# Research Papers — 2026-09 · bundle page 229 of 266 > Full-text markdown of papers #11401–11450 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/229 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/228.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/230.md) ### Functional Attentive Interpretable Regression Authors: Haixu Wang, Tianyu Guan, Jiguo Cao Published: 2026-09-05 Categories: stat.ML, cs.LG arXiv: https://arxiv.org/abs/2609.05846v1 In function-on-function regression, the coefficient surface $β(s,t)$ may exhibit complex support structure---from localized patches to global patterns such as disconnected regions, bands, or rings---where effect similarity does not align with Euclidean proximity. Projection-based methods that rely on fixed basis expansions can obscure such structure, while direct smoothing approaches risk oversmoothing the surface and its boundaries. We propose Functional Attentive Interpretable Regression (FAIR), which represents $β(s,t)$ directly through coordinate features and uses self-attention to learn effect-adaptive neighborhoods, enabling information sharing at both local and global scales. A scalar compression network maps these learned representations to the coefficient surface. Sparsity and smoothness penalties applied over these neighborhoods promote localized support with coherent boundaries. We establish a sieve equivalence to tensor-product spline spaces and derive convergence rates. Simulations and applications to oceanographic and hydrological data demonstrate that FAIR recovers support geometry more accurately than existing methods while achieving superior prediction, particularly under sparse sampling. --- ### SinoGlyphBench: A Diagnostic Benchmark for Chinese Glyph-Level Obfuscation in Language-Model Moderation Authors: Yifan Wang, Zimu Wang, Suliu Qin, Changyu Zeng, Tong Chen, Siqi Chen, Yijie Lin, Lingyu Jiang, Jionglong Su, Yushan Pan, Haiyang Zhang, Wei Wang et al. Published: 2026-09-05 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.05843v1 Glyph-level obfuscation can leave harmful Chinese content readable to humans while degrading automated moderation. We introduce SinoGlyphBench, a diagnostic benchmark that identifies label-critical semantic anchors and creates matched original and glyph-obfuscated inputs in text and image modalities. By perturbing anchors, background context, or both, this design distinguishes corruption of moderation-relevant evidence from general surface variation. Across 176,916 paired evaluations of 12 LLMs and MLLMs, obfuscation increases harmful false-negative and false-positive rates by 6.1 and 4.7 percentage points, respectively, and reduces four-way accuracy by 5.0 points. Models retain 75.7% of the decisions that were correct on the matched original inputs. Full-scope perturbations cause the largest degradation, anchor-only perturbations are more damaging than background-only perturbations, and cross-script substitution is particularly difficult in the text modality. Analysis of structured outputs identifies observable mismatches in visible-form reading, intended-message recovery, and final safety judgment. The evaluated models, therefore, remain brittle to Chinese content written with non-canonical glyphs. Resources are available at https://github.com/fengshun124/SinoGlyphBench. --- ### Do Quantum AIs Dream in Paths? Path-Integral Slow Thinking through Grover Interference Authors: Xiansheng Cai, Xiu-Hao Deng, Kun Chen Published: 2026-09-05 Categories: quant-ph, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.05842v1 Reinforcement learning with verifiable rewards enables large language models to think slowly, but the same training can induce policy collapse: probability concentrates onto a few successful trajectories and exploratory diversity erodes. We ask whether quantum AI can realize slow thinking differently. We formulate slow thinking as coherent dynamics over reasoning trajectories, a discrete path integral in which action sequences coexist in superposition and recombine before measurement. In our trainable realization, an exact verifier partitions the ensemble into collective accepted and rejected components that interfere under Grover amplitude amplification. A finite Grover evolution is maximized when the pre-amplification success probability lies at an analytically determined value below one, so inference itself defines an interior training target and removes the monotonic pressure toward unit success. In exact statevector simulations of a 2x3 sliding puzzle, Grover training reaches accuracy 0.95 on a 32-question training set at one round, against 0.73 for the strongest classical control. On held-out questions specialization has a cost: an untrained uniform policy read out through the same amplification remains the strongest reference on this solution-dense benchmark, and quantum training preserves far more held-out accuracy than classical training - at four rounds with matched circuit applications the two quantum models reach 3.2 and 3.9 times the strongest classical controls. The number of training questions supported by fixed-size policies trained at each amplification budget also grows faster with the budget than with matched classical repetition. These results establish a Grover-based realization of path-integral slow thinking: the interior target preserves exploratory path diversity, and ensemble-level interference converts it into verified performance. --- ### Map the Possibilities: Spatial Belief Fields for Language-Goal Aerial Navigation Authors: Haotian Xu, Yue Hu, Zhengqiu Zhu, Yong Zhao, Wenguan Wang, Quanjun Yin Published: 2026-09-05 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05841v1 Language-goal aerial navigation requires an agent to local- ize a potentially unobserved target from relational instruc- tions and partial observations, and translate this inference into metric actions in large-scale continuous environments. Existing methods often reduce language grounding to one single waypoint or action, prematurely collapsing the spatial uncertainty inherent in incomplete evidence and ambiguous relations. To address this limitation, we introduce SBFNav, a closed-loop navigation framework centered on a language- conditioned Spatial Belief Field (SBF). Unlike ego-centric maps that primarily record what has been observed, SBF rep- resents a task-conditioned distribution over plausible target locations, preserving multiple spatial hypotheses under par- tial evidence. At each step, this distribution is updated from accumulated observations as new evidence becomes avail- able. Built on this representation, SBFNav selects the goal that best aligns with the instruction and observations as a met- ric waypoint for control. Experiments on both the original and revised CityNav benchmarks achieve the best reported overall performance. On the Test Unseen split, our method improves SR from 25.91% to 32.29% and SPL from 19.63% to 30.43%. Ablation studies further confirm the advantages of spatial-belief modeling over single-point prediction. --- ### AgentBrew: Offline Tool-Use Agent Learning from Raw Real-World Trajectories Authors: Zhiyi Lyu, Yewen Li, Longtao Zheng, Shengtian Yang, Lang Feng, Lei Feng, Peng Jiang, Kun Gai, Qingpeng Cai, Bo An Published: 2026-09-05 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05837v1 LLM-based agents are increasingly deployed in real-world applications through tool-use APIs, yet training them for specific environments remains fundamentally difficult: real-world applications provide no pre-defined tasks or verifiers, no faithful simulators, and limited budget for large-scale environment interaction. In this paper, we propose \textbf{AgentBrew}, an offline training framework that learns effective tool-use policies from a single batch of raw interaction trajectories, without task verifiers or iterative on-policy rollouts. The agent first explores the target environment to collect a raw trajectory corpus without quality filtering. To extract training signal from this noisy corpus, \emph{retrospective task inference} reconstructs an aligned instruction for each trajectory based on its actual outcome, and \emph{PMI-Based credit assignment} decomposes the trajectory's total information about the inferred instruction into additive per-action credits via pointwise mutual information (PMI). These credits weight the policy training objective, amplifying informative actions while suppressing ineffective ones. On three real-world MCP applications (GitHub, Notion, PostgreSQL), AgentBrew improves Qwen3-32B by +8.7 Acc / +9.7 Score on average, surpassing Qwen3-235B (+2.3 / +4.4) and outperforming rejection sampling (+5.9 / +10.3). These results demonstrate that fine-grained offline learning can recover useful supervision from raw trajectories that filtering-based approaches would discard. The code is available at https://github.com/alphatogo/AgentBrew --- ### Learning Counterfactual World Models for Embodied Reasoning under Partial Observability Authors: Todd Y. Zhou, Daniel Zhang Published: 2026-09-05 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05834v1 World models promise a general route to embodied intelligence: learn predictive dynamics once, then reason, plan, and act with them. Increasingly, the representations beneath such models are pretrained on large-scale video, interaction, and multimodal corpora, which raises a question prediction quality alone cannot answer: when is a learned representation actually actionable? We identify a failure mode we call counterfactual collapse: a model predicts visually plausible futures while failing to distinguish interventions with different behavioral consequences. This arises whenever a representation is optimized for perceptual similarity rather than intervention structure, which is precisely the objective under which most large-scale pretrained encoders are learned. We introduce Counterfactual Latent World Models (CLWM), which combine a recurrent belief-state encoder, action-conditioned latent dynamics, and a contrastive counterfactual objective that separates futures induced by distinct interventions even when their observations look alike. Across occluded manipulation, aliased navigation, and long-horizon manipulation, CLWM improves planning success over the strongest baseline (65.1% $\to$ 74.6% on Occluded Push and 67.3% $\to$ 78.9% on Aliased Maze) and reduces exploitative planning failures (18.4% $\to$ 9.7% on Deferred Kitchen), with ablations attributing the gains to hard counterfactual negatives, especially perceptual-alias negatives. Finally, our counterfactual separability metric, which tracks planning success across the five baseline model classes ($r \ge 0.94$), is representation-agnostic: given intervention-outcome labels, it can audit any encoder, pretrained or trained from scratch, before a planner trusts it. We do not yet measure it on large-scale pretrained encoders. Here we establish the metric and its relationship to planning success for world models trained from scratch. --- ### Scaling Optimal Classification Trees via Adaptive Feature and Sample Reduction Authors: Jiancheng Tu, Wenqi Fan Published: 2026-09-05 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.05826v1 Dynamic programming for optimal classification trees becomes computationally expensive as the numbers of features and training samples increase. We develop a joint feature- and sample-space reduction framework based on STreeD. Weighted STreeD merges duplicate records created after projection onto a fixed candidate set into weighted representatives. This reduces sample-dependent computation without changing the fixed-candidate optimization problem. Adaptive STreeD repeatedly refines a bounded candidate set, retains features used by the incumbent tree, rebuilds the weighted representation, and solves the resulting reduced problems. Each certified Weighted STreeD solution is optimal for its current candidate set, while the outer feature search remains heuristic over the full feature space. Experiments on five data sets show that Weighted STreeD achieves speedups of up to 121.41 times over standard STreeD. Adaptive STreeD reduces runtime in matched comparisons at depths 2 to 4 and continues to return feasible trees at greater depths where full-feature methods are limited by time or memory. Under the same computational budget, its predictive performance remains comparable to the evaluated optimal classification tree baselines and is higher in some comparisons. These results show how joint feature- and sample-space reduction can scale dynamic-programming-based optimal-tree learning to more demanding instances. --- ### Beyond Top-$k$ Skill Retrieval: Diversity-Aware Skill Routing for LLM Agents Authors: Wang Wei, Tiankai Yang, Samyadeep Basu, Hongjie Chen, Yue Zhao, Zhengzhong Tu, Xiyang Hu, Franck Dernoncourt, Ryan A. Rossi, Hoda Eldardiry Published: 2026-09-05 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05824v1 Large language model (LLM) agents increasingly rely on external skills, but routing user requests over large skill registries is difficult because many skills are functionally redundant while complex tasks often require complementary skill sets. Existing skill routers typically rank candidates independently by query relevance, which can waste context budget on redundant skills. We propose Diverse Skill Routing (DSR), a diversity-aware reranking framework that uses a Determinantal Point Process to balance relevance and non-redundancy. DSR introduces a query-residual diversity kernel that penalizes redundant skill overlap while reducing penalties caused only by shared query relevance. On the SkillRouter benchmark, DSR improves recall and full coverage over a strong pointwise reranking baseline, with larger gains on multi-skill queries. These results suggest that skill routing should be treated not only as relevance ranking, but also as complementary set selection. --- ### Generalizing HVAC Control With Domain Randomized Reinforcement Learning Authors: Pablo Boitel, Kun Zhang Published: 2026-09-05 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.05822v1 Deploying advanced HVAC (Heating, Ventilation and Air Conditioning) controllers at scale remains difficult because performance often depends on accurate building models or per-site retuning. We propose NOMAD-RL (Neural Online Meta-Adaptation for Dynamics), a general-purpose Reinforcement Learning (RL) controller designed to transfer across heterogeneous thermal zones through a universal, non-invasive thermostat interface. The controller acts on temperature setpoints from zone measurements and forecasts, while a recurrent policy supports online adaptation under partial observability. Our main contribution is an adaptive domain randomization scheme based on physics-informed normalizing flows, which models correlated and multimodal distributions of thermal-zone parameters while maintaining physical plausibility and controllability. This produces a realistic and progressively adaptive training curriculum that improves transfer across buildings. We evaluate NOMAD-RL against a constant-setpoint PID controller, RL without domain randomization, and MPC in single- and multi-zone settings. NOMAD-RL consistently outperforms the PID and non-randomized RL baselines, and approaches the performance of a well-tuned MPC, especially in the more challenging multi-zone case. These results highlight the potential of adaptive, physics-informed domain randomization for robust and transferable HVAC control. --- ### CONDUIT: A Unified Residual-Stream Restoration Framework for KV Cache Reuse in Vision-Language Models Authors: Pengan Chen, Kaisheng Zheng, Liang Hong, Lixia Yi, Jiyue Jiang, Jiayang Chen, Yixuan Wang, Yimin Fan, Xinyuan Liu, Jiayi Li, Zhanqiu Zhang, Yiwen Guo et al. Published: 2026-09-05 Categories: cs.CL, cs.CV arXiv: https://arxiv.org/abs/2609.05821v1 Vision-language models (VLMs) often answer new questions about recurring visual content, where reusing the key-value (KV) cache can avoid re-encoding expensive visual prefixes. Exact-prefix reuse, however, fails when the same visual content appears under a changed prefix. Selective recomputation can recover quality under a small visual-token budget, but only when the right stale tokens are refreshed. Raw-attention selection can waste budget on high-attention tokens with small value-norm proxy scores and on query-irrelevant images. To address these failure modes, we propose CONDUIT, a training-free refresh policy that unifies single- and multi-image reuse as residual-stream restoration. Building on norm-weighted attention, CONDUIT ranks cached visual tokens using cached-key query attention and an accessible pre-output cached-value-norm proxy, then applies empirical image-level relevance amplification before one global selection. With one image, the coefficient is one and the rule reduces to intra-image token selection. The method preserves model architecture and weights, adding only a single query-conditioned scoring pass at inference. At