# Research Papers — 2026-09 · bundle page 83 of 266 > Full-text markdown of papers #4101–4150 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/83 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/82.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/84.md) ### Anatomy-Aware Dexterity-Driven Design Optimization of Surgical Continuum Robots Authors: Tony Qin, Peter Connor, Khoa Dang, Carter Hatch, Caleb Rucker, Robert J. Webster, Ron Alterovitz Published: 2026-09-25 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.30745v1 Performing complex medical procedures with continuum robots requires careful selection of their geometric design parameters. The robot should have high dexterity in the specific anatomical environment of its procedure. This work presents a design optimization method that considers both dexterity and anatomy. We introduce the Reachable Volumetric Dexterous Solid Angle (RVDSA) metric as our objective, which measures the ability of a robot's end effector to reach the points in a goal volume from different directions via collision-free paths from a start configuration. We present a computationally efficient motion planner to compute this objective function for a given robotic design, and we use an asymptotically optimal simulated annealing optimizer to compute an optimized design. We applied our new method to optimize the design of a bimanual dexterous sheaths robot for performing procedures on cancerous polyps in colon anatomies, achieving a 78% higher RVDSA on average than optimizing for 3D voxel coverage alone. --- ### From S3Q Theory to Implementation: Towards an Architecture for Machine Qualia Authors: Tetiana Grinberg, Katrina Schleisman, Patryk Laurent, Bogdan Udrea, Minda Myers, Brian Aufderheide, Luis El Srouji, Doyle Groves, Kevin Schmidt Published: 2026-09-25 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.30743v1 A key challenge in machine consciousness research is translating theoretical models into computational-level implementations. In this paper, we address this challenge by proposing a five-layer implementation architecture for the S3Q (Simulated, Situated, Structurally Coherent) theory of consciousness. Rather than introducing novel formalisms, the architecture composes published computational primitives into a single pipeline. S3Q identifies three jointly necessary conditions for qualia: (1) grounded sensorimotor situatedness, (2) internal simulation via a world model, and (3) structural coherence between predictions and observations. No existing computational system implements all three simultaneously. We map each S3Q tenet to specific, compatible computational machinery and specify how these components interface within a single representation pipeline that operates on continuous, differentiable, per-object slot vectors, along with a developmental bootstrap sequence and falsifiable predictions for the composed system that no subset of the architecture produces in isolation. The model suggests that a basic sense of "self" develops by linking actions to their outcomes, and that behavior falls into three patterns (hesitation, curiosity, or avoidance) depending on how unexpected an outcome is and whether it is experienced as positive or negative. Each prediction is individually falsifiable, providing the field with a testable framework to advance our understanding of machine consciousness. --- ### From Mono to Stereo: Accelerating Binocular Gaussian Splatting via Reprojection and Selective Patching Authors: Hongfei Zhu, Ling Zhou Published: 2026-09-25 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.30741v1 Binocular rendering requires two nearby views of the same scene and therefore repeats substantial visibility and shading work. We present a 2D Gaussian Splatting (2DGS) pipeline that fully renders a dominant-eye RGB image and an alpha-weighted depth proxy, reprojects that image to the affiliated eye, and repairs uncovered pixels. Small interior gaps are interpolated, whereas larger disoccluded regions are identified as regions of interest (ROIs) and selectively re-rendered. The depth proxy reuses the alpha-blending weights computed during dominant-eye rasterization, avoiding a separate depth-rendering pass. An adaptive ROI generator localizes the required updates using reprojected image boundaries and optional connected center-hole detection. On DTU, Tanks and Temples, and MipNeRF-360, the method reduces the measured time of a sequential two-pass binocular reference by 15.5\% to 28.8\% and peak GPU memory by 6\% to 11\%. The corresponding affiliated-eye quality degradation is at most 1.3 dB PSNR, 0.02 SSIM, and 0.02 LPIPS, representing a measurable trade-off that requires application-specific perceptual validation. These results establish a practical efficiency-quality trade-off for controlled static-scene stereo rendering and motivate future evaluation under continuous motion and on physical VR hardware. --- ### SEA-CLIP-Tiny: Efficient Multilingual Text-Vision Embedding for Southeast Asian Languages Authors: Puja Ahmad Habibi, Faiz Assabil Firdaus, Ashvanth S, Ekapol Chuangsuwanich, Pume Tuchinda, Peerat Limkonchotiwat Published: 2026-09-25 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.30739v2 Multilingual text-vision embedding models are essential for cross-lingual image-text retrieval, but Southeast Asian languages remain poorly supported due to the region's linguistic diversity and limited data and computing resources. In this paper, we introduce SEA-CLIP-Tiny, a compact multilingual text-vision embedding model for Southeast Asia with fewer than 50M parameters. Our model adapts a CLIP-KD-style framework to Southeast Asian multilingual settings through regional data curation and multilingual teacher guidance. Experiments across seven Southeast Asian languages show that SEA-CLIP-Tiny achieves the strongest average retrieval performance among the evaluated student models, reaching 12.9%, 31.5%, and 42.2% at R@1, R@5, and R@10, respectively. Compared with MobileCLIP2, it improves average R@10 by 12.1 points while using 38.4% fewer parameters and lower measured CPU latency. These results highlight the importance of region-aware training for efficient multilingual text-vision models in Southeast Asia. --- ### Beyond Mean Attention: Diversity-Aware, Layer-Wise Scoring for KV Cache Eviction Authors: Tianfang Xie, Wei Zhu Published: 2026-09-25 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.30738v1 KV cache eviction methods such as SnapKV and PyramidKV rank tokens solely by mean attention over a small observation window. We study a unified score, $μ_i+λ_1σ_i+λ_2\mathrm{corr}(i,S)$, adding attention dispersion across window queries and redundancy relative to selected tokens. For $λ_2<0$, the score penalizes similarity to selected tokens as in maximal marginal relevance (MMR), without extra forward passes. To test whether this relevance-diversity balance should vary with depth, we compare fixed global coefficients with three-segment and quadratic profiles. Only these depth profiles are searched on a development split under a $\sinh$ reparameterization. On all 16 English LongBench datasets with Mistral-7B at a budget of 64 entries per layer, a single global diversification constant improves 13 of 16 datasets (macro +1.1); the gain holds at budget 32 and narrows at 128. Per-dataset search finds no detectable layer structure on most datasets; on passage retrieval it finds a large one: a mid-layer sign flip that rewards similarity and is worth +9.6 over the baseline at budget 64 and, without re-tuning, +13.2 over the global constant at budget 128. Ablations attribute the gain to the redundancy term; replaying every accepted search state on the held-out test set separates genuine structure from tuning noise. --- ### Learning What to Skip: Counterfactual Credit Assignment for Efficient Multi-Agent LLM Workflows Authors: Jinfeng Xu, Zheyu Chen, Ziyue Peng, Zheng Lin, Shuo Yang, Jinze Li, Zheng Xing, Mengran Li, Victor C. M. Leung Published: 2026-09-25 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.30734v1 Multi-agent LLM workflows use planning, execution, verification, and summarization to improve task performance, yet the value of each component depends on the state already produced. Executing every component can waste computation or overwrite a correct intermediate answer. We formulate component omission as counterfactual credit assignment: full-workflow logs reveal the executed trajectory's reward, while controlled skip interventions reveal the consequences of omitting a future step. We introduce Learning What to Skip (LW2S), which learns action-specific safety models from these interventions and combines held-out calibration with domain-native guards to select skips. When an early skip is rejected, the controller can continue execution and reconsider a later component. Across mathematical reasoning, multiple-choice QA, and code generation with two instruction-model families, LW2S reduces recorded token cost while matching or improving aggregate full-workflow accuracy in the evaluated settings. Scale-up and second-topology experiments further examine component redundancy, while shared-error cases reveal why agreement alone is insufficient for skip selection. These findings connect efficient workflow execution to learning the conditional utility of individual components. --- ### Amplify What You Gaze At: Target Saliency Boosting in Text-to-Image Generation Authors: Shengqi Dang, Zhengxi Yu, Feilin Han, Xingyu Lan, Nan Cao Published: 2026-09-25 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.30733v1 Text-to-image generation has advanced in controlling what, where, and how objects appear, yet how visual attention is distributed among objects remains largely unexplored. In this paper, we introduce Target Saliency Boosting, a new task aimed at boosting the visual saliency of a specific object during text-to-image generation without requiring any visual priors. Our key insight is that visual saliency is inherently relative: boosting the saliency of a target object also depends on the global saliency distribution across all objects in the scene. Based on this insight, we propose GazeME, a lightweight framework that uses saliency-marked prompts, inserting learnable marker tokens around object descriptions to indicate which objects to visually emphasize or suppress. To learn these markers, we construct a saliency-semantics dataset that associates objects in image--prompt pairs with object-level saliency scores, and propose Saliency Prior Marker Activation (SPMA), a saliency-aware stochastic marker activation strategy that exploits relative saliency relationships