective stream counts, cross-stream residual weights, and inter-stream cosine similarity. Read/write routing is concentrated but varies across depth: a typical attention or FFN site effectively uses about two streams, while the dominant stream changes across layers and the representations remain directionally distinct. Residual mixing is modest and occurs primarily in early layers; in layers 22-42, the pathway mostly carries each stream forward separately. Targeted interventions establish the functional significance of these patterns. Replacing the late mixers by identity increases C4 perplexity by only 1.9% and preserves the six-task average score, whereas replacing the early mixers increases perplexity by 41%. Fixing each early mixer to its C4 diagnostic mean increases perplexity by only 0.2% and reduces the average score by 0.25 percentage points, showing that its site-specific structure matters more than its token-wise variation on the evaluated metrics. Likewise, retaining the three largest routing weights per token at every site increases perplexity by at most 2.7% and changes the average score by at most 0.4 points. Thus, the studied model realizes only part of the flexibility afforded by four-stream mHC: individual blocks rarely require all four streams, and late residual mixing provides little measured benefit. --- ### What Makes a Redundant Representation Remember? Lineage Isolation, Not Masking Authors: Jia Huang, Yangjun Ou Published: 2026-09-04 Categories: cs.NE arXiv: https://arxiv.org/abs/2609.05304v1 Memory-based evolutionary algorithms for dynamic optimization often carry a redundant second copy of the genotype and expose only one copy to the objective, on the assumption that the shielded copy accumulates information about past optima. We show this assumption is false as usually implemented, and identify the structural property that actually determines whether the shielded copy retains information. We formalize such methods as a gated dual-copy representation with two independent design axes: a gating rule deciding which copy is evaluated, and an inheritance rule deciding whether the two copies mix across generations. A ablation shows retained information is governed almost entirely by the inheritance rule (21.4 vs. 1.3 bits) and is nearly invariant to the gating rule. Per-locus independent inheritance reshuffles cross-locus structure every generation, so shielding preserves the variance of the hidden copy while destroying the pattern that constitutes a memory. Under isolated inheritance the memory effect is real: against a single-copy baseline matched for representation budget, the method gains +0.010 AUC when optima recur periodically and loses 0.078 when they drift unidirectionally---a 0.089 separation under otherwise identical settings, which excludes explanations based on added capacity. We show the readout rate is also the corruption rate, predicting and confirming an interior optimum replicated across two implementations. We report one negative result with a mechanism: dual-copy representations lower the mutational error threshold, because gated expression is a selector rather than a joint decoder and therefore provides no coding gain. Finally, we document a benchmarking hazard: on dynamic benchmarks the choice of recombination operator alone shifted our baseline by 0.062 AUC, six times the effect size under study. --- ### Learning Spatial-Spectral Refinement and Calibrating Complementary Observations for Hyperspectral Image Super-Resolution Authors: Liqian Yang, Xingchi Chen, Xinfeng Gui, Xiangyong Cao, Qianxin Yi Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05303v1 Hyperspectral and multispectral image fusion (HMIF) aims to reconstruct a high-resolution hyperspectral image (HR-HSI) by combining the fine spatial details of a high-resolution multispectral image (HR-MSI) with the rich spectral information of a low-resolution hyperspectral image (LR-HSI). Recent advances in implicit neural representations (INRs) have enabled flexible coordinate-based modeling for HMIF; however, existing INR-based approaches may not fully capture fine-grained spatial structures and rich spectral dependencies. Moreover, the LR-HSI and HR-MSI are primarily incorporated through degradation-consistency constraints, leaving their complementary information underexploited. To address these limitations, we propose Two-Stage Reconstruction with Implicit Tensor Neural Representation (TSR-ITNR), a unified self-supervised framework integrating representation refinement and observation-guided calibration. In Stage 1, TSR-ITNR learns an implicit Tucker representation and refines its low-rank spatial coefficient tensor and spectral basis to better capture fine spatial structures and interband correlations. A fixed pretrained denoiser further provides a deep prior for the preliminary reconstruction. In Stage 2, parameter-free calibration derives complementary and noninterfering corrections from both observations to recover information insufficiently captured in Stage 1. Theoretical analysis establishes the geometry-preserving property of spectral refinement and the orthogonal complementarity of calibration. Extensive experiments on multiple