# Research Papers — 2026-09 · bundle page 100 of 266 > Full-text markdown of papers #4951–5000 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/100 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/99.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/101.md) ### Spiking Neural Network Predicting Sequence of the External Worlds States in Model-Based Reinforcement Learning Authors: Mikhail Kiselev Published: 2026-09-23 Categories: cs.NE arXiv: https://arxiv.org/abs/2609.27459v1 This paper presents a spiking neural network (SNN) designed to predict the sequence of the external world states starting from the current world state. This SNN does not create the world dynamics model - instead it incorporates the SNN trained to predict the next world state and provides all mechanisms necessary to make the chain of predicted world states. These mechanisms are entirely spiking - they are implemented as spiking neuron ensembles. The present article describes this neuronal structure and tests its operation on a classic RL benchmark - ATARI ping-pong. --- ### Beyond Balanced Accuracy: A Resolution and Parity-Controlled Benchmark for Vision-Language and Vision-Only Defect Assessment in UAV Power-Line Inspection Authors: Linghao Zhang, Siyu Xiang, Junwei Kuang, Peiyu Yi Published: 2026-09-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.27457v1 Vision-language models (VLMs) are often reported to outperform task-specific vision backbones for unmanned aerial vehicle (UAV) power-line defect assessment. We test that claim on ElecVQA-Bench, a 56,972-item benchmark derived from the public InsPLAD dataset, across six evaluation choices: partition, evaluated item set, label space, replication, input resolution, and side information. On a matched partition, a Swin Transformer and the strongest adapted VLM differ by only 0.03 points at binary screening. At seven-way defect typing, increasing the vision backbones from 224 px to the measured pixel budget of the VLM preprocessor narrows the gap against InternVL3.5-8B from +20.53 to -0.57 points for ResNet-50 and from +23.67 to +4.70 points for Swin-T. A pixel-budget audit shifts Qwen3-VL-8B macro recall by 10.78 points, yet a source-pixel-matched InternVL control still leaves Qwen ahead by 7.43 to 13.61 points while using 56% fewer visual tokens, so neither source pixels nor token budget explains the difference between the two VLMs. A two-seed global replication changes Qwen binary accuracy and seven-way macro recall by 0.86 and 1.02 points. After split-specific retraining, Qwen does not lead at crop or image level, and a 14-tower, three-seed replication reverses the sign across seeds, giving mean common-six macro recall of 0.9085 for Qwen against 0.9509 for ResNet-50. No split regime yields a family-level advantage that survives multiple-comparison correction. The study supports a benchmark-audit contribution rather than a general claim of VLM superiority. --- ### Latent evolving World Action Model Authors: Xueji Fang, Boqiang Duan, Hua Wu, Jingdong Wang, Guo-Jun Qi Published: 2026-09-23 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.27455v2 World Action Models (WAMs) jointly model action generation and environment dynamics and are mostly built on pretrained Video Diffusion Models (VDMs). In VDM-based WAMs, observations are first encoded by a VAE, and the resulting compressed latents are then processed by large video diffusion backbones to extract effective features for action generation. However, this paradigm ties WAM performance and training cost to large-scale video generation pretraining, limiting WAM efficiency and scalability. In this paper, we theoretically and empirically investigate how visual representations affect action generation in WAMs. Our results show that predictive embeddings from Joint-Embedding Predictive Architecture (JEPA) encoders better support action generation than compressed VAE latents, with I-JEPA performing best in our encoder comparison. Based on these findings, we propose LeWAM, which conditions action generation on JEPA embeddings and models environment evolution by predicting future embeddings in the same space, without relying on a video diffusion backbone. We further find that imitation learning matches demonstrated actions but does not distinguish better actions from worse ones, even though small action deviations can greatly affect task success. To address this limitation without additional environment interaction or the human oversight required for resets and safety, we introduce Demonstration-Guided DPO (DemoDPO), an offline preference refinement stage that derives preference supervision directly from demonstrations. With only 0.4B trainable parameters, LeWAM achieves an average success rate of 92.28\% on RoboTwin 2.0, comparable to that of state-of-the-art VLAs and WAMs, and maintains practical effectiveness on real-world manipulation tasks. --- ### Issuer-Sovereign Agentic Payments Authors: Dishant Sharma, Rajneesh Kaushal, Ashu Kanaujia Published: 2026-09-23 Categories: cs.CR, cs.AI, cs.SE arXiv: https://arxiv.org/abs/2609.27452v1 AI agents are beginning to make real payments. Current approaches let an agent pay by relying on a credential provider that, in the approaches deployed today, typically sits outside the cardholder's bank. The spending rules are then enforced by the card network or that provider, and not by the bank itself. This leaves the issuing bank, which carries the financial risk, with little direct control at the moment a payment happens. This paper describes Issuer-Sovereign Agentic Payments, a method that keeps that control with the issuer. The cardholder approves a spending rule once, and the bank's own authentication component records it. Later, when the agent pays a specific merchant, the bank checks the merchant against the approved rule and generates the card authentication value only if the merchant is allowed. The payment then travels the normal card rails and is validated by the issuer, with no extra dependency introduced at execution. --- ### BEE: Intervention-Adaptive Real-World Reinforcement Learning with Vision-Language-Action Models Authors: Weihui Zhao, Xiaohan Yan, Zunian Wan, Xuan Du, Zhaozhan Chi, Jianbo Mao, Ruipu Wu, Rushuai Yang, Houlin Li, Shukai Yang, Jing Wu, Yuxiang Yan et al. Published: 2026-09-23 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.27450v1 Vision-language-action (VLA) models handle long-horizon manipulation, yet success hinges on a few precision-critical phases where millimeter-scale errors undo all prior progress. Online reinforcement learning (RL) can optimize exactly these actions, but free exploration is far too costly on real robots, which makes human corrections indispensable. However, existing online RL methods for VLAs either cannot incorporate such corrections or fold them into undifferentiated supervision. Yet human corrections are not uniformly noisy but reliable along some action dimensions and variable along others. Building on this, we introduce BEE, an intervention-adaptive framework for real-world RL on a frozen VLA that lets the policy go BEyond Expert imitation. We formulate human corrections not as actions to reproduce but as evidence about a constraint: a Correction Model predicts how a human would correct a given VLA proposal and how consistent the correction is along each action dimension. This predicted consistency sets the per-dimension tightness of a constraint on policy optimization. Where corrections are consistent the policy stays close to the human, and where they vary, the constraint relaxes. We evaluate BEE on three real-world manipulation tasks and one LIBERO-Pro simulation task at a matched online-data budget. BEE attains the highest success rate on every task, 91.2% on average against 57.5% for RLT and 42.1% for DSRL, and the lowest human intervention rate on all real-world tasks. --- ### Quantum Reinforcement Learning for Cost and Delay Tradeoffs in Quantum Cloud Orchestration Authors: An N. H. Phan, Dang Van Huynh, Muhammad Usman, Hoa T. Nguyen Published: 2026-09-23 Categories: cs.LG, cs.AI, cs.DC, cs.ET, quant-ph arXiv: https://arxiv.org/abs/2609.27446v1 Quantum cloud computing, delivered through the quantum-as-a-service (QaaS) model, provides access to quantum computing resources. However, applying uniform time-based pricing across fundamentally heterogeneous quantum resources significantly complicates task orchestration, particularly when addressing the tradeoff between execution costs and system performance. While heuristic methods rely on predefined scheduling rules, classical deep reinforcement learning (DRL) models may require more trainable parameters in this setting. Motivated by the potential of parameterised quantum circuits (PQCs) as compact function approximators, we propose QRLQ, a cost-delay-aware quantum cloud scheduling framework integrating PQCs with a dueling double deep Q-network (D3QN) to dynamically account for both cost and delay. Our simulation results show that QRLQ achieves lower mean cost and delay than the heuristic baselines, achieving a 5-11% lower mean cost relative to availability-based and rotation-based heuristics and reducing mean delay by 17% and 82% relative to the strongest and weakest heuristic baselines, respectively, while retaining execution fidelity within 2% of a fidelity-greedy policy. Compared with the classical DRL baseline, QRLQ achieves comparable scheduling performance while using 72% fewer trainable parameters. This work explores the feasibility of using QRL for task orchestration in quantum cloud environments and demonstrates its potential for cost-delay-aware quantum resource management. --- ### SatUnreal: A High-Precision Synthetic Dataset for Satellite Stereo Matching via Unreal Engine Authors: Han-Gyeol Kim, JaeWan Park, Junmin Park, Darongsae Kwon Published: 2026-09-23 Categories: cs.CV, eess.IV arXiv: https://arxiv.org/abs/2609.27442v1 3D reconstruction from satellite imagery is essential for large-scale topographic analysis, yet the lack of high-fidelity training datasets with accurate occlusion labels remains a primary bottleneck. Existing benchmarks, such as US3D and WHU-Stereo, face inherent challenges in spatio-temporal mismatch -- environmental changes and shadow displacements between multi-view