for robust training. During inference, GazeME automatically inserts appropriate markers into the prompt, thereby directly enhancing the visual saliency of the target object. Extensive experiments demonstrate that GazeME effectively boosts target saliency while preserving both semantic alignment and image quality. --- ### TR-SSQP: A Trust-Region Method for Constrained Stochastic Optimization under Heavy-Tailed Noise Authors: Haoxuan Wang, Yuchen Fang, Sen Na Published: 2026-09-25 Categories: math.OC, cs.LG, stat.CO, stat.ML arXiv: https://arxiv.org/abs/2609.30732v1 We consider stochastic nonlinear optimization problems with deterministic equality constraints. While unconstrained stochastic optimization is well understood, the interplay between optimality and feasibility in the constrained setting poses significant challenges. Moreover, existing theoretical guarantees for constrained stochastic methods predominantly rely on bounded-variance assumptions, leaving the heavy-tailed noise regime largely unexplored. To address this gap, we propose a novel trust-region method within the stochastic sequential quadratic programming framework, termed TR-SSQP. Our method employs a normal-tangential decomposition in the step computation to balance optimality and feasibility. In addition, we incorporate a normalization mechanism in the design of the trust-region radius, together with Polyak momentum for gradient estimation, ensuring stable updates without gradient clipping. When the trust-region radius and the momentum parameter decay at appropriate rates, we establish global almost-sure convergence of the method. To the best of our knowledge, this is the first asymptotic convergence result for constrained stochastic optimization under heavy-tailed noise. We demonstrate the promising performance of the proposed method through extensive numerical experiments, including comparisons among its variants and with existing constrained stochastic optimization methods. --- ### Structured Bayesian Modeling of Dynamic Receptive1 Fields in Salamander Retinal Ganglion Cells Authors: Alokesh Manna Published: 2026-09-25 Categories: cs.NE, stat.ME arXiv: https://arxiv.org/abs/2609.30731v1 Neurons in the visual system are selective for specific spatial and temporal stimulus features, described by their \emph{receptive field}. Estimating one means a coefficient per pixel per time bin from few trials -- a high-dimensional problem requiring regularization. Sparse regularizers such as the LASSO handle the dimension but select pixels independently at each time point, with nothing to keep the region coherent in space or smooth in time; it can fragment or reorganize discontinuously even when the true response evolves smoothly, a failure since this evolving pattern is what a receptive-field estimate should capture. We formulate dynamic receptive-field estimation as a high-dimensional Bayesian problem: a Poisson model combining a Gaussian Markov random field in space with an autoregressive process in time, so the estimated field is smooth and coherent across space and time. On recordings from $155$ salamander retinal ganglion cells, fitting this model independently per neuron recovers a coherent surface, where a pixel-level Poisson-LASSO comparison instead returns a fragmented one. Summarizing each neuron's surface by its space-averaged temporal response and clustering these curves with a model-based functional-clustering procedure, BIC selects three balanced temporal-response phenotypes ($85$, $32$, $38$ neurons), against a degenerate grouping from clustering the raw surfaces. A simulation study with known ground truth confirms the same pattern, with the model beating an unregularized Poisson GLM, LASSO, and the elastic net on recovery and estimation accuracy, though LASSO controls false positives better. The per-neuron field identification, its contrast with LASSO, and the functional-clustering population typing constitute this paper's contribution. --- ### Input-Layer Starvation: Why Per-Layer Pruning Breaks IoT Intrusion Detectors Authors: Md Anas Biswas Published: 2026-09-25 Categories: cs.CR, cs.LG arXiv: https://arxiv.org/abs/2609.30729v1 Intrusion detectors for small Internet-of-Things (IoT) devices are usually compressed by pruning and judged by overall accuracy. We show that this hides a severe class-level failure, find its cause, and give low-overhead prevention and repair. On CICIoT2023, a two-layer convolutional detector pruned with uniform layer-wise magnitude pruning at 80% sparsity loses 16 points of accuracy but half of its macro-F1, the mean per-class F1 (0.542 to 0.271 over five independently trained models); 17 of 34 classes are materially damaged. Remaining weight count does not explain it: a perceptron and a transformer pruned to the same or fewer weights lose at most 0.096. The first layer does. It has 192 weights; uniform pruning leaves 38, 46% of its 64 filters lose every input weight, and fine-tuning under that starvation leaves the running means of the first normalisation layer displaced by up to 0.8 standard deviations in a few surviving channels, on which the deployed model collapses. Protecting those 192 weights, or pruning globally at the same sparsity, prevents the collapse (loss 0.013); recomputing the normalisation statistics on unlabelled training data, with no weight changed, repairs it (loss 0.039) and returns the false-alert rate to 33% (dense 29%). Damage shows a strong increasing dose-response in first-layer sparsity, starving a perceptron's input layer reproduces the collapse, and the pattern holds on TON_IoT. The failure is misattribution and false alerts, not silent evasion: on validation-selected blind spots, uniformly pruned detectors misattribute 72% of the traffic, against 50% with the first layer protected and 47% for the dense model. --- ### Learning Polarization Image Restoration with General Restoration Priors Authors: Chenggong Li, Jinhao Liu, Caiyun Wu, Yidong Luo, Junchao Zhang, Degui Yang Published: 2026-09-25 Categories: cs.CV, eess.IV arXiv: https://arxiv.org/abs/2609.30728v1 Polarization imaging captures distinctive surface and geometric cues that benefit a wide range of vision tasks. However, real-world polarization acquisition is often affected by multiple coupled degradations, making image restoration essential for practical polarization vision. Existing methods are largely tailored to specific degradations and remain constrained by the limited scale and quality of polarization data. To address these limitations, we develop an all-in-one polarization restoration framework for diverse and composite degradations. We first study the impact of different polarization representations on restoration performance and identify the normalized Stokes representation as an effective choice for separating intensity and polarization information. Accordingly, we devise a dual-branch architecture that separates intensity and polarization modeling. To overcome the limitations of polarization-specific training, the intensity branch leverages pretrained general restoration priors and a mixture-of-experts extension for composite degradations, while its restoration knowledge is adaptively distilled into the symmetric polarization branch via a cross-domain feature transform. In addition, we establish a composite-degradation polarization benchmark to support all-in-one restoration research. Extensive experiments on public datasets and our proposed benchmark demonstrate the effectiveness of the proposed method. --- ### Analyzing and Mitigating Cost-Inefficient Behaviors in Coding Agents Authors: Yiran Hu, Nan Jiang, Shanchao Liang, Anik Dey, Yi Wu, Lin Tan Published: 2026-09-25 Categories: cs.AI, cs.SE arXiv: https://arxiv.org/abs/2609.30725v2 Although effective, coding agents often incur substantial monetary costs. Their recurring cost-inefficient behaviors remain underexplored. We conduct the first study of behavioral cost inefficiencies in coding agents, analyzing 1,200 trajectories from Claude Code and Mini-SWE-Agent across four configurations on SWE-bench Verified. We identify three cost-inefficient behaviors: subsumed retrieval, similar script generation, and test re-execution. We then evaluate three mitigation strategies: structure-aware retrieval, agent-synthesized skills, and developer-designed skills, over 10k trajectories on held-out SWE-bench Verified and Pro tasks. Our main findings are: (1) The three behaviors affect 79.00%--98.00% of coding tasks and account for up to 22.75% of task cost. (2) Structure-aware retrieval can introduce retrieval overhead and alter agent delegation, causing inconsistent improvements in retrieval efficiency and cost increases of up to 28.14%. (3) Agent-synthesized skills tend to produce low-level, trace-specific guidance, limiting their effectiveness and generality. (4) In contrast, developer-designed skills provide high-level, trace-agnostic guidance, reducing cost by up to 41.73%, roughly twice the maximum gain from agent-synthesized skills. --- ### EviDETR: Preserving Query-Relevant Temporal Evidence for Moment Retrieval and Highlight Detection Authors: Haoran Sun, Yufan Li, Qichen Zhang, Haoran Zhao, Shuqi Wang Published: 2026-09-25 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.30724v1 Joint video moment retrieval and highlight detection requires identifying query-relevant temporal segments while estimating clip-level saliency, yet DETR-style pipelines do not explicitly preserve query-relevant evidence throughout encoding, decoding, and cross-task prediction. We propose EviDETR, an evidence-preserving framework with three components. Semantic-aware Feature Reweighting (SFR) enhances query-relevant clip representations through saliency estimation and cross-modal interaction. A Temporal Top-2 Mixture-of-Experts (TTop2MoE) decoder performs query-adaptive refinement via sparse expert routing. MR-to-HD (MR2HD) fusion transfers span-level retrieval evidence to clip-level highlight prediction through confidence-weighted multi-scale aggregation. Using CLIP+SlowFast features, EviDETR achieves 69.29 R1@0.5, 54.77 R1@0.7, and 48.41 Avg. mAP for moment retrieval on QVHighlights, together with 41.83 HD-mAP and 68.33 HIT@1. Strong results on TACoS and Charades-STA further demonstrate cross-dataset transferability. --- ### TrafficImag: A Benchmark for Counterfactual Roadside Traffic Video Generation Authors: Xiangyu Li, Tianyi Wang, Zhihao Dou, Christian Claudel, Zhaomiao Guo Published: 2026-09-25 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.30722v1 Existing roadside traffic datasets support perception, forecasting, and visual question answering, but they do not evaluate counterfactual