a 10% refresh budget, CONDUIT achieves 97.0-99.5% of the corresponding full-prefill five-dataset average across three VLM backbones and leads budgeted methods on average; on the MMLongBench-Doc latency subset, it uses 13.5% of full-prefill FLOPs and achieves a 2.99x time-to-first-token speedup. --- ### Online Learning with LLM Experts from Limited Feedback Authors: Wang Wei, Soumyabrata Pal, Koyel Mukherjee, Franck Dernoncourt, Ryan A. Rossi, Branislav Kveton, Hoda Eldardiry Published: 2026-09-05 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.05820v1 We study adaptive routing of prompts to large language model (LLM) experts to maximize response quality in an online setting with limited feedback. We formulate it as a bandit problem with $K$ actions that represent experts and $d$ features that encode prompts, over a horizon of $T$ rounds. We propose algorithms that strategically select and observe rewards to minimize regret. In the full-information setting, we achieve a regret of $\tilde{O}(d T / \sqrt{m})$, while in the bandit setting we achieve $\tilde{O}(d T \sqrt{K / m})$, where $m \ll T$ is a budget on feedback. Our experiments show that we efficiently learn high-quality routing strategies across diverse LLMs from limited feedback. --- ### Agentic BAIM-LLM Evaluation (ABLE): Benchmarking LLM Use of Protein Design Tools Authors: Bryce Cai, Geetha Jeyapragasan, Samira Nedungadi, Jake Yukich, Seth Donoughe Published: 2026-09-05 Categories: cs.AI, cs.CY, q-bio.QM arXiv: https://arxiv.org/abs/2609.05818v1 We introduce ABLE, a benchmark for evaluating LLM agents' ability to use biological AI models (BAIMs), such as ProteinMPNN and AlphaFold3, in dual-use protein design workflows. ABLE assesses agent performance through a set of tasks spanning structure retrieval, sequence generation, and design validation. We evaluate 15 frontier models and find that seven refuse all tasks, while the remaining models exhibit substantial performance differences. Claude Sonnet 4 and Gemini 3 Pro achieve the highest scores across information retrieval, tool selection, and tool use. We further compare model performance on a subset of tasks against an expert human baseline. Our results suggest that current LLMs can substantially lower barriers to protein design, but remain inconsistent in planning, strategy generation, and integrating biological knowledge with tool use. --- ### CoRe-SAM3: Conditional Semantic--Visual Reconciliation for SAM3 Crack Segmentation Authors: Shipeng Liu, Liang Zhao, Dengfeng Chen Published: 2026-09-05 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05816v1 Crack segmentation requires a model to recognize target semantics while accurately recovering thin, low-contrast, and topologically continuous local structures. Although SAM3 provides strong open-concept segmentation, its direct application to the crack domain still misses weak cracks, activates crack-like background regions, and produces local boundary errors. We first diagnose the functional differences between the internal prompt-conditioned semantic representation and native visual representation of SAM3 on five crack datasets. The results show that the semantic representation already carries most task information for crack prediction, whereas the utility of the visual representation depends on the current semantic state. Directly combining the two representations does not yield consistent gains. Based on this finding, we propose Conditional Semantic--Visual Reconciliation, termed CoRe. CoRe retains semantic prediction as the primary decision path, applies lightweight semantic calibration to adjust the target-domain decision mapping, and uses spatially aligned native visual evidence to generate a zero-initialized, bounded, and regularized conditional residual that selectively corrects existing predictions. Across five domains, CoRe-SAM3 improves the average Crack IoU from 62.34% to 70.47% and clDice from 81.98% to 89.24%, while introducing only 18.914 K trainable parameters. Prediction-transition analysis further shows that CoRe corrects an average of 34.38% of native errors, with a damage rate of only 0.23% on pixels correctly classified by native SAM3. These results demonstrate that constrained prediction correction based on the functional differences between internal representations provides an effective and parameter-efficient target-domain adaptation strategy for vision foundation models with strong task-specific semantic priors. --- ### Physical policy gradient theorem for in situ stochastic-adjoint training Authors: William Tuxbury, Zin Lin Published: 2026-09-05 Categories: physics.optics, cond-mat.dis-nn, cs.LG arXiv: https://arxiv.org/abs/2609.05808v1 In situ adjoint training extracts parameter gradients directly from measurement, but has so far been limited to reciprocal or restricted systems. Here, we introduce the physical counterpart of the policy gradient theorem: a stochastic-adjoint gradient estimator that lifts these constraints by trading reciprocity for nondegenerate diffusion. As validation, we train a nonlinear resonator network, whose own dynamics supply the policy, against antagonistic temporal modulations with gradients from measured stochastic trajectories alone, without finite differences or a separate adjoint experiment. --- ### MolParser-Mobile: Ultrafast OCSR System for Large-Scale Chemical Literature Mining Authors: Xi Fang, Haocheng Lu, Han Lyu, Chengxiang Luo, Linfeng Zhang, Guolin Ke Published: 2026-09-05 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05807v1 Optical Chemical Structure Recognition (OCSR) is a fundamental component of chemical literature mining, enabling molecular database construction, reaction extraction, and AI-driven scientific discovery. Despite substantial progress in recognition accuracy with recent deep learning-based methods, inference throughput remains a critical bottleneck that limits web-scale deployment. To address this challenge, we propose MolParser-Mobile, an AutoML-optimized lightweight end-to-end OCSR framework. MolParser-Mobile contains only 9.98M parameters, while reaching a throughput of 1,520 molecules per second on a single NVIDIA RTX 4090D GPU. Despite its compact design, it maintains competitive and, on several benchmarks, superior recognition accuracy. --- ### Exposing Weaknesses in Emotion Recognition in Conversations Authors: Amir Ben Khalifa, Fanny Bezancon, Bessam Abdulrazak, Amine Trabelsi Published: 2026-09-05 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05806v2 Emotion Recognition in Conversations (ERC) aims to identify speakers' emotions in multi-turn dialogue. Accurate emotion recognition can support a wide range of applications, including empathetic conversational agents, mental health support, and educational technologies. While many recent approaches rely on task-specific fine-tuning, such models may exploit dataset-specific cues. A central yet rarely questioned assumption in ERC is that each utterance can be assigned a single unambiguous emotion label. To investigate this assumption, we study ERC using Large Language Models (LLMs) in a zero-shot setting while incorporating preceding conversational turns as context. We show that aggregate metrics mask systematic failures. Errors concentrate around utterances containing negations, exclamations, and interjections. This pattern is consistent across all evaluated models, suggesting limitations in the benchmarks rather than model-specific weaknesses. A controlled re-annotation study involving four human annotators supports this finding: strong agreement is observed in only 35 percent of cases, with neutral utterances dominating high-agreement instances, while many emotional categories fall into low-agreement regimes. These findings suggest that many apparent model errors reflect genuine annotation ambiguity rather than poor emotion understanding. Standard single-label evaluation is therefore insufficient. To address this limitation, we introduce an LLM-as-Judge framework that evaluates each emotion independently according to its plausibility in the conversational context rather than enforcing a single-label decision. --- ### AtomCite: Verification and Correction of Supplied Page-Level Citations in Multi-Page Documents Authors: Chen Qian, Yimeng Wang, Yu Chen, Lingfei Wu, Andreas Stathopoulos Published: 2026-09-05 Categories: cs.CL, cs.AI, cs.IR arXiv: https://arxiv.org/abs/2609.05802v1 Large language models answering questions over multi-page documents are expected to cite the supporting pages, yet