acquisitions -- and provide ambiguous ground truth in occluded regions due to LiDAR sparsity. In this paper, we propose SatUnreal, a high-precision synthetic dataset designed to fundamentally overcome these limitations through an Unreal Engine-based simulation pipeline. SatUnreal provides 10,000 stereo pairs with high resolution (0.3m GSD) and is characterized by: (1) Physical Geometry Simulation, replicating realistic satellite orbits by systematically varying baselines and azimuths; (2) Spatio-temporal Consistency, eliminating temporal noise through fixed virtual environments; (3) Topographic Diversity, spanning dense urban canyons to low-texture natural terrains; and (4) Mathematical Label Integrity, utilizing a novel two-step linetrace algorithm to generate flawless occlusion masks. Experimental results using SOTA iterative models demonstrate that models trained exclusively on SatUnreal achieve superior zero-shot transfer performance on real-world benchmarks (US3D, WHU-Stereo) compared to those trained on real datasets. Our findings prove that physically accurate synthetic data provides a more effective supervisory signal for learning geometric features than complex real-world observations, establishing a new paradigm for Sim-to-Real transfer in Earth Observation. Code and dataset are available at https://github.com/jmp-Telepix/SatUnreal_A_High-Precision_Synthetic_Dataset_for_Satellite_Stereo_Matching_via_UnrealEngine --- ### Stable Neural Decoding Across Sessions via Task-Conditioned Latent Alignment for Brain-Machine Interfaces Authors: Canyang Zhao, Bolin Peng, J. Patrick Mayo, Ce Ju, Bing Liu Published: 2026-09-23 Categories: cs.LG, eess.SP arXiv: https://arxiv.org/abs/2609.27441v1 Achieving stable long-term neural decoding in invasive brain-machine interfaces (BMIs) remains challenging due to variations in recorded neural populations across sessions. Current latent alignment approaches may overlook task-dependent structure during cross-session adaptation. We propose Task-Conditioned Latent Alignment (TCLA), a framework that stabilizes neural decoding by learning a shared latent space. TCLA learns a low-dimensional source representation using neural reconstruction and continuous behavioral supervision. During target-session adaptation, the shared representation is fixed, while target neural activity is mapped into the source latent space by aligning source and target distributions separately for each task condition. We evaluated TCLA on seven nonhuman primate datasets spanning multiple tasks. In long-term cross-session evaluation, TCLA achieved a mean $R^2$ of $0.476\pm0.014$ with a negative $R^2$ failure rate of only 6.8\%. Across 1,356 within-subject session pairs, TCLA achieved a mean $R^2$ of $0.371\pm0.009$ with a failure rate of 6.8\%. Across 2,134 cross-subject session pairs, TCLA achieved a mean $R^2$ of $0.218\pm0.004$ with a failure rate of 12.9\%, substantially better than those of the comparison methods. These results demonstrate that by preserving behaviorally relevant and task-dependent latent structure, TCLA improves the robustness of neural decoding across recording sessions and subjects. The source code is publicly available at \href{https://github.com/FAMD-CASIA/TCLA}{https://github.com/FAMD-CASIA/TCLA}. --- ### An Unbounded Archive-based Transfer Strategy for Dynamic Multi-Objective Optimization with a Changing Number of Objectives Authors: Zhiyun Xiao, Ke Shang, Yajun Liu, Jianguo Li, Shaojiang Wang, Wei Sun Published: 2026-09-23 Categories: cs.NE arXiv: https://arxiv.org/abs/2609.27430v1 Dynamic multi-objective optimization with a variable number of objectives is difficult because objective-dimensional variations may significantly change the Pareto front and degrade algorithm adaptability. This paper proposes an unbounded archive-based transfer strategy (UATS), which maintains an unbounded archive of offspring solutions within each environment stage and extracts feasible nondominated solutions as transferable elites when objective changes occur. UATS is embedded into SPEA2SDE to construct UATS-SPEA2SDE, enabling the algorithm to reuse historical evolutionary information while retaining the convergence and diversity advantages of shift-based density estimation. Experiments are conducted on four benchmark problems under three objective-changing settings, where UATS-SPEA2SDE is compared with a restart-based SPEA2SDE baseline and four representative dynamic multi-objective optimization algorithms. The results indicate that the archive-guided transfer improves recovery after environmental changes and enhances adaptability to objective-number variations. --- ### Overlapping Visual Grouping Without Semantic Priors Authors: Teemu Saukkio, Hashem Haghbayan, Juha Plosila Published: 2026-09-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.27423v1 Most computer-vision systems organize visual input toward a predefined interpretation, such as semantic categories, prompted regions, learned object-like representations, or a single spatial partition. This work considers an earlier stage of visual organization: the formation of candidate perceptual units directly from sensor measurements before their identity, meaning, or task relevance is known. We introduce Domain Parent Grouping (DPG), a sensor-grounded grouping method in which complementary measurement relationships are represented in separate processing domains. Spatially connected groups formed within these domains are related through cross-domain overlap, yielding a non-exclusive grouping representation rather than a single mutually exclusive segmentation. This representation retains broader and more localized groups, as well as alternative grouping boundaries over the same image locations, simultaneously available. DPG also includes a native mechanism for reprocessing selected group content, in which input-relative measurement ranges allow the observational resolution to change while preserving previously formed groups. DPG is implemented using three domains representing locally contextualized luminance, direct chromatic relationships, and contextual chromatic relationships. Experiments on the BSDS500 dataset demonstrate the benefit of combining the three domains. The results further show that DPG forms measurement-supported groups corresponding to low-level image structure, and that these groups exhibit measurable correspondence with human-annotated regions and boundaries. This demonstrates that structured visual organization can emerge directly from relationships among sensor measurements. --- ### Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following Authors: Yanzhao Zheng, Yuanqiang Yu, Tianze Xu, Chao Ma, Zhentao Zhang, Jihuai Zhu, Baohua Dong, Hangcheng Zhu, Ruohui Huang Published: 2026-09-23 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.27421v1 Multi-constraint instruction following requires a model to respond to a query under many simultaneously active constraints. Even strong instruction-tuned models still routinely violate some of them. Existing approaches either augment supervision with sequence- or token-level RL rewards from external verifiers or learned graders, or use on-policy distillation (OPD) against a single full-context teacher whose probability mass becomes diluted as more constraints become simultaneously active. We propose CC-OPD (Counterfactual Constraint-Conditioned On-Policy Distillation), which inverts the standard supervision-generation direction in distillation. Rather than enriching the teacher with information beyond what the student sees, CC-OPD ablates each constraint from the teacher's conditioning in turn, and constructs the per-constraint signal from the resulting per-token probability differentials. The resulting per-token leave-one-out log-likelihood shifts are summed, clipped, and added to the vanilla OPD reward as a token-level shaping term. All shaping terms are obtained from the frozen teacher, without an external verifier during distillation, and the reward equals vanilla OPD wherever the aggregate shift is zero. Across two Qwen model pairs and seven benchmarks, CC-OPD achieves the highest average among all evaluated student-training methods. A 1.5B student trained with CC-OPD surpasses its own 7B RL-trained teacher on the MulDimIF benchmark. --- ### EviStreams: Human-in-the-Loop AI Data Extraction for Systematic Reviews in Medicine Authors: Sai Karthik Kosuri, Ankita Shashikant Bhosale, Michael Glick, Alonso Carrasco-Labra, Chris Callison-Burch Published: 2026-09-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.27418v1 Systematic reviews underpin clinical guidelines, yet their data-extraction step is a major expert-labor bottleneck bound by a protocolized workflow: two reviewers extract each study independently, an adjudicator resolves disagreements, and the team keeps an auditable record of how every value was produced. Large language models can assist with extraction, but that assistance must fit established review protocols and preserve reproducibility. We present EviStreams, a live, open-source, no-code web platform that puts review teams in control of AI-assisted extraction at three key stages: program design (a structured decomposition approved before any code runs), field specification (typed field definitions calibrated from a pilot), and extracted predictions (reviewer-blinded dual review with adjudication). Working through a form builder, a domain expert defines typed fields rather than prompts, runs extraction over uploaded PDFs, inspects every value alongside the supporting passage it came from, and resolves a reviewer-blinded dual review into an auditable consensus export. An evaluation across four clinical corpora and three frontier model families, released with the system, shows that extraction quality is shaped far more by the field specification than by the choice of model. EviStreams is live at https://evistreams.com/demo and released under Apache-2.0. --- ### Emergi-PersonaOS: A Persona Agent Operating System for Situational Adaptation and Controllable Evolution Authors: Haoluan Fu, Keni Chen, Xinyu Jia, Jinpeng Wang, Yuyu Yin, Yubiao Hu Published: 2026-09-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.27417v2 Symbiosis between humans and digital beings offers a vision for the future of