video generation, in which a selected actor is modified and the generated future should remain consistent with road topology and unrelated traffic. We introduce TrafficImag, the first benchmark for counterfactual roadside traffic video generation. TrafficImag combines a large-scale roadside dataset (9,022 annotated images, 7,043 deduplicated video clips, and 31,145 actor-centered history-future samples) with an executable protocol that supports behavior reasoning, intervention-aware image editing, and conditional video generation. Each intervention is represented as an actor-level program describing the target actor, intended behavior, legal route, interaction order, and temporal constraints, enabling a unified evaluation interface across heterogeneous foundation models. TrafficImag evaluates four complementary validity dimensions: initial-state correctness, route and behavior validity, interaction consistency, and non-target preservation, and considers an end-to-end counterfactual successful only when all four are satisfied. Across state-of-the-art foundation models, the strongest reasoner reaches 80.4% macro F1, the complete condition interface raises end-to-end success from 23.3% to 55.0% for the best generator. Oracle studies further show that conditional video execution is the primary remaining bottleneck. TrafficImag provides a reproducible benchmark for evaluating and diagnosing counterfactual traffic video generation beyond perceptual video quality. --- ### When 10,000 Windows Are Not 10,000 Tests: Auditing Statistical Confidence in Sliding-Window Time-Series Classification Authors: Xinze Shi, Litian Zhang, Binrui Shi Published: 2026-09-25 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.30721v1 Sliding-window classifiers are often evaluated on thousands of overlapping test windows, even though neighboring predictions share observations and remain nested within recordings and subjects. Subject-disjoint evaluation prevents one form of leakage but does not make those test windows independent. We present a practical audit that maps three claims - performance on observed recordings, future recordings from observed subjects, and unseen subjects - to explicit aggregation rules and established dependence-robust inference. At 75% overlap, controlled simulations give 16.9% Type-I error for IID observed-record inference and 7.2% for session-centered Bartlett-HAC: a substantial improvement with residual miscalibration. Audits of frozen WISDM and HARTH predictions show that nearly fourfold growth in test rows yields only 1.75-1.94-fold variance-equivalent information growth. At that overlap, fixed-record paired Accuracy-difference intervals are 1.22-1.66 times the IID widths; this inflation is not universal at zero overlap. On HARTH, paired Accuracy-difference intervals include zero across three overlap settings, whereas Macro-F1 favors MiniROCKET. Independent recomputation, common-session checks, class-level results, and separately seeded calibration make the audit's scope and limitations inspectable. The resulting workflow distinguishes additional predictions from additional independent evidence. --- ### Werracle: Sub-Cent Intra-Block AI Reflex Oracles and Flash-Loan Circuit Breakers for EVM Smart Contracts Authors: Volkan Dağlı, Zerrin Dağlı, Dağhan Dağlı Published: 2026-09-25 Categories: cs.CR, cs.AI, cs.DC arXiv: https://arxiv.org/abs/2609.30719v1 Contemporary on-chain artificial intelligence (AI) encounters an intractable Von Neumann memory and latency wall. Storing static floating-point neural weight matrices inside Ethereum Virtual Machine (EVM) storage costs millions of gas, rendering direct on-chain inference impossible. While Zero-Knowledge Machine Learning (ZK-ML) offloads matrix tensor multiplications to off-chain provers, it introduces fatal constraints: 10 to 300 seconds of SNARK proving latency and 250,000 to 500,000 gas per proof verification. Because decentralized finance (DeFi) exploits - such as uncollateralized flash-loan attacks, predatory sandwich MEV, and toxic loss-versus-rebalancing (LVR) flow - occur atomically inside a single block, ZK-ML oracles cannot react in time. Here, we present Werracle, a production-grade, zero-storage on-chain AI decision oracle fitting inside a single 32-byte EVM storage slot (bytes32). Leveraging foundational procedural Mandelbrot escape dynamics (z_{n+1} = z_n^2 + c) established by Dagli et al. (arXiv:2609.25498), Werracle derives continuous non-linear decision hyperplanes from a 24-byte coordinate triplet Theta = (c_x, c_y, zoom). Implemented in pure Solidity bytecode using fixed-point Q16.16 arithmetic (WerrMath.sol), Werracle evaluates a 16-point Pareto micro-grid in only 21,438 gas (under 0.0005 USD on Layer-2 rollups like Base and Arbitrum) with sub-millisecond execution latency. We demonstrate real-world DeFi efficacy via WerracleFeeHook.sol, a Uniswap v4 dynamic swap fee governor that measures orderbook turbulence on-the-fly and atomically adjusts liquidity provider fees between 0.05% and 0.50%. The protocol is formally verified against a 1,000-test cryptographically sealed deterministic verification suite (100.0% pass rate) with telemetry permanently disabled, operating live on a dedicated EVM devnet sandbox (Chain ID 4242). --- ### NEMSim: Learning Control-Conditioned Multi-Event Physical Dynamics via Executable Event-Mechanism Priors Authors: Junsong Yu, Junjie Xie, Pengwei Liu, Dong Ni Published: 2026-09-25 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.30718v1 High-fidelity simulation of control-conditioned multi-event physical systems is computationally expensive, especially across broad control spaces and long trajectories. In these systems, macroscopic evolution emerges from localized discrete events whose intensities and effects depend on process controls and evolving local states, while the available system knowledge is typically expressed as event-attribute descriptions. Purely data-driven surrogates must infer these event effects from limited trajectory coverage, which can hinder generalization to unseen control regimes. Physics-guided methods instead primarily build on equation-level constraints or differentiable solvers rather than discrete event-rule priors. We therefore propose NEMSim (Neural Event-Mechanism Simulator), which compiles predefined event-attribute descriptions into an executable transition structure linking control-dependent event intensities, prior-guided mechanism attribution, and state-dependent responses. To enable evaluation of control-conditioned multi-event dynamics with explicit system knowledge, we construct a 3D KMC-based benchmark pairing high-fidelity trajectories with explicit event rules, standardized splits, and evaluation protocols. Across three settings, NEMSim reduces Avg. RMSE by 58.9%-81.3% relative to the strongest baseline in each setting. It also remains best in the data-efficiency study with training-data fractions down to 10%. Mechanism analyses further show that these gains arise from executable rule integration rather than prior access or architecture alone. --- ### Words Speak Louder Than Order: A Behavioral Evaluation of Gemma 4 Authors: Amanda Fitch Published: 2026-09-25 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.30716v1 When a language model receives two conflicting documents as input, how does it decide which one to prioritize? Does it rely on how the sources are framed or the presentation order of the documents? We evaluated this behavior on Google's pre-trained Gemma 4-e4b model across a targeted behavioral suite (n = 13 items, 784 forward passes in short, single-turn contexts) using a completely counterbalanced experimental design. This setup allowed us to mathematically isolate the specific effects of source framing and reading position, while ensuring the model's natural vocabulary biases were canceled out. Across ten test conditions, we discovered the following: 1. Source framing heavily overpowers reading position. When directly competing, the semantic framing of a source (such as presenting it as an official guideline or a fresh update) had a significantly stronger impact on the model's final answer than the presentation order of the document. 2. The model favors the first document it reads, but this bias is highly variable. While the model consistently demonstrated a primacy effect (preferring the first document presented), the actual strength of this bias fluctuated by at least a factor of 5 based solely on the surface wording. 3. Overall structural repetition, not short copy-cues, drives positional bias. The model's preference for the first document is not a mechanical reaction to short, repetitive trigger phrases, such as "is (Answer)". However, the primacy effect does increase significantly when the two competing documents are structurally identical, using word-for-word verbatim templates. Introducing variation in the overall wording between the two sources reduces this positional bias. --- ### CRC-Router: Risk-Constrained Routing for Medical Agentic AI Systems Authors: Xueyang Li, Mingze Jiang, Gelei Xu, Jun Xia, Ching-Hao Chiu, Mengzhao Jia, Danny Z. Chen, Yiyu Shi Published: 2026-09-25 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.30714v1 Agentic AI systems are increasingly being explored in medical imaging to improve throughput and reduce clinician workload; however, safe deployment remains challenging because autonomous errors may propagate into downstream clinical decisions. A central requirement is therefore not only strong predictive performance, but also a reliable routing mechanism that determines when the system should proceed autonomously and when a case should be escalated for further review. To address this gap, we propose CRC-Router, a risk-constrained, uncertainty-aware routing module that is applicable to both conventional medical prediction models and agentic medical AI systems. CRC-Router combines multiple complementary uncertainty signals with the predictive score to construct a per-finding routing feature vector, maps this vector to an estimated wrong-accept risk using a lightweight per-finding risk model, and then applies Conformal Risk Control (CRC) to calibrate acceptance thresholds under a user-specified risk target. Instantiated on chest X-ray multi-finding triage using the NIH ChestX-ray14 dataset, CRC-Router achieves the strongest empirical risk--coverage trade-off among the evaluated baselines, both as a standalone routing layer and as a plug-in module integrated with the state-of-the-art MedRAX agent. These results demonstrate both the effectiveness of CRC-Router in selective medical automation and its modular, model-agnostic compatibility