supplied citations are sometimes inaccurate, and current evaluations score citations at generation time or against text passages: no existing benchmark evaluates whether a system can verify and correct a page-level citation already attached to an answer. We propose AtomCite, an agentic framework that parses an answer into claims, checks each claim against the image of its cited page, and applies a deterministic repair policy. To evaluate it, we introduce DocCite, to our knowledge the first benchmark for systems that verify and correct page-level citations in document images. Built on MP-DocVQA and DUDE, it combines 928 validated injected instances with 2,468 candidate natural errors harvested from frontier- and efficiency-tier models, of which a two-annotator audit confirms 1,909 as genuine errors. Primary labels are assigned deterministically, not by LLM judges, with the human audit as a separate validation layer. Across three model families (Gemini, Claude, and GPT), AtomCite reaches around 93% binary verification accuracy on the injected benchmark, significantly outperforming every OCR-only condition, including a compute-matched control, and exceeding every prior text-based baseline given the same OCR text. Its repair policy lifts citation precision on the injected mix from a constructed 34% to 87-90% while retaining over 90% of correct claims. AtomCite also transfers: with frozen prompts and zero training, it raises the hallucination-detection scores of two open 7-8B models on five public benchmarks above the same models prompted as direct judges. Finally, the audit shows that noise in automatic labels biases measured verifier accuracy and can reverse system rankings, so evaluations relying only on synthetic or automatic labels risk mismeasuring verification capability. --- ### Evidence-Aligned Local Composition of Discrete Experts for Sequence Restoration Authors: Mohammad Panahazari, Usman A. Khan, Shuchin Aeron Published: 2026-09-05 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.05801v1 A document modeled as a discrete sequence of tokens can be thought of as being generated from a composition of texts from different domains; a README file, for example, moves between prose, code, and configuration. When such a document is corrupted and only frozen domain experts are available, restoring it requires deciding both what is missing and which expert to trust at each position, at test time and without region labels or a trained router. We introduce evidence-aligned local composition, which infers a soft, position-wise weighting over the experts from the marginal evidence of the corrupted observation under a given corruption model, estimating the evidence from the experts' own denoising losses and smoothing the weights across positions. Because the weighting is soft, it recovers a mixture when the true composition is mixed and concentrates on one expert when that suffices. Across a categorical simulator, byte-level experts, and experts fine-tuned from a $1.3$B discrete flow-matching model, the inferred weights track the true regions at $0.85$ field accuracy on naturally mixed scientific documents, and at $0.98$ on constructed mixtures whose regions are lexically disjoint. Restoration improves over a single global weight when the experts are genuinely distinct and reduces to it when they converge, tracking a measure of expert separation. --- ### Spillover-Aware Multi-Value Steering for Pluralistic LLM Alignment Authors: Weici Pan, Xander Barron, Jiawei Zhou, Zhenhua Liu Published: 2026-09-05 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05800v1 Activation steering controls LLM behavior at inference time by adding learned directions to hidden states, but existing methods handle one concept at a time. Pluralistic alignment, where different stakeholders need different value emphases, requires steering multiple dimensions simultaneously. We show that naive steering produces substantial spillover: the effect intended for one value leaks into others. This parallels the treatment-versus-spillover decomposition in causal inference. We trace spillover to geometric entanglement of steering directions, captured by their Gram matrix, and derive a zero-cost correction from an activation-norm-penalized objective that decouples each direction's contribution exactly. Our end-to-end pipeline requires no fine-tuning, no reward model, and no manual prompt engineering: given only domain questions, it automatically discovers value dimensions, extracts directions, diagnoses entanglement, and applies corrected steering. On climate discourse, the correction improves the net steering effect from +5.9% to +14.0%, validated over 100,000 pairwise judgments. --- ### Dynamic Lagging using Stable-Prefix Training for Simultaneous Translation Authors: Hieu Hoang, Amittai Axelrod, Matt Post Published: 2026-09-05 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.05799v2 In streaming simultaneous speech translation, the speech translation system is trained to learn a read-write policy that alternates between consuming source words and generating target ones. In a cascaded setting, the output from the speech recognizer is passed to a separate machine translation component, making it more difficult to learn such a policy. Approximations such as fixed wait-k strategies or target-suffix deletion can be employed, but these approaches do not provide the model with a streaming system's flexibility to make contextual read-write decisions. This paper presents a training strategy for a cascaded machine translation system that enables it to dynamically decide how much of the growing source prefix to translate. We achieve this by fine-tuning a large language model (Qwen3-8B) on stable prefixes of the training data, which are produced by pairing every source sentence prefix in the training data with the longest translation of that prefix that is shared with the full source sentence translation. We fine-tune variants of the model on different subsets of the prefixes and compare against wait-k and target-suffix deletion. We also investigate the effect of fine-tuning the target-token generation confidence. Our experiments show that stable prefixes improve the quality-latency tradeoff when translating from English into German, Japanese, and Chinese across a range of test sets. --- ### Recall Is Not Protection: Evaluating Safety Monitors Against Model Compliance Authors: Sripad Karne Published: 2026-09-05 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.05797v2 Safety monitors screen prompts sent to deployed language models, flagging harmful requests so they are never answered. They are evaluated by recall against harmfulness labels, but a catch only prevents harm if the model would otherwise have complied. We measure the difference directly: we sample repeated responses from the target model, call a harmful prompt \emph{elicitable} if the model complies at least once, and report monitor recall separately on elicitable and non-elicitable prompts. Across six monitor configurations and three model families, spanning activation probes, fine-tuned text guards, and a 120B policy-conditioned reasoning classifier, recall on elicitable prompts falls 0.22 to 0.38 below recall on non-elicitable prompts at a fixed false positive rate. The prompts a monitor misses are 2.8 to 5.6 times more likely to be complied with than the prompts it catches. The gap replicates across three model families and appears also in text-only monitors entirely independent of the target model. This suggests that standard recall may overstate the protection monitors provide in practice, and that monitors should be evaluated against what their models will actually answer. --- ### Diagonal Attenuation: A Finite-Sample Correction for PCA Authors: Qiang Sun Published: 2026-09-05 Categories: stat.ML, cs.LG, math.ST, stat.ME arXiv: https://arxiv.org/abs/2609.05796v1 Principal component analysis (PCA) can rotate away from its population target when a covariance matrix is estimated from limited data. We introduce diagonal attenuation, which preserves sample cross-covariances while reducing coordinatewise sample variances. The method is revealed exactly by averaging a linear full-output reconstruction loss over random input masks; studying the correction directly extends it beyond the range attainable by masking. We isolate the part of the random coupling between retained and omitted population directions that is contributed by sample-variance errors, and show how attenuation can reduce