human--machine interaction. In enduring human--machine relationships, personality provides a foundation for continuity of identity, individuality in interaction, and development through experience. We investigate this capacity through persona agents as computational implementations and introduce Emergi-PersonaOS, a psychology-grounded operating system for managing persona objects throughout their lifecycle. The system organizes dispositional traits, characteristic adaptations, and narrative identity into a three-layer persona representation, distinguishing relatively enduring persona beliefs from their activation in the current persona state. During situational adaptation, it integrates the current interlocutor, relationship, event, and retrieved memories to infer a persona state and generate actions and replies; during long-term development, it records experiences and outcomes, and develops and evaluates revision candidates through change attribution, meaning-making, and behavioral testing. Belief updates are managed through explicit review, traceable evidence and version records, and the ability to reject candidates, making persona evolution controllable. Using television-character dialogue as longitudinal material, we demonstrate long-horizon system operation and examine its principal mechanisms in a concrete implementation. This work provides a computational framework for persona agents to maintain individual continuity, produce situation-specific expression, and develop through experience over sustained interaction. --- ### S2A:Semantic-to-Spatial Alignment for Alignment-Free RGB-T Salient Object Detection Authors: Qiangqiang Zhou, Yang Luo, Yong Chen, Jiawei Xu Published: 2026-09-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.27413v1 Alignment-free RGB-T salient object detection (RGB-T SOD) aims to identify salient objects from unregistered RGB and thermal image pairs without costly pre-alignment. However, spatial misalignment breaks pixel-wise correspondence and causes feature contamination during cross-modal fusion. To address this issue, we propose S2A, a semantic-to-spatial alignment framework for alignment-free RGB-T SOD. Specifically, a global-guided hierarchical fusion module (GGHF) first exploits global semantic guidance to suppress background interference and refine hierarchical intra-modal features. Subsequently, the alignment-free cross-modal channel attention module (AFCA) globally exchanges complementary semantic information through channel-wise interaction, effectively overcoming the interference caused by local spatial misalignments. Finally, a spatial deformable cross-attention module (SDCA) predicts adaptive sampling offsets to recover local cross-modal spatial correspondence. Through this semantic-to-spatial paradigm, S2A first enables reliable cross-modal semantic interaction and subsequently performs local spatial calibration, effectively reducing misalignment-induced feature contamination. Without bells and whistles, S2A achieves highly competitive performance on multiple public alignment-free RGB-T benchmarks, demonstrating its effectiveness in alleviating misalignment-induced feature contamination. --- ### When Labels Are Scarce: An Oscillatory State Space Model for Vibration Diagnosis Authors: Mainak Mallick, Seung-Kyum Choi Published: 2026-09-23 Categories: cs.LG, eess.SP arXiv: https://arxiv.org/abs/2609.27411v1 Machine fault diagnosis from vibration requires learning from scarce labelled fault recordings while meeting the computational constraints of edge devices for local inference. We introduce DualRes, a compact oscillatory state-space model that combines two complementary spectral views of vibration, capturing rapid changes and fine frequency structure. Time-aligned views are processed by selective oscillatory memory, which learns how long to retain temporal patterns. The encoder contains 39,528 parameters. We evaluate supervised learning across six bearing datasets and a gearbox benchmark, with an additional gearbox pilot. Recording-level splits and explicit accounting of labelled duration distinguish data efficiency from repeated exposure to correlated samples. On the main gearbox benchmark, DualRes achieves state-of-the-art performance among the nine evaluated methods at six of seven label budgets. With about six labelled seconds per class, it improves macro-F1 by 16.1 percentage points over the next strongest comparator. On the same benchmark, DualRes achieves a 1.44-fold recording-level speedup and a 24.8-fold reduction in checkpoint storage relative to a selective state-space baseline under matched hardware and runtime conditions. Bearing results reveal task-dependent trade-offs. These findings support oscillatory memory as a compact approach to vibration diagnosis under limited labelled exposure. --- ### Active Learning for Biodiversity Monitoring: From Label Efficiency to Reliable Ecological Inference Authors: Ben McEwen, Shiqi Zhang, Dan Stowell Published: 2026-09-23 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.27409v1 Limited expert annotation capacity is a pervasive constraint in biodiversity monitoring. Passive acoustic recorders and camera traps generate data faster than experts can analyse them. Machine learning (ML) models can process these data at scale, but their reliability depends on the quality, quantity, and coverage of labelled samples, so expert time remains a constraint. Active learning (AL) eases this bottleneck by selecting, under a fixed annotation budget, the samples expected to improve a model most, and published evidence shows it can reduce the labels needed to reach a target performance. Monitoring programmes, however, face a broader question: how should a limited expert budget be divided so that model training, validation, and the ecological estimates built on model outputs all remain reliable? Because AL selects samples non-randomly, its labels are unsuitable for validation, calibration, or threshold selection, a tension rarely acknowledged. We synthesise AL research across acoustic and image modalities and identify gaps and opportunities. Most studies evaluate query strategies on pre-labelled benchmarks with simulated annotators; deployments in real monitoring workflows are rare and concentrate on birds and cetaceans. Bats, insects, amphibians, and fish are underrepresented, and multimodal applications remain largely unexplored. Evaluation centres on headline reductions in annotation effort, often without random-sampling baselines, per-class results, or calibration analysis, and rarely accounts for the labels required for validation. We provide a tutorial treatment of the AL loop that makes these budget decisions explicit, and a roadmap towards AL methods that support label-efficient training, validation, and trustworthy downstream ecological inference. --- ### What Looks Like a Capability Limit in Vision-Language Models Is a Readout Limit Authors: Alfredo F. Frontera Del Valle Published: 2026-09-23 Categories: cs.CV, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.27408v1 Benchmarks for vision-language models offer their answer choices in some convention: a letter, a color name, a pixel coordinate. That convention is treated as neutral. We find it is not, and that the limits a benchmark reports can belong to the readout rather than to the model. On 200 COCO photographs, Qwen3-VL-4B picks the correct one of nine locations for a named object 68.5% of the time when the locations are given in English and 20.0% when the same locations are given as pixel coordinates. Chance is 11.1%. The cost arises when the answer options are coordinates; giving the model a coordinate in the question instead costs 3.5 points and is not significant. The gap holds on a 4x4 grid, under 8-bit rather than 4-bit quantization, and in every slice by object size, boundary distance and category. It also decides which model wins. Two models that tie under English names differ by 39 points in one coordinate system and by 54 in the other, in opposite directions. On the color task, three of the four open models capable of the task show the penalty; on photographs, two of three open models do, and so does Gemini, at 11.1 points on parseable answers (p = 1e-4). GPT-4o does not. To ask whether a model reads a coordinate at all, we attach the wrong name to each one and record which the model follows. Color options written as hue angles are followed below chance; a normalized pixel convention is followed at four times chance. This tells apart conventions a model can use from ones it cannot, though it did not predict accuracy on two untried conventions. Five models also name the same color wheel five different ways, so a fixed answer vocabulary is not neutral across models either. Five times during this work we measured a capable model as incapable because our scorer and the model disagreed about what an answer looks like. We report each case. They are the phenomenon in miniature. --- ### Forget who you Forgot: Speaker Unlearning to Prevent Re-Identification in Zero-Shot Text-to-Speech Authors: Hyoeun Kim, Yujun Lee, Kyuhong Shim Published: 2026-09-23 Categories: cs.SD, cs.AI arXiv: https://arxiv.org/abs/2609.27399v1 Recent zero-shot text-to-speech (ZS-TTS) systems can reproduce a speaker's voice with high fidelity from only a few seconds of reference speech, raising concerns over unauthorized voice cloning and impersonation. Speaker identity unlearning has recently emerged as an approach to selectively suppress this capability for speakers who opt out while preserving synthesis capability for other speakers. Although existing approaches reduce speaker similarity, preventing re-identification often faces severe degradation of speech quality. Motivated by this observation, we propose GUARD, a lightweight speaker identity unlearning framework that combines a learned speaker gate with speaker-agnostic activation steering on a frozen TTS backbone. The steering vectors are optimized using group-relative reward optimization to shift outputs from forget speakers toward population-level impostor similarity while preserving intelligibility and speech naturalness. On CosyVoice2, GUARD reduces forget-speaker similarity from 0.541 to 0.103 and re-identification accuracy in a 150-speaker gallery from 73.5% to 