benchmark datasets demonstrate strong quantitative, visual, and spectral reconstruction performance without ground-truth HR-HSI supervision. Beyond conventional reconstruction metrics, we further assess the effectiveness of TSR-ITNR using downstream semantic segmentation accuracy. --- ### Online Change-point Detection for Cooperative Multi-Agent Reinforcement Learning Authors: Fatemeh Saberi Khomami, Julita Vassileva Published: 2026-09-04 Categories: cs.MA, cs.LG arXiv: https://arxiv.org/abs/2609.05298v1 Cooperative multi-agent reinforcement learning (MARL) systems rely on past experience for learning coordinated behaviour, but this experience may become unreliable if the environment or task objective changes during training. In such cases, agents first need a way to recognize that the situation has changed before deciding how to adapt. This paper studies online change-point detection for cooperative MARL using reward-derived signals. We propose \emph{Patterns of Past Rewards} (PPR), a lightweight algorithm-agnostic detector that smooths agents' return streams, highlights recent changes, and applies a statistical drift detector to flag significant shifts. We evaluate PPR in a custom Speaker-Listener environment based on the Multi-Agent Particle Environment under two controlled non-stationarity scenarios. Our results show a trade-off between detection speed and alarm stability. A smoothed-return baseline detects earlier but produces many repeated alarms. In contrast, applying the detector directly to raw returns often misses the shift. PPR offers a more balanced approach by limiting redundant detections while still identifying the controlled shifts. These findings highlight PPR as a lightweight, reward-based monitoring tool that enables cooperative MARL systems to reliably identify major changes during training. --- ### LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics Authors: Gaurab Baral Published: 2026-09-04 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.05296v1 Does a simplified legal clause still say what the original said? The checks in current use cannot establish that it does: requiring an identical pair to score highest and an unrelated pair lowest moves lexical overlap and legal force together, so any monotone function of token overlap satisfies both. Our remedy is a dissociation, an item holding surface form fixed while legal force moves. We release LexFlip, 373 minimal perturbations of Quebec statutory French that reverse legal force while preserving 0.93 of the tokens, with a harness scoring metrics, regressors and prompted judges alike. The seven embedding and BERTScore metrics we test spend only 0.022 to 0.039 of their identical-to-unrelated range on such an edit, against 0.670 for bidirectional NLI, the one family the identical-pair check would disqualify. On FrJudge, against a measured human ceiling of r=0.597, a bare length feature outscores every semantic metric and has the lowest margin we measure. --- ### RISE: Recursive Improvement via Self-Extrapolating Policy Distillation Authors: Yang Li, Semih Yavuz, Shafiq Joty Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05295v1 On-policy distillation (OPD) provides dense, per-token supervision for language model post-training, but its effectiveness is bottlenecked by teacher quality: external teachers suffer from distribution mismatch, while self-distillation with privileged conditioning is limited by in-context learning capacity. We propose \textbf{RISE} (\textbf{R}ecursive \textbf{I}mprovement via \textbf{S}elf-\textbf{E}xtrapolating Policy Distillation), which constructs a synthetic teacher directly from the model's own RLVR training trajectory. By extrapolating the displacement between the current checkpoint and a trailing anchor---in parameter space or output logit space---RISE converts a sparse outcome-induced parameter update into a dense token-level target, without any external model or privileged conditioning. RISE combines RLVR and OPD in a complementary loop: outcome rewards ground the extrapolation toward correct reasoning, while the extrapolated teacher refines token-level decisions. Moreover, since the teacher is refreshed every iteration as the student improves, distillation becomes a recursive improvement mechanism rather than a one-shot compression step. Experiments spanning mathematical reasoning, multi-domain STEM, code generation, and multi-turn agentic tasks show that RISE outperforms RLVR-only training and on-policy self-distillation across all settings. --- ### Learning from VAE Errors to support ECG-based Differential Diagnosis of Myocardial Scar Authors: Shayan Sharifi, Riccardo Treu, Ilaria Gandin, Federico Garoia, Marco Merlo, Giulia Cisotto Published: 2026-09-04 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.05294v1 Late Gadolinium Enhancement (LGE) on cardiac magnetic resonance is a key marker of myocardial scar, but its limited accessibility motivates routine ECG-based screening. We evaluated whether $β$-variational autoencoder (VAE)-derived ECG representations can discriminate LGE+ from LGE- cardiomyopathic patients in a local cohort of 300 subjects. We compared 32-dimensional features from the foundation ECGx.AI