the resulting rotation. Under balanced marginal variances, we derive an explicit expected-risk theorem, uniform over the attenuation path for all sufficiently large finite samples, and obtain the asymptotically risk-minimizing strength. For general covariances, we characterize when attenuation leaves the population PCA subspace unchanged and give a risk theorem that also accounts for changing eigengaps and the population cost when the target moves. Simulations track this tradeoff from exact preservation back to PCA. Across local image patches, speech spectra, and smartphone acceleration, both mask-derived and direct attenuation improve PCA under two fitting-sample budgets, and one of them has the largest mean gain among seven methods in every data--budget cell. The full path selects strengths beyond the mask-derived boundary on $63\%$--$95\%$ of the subsamples. --- ### Bait-and-Recover: Poisoning Internal Refusal Signals to Defend LLMs against White-Box Editing Jailbreaks Authors: Tian Gao, Zhipeng Xie, Yuhao Wu, Junhua Liu, Xin Fang Published: 2026-09-05 Categories: cs.CR, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.05794v1 Open-weight large language models face a low-cost white-box threat from representation engineering attacks. Attackers can estimate refusal directions and search for projection-matrix edits that suppress safety alignment while preserving general capabilities, within minutes on a single GPU and without gradient-based training. We propose Bait-and-Recover, a weight-level defense that places a bait adapter where attackers read activations and a paired recovery adapter at the subsequent layer. Trained via gradient routing, this decouples the observation path from the behavior path. By actively poisoning the residual signal used for measurement, Bait-and-Recover disrupts the attacker's edit search, while the recovery layer restores clean downstream computation. Across four open-weight models, our defense raises the minimum refusal rate against white-box edit searches from 16.25% to 71.75% under a strict behavior-preservation budget (KL <= 0.10), with negligible impact on general benchmarks. By invalidating the core measurement assumption of these attacks, observation-path poisoning offers a practical complement to behavior-level safety training. --- ### More Than Mimicking Reviewers: Evaluating LLMs for Pre-Submission Peer Review Authors: Pouya Parsa, Amin Rezaei Published: 2026-09-05 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05788v1 Peer-review feedback often arrives too late for authors to make meaningful revisions. We study an author-facing LLM system that moves part of this stress test before submission: it generates a broad pool of atomic concerns and compresses them into a short report. We evaluate agreement with historical reviews and, separately, the possible validity of concerns they omit. From 10,000 ICLR 2026 submissions, we use 3,398 manuscripts with accessible versions that predate review. On a ten-paper diagnostic, independent sampling covers 44.9% of historical issues; deduplication and refill reaches 78.7% strict and 84.9% seriousness-weighted coverage, at 3.6$\times$ more requests and 5.2$\times$ more tokens. A hidden Top-32 Oracle preserves the full 79.3% weighted coverage of a 256-candidate pool, but paper-only selectors retain only 40--44%. LLM review therefore provides broad coverage with a large candidate pool but compresses poorly; ablations identify representative selection and matcher sensitivity as the main sources of this gap. --- ### Closed-Loop Evaluation of Bird's-Eye-View Maps from Cross-View Transformers as Inputs to Behavior-Cloning Policies Authors: Felipe Carlos dos Santos, Eric Antonelo, Gustavo Claudio Karl Couto Published: 2026-09-05 Categories: cs.RO, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.05783v1 In autonomous driving, Bird's-Eye View (BEV) representations provide a structured, top-down abstraction of the vehicle's surroundings and have become a key input modality for Behavioral Cloning (BC) policies. While ground-truth BEV maps are readily available in simulation, real-world deployment requires replacing them with camera-predicted counterparts - a substitution that introduces perceptual errors whose downstream impact on closed-loop driving performance is not well understood. In this work, we investigate the use of Cross-View Transformer (CVT)-predicted BEV maps as direct policy inputs for a BC agent in the CARLA simulator. We propose a six-channel BEV representation covering road surface, planned route, lane boundaries, vehicles, pedestrians, and traffic lights, and introduce a Kernel Density Estimation (KDE) weighting scheme that rebalances the segmentation loss towards underrepresented driving maneuvers such as curves and intersections. Closed-loop evaluation across two CARLA towns shows that the KDE-weighted model is the only predicted-BEV agent to complete a full episode without infractions, despite not achieving the highest aggregate IoU. This discrepancy reveals that global segmentation metrics are poor proxies for driving performance: what determines navigation success is prediction quality at geometrically critical locations, and the route channel emerges as the primary bottleneck for reliable agent navigation under predicted BEV inputs. --- ### Distilling Vision-Language Models for On-Device Fire Understanding Authors: Mohammad Kazzazi, Zixuan Liu, Siavash Khajavi Published: 2026-09-05 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05782v1 Vision-language models (VLMs) offer a promising alternative to conventional fire detection systems by reasoning about the semantic context of a scene and thus reducing false alarms, yet their large model size makes deployment on embedded fire sensors impractical. In this paper, we study how domain-specialized VLMs can be compressed for fully on-device deployment without losing the safety-critical behavior required for fire detection. We develop a teacher-student knowledge distillation framework in which large VLMs fine-tuned for fire understanding can be distilled into lightweight students. Experiments across multiple VLM families and model scales show that compact students preserve most of their teachers' fire-understanding capability. We further deploy the distilled models on our commercial Detectium fire detection sensor and jointly evaluate reasoning accuracy, latency, and memory usage. The results show that compression and deployment affect not only accuracy but also model failure modes, with Qwen2.5-0.5B providing the strongest overall deployment trade-off. Our findings provide broader guidance for deploying domain-specialized VLMs in resource-constrained, safety-critical settings. --- ### Diffs vs. Whole Files: An Empirical Comparison of Iterative Edit-Based and Direct Generation for Flutter/Dart Code Models Authors: Andrej Andrejev Published: 2026-09-05 Categories: cs.SE, cs.AI arXiv: https://arxiv.org/abs/2609.05779v1 Large language models used for code editing can be trained and deployed in at least two output regimes: direct generation, where the model emits the entire modified file in one shot, and iterative diff-based generation ("steps"), where the model emits a sequence of localized search/replace edits applied one at a time until it signals completion or a step budget is exhausted. The diff-based regime is attractive because it mirrors how developers edit code and should require far fewer generated tokens per turn. We train two code models - a 100M-parameter model trained from scratch (Rainbow-Pony-100M) and a fine-tuned Qwen2.5-Coder-0.5B - in both regimes on a shared Flutter/Dart dataset, and evaluate all four resulting models on a held-out set of approx 1,790 tasks per model. Direct generation substantially outperforms diff-based generation on every metric we measure - compilation/static-analysis pass rate, bits-per-byte, character-level similarity to the reference, and blinded LLM-judge ratings of goal fulfillment, correctness, and code quality - and the gap persists after controlling for task difficulty via a matched-ID comparison and when restricting to code that compiles on both sides. We then identify a single, architecture-independent mechanism behind the conditions where diff-based generation does win: it is competitive on short, spatially localized edits, and its category-level wins concentrate in exactly the two task categories - refactoring and error-handling/edge-case fixes - with the lowest mean edit-step count in our dataset. We term this