0.5%, while preserving retain-speaker reproduction. The results demonstrate that similarity reduction alone may not fully characterize successful speaker identity unlearning and highlight re-identification as a complementary criterion for its evaluation. --- ### When Parallel Drafter Meets Parallel Speculative Decoding Authors: Fuliang Liu, Xue Li, Kun Qian, Zhibin Wang, Wanchun Dou, Wenyuan Yu, Chen Tian Published: 2026-09-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.27396v1 DSpark-style parallel drafters have made speculative decoding highly effective, yet their draft phase remains serialized on the critical path of every round. Parallel speculative decoding (PSD) overlaps drafting with verification, yet existing methods must guess the accepted prefix and bonus token in advance: a wrong guess reverts the whole batch to serial drafting. We present DPara, a PSD framework that reuses effective parallel drafters yet guarantees backbone--verification overlap in every round, thereby eliminating this probabilistic fallback altogether. While the target verifies, DPara's diffusion backbone precomputes draft representations for every acceptance boundary with the bonus left unspecified; a lightweight autoregressive head then combines the revealed verification outcome with the matching precomputed representation to emit the next round's draft tokens almost instantly---fully parallelizing the dominant backbone forward with verification and leaving only the negligible head cost serial. Experiments on Qwen3-8B and Qwen3-14B across seven math, coding, and chat benchmarks show that DPara achieves average speedups of $3.21\times$ and $3.52\times$ over autoregressive decoding, surpassing the strongest serial and parallel speculative decoding baselines alike. --- ### PRISM-VLM: A Multi-Axis Discriminative Benchmark for Compact Vision-Language Models Authors: Sanghee Park, Kee-Eung Kim Published: 2026-09-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.27395v1 Compact vision-language models (VLMs) now power a growing share of multimodal applications. The benchmarks used to compare them, however, inherit a frontier-centric design: each model is reduced to a single accuracy number, narrowing the inter-model gap on saturated suites and pressing models into low-score bands on harder ones. We introduce PRISM-VLM, a multi-axis discriminative benchmark that scores every item along seven axes covering the recurring failure modes (task quality, behavioral robustness, and capability bottlenecks) and combines them into a single PScore, with items recycled from fifteen public benchmarks. Across compact VLMs from the past two years, PScore separates model pairs more reliably than prior single-axis benchmarks under an item-level paired bootstrap, and surfaces behavioral differences these benchmarks average away. Even models with statistically indistinguishable PScores diverge sharply along the per-axis profile, particularly on sycophancy, which is nearly orthogonal to single-prompt accuracy. We will release the full pipeline, prompts, and per-item annotations. --- ### Automotive mmWave Spinning Radar Place Recognition with Spatially Gated Feature-Correlation Representation Authors: Saimunur Rahman, Sagun Singh Shrestha, Abdelwahed Khamis, Peyman Moghadam Published: 2026-09-23 Categories: cs.RO, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.27394v1 Automotive spinning FMCW radar provides dense, $360^\circ$ sensing and remains reliable under poor illumination and adverse weather, making it well-suited to autonomous navigation. Place recognition uses these observations to identify previously visited locations for re-localization and long-term navigation. However, heading changes appear as circular shifts in the polar radar representation, and conventional global aggregation can lose relationships among radar responses that are important for distinguishing similar places. We propose SGCA-Net, a spinning radar place recognition framework that combines rotation-robust feature extraction with Spatially Gated Correlation Aggregation (SGCA). SGCA learns spatial weights to reduce the influence of unstable and ambiguous radar regions, while aggregating pairwise correlations among local responses to preserve informative feature relationships. Experiments on the MulRan dataset show that SGCA-Net consistently outperforms SOTA methods across urban, campus, and open-road environments, while remaining robust to substantial heading variation. Evaluation on the HeRCULES dataset further demonstrates that SGCA-Net generalizes to unseen environments and radar sensors without fine-tuning. --- ### EvoAudio: Recursive Self-Improvement for Audio Understanding Authors: Yuxiang Wang, Shengbo Cai, Yingda Shen, Ming-Hao Hsu, Qinke Ni, Liqiang Zhang, Teddy Sun, Steve Yevs, Zhizheng Wu Published: 2026-09-23 Categories: cs.SD, cs.LG arXiv: https://arxiv.org/abs/2609.27389v1 Audio language models understand what is said far better than how it sounds. Closing this gap takes more than data. Detailed acoustic annotation is costly, labels from stronger models inherit their errors and limits, and fixed data cannot adapt as the learner improves. We therefore propose EvoAudio, a recursive self-improvement system for audio understanding. To our knowledge, it is the first to evolve the model, waveforms, questions, and difficulty in one closed loop. EvoAudio uses the current model's performance to set the focus and difficulty of the next training data. A library of audio tools then constructs questions whose answers follow from how the audio was made, providing verifiable supervision without new human annotation. Reinforcement learning updates the model, and validation decides whether it enters the next evolution round. Across 13 rounds, EvoAudio improves five models with different audio encoders and language backbones on MMSU, MMAU-Pro, and MMAR. It achieves the highest average for every backbone, raising overall performance by up to 6.3 points. The improvement unfolds over successive rounds, with each stronger model starting the next round. --- ### AraGenre 2026: A Hierarchical Definition-Guided Arabic Genre Classification Shared Task Authors: Mo El-Haj, Saad Ezzini, Shadi Abudalfa, Mustafa Jarrar, Nguyen Minh Chi, Nguyen Minh Quan Published: 2026-09-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.27387v1 AraGenre is a shared task on hierarchical, definition-guided Arabic genre classification, motivated by the limited availability of annotated data in Arabic and other low-resource languages. Systems assign each Arabic text segment both a broad communicative genre and a fine-grained specific genre. The released training and development sets contain limited, primarily synthetic and controlled examples, whereas the hidden final benchmark contains noisier naturally occurring text spanning Modern Standard Arabic, Classical Arabic, and multiple dialects. Participants received natural-language definitions for 74 previously unseen specific genres, creating a zero-shot label generalisation setting in which systems had to infer class semantics rather than memorise fixed label-feature associations. The task attracted 46 registrations and 373 submissions, with 17 teams completing the final evaluation. Thakaa ranked first with a Hierarchical Macro F1 of 0.7352, followed by HoangPhong (HP) with 0.7169 and NAMAA with 0.7013. The results show strong broad-genre recognition but a substantial gap in fine-grained classification under linguistic and domain variation. --- ### Forecast Workflow Bench: Evaluating Language-Model Decisions with Budgeted Forecast Tools Authors: Shunya Nagashima Published: 2026-09-23 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.27385v1 Time-series foundation models (TSFMs) provide forecasts for operational decisions, but accuracy alone does not determine their value. Evaluating agents that use these models requires measuring decision quality and forecast cost. FWBench evaluates this capability on 1,251 electricity and cycle-hire cases using fixed forecast tools and simulated capacity contracts. Agents select models, histories and horizons, then submit capacities to minimize a stated loss-cost objective. We evaluated two hosted and eight local configurations, including small language models, and tested local models with and without TSFMs. GPT-6 Astra bought inexpensive short-horizon forecasts selectively, using 2.5% of the budget, and outperformed fixed policies when the saved decisions were scored with three loss-cost weightings. FWBench enables reproducible evaluation of how language models select and use time-series forecasts to make decisions under cost constraints. --- ### When Entanglement Lower-Bounds Disparity: Auditing and Repairing Demographic Fairness in Audio Understanding Models Authors: Kian Shamsaie, Iman Modarressi Published: 2026-09-23 Categories: cs.SD, cs.CL arXiv: https://arxiv.org/abs/2609.27382v1 Speech technology penalizes some voices: recognition errs nearly twice as often for Black speakers, and accuracy declines for second-language accents and older speakers. We introduce TRIAD, an audit grid crossing 120 texts, 24 rendered demographic voice profiles (gender, age band, accent), and ten expressive styles via controllable text-to-speech, isolating perceived demographic attributes from content and affect. For ten open-weights encoders we define axis-fidelity functionals, principal-angle leakage between axis subspaces, and group-conditional gaps; a proposition proves that average probe disparity grows with the same aggregate voice-semantic leakage $Λ$ we measure, and a corollary shows that peak leakage forces worst-case disparity inside an active region. The measured mean-square probe disparity tracks $Λ$ (Pearson r = 0.93), and a black-box protocol exposes the same signature in two closed-source models. ORCA, an adapter combining axis-specific contrastive heads, an orthogonality penalty, and group-balanced sampling, cuts leakage 72% and roughly halves the gaps. --- ### MORSE: Multi-Context Ordering via Reverse Scoring for Evidence-Preserving Compression Authors: Ke Wan, Yifan Wang, Liheng Lai, Chen Chen Published: 2026-09-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.27380v2 Retrieval-augmented generation often relies on multiple retrieved contexts that