model with those from a shallower $β$-VAE trained on normal PTB-XL ECGs, evaluating downstream classification and Dynamic Time Warping (DTW)-based reconstruction errors. ECGx.AI reached an area under ROC of 0.686 with Random Forest, while the proposed $β$-VAE reached 0.577 with sensitivity of 0.775 with Gradient Boosting. Notably, DTW-reconstruction errors significantly differed between classes in 10 out of 12 leads according to Mann-Whitney U test and help in classification, leading to an area under ROC of 0.643 with Logistic Regression, supporting their potential as markers of scar-related ECG alterations. --- ### Beyond Aggregate Scores: Behavioral Correctness Assumptions for Assessing Reference-Based Automatic Evaluation Methods Authors: Maria Mahbub, Ashley Rice, Michael R. Munroe, Amidu Kamara, Amir Sadovnik Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05289v1 Automated reference-based evaluation methods play a critical role in assessing natural language generation systems. Existing meta-evaluation primarily measures agreement with human judgments or benchmark labels, providing limited insight into evaluator behavior under controlled conditions. We introduce behavioral correctness assumptions, a complementary framework for evaluating reference-based automatic evaluation methods. We define a taxonomy of correctness-preserving and correctness-altering assumptions and operationalize them through controlled response transformations that specify expected scoring behaviors. We evaluate diverse lexical, character-level, semantic, LLM-based, and hybrid evaluators and analyze their assumption-level behavior, stability, sensitivity, repeat-run variability, configuration sensitivity, and reproducibility. Our experiments reveal distinct behavioral trade-offs across evaluation paradigms: no evaluator satisfies all proposed correctness assumptions, and evaluators with similar aggregate performance can exhibit substantially different behavioral profiles. These findings demonstrate that behavioral correctness assumptions provide diagnostic information obscured by conventional aggregate meta-evaluation. --- ### GUT: Quantifying and Optimizing the Reasoning Uncertainty of LLMs via Graph Complexity Authors: Shuang Liang, Xin-Yu Hu, Xiang-Jun Ou, Shao-Qun Zhang Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05284v1 Recent years have witnessed great advances in the reasoning ability of Large Language Models (LLMs). However, the reasoning processes of LLMs often exhibit uncertainty, where LLMs often produce a proliferation of divergent branches at each reasoning step even when fed the same prompting inputs, and certain branches exhibit evidently incredible, even nonsensical, reasoning chains and results. In this paper, we propose the Graph-complexity-based UncerTainty (GUT) method for investigating the reasoning uncertainty of LLMs. The key idea of GUT is to characterize the potential branches of each reasoning chain with a directed acyclic graph, thereby ensuring that all potential branches are comprehensively covered within the graph space. Building upon this recognition, we further build two modules of GUT, that is, a Quantification (GUT-Q) module and an Optimization (GUT-O) module, for quantifying and reducing the reasoning uncertainty of LLMs, respectively. GUT-Q measures LLM reasoning uncertainty by approximating the reasoning space complexity with graph complexity. GUT-O implements uncertainty optimization by treating negative uncertainty as the reward function in reinforcement learning. Experimental results conducted on four LLMs and five datasets validate the effectiveness of GUT. --- ### Testing Interchangeability in LLM Agent Teams Authors: Jianxin Gao, Tianyi Yu, Linna Deng, Runze Li, Zining Wang Published: 2026-09-04 Categories: cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.05279v1 Production multi-agent systems replace agents constantly, on the assumption that an agent filling a role is interchangeable with any other agent that can do the job. We test that assumption. Eight teams per setting are formed independently from one base model on the same tasks, each agent keeping a private notebook across ten formation episodes; we then trade role-matched agents between teams and measure what changes on held-out tasks. Against a placebo that reproduces the disruption of a roster change without changing who occupies the seat, a swap costs little in task score but raises the communication a team spends per unit of progress by 16 to 63 percent, and in Hanabi a swapped agent is more expensive than an inexperienced one, consistent with interference from conventions learned with its former partner. In Collab-Overcooked, when the agent that sets the agenda is replaced, most of the extra communication comes from the agent that stayed. Three ablations, over base models, decoding temperature and formation length, move the swap penalty alongside one other quantity: how far independently formed teams drift apart. Greedy decoding lowers both; doubling a team's history raises both. In these settings, agents are more fungible in task outcome than in coordination efficiency, with larger swap effects after longer formation histories. ---