task locality and discuss its implications for when an edit-based training regime is and is not the right choice for a code-editing model. --- ### Nonlinear elliptic homogenization with the parametric Deep Ritz method Authors: Conor Rowan Published: 2026-09-05 Categories: cs.LG, math.NA arXiv: https://arxiv.org/abs/2609.05778v1 Elliptic homogenization is used to determine coarse-grained properties of materials with features on small scales. When these small scale features have rapid, periodic fluctuations, the solution field corresponding to a homogenized constitutive relation closely resembles the true solution based on the heterogeneous material. This homogenized behavior of the material is computed from a cell problem, where a cell is defined to be one period of the fluctuating material. In the context of linear elliptic partial differential equations, the homogenized constitutive relation is defined simply by a constant coefficient tensor, but for nonlinear problems, the homogenized response depends on the macroscopic state and/or its gradient, thus requiring solutions to parametric cell problems. When computing a numerical solution with the homogenized constitutive relation, it is useful to have a differentiable representation of the solution to the cell problem, as derivatives of the homogenized constitutive relation are required in Newton iterations for the macroscopic state field. In this work, we use the Deep Ritz method to solve the parametric cell problems that arise from nonlinear homogenization. First, we exploit the variational structure of the cell problem, then we discretize the dependence of the cell response on both space and the macroscopic state with a neural network. Enforcing boundary conditions on the cell response strongly, we next use the parametric Deep Ritz method to simultaneously solve the cell problem over a range of macroscopic states. We show that this method is accurate, efficient, and offers a continuous and differentiable representation of the cell response over the macroscopic state and gradient. We then show that our parametric representation of the cell response significantly expedites macroscale solutions when compared to a traditional $\text{FE}^2$ scheme. --- ### Backdoors Leave Structural Traces: FedMAST for Backdoor Detection and Containment in Federated Learning Authors: Srinivasan Subramanian, Md. Abdullah Al Hafiz Khan, Kazi Aminul Islam Published: 2026-09-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.27760v2 Federated learning enables distributed training without requiring clients to share their raw data. However, its reliance on the integrity of the client-submitted updates exposes the global model to stealthy backdoor poisoning. Existing defenses often rely on individual evidence sources, but stealth-constrained attacks can adapt to these signals. Such attacks can suppress anomaly signals they are optimized to evade, yet their poisoned updates still leave residual structural traces. We propose FedMAST, a Federated Multi-Axis Structural Tracing defense for backdoor detection in federated learning. FedMAST scores client updates using complementary structural, spectral, and historical evidence and then applies tiered filtering and round-level containment to limit adversarial influence. To capture traces that isolated signals may miss, FedMAST uses squeeze-pair coherence scoring to expose coupled feature distortions and signed spectral-drift tracking to reveal persistent directional changes over time. Across six backdoor attacks, FedMAST achieves lower attack success rate (ASR) than baseline defenses in all nine evaluated comparisons, averaging 1.51% ASR and 94.84% main-task accuracy (MTA) across the complete 200-round runs. Over the full 200-round method-aware CovertLayers run, FedMAST achieves 1.53% ASR and 92.26% MTA, compared with ASRs of 100.00%, 99.67%, 99.53%, and 32.84% for FedAvg, MultiKrum, AlignIns, and FLAME, respectively. --- ### H2LooP Telecom Model v1: From Telecom Comprehension to Autonomous Issue and PR Resolution Authors: Amit Singh, Vedant Nipane, Mayank Goel, Pulkit Agrawal, Sairanjan Mishra Published: 2026-09-04 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.22241v1 We present H2LooP Telecom Model v1, a domain-specialized large language models fine-tuned for the telecommunications industry. We release two domain-adapted model variants serving complementary use cases: a comprehension-focused variant for telecom domain question answering and reasoning, and an agentic variant for autonomous telecom code generation, pull request resolution, and code commits on production repositories. H2LooP Telecom achieves strong results on the GSMA Open Telecom Lite (OT-Lite) benchmark and a proprietary telecom code generation benchmark, outperforming frontier closed-source models such as GPT-5 and Claude Opus on independent leaderboard evaluation, while preserving general-purpose capabilities. The Comprehension variant achieves 81.8% weighted average on OT-Lite Pass@3, and, independently, ranks 5th overall on the official community-run Open Telco AI Leaderboard* at only 31B parameters-ahead of frontier closed-source systems including Claude Opus 4.6, GPT-5, Gemini 3 Flash, Grok-4-fast, and Kimi K2.5. Our agentic variant obtains a relative improvement of +8.8% in AST Similarity and +20.0% in Location IoU over the base model on telecom code generation, while maintaining identical MMLU (74.0%) and BFCL v3 multi-turn function calling (79.0%) performance, indicating zero catastrophic forgetting. Domain specialization on curated telecom corpora, spanning 3GPP standards, O-RAN specifications, network telemetry, and real repository commits, yields substantial improvements over general-purpose models of equivalent scale, approaches frontier closed-source models on domain-specific evaluation, and is independently corroborated by our official leaderboard standing. --- ### Statistical Inference for Adversarial Training: Central Limit Theorems via Optimal Transport Authors: Kim Jakwang, Kwon Dohyun Published: 2026-09-04 Categories: cs.LG, math.PR, math.ST, stat.ML arXiv: https://arxiv.org/abs/2609.22240v1 The purpose of this paper is to rigorously quantify the statistical and learning-theoretic properties of adversarial training models for classification. Equivalently, we establish the statistical properties of empirical optimal partial transport. Precisely, first we provide two types of central limit theorems (CLT): CLT centered at the expected empirical value, and CLT centered at the population one with smoothing. These results are based on the uniqueness of optimal potential for various equivalent optimal transport formulations, and the empirical process theory argument. For the binary setting, we indeed prove the uniqueness of optimal potential by leveraging the connection between optimal partial transport and the derived multi-marginal optimal transport formula. As byproducts, we also obtain the stability of a saddle point of the adversarial training model, and the sample complexity and concentration probability of the generalization error. --- ### Knowledge Graph-Augmented Ambient AI for Clinical Note Generation Authors: Jakir Hossain, Yi-Fei Zhao, Hongjian Wang, Minmei Shih, Katie Leigh Mullen, Ahmad P. Tafti, Leming Zhou, Manoj Purohit, William Hogan, Jay Zeng, Elizabeth Skidmore, Yanshan Wang Published: 2026-09-04 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.22239v1 Ambient AI is increasingly adopted in healthcare to automatically generate clinical notes from patient-clinician conversations, with the potential to substantially reduce clinician documentation burden. However, generated notes may omit clinically relevant information discussed during the encounter, creating information gaps that can affect downstream care. Knowledge graphs (KGs) constructed from encounter transcripts can provide a structured representation of what was discussed and enable systematic identification of missing information from generated notes that are critical for patient care. In this study, we introduce Coverage-Directed Revision (CDR), a model-agnostic framework that constructs a KG from the encounter transcript, identifies medical concepts absent from an initially generated note, and directs large language models (LLMs) to restore the missing information without modifying the underlying note-generation system. We evaluate CDR on two datasets: 1) Pitt-Bench, a local dataset comprising rehabilitation sessions, and 2) ACI-Bench, a public dataset for benchmarking