contain substantial redundancy, motivating context compression to preserve useful information under limited input budgets. Likelihood-based compressors can account for cross-context redundancy through sequential scoring, but this makes evidence scores dependent on context order. We show that permuting the same contexts under an unchanged compressor can substantially change which supporting evidence survives compression. We attribute this sensitivity to information preemption: earlier, partially relevant contexts can absorb credit for shared information, reducing the incremental scores of later, stronger evidence and increasing its risk of removal. Controlled pair-swap interventions provide direct empirical support for this mechanism by showing that placing stronger evidence before overlapping, partially relevant contexts can improve its survival. Based on this insight, we introduce MORSE, a compression-aware method for evidence-preserving context ordering. MORSE uses reverse query likelihood to construct an evidence-first anchor and to evaluate compressed candidate outputs, enabling compression-aware selection among alternative permutations. Across multi-hop Question Answering (QA) benchmarks, compression procedures, budgets, and scoring models, MORSE improves evidence retention over reverse ordering and generally outperforms matched random search, with downstream QA gains. Our code is available at https://github.com/tbn5pj/MORSE_code --- ### Psychoacoustically Aligned Latent Smoothing for Adversarial Robustness of Full-Duplex Speech-to-Speech Dialogue Models Authors: Kian Shamsaie, Iman Modarressi Published: 2026-09-23 Categories: cs.SD, cs.AI, cs.CL, cs.HC arXiv: https://arxiv.org/abs/2609.27378v1 End-to-end speech-to-speech dialogue models listen and speak simultaneously, so a continuously open acoustic channel is exposed to adversarial manipulation. We formalize imperceptible attacks on full-duplex agents as optimization over additive perturbations confined beneath the psychoacoustic masking threshold of the carrier speech, under three goals: targeted semantic hijacking, response suppression, and policy jailbreaking. Against an undefended Moshi-style agent, white-box attacks succeed in up to 91.7% of trials. We then introduce psychoacoustically aligned latent smoothing (PALS), which injects anisotropic Gaussian noise shaped by local codebook covariance at the residual-vector-quantized latent interface, with input noise shaped by the masking threshold constraining the attacker and trained by a Kullback--Leibler consistency objective. Deployed with no inference-time cost, PALS reduces hijack to 8.3%, mute to 11.2%, and jailbreak to 9.1% at clean quality within 2.3%. A Monte Carlo-smoothed variant certifies an ellipsoidal latent radius up to 0.616, a guaranteed floor that the empirical robustness far exceeds. --- ### Cross-Lingual Legal QA for Vietnamese Labour Law: Retrieval, Translation, and Verifier-Guided Correction Authors: Nguyen Minh Chi, Mo El-Haj, Nguyen Ha Thanh, Dawn Knight, Paul Rayson Published: 2026-09-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.27376v1 Cross-lingual legal question answering must retrieve statutes across languages while preventing unsupported legal claims. We introduce a bilingual evaluation suite of 231 Vietnamese--English question--answer pairs from Vietnamese labour law. Of these, 75 are additionally annotated for five challenging legal reasoning phenomena. We evaluate a verifier-guided pipeline that decomposes answers into claims, checks citation reachability and entailment, and corrects citation failures and contradictions. We also introduce six automatic diagnostics for faithfulness to retrieved evidence, covering citations, modality, exceptions, procedures, conclusions, and evidential support. Experiments show that learned-sparse retrieval performs poorly for English-to-Vietnamese retrieval (R@5~=~0.032), whereas dense retrieval reaches 0.358 and slightly outperforms hybrid retrieval. Translation placement has no statistically detectable effect on these automatic diagnostics in our controlled comparison and supporting sensitivity analyses. Verifier-guided correction improves citation preservation by $0.022$--$0.034$ at the system level but produces no reliable gains in the remaining dimensions. Human evaluation further shows that the automatic diagnostics do not fully align with human judgements of answer quality. --- ### Planned Test-Time Scaling with Coordinated Reasoning Paths Authors: Xueqing Wu, Langxing Bai, Hritik Bansal, Po-Nien Kung, Shuo Li, Hao Liu, Nanyun Peng, Kai-Wei Chang Published: 2026-09-23 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.27374v1 Test-time scaling with parallel branches is widely adopted to improve performance on challenging reasoning tasks. The predominant approach, repeated sampling, draws branches independently from a single policy, which can produce redundant attempts and thereby limit the gains from additional inference compute. To address this limitation, we propose Planned Test-Time Scaling (PTTS), which replaces independent sampling with a coordinated joint policy: a planner generates a solution outline for each branch, steering the branches toward distinct reasoning paths, and an executor produces a full solution conditioned on each outline. Formally, we show that PTTS strictly generalizes repeated sampling and, in a stylized setting, provably promotes coverage of complementary reasoning modes and yields better pass@k scaling. We instantiate PTTS on top of strong reasoning models, keeping them fixed as executors while replacing repeated sampling with PTTS inference to further enhance test-time scaling. Concretely, we develop two variants: PTTS-ZS prompts a model to jointly generate outlines for all branches in a single autoregressive pass, while PTTS-RL directly optimizes the planner against the pass@k reward using truncated execution rollouts for efficient training and a sharper reward signal. Across five mathematical reasoning benchmarks with Qwen3-1.7B and 4B, PTTS-ZS improves pass@64 over repeated sampling by up to 6.7 points, while PTTS-RL further increases the gain to up to 13.4 points. Further analysis indicates that broader coverage of distinct reasoning paths contributes to these gains. Overall, PTTS provides a general framework for improving test-time scaling by coordinating reasoning branches, with zero-shot and trainable instantiations that yield substantial performance gains. --- ### Attention Routing Stabilizes Early: Working-Set Inference for Recurrent Language Models Authors: Ke Wan, Chen Chen Published: 2026-09-23 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.27373v2 Recurrent-depth language models, such as looped Transformers, repeatedly apply shared network blocks to refine latent representations without generating explicit intermediate reasoning tokens. However, each step recomputes full attention over the entire context, repeating costly global routing. We study how attention routing evolves across recurrent depth and find a consistent separation in convergence timescales: attention support and distributions stabilize substantially earlier than hidden states and attention outputs. This suggests two stages of recurrent inference: early discovery of a sparse working set, followed by representation refinement over largely stable routing support. Motivated by this finding, we introduce WISE (Working-set Inference with Support Exploitation), a training-free method that uses unrestricted attention during early recurrent steps to discover a block-structured working set, then reuses its support in later steps while keeping attention weights and recurrent refinement dynamic. Controlled interventions show that multi-step discovery yields more effective working sets than first-step selection, and that support reuse better preserves model behavior than more restrictive forms of attention reuse. Across multi-hop QA benchmarks, WISE largely preserves full-attention performance. Matched context-scaling experiments reveal an increasingly favorable quality-efficiency tradeoff as routing support becomes sparser with longer contexts. A sparse-attention implementation achieves up to a 1.76x late-step attention speedup over native FlashAttention at 4K context. Code: https://github.com/tbn5pj/WISE_code. --- ### Neither Silence nor Overlap Is Failure: Intent-Conditioned Evaluation of Turn-Taking in Full-Duplex Spoken Dialogue Models Authors: Kian Shamsaie, Iman Modarressi Published: 2026-09-23 Categories: cs.CL, cs.AI, cs.HC, cs.SD arXiv: https://arxiv.org/abs/2609.27372v1 Benchmarks for full-duplex spoken dialogue models score turn-taking with binary fixed-window rules that reward immediate response or silence by completeness of the prior turn. We argue that the appropriateness of a response offset, whether delayed silence or anticipatory overlap, is conditional on the speaker's latent intent, identifiable only from that speaker's behavior. We introduce TACT, a benchmark of 9,728 episodes and 73.2 hours from five dyadic corpora; each episode carries dialogue history, a per-speaker memory profile, and an annotator-derived posterior over six intent classes. Scoring replaces binary windows with a strictly proper threshold-weighted continuous ranked probability score whose weights are intent-conditioned timing kernels fitted to human floor-transfer-offset distributions, proving boundedness, consistency, and binary reduction. Across eleven systems the best model reaches 0.47 against a human topline of 0.86, is nearly invariant to speaker profiles, and TACT agrees with human judgments at Spearman 0.81 versus 0.46 for binary metrics. --- ### ASAP: Visual Analytics for Identifying and Analyzing Image Patterns in AI-generated Images Authors: Jinbin Huang, Yuki Ueno, Chen Chen, Aditi Mishra, Bum Chul Kwon, Zhicheng Liu, Chris Bryan Published: 2026-09-23 Categories: cs.HC, cs.CV arXiv: https://arxiv.org/abs/2609.27371v1 Generative image models can produce highly realistic images, raising concerns about potential misuse in creating deceptive content. Current deepfake approaches face several challenges, including limited generalizability, lack of interpretability, and poor actionability. To help address these, we present ASAP, an interactive visualization system