with existing predictive and agentic medical pipelines. Code is publicly available at https://github.com/XLIAaron/CRC-Router --- ### Parameter Estimation for Unnormalized Discrete Models via Empirically Localized Deformed Bregman Divergence Authors: Takashi Takenouchi Published: 2026-09-25 Categories: stat.ML, cs.LG arXiv: https://arxiv.org/abs/2609.30713v1 Estimation of parameter of probabilistic models is an important task in the field of machine learning.For models of discrete variables, calculation of the normalization constant of model is sometimes difficult and a lot of researches have been done to avoid the calculation of the normalization constant. In this paper, we tackle with the difficulty by combining a technique of empirical localization and a deformed Bregman divergence.The technique of empirical localization makes it possible to drastically reduce computational cost of the calculation of the normalization constant, and in addition, appropriate choice of the deformation for the Bregman divergence can invest the proposed estimator with various kinds of favorable statistical properties, such as efficiency or robustness against outlier noise. --- ### VLALight: Lightweight Vision-Language-Action Models for Emergency-Aware Traffic Signal Control Authors: Kemou Jiang, Maonan Wang, Xingchen Zou, Jiayue Zhu, Yuhang Fu, Sicheng Wang, Xi Chen, Yirong Chen, Zhiyong Cui Published: 2026-09-25 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.30709v1 Traffic signal control (TSC) is essential for mitigating urban congestion. Recent advances in vision-language models (VLMs) enable richer interpretation of intersection scenes, opening new opportunities for visual-context-aware TSC. However, the loose coupling and repeated information conversion between modules can lead to the loss of fine-grained visual details, while sequential inference introduces substantial latency. To address these limitations, we propose VLALight, a lightweight end-to-end vision-language-action framework that directly maps intersection observations and signal-phase information to discrete signal actions. To handle the multi-view nature of TSC, VLALight combines multiple directional camera views into a unified visual input and uses textual instructions to establish their correspondence with traffic movements and signal phases. This design enables direct action prediction with a compact 0.5 B-parameter model, without intermediate image-to-text descriptions or handcrafted traffic-state representations. Experiments show that VLALight delivers the best emergency-vehicle service of all compared methods, reducing pooled emergency waiting time by 21.1% over the cascaded VLMLight while running in real time on local hardware and generalizing to unseen intersection topologies and traffic-flow patterns. --- ### Combining General and Domain-Specific Pretext Tasks for Brain MR Image Segmentation Authors: Tasneem Nasser, Susanne Schmid, Roberto Souza, Naser El-Sheimy Published: 2026-09-25 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.30708v1 A key challenge in medical image analysis is the scarcity of large annotated datasets for specific populations and diseases. As deep learning models rely heavily on labeled data, effective transfer learning strategies are needed to reduce the dependence on manual annotations. Self-supervised learning has emerged as a promising approach for developing foundation models by enabling the learning of transferable feature representations from large-scale unlabeled medical imaging datasets. In this study, we investigate voxel-level brain age prediction as a domain-specific self-supervised pretext task and compare it with image inpainting, a widely used non-domain-specific alternative. We further propose a multitask self-supervised pretraining framework that jointly optimizes both objectives to learn complementary neuroimaging representations. The pretrained models are evaluated on three downstream magnetic resonance image segmentation tasks: multiple sclerosis lesion segmentation, ischemic stroke lesion segmentation, and cortical brain structure segmentation. Overall, the proposed multitask pretraining framework consistently outperformed the single-task pretrained models and training from scratch across most experimental settings, demonstrating the benefit of combining domain-specific and general self-supervised learning pretext tasks for the development of generalizable neuroimaging foundation models.\ Code Availability: The source code used in this study is publicly available at https://github.com/TasneemN/Combining-General-and-Domain-Specific-Pretext-Tasks-for-Brain-MR-Image-Segmentation/ --- ### LAVOIR: Teaching a Single-Pass Decision Encoder When and What to Ask with Amortized Value of Information Authors: Furkan Yilmaz, Habibe Aleyna Tasdemir, Muhammed Faruk Gozay Published: 2026-09-25 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.30706v1 "System One" decision models such as TypeSafe's Jev and its open counterpart Laya answer typed questions about a text in a single forward pass with calibrated probabilities, but they cannot ask for missing information: when a first message does not say what separates two departments, they guess. We present LAVOIR (Laya with Value-Of-Information Routing), which places the candidate pieces of missing information (slots) in the input next to the answer options, so that one forward pass returns both the decision distribution and, for every slot, the expected gain in the probability of the correct decision if the user were asked about it. VOI targets need no human labels: gold decisions come from schema rules, an LLM only verbalizes messages and answers, a model from another family checks every text, and pairing each message with several profiles makes regression on realized gains estimate the expected gain. A Gini-impurity cap bounds the predicted value by what a calibrated model can still gain. In a controlled study, decisions on seen schemas are statistically indistinguishable from the Bayes ceiling. The final model's question policy matches a greedy oracle VOI policy on seen schemas (AUC 0.799 vs. 0.797), and with at most 0.5 questions per conversation it is 14.1 points more accurate than never asking. On real ABCD conversations, one real exchange raises accuracy by 8.3 points where LAVOIR asks and leaves it unchanged where it does not; on SGD the cap lowers the asking rate from 93% to 8.6%. On Laya's twelve benchmarks LAVOIR is above Laya's reported scores on seven, and it answers a question in 31 ms (median, GH200). --- ### The Price of Thought: Does Test-Time Reasoning Pay in LLM Trading? Authors: Jiayi Chen, Guiling Wang Published: 2026-09-25 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.30705v1 While inference-time reasoning in large language models (LLMs) promises better decision making, its higher computational cost may not yield better economic outcomes. Yet reasoning controls are rarely evaluated as economic interventions, where changes in model outputs must translate into better portfolios after trading costs. We conduct a controlled study of representative LLMs from the DeepSeek, GPT, and Gemini families. We vary reasoning effort while holding information available at each formation date, prompts, output formats, and portfolio construction fixed. Our evaluation covers a full year of U.S. equities under three input conditions: numerical, identifiable news, and masked news. It includes more than 800,000 asset predictions and repeated model generations. Across all three model families, additional reasoning does not produce a reliable improvement in net portfolio returns. For DeepSeek, where we examine the full progression from no reasoning to maximum reasoning, performance is nonmonotonic. Repeated generations also produce unstable treatment effects and portfolio selections, even when overall scores remain similar. These findings show that additional reasoning can change financial decisions without reliably improving their economic value, motivating validation for each task before deployment. --- ### SAGE: Source-Anchored Guidance via Frequency Equalization for Hierarchical RGB-T Alignment and Fusion Authors: Timing Li, Yiming Sun, Boan Tao, Xiyuan Gao, Haifang Cao, Pengfei Zhu Published: 2026-09-25 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.30703v1 Spatial misregistration and cross-modal discrepancies often cause ghosting, structural blurring, and content imbalance in RGB-T fusion. Existing methods typically decouple appearance adaptation, geometric alignment, and information fusion, limiting dependency propagation across stages. We propose Source-Anchored Guidance via Frequency Equalization for Hierarchical RGB-T Alignment and Fusion (SAGE), a unified framework integrating frequency equalization, hierarchical alignment, and subband fusion. SAGE employs invertible joint encoding and source-specific low-frequency modulation to derive structural and gain guidance while preserving source information. Hierarchical frequency collaborative alignment estimates global affine geometry from low-frequency approximations and transfers geometric and contextual cues to high-frequency correlation reasoning for reliability-aware residual refinement. Guided subband fusion jointly aggregates the aligned frequency coefficients under propagated source and alignment guidance, coordinates complementary low- and high-frequency information, and reconstructs the fused image through the inverse wavelet transform. Extensive experiments on RGB-T datasets with real-world and synthetic misalignments demonstrate consistently competitive performance in alignment and fusion, validating the effectiveness of source-anchored guidance for weakly registered RGB-T images. --- ### MM-VeriAgent: Learning to Use Extensive Tools to Verify Multimodal Misinformation with Reinforcement Learning Authors: Peipei Li, Shuhan Xia, Shengyang Liu, Zekun Li, Ran He Published: 2026-09-25 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.30698v1 Real-world multimodal misinformation often involves mixed forgery sources, requiring sample-specific detection strategies. Existing tool-augmented methods rely on predefined workflows or inference-time planning, limiting adaptability or increasing inference cost. To address this issue, we introduce \textbf{MM-VeriAgent}, which learns to verify mixed-source multimodal misinformation with tools. We first build \textbf{MM-VeriTools}, a specialized toolkit for misinformation detection agents. By benchmarking various candidate models and methods on the sub-tasks required by mixed-source detection, we select the strongest for textual, visual, and cross-modal forgery