clinical note generation. We tested four underlying LLMs widely used in ambient AI systems. The results show that CDR consistently improves content recall across all evaluated conditions. Our study provides a practical approach for improving the completeness of ambient AI-generated clinical documentation. --- ### Task-Aware QUBO Allocation for Mixed-Precision Quantization Authors: Osama Orabi, Artur Zagitov, Hadi Salloum, Viktor A. Lobachev, Yaroslav Kholodov Published: 2026-09-04 Categories: cs.LG, cs.CV arXiv: https://arxiv.org/abs/2609.22238v2 Mixed-precision quantization requires discrete allocation of weight and activation bit-widths, followed by recovery of the selected network. We develop a task-aware quadratic unconstrained binary optimization (QUBO) surrogate with separate weight and activation profiles, a bit-operation (BOP) cost, and selected structural priors. QUBO provides a network-wide allocation that can be refined through direct validation-based PROTES search. On a compact NAFBlock-based denoiser, the refined route achieves 37.192 dB after LSQ+ at 4.035\% routed-layer BOPs, versus 37.092 dB at 4.101\% for a HAWQ-style baseline. The repeated-search primary experiment shows that LSQ+ largely closes the quality gap between QUBO allocation and expensive direct refinement. An additional restoration architecture retains a larger recovered gain, indicating that refinement's value depends on architecture and recovery. We evaluate quality, achieved cost, routing stability and optimization expense together. Deployment measurements characterize a fake-quantized floating-point implementation; BOP reductions describe analytical allocation savings. --- ### Evaluation of MLLM-Agnostic Plug-and-Play Keyframe Selection Methods for Long Video Understanding Authors: Dilip Sarkar, Md. Safayet Islam, Liang Liang Published: 2026-09-04 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.13250v1 Multimodal large language models (MLLMs) cannot process every frame of a long video because of limitations in visual-token and computational budgets. Three primary approaches have been proposed to enhance their long-video understanding capabilities: (i) Retraining an MLLM on a large video corpus and/or extending its input length; (ii) Training an adapter for a specific MLLM that takes the entire video and the query as input and selects the most relevant video frames; and (iii) Developing a training-free, plug-and-play (PaP) adapter that is MLLM-agnostic. We refer to the third approach as PaP keyframe selection. A PaP method may use only candidate video frames without considering the query, or it may use both candidate video frames and the query. The first approach is prohibitively expensive. The second approach requires substantial training time and computational resources, but it is accessible to many because an adapter contains significantly fewer trainable parameters than an entire MLLM. The third approach has the lowest computational cost and is therefore broadly accessible. To the best of our knowledge, only five PaP methods have been reported within the past year. All of these methods have been evaluated on one or more video question-answering benchmarks and have demonstrated improvements in long-video understanding. However, the methods were evaluated on different benchmarks using different MLLMs. We present a comprehensive evaluation of these five methods using three MLLMs across three long-video understanding benchmarks. Our results show that QAaF achieves the best performance in 13 of the 15 aggregate evaluation settings, while FOCUS ranks second overall. These results provide a common experimental reference for comparing training-free keyframe-selection methods for MLLMs. --- ### Machine learning-assisted calibration of Agent-based Models: surrogate-based optimization with Genetic Algorithm and Particle Swarm Optimization Authors: Duguma Yeshitla Habtemariam, Jihwan Lee Published: 2026-09-04 Categories: cs.MA, cs.LG, cs.NE arXiv: https://arxiv.org/abs/2609.13247v1 Calibrating an agent-based model (ABM) is difficult because its objective landscape is stochastic and rugged, and can be evaluated only through costly black-box simulations. This study adapts inner-loop surrogate-assisted evolutionary computation (SAEC) to ABM calibration by embedding a machine-learning surrogate within genetic algorithm (GA) and particle swarm optimisation (PSO). At each iteration, the surrogate screens the candidates and the simulator validates only the top 50%, reducing simulation demand while correcting surrogate errors. We evaluate a full factorial of 48 configurations combining two optimisers, five surrogates, and four calibration objectives on two contrasting ABMs, the Brock-Hommes asset-pricing model and the Island growth model. Parameter recovery is measured against pseudo-true values. Relative to the strongest pure-optimiser baseline, the best ML-assisted configurations reduce RMSE by 20.0% on Brock-Hommes and 63.8% on Island, while reducing computation time by 32.1% and 61.1%, respectively. ANOVA with Dunnett's post-hoc tests confirms significant time savings for every surrogate under both Brock-Hommes optimisers and under GA on Island. No surrogate differs significantly from the pure-optimiser baseline in parameter-recovery accuracy on Island; the reported accuracy gains are therefore best-configuration outcomes rather than average effects resolved at this sample size. The best surrogate-optimiser-objective combination changes with ABM complexity, indicating that surrogate-assisted calibration depends on the model and does not admit a universal default recipe. --- ### Pixel-wise Planarity for High-Precision Monocular Plane Segmentation Authors: Ahmetcan Yavuz, Alpay Ozkan, Rémi Pautrat, Shaohui Liu, Marc Pollefeys Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.13246v1 Plane segmentation from a single RGB image remains challenging due to imprecise region grouping and geometrically inconsistent supervision, often leading to over-segmentation and false planar detections. We propose instead a pixel-wise planarity prediction framework for robust monocular plane segmentation. Building on a pretrained monocular geometric backbone predicting depth and surface normals, we introduce a dedicated planarity head that estimates per-pixel planarity confidence. During inference, predicted depth, normals, and planarity are combined in a lightweight region-growing procedure that enforces geometric consistency when forming plane segments. We further analyze existing plane ground-truth annotations and demonstrate substantial geometric inconsistencies under strict distance thresholds. Across multiple datasets, our method achieves improved geometric precision and segmentation quality compared to prior state-of-the-art approaches, while improving computational efficiency. Our code and models are available at https://github.com/alpayozkan/PixelwisePlanarity. --- ### SJD-SV: Speculative Jacobi Decoding with Semantics Verification for Autoregressive Image Generation Authors: Baoquan Zhang, Bingqi Shan, Shihao Fang, Kenghong Lin, Xutao Li, Yunming Ye Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.13245v1 Speculative Jacobi Decoding (SJD) is an important approach for accelerating autoregressive image generation. Although SJD has shown superior performance, recent studies point out that it usually suffers from a token ambiguity issue during token verification but its reason can not be well explained. To figure out this reason, in this paper, we conduct a visualization analysis on vision token and find that different from text tokens, vision tokens generally corresponds to some local, small, and unclear vision details, which means only using single token is difficult to accurately express a certain semantic, thereby causing token ambiguity issue. To this end, we propose a novel Speculative Jacobi Decoding with Semantics Verification (called SJD-SV), for accelerating autoregressive image generation. The key idea is that leveraging the strong correction characters between tokens to recognize semantic-aware token subsequence and then instead of perform token-by-token verification, turning to perform verification on semantic-aware token subsequence level for accelerating image generation. In particular, our method is plug-in, which can be directly integrated into existing SJD and its variants. Extensive experiments on various datasets show that existing SJD methods achieve significant performance improvement after integrating our SJD-SV method. --- ### Physical