designed to empower users in the analysis and summarization of deceptive patterns in AI-generated images. ASAP introduces a novel CLIP-adapted image encoder that generates interpretable representations, enabling the extraction of influential pixel regions via calculated masks. This approach facilitates the identification of key deceptive features through influence measurement techniques. These backend techniques are integrated into a visual analytics dashboard that allows users to quantify and analyze authenticity-indicative patterns in image collections containing both authentic and AI-generated images. This approach also supports the comparative analysis of various generative models, including GANs and diffusion models. We demonstrate ASAP's efficacy through a user study and two application scenarios using established fake image detection benchmarks, showcasing its ability to effectively extract and quantify deceptive patterns. --- ### Geometry-Conditioned Visual Place Recognition in Natural Environments Authors: Walter Nedov, Saimunur Rahman, Kavindie Katuwandeniya, David Hall, Kaushik Roy, Peyman Moghadam Published: 2026-09-23 Categories: cs.CV, cs.AI, cs.RO arXiv: https://arxiv.org/abs/2609.27370v1 Visual Place Recognition (VPR) in natural environments remains challenging due to repetitive vegetation, sparse distinctive landmarks, and substantial appearance and viewpoint variation across traversals. While visual observations of the same place can change considerably, their underlying spatial structure is often more persistent. We exploit this complementary geometric consistency through Depth-Aware Distillation (DAD), which conditions the token representations of a pretrained Vision Foundation Model (VFM) on geometry inferred by a Geometric Foundation Model (GFM), without any depth sensor. Rather than treating geometry as an additional input modality, DAD projects image-aligned depth into the VFM token space and selectively modulates visual representations through channel-wise geometric conditioning. A two-stage teacher-guided learning strategy first anchors the geometry-conditioned representation to the pretrained appearance space, before refining it for place discrimination. Evaluated on the WildCross benchmark, DAD improves average inter-sequence Recall@1 from 61.41% to 66.37% and Recall@5 from 65.86% to 72.49% over a matched appearance-only baseline, with the largest gains under reverse traversal and long-term appearance variation. These results show that GFM-derived geometry can provide a persistent structural prior for VPR when visual appearance becomes unreliable. --- ### Anomaly-Free Self-Optimization via AUC Bounds Authors: Kevin Wilkinghoff, Zheng-Hua Tan Published: 2026-09-23 Categories: cs.LG, eess.AS arXiv: https://arxiv.org/abs/2609.27362v1 Anomalies are rare, and anomalous data are often unavailable during development, making it difficult to determine which anomaly detection models and configurations will generalize to unseen anomalies. Recent approaches address this challenge by generating pseudo-anomalies and using bounds on the achievable area under the ROC curve (AUC) to select the optimal configuration from a finite set of candidates. Instead, we use the AUC bound as a differentiable, anomaly-free objective for directly optimizing continuous parameters of anomaly detection systems. We demonstrate this framework by optimizing ensemble weights and introducing a learnable score-rescaling mechanism that adapts pseudo-anomaly scores, enabling optimization beyond a predefined candidate set. Experiments across multiple datasets and embedding models show that AUC-bound optimization achieves significant performance gains over conventional model selection and prior development-set-based parameter selection. The results further show that direct optimization is less sensitive to the choice of pseudo-anomaly construction. --- ### Automated Extraction of Records of Processing Activities (RoPA) Using Hybrid RAG and Locally Deployed Large Language Models Authors: To Duy Hinh, Nguyen Le Quoc Anh, Phan Van Tri, Khuong Nguyen-An Published: 2026-09-23 Categories: cs.CL, cs.CR, cs.IR arXiv: https://arxiv.org/abs/2609.27359v1 Vietnam's Personal Data Protection Law (Law No. 91/2025/QH15) and Decree No. 356/2025/ND-CP, effective January 1, 2026, require organizations to establish and maintain Records of Processing Activities (RoPA). Manual RoPA preparation is labor-intensive, while cloud-hosted large language models (LLMs) may conflict with data-sovereignty requirements. We propose RoPA Manager, a system for automated RoPA information extraction using hybrid retrieval that combines lexical ranking over tsvector, dense-vector search, Reciprocal Rank Fusion (RRF), and locally deployed LLMs. We introduce a Vietnamese RoPA benchmark with 32 organizations, 77 processing activities, 12 field groups, and 4,338 reference values. Evaluation is reported at three distinct levels. The automated scorer, tested on perturbed data without invoking an LLM, achieved F1 = 0.9493 (0.9436, 0.9548); this measures scorer robustness rather than end-to-end extraction accuracy. End-to-end extraction achieved token coverage of 50.04-55.25% against the reference labels. Two independent experts reviewed 1,558 reference values (35.9% of the benchmark), found no incorrect values, and achieved 99.68% agreement with PABAK = 0.9936. Value-level precision was not measured. Across 32 paired scenarios on a 24 GB GPU, locally deployed Qwen3.5-27B-GPTQ-Int4 showed no statistically significant difference from cloud-based DeepSeek-V4-Flash (difference 0.20 percentage points in favor of DeepSeek, 95% CI (-0.93, 1.32), p = 0.72), while Gemma-4-31B performed significantly worse (p < 0.01). --- ### Beyond Mean Foils: Auditing Worst-Foil Specificity in Frozen CLIP Region Explanations Authors: Kaixin Liu, Zhipeng Ye, Feng Jiang, Zhenghao Wang, Qihang Wu Published: 2026-09-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.27356v1 A region can overlap a target object yet contribute more to another class. We test regions selected by Cluster-based Concept Importance (CCI) in frozen CLIP. Across COCO and VOC with two checkpoints, 41.08-64.78% of regions that pass overlap and mean-contrast checks fail against the strongest competing class. Removing competitors annotated in the image leaves 39.69-63.64% failing. We then test all eight candidate regions per image. An alternative passes the test for 6.25-7.84% of failures on COCO and 27.40-31.15% on VOC. Requiring it to preserve the original target-score drop within $ε= 0.02$ reduces these rates to 0.16-0.98%. Available regions and target-drop tolerance constrain repair; relaxing the tolerance increases repair opportunities. --- ### Quantization-Robust Unlearning through the Lens of Retain-Forget Loss Landscapes Interaction Authors: Jialu Wang, Jianing Deng, Shuqing Luo, Yuanzhe Li, Dongwei Wang, Jingtong Hu, Huanrui Yang, Song Wang, Tianlong Chen Published: 2026-09-23 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.27355v1 Unlearning ensures LLM compliance by removing the influence of private or copyrighted training data. However, since LLM models typically undergo post-training compression, like quantization, in practical deployment, it has been observed that the unlearning effect can be substantially weakened, with the forgetting behavior degrading more severely than that of model utility. This paper proposes a quantization-robust unlearning framework that makes forgetting robust to quantization while maintaining overall model utility. We analyze this gap through the lens of loss landscape. Specifically, our analysis reveals a curvature-based criteria that pinpoints sensitive weights in the unlearned model that leads to both non-robust forgetting and reduced utility. We therefore propose sensitivity-guided noisy regularization, which is applied on the sensitive parameters to steer the model convergence towards a smoother minima of uniformly low forget and retain losses. Balancing unlearning and utility, we further propose forget-critical optimization, which updates only forget-critical layers, preserving most of the network to retain useful knowledge. Extensive experiments on the MUSE and TOFU benchmarks across multiple LLM unlearning algorithms show that our approach achieves substantially more quantization-resilient forgetting while maintaining utility. --- ### Constraint-Driven Context Engineering: Designing Domain Interfaces for AI Systems Authors: Xiwei Xu, Chen Wang, Mengmeng Yang, Yipeng Zhang, Jacky Jiang, Suyu Ma, Youyang Qu, Ming Ding, Liming Zhu Published: 2026-09-23 Categories: cs.SE, cs.AI arXiv: https://arxiv.org/abs/2609.27354v1 Generative AI systems are increasingly deployed to address domain problems. These systems operate under technical, regulatory, institutional, and normative constraints that define acceptable AI behaviour and outcomes within their domains. We observe a recurring pattern in our industry engagement: partners often arrive with a functioning but relatively generic AI solution. The challenge is no longer to build an AI system from scratch, but to improve the quality and domain appropriateness of an AI-generated solution. In these settings, the limiting factor is often the quality, scope, and structure of the context available to the system. Yet, existing context engineering approaches primarily focus on supplying domain knowledge through retrieval, memory, and tools, with limited support for systematically identifying and operationalising the constraints that govern AI systems in their operational environments. This paper proposes Constraint-Driven Context Engineering (CDCE), a design approach for engineering domain interfaces for AI systems. Drawing on software architecture design and Domain-Driven Design (DDD), CDCE treats domain constraints as first-class design drivers. It identifies and characterises constraints, determines the required context assets, and designs representations through which these assets are made available to AI systems. We conducted a comparative multiple-case study with industry and public-sector partners across educational assessment, healthcare decision support, and financial-distress prediction. Depending on their characteristics, constraints can guide AI behaviour, enforce