analysis and encapsulate them as callable tools with a unified interface. On top of this toolkit, we train the LVLM agent with reinforcement learning to teach it how to use these tools to better solve mixed-source detection. Since many of the tools are specialized models whose online execution at every rollout severely limits RL efficiency, we further introduce \textbf{Tool-Execution Cache}, which pre-executes candidate tool calls and reuses their cached outputs during training. This preserves multi-step rollouts while reducing online tool execution, largely improving the training efficiency.Experiments on MMFakeBench demonstrate substantial accuracy gains over the base model without explicit tool search at inference time. Ablation and efficiency analyses further validate the learned tool-use policy and show that Tool-Execution Cache reduces online tool executions during training. --- ### LUMO (Lightweight Unified Multilingual Orchestrator): A Privacy Preserving Offline Voice Assistant Authors: Md. Mehedi Hasan Naeem, Mst. Kamrunnahar Ruma, Nafiza Anjum, Shakila Sultana, Md. Sujan Ali Published: 2026-09-25 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.30692v1 Reliable voice interaction is essential in environments with limited internet connectivity and strong privacy. However, most existing voice assistants depend on cloud-based services, which leads to latency issues, dependency on internet access, and privacy vulnerabilities. This research presents LUMO (Lightweight Unified Multilingual Orchestrator), a privacy preserving offline voice assistant designed for edge computing environments. This system integrates local Automatic Speech Recognition (ASR), locally deployed quantized Large Language Model (LLM), and Text-to-Speech (TTS) synthesis into a fully offline pipeline running on a Raspberry Pi 5 with 8 GB RAM. To enable efficient operation on resource constrained hardware, the language model is compressed using 4-bit GGUF quantization, which reduces memory usage while preserving practical conversational capability. Existing edge based voice assistants Mycroft provides partial offline functionality without a generative LLM, with an approximate latency of ~5 s and power consumption of ~12 W, while Rhasspy supports full offline operation but lacks generative capabilities, with ~3 s latency and ~11 W power usage. In contrast, LUMO achieves a Word Error Rate (WER) of 6.8% for short English utterances in low noise conditions, an end-to-end response latency of 2.0-4.0 s, and a lower peak power consumption of approximately 9.0 W. The system also achieves effective offline recognition for Bangla speech, supporting multilingual accessibility in low resource settings. By operating entirely offline, LUMO provides strong data privacy, reduced need for cloud connectivity, and suitability for privacy sensitive edge execution such as rural healthcare, education, and disaster response scenarios. --- ### Threat-Aware Energy-Efficient Deployment for Dynamic UAV Networks: A Multi-Agent RL Approach Authors: Faisal Al-Kamali, Hussein A. Ammar, Francois Chan, James H. Bayes, Yasser Gadallah, Mohamed H. Ahmed Published: 2026-09-25 Categories: cs.IT, cs.AI, cs.LG, eess.SP arXiv: https://arxiv.org/abs/2609.30690v1 Ensuring operational safety in threat-prone environments remains a critical challenge for multi-UAV networks serving as aerial base stations. This paper proposes an efficient framework to maximize global energy efficiency (EE) while promoting safe operation through threat-aware clustering and reward-based safety enforcement. The proposed framework is executed in three steps. First, a threat-aware K-means (TAKM) algorithm determines the minimum required UAVs and computes safe initial placements. Second, an optimal matching stage assigns physical UAVs to these centroids to minimize energy expenditure. Third, a threat-aware multi-agent twin delayed deep deterministic policy gradient (MATD3) algorithm dynamically optimizes trajectories, power, and user associations. Simulation results show that the proposed framework achieves zero observed safety violations in the considered scenarios while achieving superior EE and faster convergence than other learning methods and non-clustering baselines. Compared to heuristic optimization, the proposed framework outperforms the greedy particle swarm optimization (GPSO) and achieves performance comparable to that of the optimized PSO (OPSO), while incurring significantly lower online deployment computational complexity. Furthermore, the proposed framework demonstrates effective generalization to unseen user distributions, large UAV fleets, and different threat geometries, while maintaining zero safety violations. --- ### On the Limits of Univariate Deep Learning for Significant Wave Height Forecasting Authors: Yilin Zhai, Hongyuan Shi, Zaijin You Published: 2026-09-25 Categories: physics.ao-ph, cs.LG, physics.comp-ph arXiv: https://arxiv.org/abs/2609.30688v1 This study conducts a systematic hyperparameter search across five deep learning architectures, DLinear, LSTM, PatchTST, ResAttLstm, and Mamba2, and nine context lengths (1-168 h) for single-station significant wave height (Hs) forecasting on NDBC buoy 41009, followed by re-evaluation of the best configurations on a 47-buoy, 37-year corpus. The five families converge to a common performance level on the multi-buoy evaluation (between-family SD = 0.0014 m^2, 0.8% of the grand mean), a spread dwarfed by the 4.83x cross-dataset MSE shift between buoy corpora. All multi-buoy trials beat persistence (mean skill +0.062), but no architecture consistently outperforms the others. On the single-buoy experiment, skill peaks at 12-24 h where five trials fall below persistence, per-family Q4/Q3 test MSE ratios range from 2.4 to 2.6, and deep models underperform persistence for the most extreme 1% of waves. These findings are consistent with the interpretation that persistence already captures the dominant linear-inertial signal in univariate Hs, and that architecture engineering under this univariate input setting has reached diminishing returns: cross-buoy variance, not model class, dominates forecast error. Future work should prioritise atmospheric covariates, zero-shot cross-buoy transfer, and decomposition of Hs into swell and wind-sea components. By establishing a rigorous reference baseline for what univariate Hs models can and cannot achieve, this study provides a benchmark against which future multivariate and physics-informed approaches can be calibrated, and offers practical guidance for lightweight buoy-level forecasting in mid-latitude storm-dominated and swell-mixed environments. --- ### PixSim: a calibrated open-source simulator of instant-payment fraud, recovery and interdiction under analyst capacity constraints Authors: Bashir Zeimarani, Alireza Khatib, Somayeh Mousavinasr, Carlos Maurício Serodio Figueiredo Published: 2026-09-25 Categories: cs.LG, cs.CE, q-fin.RM arXiv: https://arxiv.org/abs/2609.30684v1 Brazil's Pix settles about 5.9 billion instant, irreversible transfers a month. A fraudulent transfer can be recovered only while the funds remain in a traceable account, and in 2025 the Central Bank's recovery mechanism (MED) returned 9% of accepted contested value. Interdiction therefore has to happen before settlement, by routing each transaction to pass, human review or block, under a finite analyst team and a regulatory hold window. To our knowledge no public simulator jointly models irreversible settlement, a regulated recovery mechanism, downstream fund dispersal and capacity-constrained review. We present PixSim, an open-source simulator of the Pix rail with these elements, calibrated to Banco Central do Brasil open data, with every parameter sourced, calibrated to one published observable, or registered as an assumption. With the model frozen, full-scale runs reproduce the 2025 recovery rate within 0.006 and its decomposition within 0.02; the February-April 2026 window is reported as a misfit and the May 2026 tracing regime as a projection. On a benchmark with a payer-side scorer, four reference policies and ten scenarios, within the simulated mule model: recovery after settlement is constrained by dispersal speed; staffing by the arrival profile cuts a fixed rule's alert expiry from 52% to 2% at constant hours; halving the team removes a fixed threshold-and-block rule's advantage over a queue-aware rule, on loss and on loss plus false-block harm (+0.106 of victim value, positive on all twenty paired seeds), while a reversal at two thirds of the team was not confirmed on independent seeds; and a synthetic scorer of held-out AUC 0.82 cuts lost value by about a quarter. Code and data: https://doi.org/10.5281/zenodo.22948895 --- ### Structure-Guided Masked Autoencoders for Ultra-High Resolution Scientific Image Understanding Authors: Enzhi Zhang, Du Wu, Rui Zhong, Cong Ma, Isaac Lyngaas, Amir Koushyar Ziabari, Xiao Wang, Peng Chen, Tao Luo, Toshio Endo, Fumiyoshi Shoji, Kento Sato et al. Published: 2026-09-25 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.30682v1 Self-supervised pre-training with Vision Transformers, including Masked Autoencoders (MAE), is difficult to apply to gigapixel scientific images. Random masking is poorly matched to the structured, multi-scale morphology of scientific data, while uniform tokenization produces prohibitively long sequences that make $O(N^2)$ attention impractical. We propose SGMA, a structure-guided masked autoencoding framework for ultra-high-resolution scientific images. SGMA couples two components: a content-adaptive quadtree tokenizer that compresses gigapixel images into a fixed-length sequence, and a structure-conditioned masking process that biases reconstruction toward spatially informative regions. To stabilize this process across scales, we introduce Damped Accumulation (DA), which aggregates signal-dependent responses across the tree into a structure canvas used to guide masking. The resulting pre-training task preserves fine microstructure while remaining compatible with standard ViT encoders and MAE-style reconstruction. Across electron microscopy, whole-slide optical microscopy, and X-ray CT datasets, SGMA consistently outperforms MAE baselines. It achieves 95.68% Dice on the 8K x 8K x 28K SpringXCT dataset, improving over the same-architecture MAE baseline by +13.00 points, and 83.21% Dice on the 32K^2 WSI PAIP dataset, improving by +16.84 points, while providing up to a 24.8x inference speedup. --- ### Design-Ignoring versus Design-Respecting World Models for Epidemiology Authors: Xiangyu Yu, Weiyu Liu Published: 2026-09-25 Categories: stat.ME, stat.AP, stat.ML arXiv: https://arxiv.org/abs/2609.30679v1 World models for epidemiology learn from records shaped by study designs, including assignment, sampling, measurement, and related processes. A model may therefore reconstruct observed trajectories while learning an intervention contrast that depends on how records were collected. We formalize study design as constraints on a world model latent-world interface, action mechanism, observation likelihood, and target readout. This distinguishes design-respecting models, which encode these constraints, from design-ignoring models, which fit selected records without relating assignment and observation to the intended intervention question. Using a large scale cluster-randomized test-negative trial, we hold the latent structure and fitting settings fixed and compare a design-ignoring case-count model with a design-respecting test-negative observation model. Both models achieve comparable factual reconstruction. Yet across 500 paired resampling experiments at different relative sampling intensity, median contrast changes are substantial for the design-ignoring model, versus merely marginal for the design-respecting model. The same qualitative separation holds when sampling also varies across clusters. Thus, factual reconstruction alone does not establish design alignment; testing whether fitted contrasts preserve design-implied observation-process invariances provides a sharper evaluation. --- ### TRACE: Temporal Audit and Condition-aware Evaluation of Streaming Video Understanding Authors: Yibo Ma, Qianqian Zhang, Peng Liu, Tiancheng Zhao Published: 2026-09-25 Categories: cs.CL, cs.CV arXiv: https://arxiv.org/abs/2609.30670v1 Streaming video understanding requires models to interpret evidence as it arrives, yet current evaluations often report task scores without specifying when evidence becomes valid, how visual history is maintained, or how responses are triggered. As a result, similar scores may correspond to different workloads, failure modes, and operational behavior. We introduce TRACE (Temporal Audit and Condition-aware Evaluation), a condition-aware benchmark and evaluation framework that makes these factors explicit. TRACE combines temporally audited visual tasks with evidence timing and instruction-dependent trigger annotations, a unified causal Core--Adapter protocol that controls information availability while recording actual history processing and response events, and multidimensional reporting of answer quality, timeliness, response-selection behavior, workload, completion, and reliability. On 1,240 records from 517 videos, we evaluate eight publicly available models or systems in eight configurations. We find that nearly identical QA accuracy can mask substantial differences in completion, answer validity, and generation workload, while proactive performance separates into response quality, response delay, false alarms (responses emitted while no target window is currently valid and a later one remains), and missed target windows. These results show that streaming-video performance should be interpreted as execution-conditioned system behavior rather than a single score. Our benchmark and code can be accessed at \href{https://github.com/om-ai-lab/trace-bench}{https://github.com/om-ai-lab/trace-bench}. --- ### StarWM: Self-Supervised Trained Attention Routing for Robust World Models Authors: Zeqiang Zhang, Fabian Wurzberger, Maximilian Otte, Daniel Schmid, Sebastian Gottwald, Arne Peter Raulf, Daniel Alexander Braun Published: 2026-09-25 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.30667v1 A robust world model must strike the balance between faithfully capturing environmental dynamics and abstracting away from irrelevant content. While reconstruction-based world models ensure faithful supervision, they misallocate representational capacity by pixel area rather than dynamics relevance for visual tasks, which can cause task-irrelevant content to dominate the learned representation. Alternatively, reconstruction-free methods avoid this bias but risk discarding possibly relevant information. We propose StarWM, which uses a cross-attention module trained on self-supervised dynamics to decide where reconstruction applies. A dual-stream decoder then restricts reconstruction to the attended regions, with stop-gradient barriers preventing interference between the two objectives. These components allows reconstruction to supervise the visual content of attended regions without contaminating the latent with non-predictive information. On DeepMind Control with dynamic video backgrounds, default (reward-free) StarWM achieves the strongest performance under random-frame distractors and substantially outperforms reconstruction-based baselines under sequential video. In addition, its reward-augmented variant matches or exceeds reconstruction-free methods on sequential video, achieving the highest overall return across all distractor regimes. Mechanistic probing confirms StarWM preserves state attributes with near-perfect fidelity through long-horizon imagination while systematically discarding distractors. --- ### LLM Parkinsonism: Executive-Control Failure, Token-Inefficient Persistence, and an Uncertainty-Aware Global Executive Control Architecture for Autonomous Language-Model Agents Authors: Dongsheng Xiao, Zeyuan Wang, Xuzhe Xia, Bo Zhao, Yankai Cao Published: 2026-09-25 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.30662v1 Large language models (LLMs) can plan, use tools, write code, and execute long-horizon workflows, yet strong local competence does not guarantee project-level executive control. Agents may continue acting after the original objective is satisfied, producing low-value refinements, repeated verification, and repairs to self-created complexity. We use LLM Parkinsonism as a narrowly defined, non-clinical metaphor for this pattern of persistent action despite diminishing task-level value. We argue that the problem is not explained by autoregressive next-token prediction alone, but more directly by concentrating proposal generation, scope interpretation, progress assessment, and stopping authority within the same self-conditioned loop. We therefore introduce Global Executive Control (GEC) v0.2, an uncertainty-aware governance architecture that separates action generation from project-level control. In a 24,000-episode matched-candidate benchmark under a common 40,000-token ceiling, a first-candidate baseline achieved 67.42% hard-goal success, a candidate-set local control achieved 96.53%, and GEC achieved 96.57%. The candidate-set control shows that access to multiple candidate actions explains most of the success gain; relative to that control, GEC preserved success while reducing mean token use from 19,782 to 12,574 (36.4%) and restricted mean tokens to completion at the 40,000-token ceiling from 16,136 to 13,114 (18.7%), while eliminating measured pre-completion drift and sharply reducing gross complexity. Governance-overhead sensitivity remained favorable through an additional 500 synthetic governance tokens per cycle. These mechanistic simulations support explicit governance of scope, evidence, resource use, and stopping, while live-model validation remains necessary. --- ### Population loss in shallow ReLU networks: Bias & families of critical points Authors: Michael Field Published: 2026-09-25 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.30661v1 The main result presented is a formula for the population loss in the student-teacher kernel model that is applicable to shallow ReLU networks with bias. This extends previous work of Choo and Saul (2009) and Brutzkus and Globerson (2017). The formula makes essential use of Owen's T-function. The necessary theory of the T-function is given and a high precision coding using MPFR for the T-function, based on an algorithm of Komelj (2023), is available on request. It is shown that various families of spurious minima described in past papers of Arjevani and the author extend to biased networks and that the loss is always strictly decreased when bias is added. The change in landscape geometry caused by adding bias appears to be relatively mild. Only the simplest examples are described in this paper where it is assumed that the number of inputs is equal to the number of neurons (this restriction is for reasons of length). A review of relevant previous results on unbiased networks is included. Aside from Gaussian statistics, the main mathematical tools and ideas come from analytic geometry (analytic and subanalytic sets, the Curve Selection Lemma). --- ### Image Reconstruction from Phase with Untrained Neural Priors Authors: Ene Meco, Ahmet Enis Cetin Published: 2026-09-25 Categories: eess.IV, cs.CV arXiv: https://arxiv.org/abs/2609.30659v1 Fourier phase encodes important spatial image structure, but recovering an image without measured spectral magnitude requires additional constraints and leaves absolute intensity ambiguous. We propose a projection-based two-stage framework that combines Fourier-phase and spatial-support constraints with an image-specific neural prior. The first stage alternates constraint enforcement with regularized neural-prior updates, while the second performs phase/support refinement alone with guaranteed convergence. We evaluate two neural-prior implementations on the same 77 microscopy images and compare them with a constraint-only baseline. After 500 final refinement passes, the best-performing variant achieves 31.41 dB pooled PSNR, 35.75 dB mean PSNR, and 0.9531 mean SSIM, improving pooled PSNR by 1.51~dB and reducing pooled MSE by 29.3% relative to the baseline. The results demonstrate the benefit of combining neural guidance with explicit constraint refinement at the evaluated iteration budget, while showing that lower phase residual alone does not guarantee greater reconstruction accuracy. --- ### DiffusionShadow: Diffusion-based Shadow Caching for Neural Volume Rendering Authors: Kai-Chen Tung, Qi Wu, David Bauer, Mengjiao Han, Silvio Rizzi, Kwan-Liu Ma Published: 2026-09-25 Categories: cs.GR, cs.LG arXiv: https://arxiv.org/abs/2609.30658v1 Implicit neural representations (INRs) have gained momentum in scientific visualization due to their compactness and scalability to large datasets, making them well suited for integration with direct volume rendering (DVR). However, real-time volume rendering of INR with advanced illumination effects, such as shadows, remains computationally expensive, as evaluating shadow terms via ray marching is costly. Alternatively, precomputing and storing shadows for many lighting directions is prohibitive in both memory and storage. To address this, we introduce a diffusion-based shadow caching framework that compresses a vast set of pre-calculated shadow INRs into a single diffusion model. Rather than focusing on generalizing to unseen directions, our method effectively memorizes and reconstructs a dense set of pre-trained lighting conditions on the fly. We first encode a collection of shadow coefficient volumes as shadow INRs, and then train a diffusion model conditioned on lighting direction to predict the corresponding shadow INR weights at inference time. This design integrates directly with standard INR renderers without additional runtime sampling. Experiments show that our approach achieves faster rendering than traditional methods while bypassing the massive storage bloat of independent INRs, producing shadows that closely match most of the reference results. --- ### Prompt Injection Detection for Email Agents Through Attack Chain Modeling Authors: Ahmad Hashmi, Dhyey Patel, Yunting Yin Published: 2026-09-25 Categories: cs.CR, cs.CL arXiv: https://arxiv.org/abs/2609.30657v1 Large language model email assistants are particularly vulnerable to indirect prompt injection because untrusted email content can be retrieved into the model context and influence subsequent tool use. Existing prompt injection detectors mainly formulate this problem as binary malicious text classification, which overlooks the important factor that harmful agent behavior often arises through a sequence of stages. We propose a detection framework that models this attack chain by combining a text detector, verifiers specific to each stage, explicit rule-based risk signals, user intent and action consistency analysis, and a logistic decision policy. To support this framework, we derive attack chain labels from prompt injection datasets, evaluate the proposed framework under random splits, temporal phase transfer, conditional stage transfer, cross-dataset transfer, and conduct ablation studies on multiple benchmarks. Results show that random train test splits substantially overestimate robustness under distribution shift, while later tool argument stages are more predictable than earlier stages in the framework. We also show that training on harmless emails that resemble attacks helps reduce false alarms while preserving the ability to detect real attacks. Across five binary benchmarks, our framework achieves a mean F1 score of 0.406 under the strict threshold setting policy, compared with 0.216 for the strongest of five pretrained detectors evaluated without additional training. These results highlight the value of combining attack stage predictions with checks for conflicts between the user's request and instructions in retrieved emails. Our experiments also demonstrate the importance of training with challenging benign examples to balance attack detection and false alarms. --- ### Recursive Self-Improvement via On-Policy Distillation for Reasoning Authors: Shangjian Yin, Zehao Zhao, Kavosh Asadi, Rui Liu, Yuchen Lu, Shike Mei, Hang Cui, Luke Simon, Zhouxing Shi, Hamed Firooz Published: 2026-09-25 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.30652v1 On-policy distillation (OPD) trains a student model by having it generate trajectories, then matching its next-token predictions with an external teacher's next-token predictions. This provides dense, token-level supervision to the student. On-policy self-distillation (OPSD) eliminates the need for the external teacher. Specifically, a second frozen copy of the student model, now given the ground truth in its context, serves as the teacher. The student model only receives the problem and learns to mimic the privileged teacher model, while the teacher remains frozen throughout training. Previous work showed that freezing the teacher is useful for training stability, but we argue that this can prevent the teacher from incorporating the improvements learned by the student during training. Our primary contribution is to address this limitation with a recursive framework built around two complementary components. First, we let the privileged teacher co-evolve with the student so that revision learned in one round can guide the next, a process we refer to as Dynamic Co-Evolution (DCE). Second, because stronger revision can also make responses too verbose and self-critical, we additionally train on shorter, verified rewrites of the model's own on-policy responses. We call this complementary objective Self-Refined Concise Learning (SRCL). Overall, our comprehensive evaluations show that DCE+SRCL outperforms OPSD across multiple model scales and four competition-level mathematics benchmarks. Specifically, on Qwen3-8B, DCE+SRCL reaches 65.97% Average@12, outperforming OPSD by 35.62 percentage points while reducing mean output length by 7.80% relative to DCE alone. --- ### Causal Retention in Interactive Agents: Interface Factorization and Selective Adaptation Authors: Shengjun Zhang, Tingyi Liu, Dong Xie, Yunlong Dong, Xiang Wang, Cheng Zeng Published: 2026-09-25 Categories: cs.LG, cs.AI, stat.ML arXiv: https://arxiv.org/abs/2609.30650v1 Task performance need not determine which intervention mechanism an agent retains. We study causal retention: whether a frozen learned state answers a mechanism-probe map fixed independently of training, including action, context, direct target, value, and delay. For finite structural causal model classes, the optimal probe error is a Bayes decision risk. It vanishes exactly when every learning-interface fiber lies within one probe-answer fiber; any state obtained by post-processing that interface inherits the same lower bound. A posterior-coverage theorem characterizes budgeted retesting, while an exact edit decomposition shows that the shifted set is the unique support of an error-free target update. Causal Core implements these conditions through evidence-gated writing, readout filtering, temporal credit, hidden-context setup, and local diagnostic updates. Experiments cover finite causal systems, continuous simulators, an official TD-MPC2 world model, and Qwen2.5-7B-Instruct. A frozen Qwen last-layer probe reaches 0.958 balanced accuracy on source mechanisms but 0.583 on changed delays; the gated mechanism state reaches 1.000 and accepts only 0.056 of synchronized-readout candidates. In TD-MPC2, five target states per actuator recover effect-sign accuracy from 0.057 to 0.948 without degrading stable responses. Causal retention is therefore distinct from task sufficiency and source-domain decodability. --- ### Conditional Predictive Sufficient Statistics for Visual Representation Learning Authors: Yuzhou Hong Published: 2026-09-25 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.30647v2 A useful visual representation is a statistic of the observed past that retains the latent factors shared with the future and discards patch-private noise. We formalize this requirement as a conditional predictive sufficient statistic (CPSS). Under a shared-factor model of image patches, the mutual information between the past and the next patch equals the information the past carries about the shared factor, up to a remainder that the next patch itself fails to reveal. Predicting the next patch embedding with a cosine loss is maximum likelihood for a von Mises-Fisher model of that embedding's direction, and is therefore a tractable surrogate for the predictive information. The same population loss is also minimized by a constant embedding, so stop-gradient does not by itself select the sufficient statistic; it only blocks the symmetric gradient that implements the constant solution in one step. The regression target is a shallow embedding, which forces the network output back into that shallow range and leaves the sufficient statistic in intermediate blocks. Small causal Transformers on MNIST and CIFAR-10 are used as diagnostics, not as a leaderboard. On MNIST the future shift and the stop-gradient move probe accuracy by tens of points, and the CPSS readout peaks before the output. On CIFAR-10, with the same short budget and no augmentation, every objective lands near a linear classifier on pixels. What still matches the derivation is the geometry: the CPSS output is a worse readout than its best intermediate block, next-pixel regression does not pay that penalty, and removing the stop-gradient collapses the effective rank of the embedding even when the pretext loss looks perfect. --- ### MARCEDES: Score-based causal discovery under non-Gaussianity with continuous optimization Authors: Anamitra Chaudhuri, Anirban Bhattacharya, Yang Ni Published: 2026-09-25 Categories: stat.ML, cs.LG, stat.CO, stat.ME arXiv: https://arxiv.org/abs/2609.30643v1 We consider the problem of learning the underlying causal directed acyclic graph (DAG) structure corresponding to a structural equation model (SEM) with non-Gaussian errors. Motivated by an intentionally misspecified non-Gaussian SEM with all Laplace errors, we first introduce the mean absolute residual risk, defined over the space of all real matrices, and show that, asymptotically, the risk of the true weighted causal DAG matrix is strictly smaller than that of any other matrix. Nevertheless, to enhance generality and account for high-dimensional and finite-sample settings, we further incorporate row-specific sparsity penalties along with a soft DAG constraint to derive a continuous score function over the space of real matrices. Accordingly, we propose a score-based DAG learning method, named MARCEDES, formulated as an unconstrained score minimization problem, which can be efficiently solved using gradient-based optimization techniques, thereby circumventing the challenges associated with constrained optimization. Furthermore, we develop a computational algorithm to handle the non-smoothness of the score objective and to enable optimal tuning of row-specific sparsity penalties under a generalized Bayes framework. Finally, we demonstrate the efficiency and improved performance of the proposed method over existing approaches through an extensive simulation study. --- ### A Framework for Identifying, Categorizing, and Explaining Bias in AI-Generated Code Authors: Manaal Basha, Aimee M. Ribeiro, Gema Rodriguez-Perez Published: 2026-09-25 Categories: cs.SE, cs.AI arXiv: https://arxiv.org/abs/2609.30642v1 As Large Language Models (LLMs) become integrated into software development workflows, concerns regarding unintentional biases in AI-generated code. Although evidence suggests these biases exist, limited research has systematically identified, categorized, and explained them. This study investigates bias in AI-generated code and evaluates whether LLMs can reliably identify and explain it through a