permissible boundaries, or support verification of AI-generated outcomes. The cases demonstrate CDCE's applicability across contrasting domains and show how constraint characteristics shape the resulting domain interfaces. --- ### Guides That Cause Actions: An Offline Study of Guide-Action Mutual Reinforcement in Multimodal Web Agents Authors: Chengguang Gan, Yunhao Liang, QingHao Zhang, Shiwen Ni Published: 2026-09-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.27353v1 Web agents are usually evaluated in live environments, where environment state and judge models drift between runs, so the same checkpoint rarely reproduces the same score, making controlled studies of training phenomena impractical. We present WebMRE, an offline benchmark of 541 tasks and 5,293 steps derived from successful WebArena trajectories, with fully audited test labels and a deterministic protocol that scores a checkpoint identically on every run without any environment. Each step pairs a human oriented guide sentence with a grounded action, enabling the first study of the mutual reinforcement effect between them in web agents. Averaged over three seeds the effect holds for both models in both decoding orders and grows with scale: jointly decoding a guide lifts element selection over an action only reference by 0.9 and 0.2 points for Qwen3.5-4B and by 1.7 and 2.2 points for Qwen3.5-9B. A mediation analysis shows that the guide is a causal channel rather than commentary: forcing the gold guide as a decoding prefix lifts action accuracy from .422 to .684, another step's guide collapses it to .055, and a paraphrase that renames the target still recovers half of the gain, so the channel carries instruction meaning and not only the label string. The same channel yields an offline reward that only a replayable protocol makes computable, though optimizing it from a strong checkpoint brings no gain yet. Our fine tuned models outperform GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash, run zero shot, on every offline metric. --- ### MolDesignBench: Evaluating LLM-based Agent for Scenario-grounded Molecular Design Authors: Yongjun Jeong, Hanbum Ko, Ye Rin Kim, Chanhui Lee, Rodrigo Hormazabal, Jaewan Lee, Sehui Han, Sungbin Lim, Sungwoong Kim Published: 2026-09-23 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.27349v1 Real-world molecular design remains challenging for large language model (LLM)-based agents. It requires them to interpret design contexts, satisfy multiple constraints, identify infeasible specifications, and reason over multi-step tool outputs. Existing benchmarks do not capture this complexity, focusing instead on explicit and narrow constraints, only feasible problems, and single-path solutions. To address this gap, we propose MolDesignBench, a scenario-grounded benchmark that more closely reflects real-world molecular design for evaluating tool-augmented LLM agents. MolDesignBench comprises 2K generation and optimization instances that combine implicit requirements embedded in design narratives with explicit property and functional-group constraints, including infeasible cases, and require the effective use of 17 specialized chemistry tools. Experiments across diverse frontier LLMs reveal low success rates--with the best achieving only $\sim43$\%--and frequent failures in implicit-constraint reasoning, infeasibility detection, and tool reasoning. The corresponding fine-grained failure-mode analysis identifies implicit constraint interpretation and infeasibility detection as the primary bottlenecks, establishing MolDesignBench as a rigorous testbed to guide future research on chemical agents. The benchmark, tool interface, and evaluation code are publicly available. --- ### Evolving Inspectable O-RAN Slicing xApps with LLMs Authors: Faezeh Dehghan Tarzjani, Bhaskar Krishnamachari Published: 2026-09-23 Categories: eess.SY, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.27337v1 Open RAN (O-RAN) slicing xApps must adapt resource allocations to changing channel conditions and traffic demands while meeting service-level agreements (SLAs). Deep reinforcement learning can produce adaptive policies, but their allocation rules remain encoded in neural-network parameters. Our goal is to retain this adaptability while making the controller's decision logic directly inspectable and editable by operators. We use a large language model (LLM) to evolve slicing controllers as compact Python programs whose decision logic remains readable and editable after optimization. The LLM proposes and revises candidates offline, while a calibrated simulator scores them, and the selected decision module runs unchanged in the O-RAN control path. On the NSF POWDER 5G testbed, the evolved controller releases resources from a guaranteed slice whose throughput target becomes unattainable under a sustained channel fade, improving best-effort throughput from 158.2 to 228.6 Mbps, a 44.5% gain over the best static allocation. Since the controllers are readable source code, their behavior can be predicted from their equations, defects can be diagnosed by reading the code, and calibration errors can be corrected with one-line edits, reducing SLA misses from 79.9% to 2.2% in one case and more than doubling fitness in another. In a four-slice trace-driven simulation calibrated to the same testbed, evolutionary search achieves higher average evaluation scores than independent prompting at a matched proposal budget, with mean normalized gains on held-out traces of 16.3% for prompting alone, 32.1% for evolution from scratch, and 51.0% for evolution from a starting program. --- ### CART: Closed-Loop Adaptive Red Teaming for Large Language Models Authors: Dongdong Zhang, Tengchao Lv, Yilin Jia, Yuzhong Zhao, Yupan Huang, Wenshan Wu, Xiangyang Zhou, Shaohan Huang, Nan Yang, Li Dong, Lei Cui, Furu Wei Published: 2026-09-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.27336v1 Automated red teaming often replays a fixed set of prompts, which measures known risks but cannot learn from failures found during testing. We present CART (Closed-Loop Adaptive Red Teaming), a framework that uses each result to guide what it tests next. CART begins with broad risk coverage, follows weaknesses that emerge, keeps new probes diverse, and records the evidence and source of every finding. It separates the Challenger that creates tests, the Target being tested, which may be a text-only model or a bounded tool-using agent, and the Judge that evaluates the results, allowing these roles to be studied independently. Across three evaluation families (Frontier, JAH, and Agentic), CART discovers more failures and higher average risk than static seed replay for every Target with an available baseline. The gains extend to tool-mediated agent tests, suggesting that contextual adaptation can reveal weaknesses that direct prompt replay does not exercise. These results describe what the test policies discover, not how often failures occur in real deployments. We also find that Challenger-Judge choices affect the evidence uncovered, highlighting the need for role separation and independent review. Overall, CART turns red teaming from a one-time checklist into a continuous, adaptive, and auditable search for model and agent weaknesses. --- ### Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents Authors: Yefan Zhou, Yang Li, Zeyu Leo Liu, Semih Yavuz, Shafiq Joty Published: 2026-09-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.27334v1 Agentic memory systems reuse past experience to improve future performance, yet most existing designs curate memory at write time: once a task is completed, its trajectory is distilled into a fixed artifact, such as a reflection, workflow, skill, or reasoning strategy, that is later retrieved by similarity. This forces the system to decide what is worth remembering before the future query is known, irreversibly discarding information and producing a query-independent summary that must serve many possible downstream tasks. Learning such a write-time curator is also difficult because the value of a storage decision may only become apparent when a relevant query arrives, potentially many tasks later, creating a long-horizon credit-assignment problem. We instead retain raw trajectories and defer curation until read time, when the current task is known. Given the retrieved traces and the new task, a memory curator synthesizes a compact, task-adaptive payload tailored to the immediate need. Because this payload is consumed on the same task, the curator can be trained directly from immediate task success, avoiding delayed utility signals and the need to artificially group related tasks. Across ALFWorld, WebShop, and $τ^2$-bench, our Just-in-Time Memory (JitMem) consistently outperforms no-memory agents as well as heuristic and learned write-time memory methods, improving over the strongest baseline by 16.2, 16.3, and 3.9 absolute success-rate points, respectively. Notably, even an untrained curator is already competitive with or surpasses these baselines, showing that task-adaptive read-time curation itself is a major source of the gain; training the curator further compounds the improvement. --- ### Alignment Inertia: Auditing the Durability of Training Data Influence Through Policy Override Resistance Authors: Renata Barreto, Markelle Roesti, Mohammad Tahaei Published: 2026-09-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.27333v1 Platform operators increasingly rely on system prompts and fine-tuning to govern model behavior, yet it remains unclear how reliably these interventions override behavior inherited from prior training. We propose Override Success Rate (OSR) and alignment inertia to measure when operator interventions succeed or fail to change prior behavior. We evaluate zero-shot prompting and LoRA fine-tuning across Llama and Mistral in medical misinformation and hate speech. Alignment inertia persists across both models but varies by model, domain, and policy direction. Notably, in Mistral's restrictive hate-speech condition, LoRA increased inertia by 46.5 percentage points, showing that fine-tuning can reinforce rather than override prior behavior. We also use TRAK to test whether inertia is associated with weaker adaptation signals. TRAK achieves AUC of at least 