taxonomy-driven framework. We extended an existing dataset of biased AI-generated Python code and manually annotated snippets with bias categories and human-authored justifications to establish a ground-truth dataset. Using this dataset, we evaluated proprietary and open-source LLMs as automated bias detection and justification systems through ICL. Finally, we analyzed similarity between LLM-generated explanations and human-authored justifications using structured justification and code identification metrics. Our findings demonstrate that LLMs can effectively support code bias identification and explanation. Gemini achieved 80.14% classification accuracy, with 84.0% precision and 95.7% recall, while the best open-source alternative, Qwen3-coder, achieved 82.45% accuracy, 68.64% precision, and 80.22% recall. Additionally, the models achieved justification similarity scores of 80.4% and 80.14%, respectively, relative to human-authored reasoning, and code identification similarity scores of 86.0% and 87.82%. These results suggest that LLMs can detect biased logic in generated Python code and produce explanations that substantially align with expert interpretations. --- ### Beyond Rule-Based Mutation Testing: Test-Aware Mutant Generation Using Large Language Models Authors: Nils Kiele, Zainab Saad, Zirui Wang, Steve Drew, Samira Ebrahimi Kahou Published: 2026-09-24 Categories: cs.SE, cs.AI arXiv: https://arxiv.org/abs/2609.35841v1 Mutation testing evaluates test-suite adequacy by injecting synthetic faults into program code. However, traditional rule-based tools often generate large numbers of trivial, redundant, or equivalent mutants that limit their practical use for identifying gaps in a test suite. While recent large language model (LLM)-based approaches generate more realistic faults, most remain test-blind: The model sees only the source code and cannot reason about what existing tests already cover. We propose test-aware mutant generation, in which an LLM receives the problem statement, canonical solution and base tests in a single prompt, and must generate a nontrivial mutant that passes the base unit tests. We evaluate this approach across a set of five LLMs -- Gemini 3.1 Pro, Gemini 3 Flash, GPT 5.1 Codex Mini, GPT 4.1 Mini, Qwen3-32B -- on the HumanEval and MBPP benchmarks. The extended EvalPlus test suites serve as an automated oracle to verify whether surviving mutants represent genuine bugs. Test-aware prompting yields verified fault rates of 87.7% (HumanEval) and 79.1% (MBPP), meaning these mutants pass all base tests but are caught by the oracle. This vastly outperforms the matched test-blind prompting (which yields only 12.2% and 23.0%, respectively) and the traditional rule-based tool mutmut (4.4% and 5.7%). While fault subtlety (the fraction of extended tests a mutant fails) remains comparable across all three methods, test-awareness minimizes the computational cost per verified fault, compared to test-blind prompting. Exposing an LLM to existing unit tests shifts mutant generation from untargeted bug injection toward effective discovery of weaknesses in an existing test suite. Our work establishes a concrete foundation for future research to scale test-aware mutant generation to production-level environments. --- ### Estimation of Room Impulse Responses from Handclaps Authors: Shih-Yu Lai, Kyung Yun Lee, Nils Meyer-Kahlen, Eloi Moliner, Bing-Yu Chen, Vesa Välimäki Published: 2026-09-24 Categories: cs.SD, cs.AI arXiv: https://arxiv.org/abs/2609.35839v1 Handclaps provide an equipment-free excitation for room acoustics, but their unknown and variable source waveform makes room impulse response (RIR) estimation challenging. In this work, we investigate whether RIRs can be estimated directly from handclaps. To this end, we introduce an anechoic handclap dataset containing 2,540 claps from 17 participants, designed to capture variability across natural claps and different hand configurations. We first establish the performance attainable when the excitation clap is known using regularized deconvolution, and show that approximating the unknown excitation by windowing the direct sound from the reverberant recording is insufficient. To estimate the RIR without a known excitation, we propose using the anechoic handclap recordings to train a deep neural network with a supervised regression objective. Evaluated on a controlled synthetic benchmark, the proposed neural regressor significantly outperforms windowing-based baselines across all instrumental metrics. Furthermore, we test the proposed method on handclap recordings measured in real acoustic spaces, showing that the inferred RIR spectra are consistent across different handclap measurements taken in the same room location. These results showcase the feasibility of directly estimating RIRs from natural handclaps without requiring knowledge of the excitation signal. --- ### Amadeus: When Models of People Meet Authors: Karl Hanna Published: 2026-09-24 Categories: cs.MA, cs.LG arXiv: https://arxiv.org/abs/2609.35835v1 With the sheer constant advancements raining down in the field of Artificial Intelligence, one particular possibility that may cross our mind is whether it is possible to model agents after humans and, in turn, use these agents to carry out synthetic interactions that predict their real counterparts, or even interactions at a larger scale such as groups or societies. In this paper, we test a more controlled version of this question through chess. We use 8 elite chess players, seal their direct pairwise games, learn each player independently using different methods, and then compose the resulting models on the withheld dyads. To evaluate the generated interactions, we use two measurements: opening-family total variation distance and win-draw-loss (WDL) total variation distance. M1 reduces WDL-TV while leaving opening-family TV largely unchanged, whereas M2 substantially reduces opening-family TV while having little effect on WDL-TV. An additional post-hoc method combining components of the other two retains improvements across both measurements. These results suggest that independently learned models can recover aspects of previously unseen interactions, and that recovery across different behavioural aspects need not be mutually exclusive. --- ### Neurosymbolic Routing for Reliable Reasoning on Resource-Constrained Edge Devices Authors: Avyay Sadhu, Alvaro Velasquez, Lekai Chen Published: 2026-09-24 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.35833v1 Running a language model on edge hardware provides private and low-latency reasoning without a network connection, and yet the small models that fit on such devices are unreliable on the tasks computers are expected to handle well, such as arithmetic, algebra, and formal logic problems. We argue that much of this unreliability is avoidable. Many queries appearing to demand reasoning are in fact structurally deterministic and permit fast and exact symbolic solutions. Therefore, forcing a probabilistic model to approximate them sacrifices accuracy and energy for little benefit. We present a neurosymbolic router that classifies each incoming query and dispatches it to the cheapest correct solver, sending structured tasks to deterministic engines and reserving the small language model (SLM) for open-ended word problems. Instead of hand-coding the routing logic, we learn a deterministic finite automaton (DFA) with the L* grammatical inference algorithm, using the SLM as a membership oracle and labeled data as an equivalence oracle. On a Raspberry Pi 4B (8 GB RAM, no GPU), evaluated on 100 untested prompts from DeepMind Mathematics, GSM8K, and RuleTaker, learned routing attains 100% routing accuracy and 98.3% overall accuracy with a 512-token reasoning budget (93.3% on word problems), compared with 72.0% for the strongest agent baseline, Program-of-Thought, and 58.7% for a tool-calling agent given the same solvers. Since formatted queries never reach the model, the router answers them in 1-11 ms and, in its 30-token configuration, runs 8.8x faster and 2.8x more energy-efficient than Program-of-Thought. --- ### Optimal transport meets speech: a tutorial review Authors: Xugang Lu, Yu Tsao Published: 2026-09-24 Categories: eess.AS, cs.AI, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.31787v1 Optimal Transport (OT) provides a principled framework for comparing and transforming probability distributions while preserving geometric structure. Recently, OT has gained significant attention in machine learning due to its ability to measure discrepancies between distributions, even when their supports do not overlap, making it effective for tasks such as generative modeling, domain adaptation, and transfer learning. Despite its success in fields such as computer vision and natural language processing, OT remains relatively underexplored in speech research. Speech signals present unique challenges, including temporal dynamics, speaker variability, noise, reverberation, and heterogeneous multimodal representations involving audio, text, and visual information. These factors often lead to distribution mismatches, where OT offers a natural framework for alignment and interpretation. This work aims to promote broader adoption of OT in speech processing by: (1) reviewing OT foundations through intuitive physical interpretations and highlighting connections to modern generative models; (2) presenting computational algorithms suitable for deep learning frameworks; and (3) demonstrating OT applications in cross-domain and cross-modal speech tasks, including speech enhancement, automatic speech recognition, language and speaker recognition, and audio spoof detection. We highlight OT's strong potential for addressing distributional variations in real-world speech applications. --- ### Convergence-Aware Pareto Selection of Covariate Scaling Transformations for Markov Deterioration Hazard Models: Evidence from Bridge Inspection Data Authors: Takato Yasuno, Keita Kobayashi, Ryuta Sakaguchi Published: 2026-09-24 Categories: stat.ME, cs.LG arXiv: https://arxiv.org/abs/2609.31786v1 Maximum likelihood estimation of Markov deterioration hazard models for infrastructure asset management is sensitive to the numerical conditioning of explanatory covariates, yet production pipelines often adopt a single scaling convention without systematic justification. We study four covariate scaling transformations---baseline max scaling, min-max scaling, z-score scaling, and Box-Cox transformation with a training-derived positivity shift---applied to an Exponential Hazard Markov (EHM) model estimated via L-BFGS-B on bridge inspection transition data. ---