0.85 in 7 of 8 conditions and outperforms model confidence, TF-IDF similarity, and embedding similarity as a predictor of inertia. These results provide an operator-facing audit of where prior training constrains downstream model governance. --- ### Stable Geometry with Divergent Task Evidence for Efficient Long-Horizon Agent Compression Authors: Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, Yanbiao Ma, Jungong Han Published: 2026-09-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.27332v1 Long horizon agents accumulate growing interaction histories that increase context and inference costs. We find that geometric redundancy alone is an insufficient criterion for safe compression. Although agent histories exhibit strong low dimensional structure, similar global geometry can preserve very different amounts of task evidence. At identical retained block counts, evidence aware selection raises next action Top 3 retention from 0.31 to 0.69, while centroid similarity remains 0.98. Controlled replacement further shows that action related information can be substantially altered while global geometric measures remain nearly unchanged. Motivated by this gap between geometry and evidence, we introduce Geometry Guided Evidence Preserving Memory (GEM), a training free compressor that protects task and execution evidence before using geometric residuals to complete coverage. GEM reduces mean combined token usage from 2.69M to 2.11M per task, a 21.4% reduction, while maintaining comparable task reward. Our results show that efficient agent history compression should optimize for preserved task evidence rather than geometric coverage alone. --- ### Can Vision-Language Models Analyze Human-Centered Video? Mapping Model Capabilities and Human-AI Collaborative Workflows Authors: Xiyuan Shen, Jiuyang Lyu, Seokhyun Hwang, Huanfen Yao, Shwetak Patel, Zhihan Zhang, Jacob O. Wobbrock Published: 2026-09-23 Categories: cs.CV, cs.HC arXiv: https://arxiv.org/abs/2609.27327v1 Video provides a rich record of human behavior, interaction, and situated contexts, offering important evidence for understanding people and conducting human-centered research. As vision-language models (VLMs) become increasingly capable of analyzing video, they offer opportunities to automate this traditionally human-intensive process. Yet a central question remains: when can VLMs analyze human-centered video independently, and when does reliable analysis still require human involvement? To address this question, we first characterize video analysis practices in human-centered research. We systematically analyze all 1,702 CHI 2026 full papers and identify 125 that annotate videos. Through iterative coding, we derive a five-dimensional taxonomy spanning analytic purpose, viewpoint, phenomenon, reasoning requirement, and annotation authority. Grounded in recurring annotation tasks captured by this taxonomy, we construct a benchmark of 15 representative tasks from open datasets to map the capabilities and limitations of a general-purpose VLM. We examine the division of labor between humans and VLMs by comparing three annotation workflows: VLM alone, human alone, and human verification of VLM outputs. Across tasks, VLM-alone annotation approaches human accuracy on average (HNS = 97.0, where 100 denotes human-alone performance), demonstrating substantial potential to automate human-centered video analysis. Human verification achieves the highest accuracy (HNS = 121.5) while reducing human annotation time by 48.9% and monetary cost by 31.3%-44.5% relative to human-alone annotation. Our findings connect real-world human-centered video analysis tasks and current VLM capabilities, and clarify how human-AI collaboration can make VLM-assisted analysis reliable and efficient. --- ### A Hybrid Iterative Deep Ritz Method for Elliptic Interface Problems Authors: Tianhao Hu, Bangti Jin, Fengru Wang, Yifeng Xu Published: 2026-09-23 Categories: math.NA, cs.LG arXiv: https://arxiv.org/abs/2609.27325v1 In this work, we propose a hybrid iterative deep Ritz method (H-IDRM) for a class of interface problems for second-order elliptic operators. It is based on a new mixed formulation of the problem and involves solving a sequence of convex minimization problems. We employ a level-set neural network architecture, featuring a level-set representation of the interface, to accommodate the piecewise smoothness of the solution and the flux. The approach involves only volumetric representations instead of duality pairing on the interface and avoids explicit interface sampling that is inconvenient for complex interface geometries. Further, we present an analysis of the method, including the errors arising from the neural network approximation, Monte Carlo approximation, iterative scheme, and penalty parameters. Numerical experiments indicate that the H-IDRM outperforms existing neural solvers on problems with high-dimensional domains, intricate interface geometries, and lower subdomain regularity. --- ### Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms Authors: Xinjie Shen, Wei Fan, Xudong Guo, Jianhong Tu, Yang Su, Chuqiao Kuang, Yinger Zhang, Dayiheng Liu Published: 2026-09-23 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.27321v1 Language-model agents increasingly face long-horizon tasks with evolving state, interdependent decisions, and delayed outcomes. Scaling their training requires diverse agentic environments, dependable outcome signals, and low extension cost. Existing generation pipelines commonly construct an environment before defining its outcome rule or annotating its trajectories, leaving dynamics and evaluation to be aligned post hoc. VHD-Play reverses this dependency by sampling and solving a mathematical model before a corpus-grounded setter renders its decision process as stateful tools. The executable dynamics and trajectory-scoring reference are inherited from the same solved model. The pipeline produces 3,300 diverse agentic environments at a cost of a few cents each. Training Qwen3.6-35B-A3B on three families raises its mean agentic score from 0.204 to 0.815 in a five-family diagnostic. Gains also appear on held-out instances from all three training families and eight unseen mechanism families, then extend beyond the generated substrate to external benchmarks for general function calling, travel planning, and 365-day e-commerce. On E-Commerce Bench, the trained checkpoint completes every run without bankruptcy and exceeds Qwen3.7-Max. We compare written-out problems with stateful versions that reveal or hide their parameters. The comparison shows that most of the learnable gap lies in stateful interaction rather than underlying problem solving. A frozen 35B setter realizes larger environments, and scale-matched training retains gains as mechanism size and horizon grow, indicating the potential for an evolving training substrate. --- ### Breaking Weather-Content Coupling: Type-Severity Guided Progressive Disentanglement for All-in-One Infrared Restoration Authors: Xinyao Wang, Lijun He, Zhihan Ren, Fan Li Published: 2026-09-23 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.27317v1 Infrared (IR) imaging is crucial for autonomous driving, remote sensing, and other perception tasks. However, adverse weather may introduce fake structural responses that are entangled with real thermal structures. Existing IR restoration methods are typically designed for a single degradation type or directly reconstruct from degradation-entangled representations. Consequently, they struggle to distinguish intrinsic thermal structures from weather-induced fake responses and to accommodate spatially varying degradation severity, leading to artifacts or the over-suppression of weak but meaningful thermal responses. To address these issues, we propose TSGPD-IR, a type-severity guided progressive disentanglement network for all-in-one infrared restoration that factorizes restoration guidance into task-level weather semantics and region-level degradation severity. Specifically, a Weather and Semantic Co-Guided Multi-Level Prompt Generation Module combines global weather semantics with stage-wise local features to generate adaptive prompts that progressively suppress degradation-induced responses while preserving intrinsic thermal structures. To complement global weather semantics with spatial restoration control, a Proxy-Supervised Regional Degradation Estimator derives severity supervision without manual annotations and predicts spatially varying degradation priors. Guided by these cues, a Multi-Source Collaborative Expert Selection Strategy uses a shared branch to preserve weather-invariant thermal structures and hierarchical routing to select weather-specific expert pools and severity-compatible regional experts. This design progressively separates degradation interference from genuine thermal content and enables region-adaptive restoration, reducing both residual artifacts and over-suppression. --- ### Turning Safety into Competence: Minimally Exploitable Robot Policies via Safety-Filtered Reinforcement Learning Authors: Ruihan Wu, Rui Yang, Donggeon David Oh, Duy Nguyen, Haimin Hu Published: 2026-09-23 Categories: cs.RO, cs.AI, cs.LG, eess.SY arXiv: https://arxiv.org/abs/2609.27312v1 Robots deployed for competitive tasks must outmaneuver their opponents without sacrificing safety. Existing approaches, including safe reinforcement learning (RL), train a single policy to achieve task success and avoid failures simultaneously. This coupling can complicate training and leave the learned policy exploitable by deliberate attacks. We propose Safety to Competence (S2C), a two-stage RL framework that separates safety synthesis from competitive task learning. We formulate competitive interactions as safety-critical Markov games and prove that perfect filtering preserves policy non-exploitability when all players commit to safe maneuvers. S2C learns a robust safety filter via adversarial RL, embeds it in the environment during task policy training, and retains the same filter at deployment. In simulated touchdown games, S2C outperforms eight safe RL baselines, achieving the highest win rate and Elo rating, and the lowest exploitability. Hardware stress tests against a human opponent confirm S2C's competence. ---