Kernel: Structured Visual Latents for Dark Manipulation Authors: Jinting Hang, Hong Li, Zhenhui Cai, Zhihao Zhao, Jian He Published: 2026-09-04 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2609.13244v1 We study dark manipulation: after a brief lit Write encodes z0 = Enc(rgb), a policy pi(z) and open-loop dynamics f(z,a) complete contact-rich skills without further pixels (dark_f). On ManiSkill StackCube (n=160; seed packs 0/1000), dark_f attains 68.1% stacked on the five-rung chain (near_A -> grasped -> lifted -> on_B -> stacked), compared with 35.6% for per-step lit_reenc and 0% for freeze/encode_black. On a shared Write->HOLD protocol (n=40), occlusion and camera-aligned GT contact-neighbor masks drive lit lift from 43% to 0%, while dark_f holds 82.5%; shuffling actions inflates dynamics MSE by ~9.4x; write-time appearance shifts break encoding (night: 0% stacked), yet the same shifts during HOLD leave dark_f lift unchanged; Write length Tw is flat once the stop phase is reached, while earlier stops and write-time blur/JPEG sharply cut stacked. A dedicated pi_write reaches 35% vision-budget stacked (n=80); matched Dreamer-style/pixel nulls without privileged geom stay at 0%. Privileged state-RSSM MPC reaches ~35% stacked with 9D dark observations -- a stronger-observation null, not a matched visual baseline. --- ### Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work Authors: Wenhui Chen, Shiwen Cheng, Hao Dong, Chenda Duan, Ruixiang Feng, Zhong Guan, Boqiang Guo, Xueyuan Han, Haojie Hao, Liangmeng Huang, Zhelong Huang, Xinke Kong et al. Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.11977v1 Co-work agents execute complex workflows that combine information gathering, tool use, coding, and file manipulation across many model invocations. Because cost and latency accumulate over the full episode, their practical value depends not only on peak capability but also on how efficiently that capability is delivered. Yet many steps in everyday work emphasize state tracking, coordination, recovery, and follow-through rather than frontier-scale reasoning. We present Occamy-1.0, a cost-efficient co-work model obtained by further training the post-trained Qwen3.6-35B-A3B checkpoint. We construct execution-grounded data and environments, capture replayable long-horizon trajectories across multiple harnesses, and use staged post-training to develop and consolidate complementary execution capabilities. Across a broad suite of co-work benchmarks, Occamy-1.0 is consistently among the strongest comparably sized models and remains competitive with substantially larger frontier systems on several tasks. Under our stated evaluation and pricing protocol, its aggregate performance across four representative benchmarks places it at the low-cost knee of the observed cost--performance Pareto frontier. Supporting evaluations in tool calling, coding, and instruction following further show that this specialization preserves broad agentic capability. We release the model weights and a subset of the training data to support research on practical co-work agents and agentic post-training. --- ### DI-Bench: Systematically Generating In-Domain Data Intelligence Benchmarks for Enterprise Agents Authors: Jiangyun Zhang, Kristen Surrao, Torpong Nitayanont, Yupei Zhang, Roopali Singh, Zhiyu Chen, Julia Huang, Zhou Tang, Shayan Ali Akbar, Omar Alonso, Erwin Cornejo, Yuan Li et al. Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05776v1 Evaluating enterprise agents on domain-specific benchmarks is critical, yet public benchmarks rarely evaluate whether agents can integrate business knowledge with analytical computation, and constructing such benchmarks manually is costly. We present DI-Bench, a pipeline for generating realistic benchmarks for data intelligence (DI), the practice of extracting insights from large volumes of enterprise data. To emulate realistic DI tasks that require both computation and knowledge retrieval, DI-Bench builds an artifact linkage graph over data tables, dimensions, metrics, and documents to form questions involving structured data and associated knowledge. Ground truth answers are derived via query execution, followed by LLM question generation and validation. Applied to two public datasets, the pipeline produces a 731-task benchmark covering knowledge retrieval, analytical computation, and rule-grounded reasoning. To show the discriminatory capability and difficulty of the benchmark, we evaluate four models, revealing a substantial finding: models achieve only 32% accuracy when doing computational tasks where retrieved business rules modify the computation. --- ### Inference-Time Graph Engineering for Multi-Agent LLM Workflows Authors: Katherine Tieu, Dongqi Fu, Yinglong Xia, Hong Li, Hong Yan, Jingrui He Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05774v1 Recent multi-agent LLM systems increasingly rely on graph-structured communication to coordinate specialized agents. We revisit multi-agent orchestration from a graph-engineering perspective: rather than optimizing a static topology, we synthesize a task-conditioned temporal workflow graph that jointly specifies agent connectivity and edge-level communication semantics. We introduce ReActNet, a training-free framework that compiles a query and a set of role-specialized agents into a sequence of directed communication graphs. Each graph snapshot corresponds to one reasoning stage, and each edge carries a natural-language instruction specifying the message that a source agent should provide to a target agent. The compiled temporal graph is then executed through structured message passing: agents update their reasoning states by integrating their previous states with messages from controller-assigned neighbors, and a final aggregator synthesizes the resulting states into the answer. This design separates graph compilation from graph execution, making multi-agent coordination explicit, inspectable, and task-conditioned without requiring reinforcement learning or gradient-based topology optimization. Across knowledge reasoning, mathematical problem solving, code generation, and GAIA-style assistant tasks, ReActNet consistently improves over fixed-topology and learned-topology baselines while maintaining competitive inference cost. These results suggest that effective multi-agent orchestration depends not only on which agents communicate, but also on engineering executable workflow graphs that encode when, why, and how information should flow during reasoning. --- ### RAPTOR: Role-Aware Private Training for Mixture-of-Experts Authors: Duc Dm, Khai Le-Duc, Nguyen Do, Minh Son Hoang, Florent Draye, Thai Hoang, Hoang Phuong Dam, Jiarui Liu, Chris Ngo, Terry Jingchen Zhang, Anh Le Duc Tran, Nhat Do Minh et al. Published: 2026-09-04 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.05770v1 Differentially private (DP) fine-tuning methods treat sparse Mixture-of-Experts (MoE) models as a single dense block, ignoring that shared layers see all data while experts only see routed records. We identify and formally characterize three resulting failure modes: global clipping suppresses expert gradients, batch-level normalization dilutes sparse expert updates, and fixed privacy noise degrades signal-to-noise ratio on low-load experts. We introduce RAPTOR - a Role-Aware Private Training framework, which alternates shared and expert optimization and targets each failure directly, using expert-specific clipping and noise together with a public expected-owner denominator and a count-independent update schedule that avoids conditioning on private, realized expert counts. We prove the resulting mechanism satisfies $(\varepsilon,δ)$-DP: because each record is assigned to exactly one owner expert, per-expert mechanisms within a layer compose in parallel, so updating all $E$ experts costs no more, in privacy terms, than updating one, with shared and expert streams composing sequentially across training. We further derive a bias-variance decomposition of the public-denominator estimator showing its bias grows predictably with routing imbalance, yielding a privacy-free rule for selecting which layer to protect from routing entropy measured on a small public corpus. Experiments on Switch Transformer and OLMoE fine-tuning across GLUE tasks, and on DeepSeek-VL2-Tiny, show consistent gains over standard DP baselines across several privacy levels ($\varepsilon$), with the largest margins typically at the tightest budgets. Code and models are publicly available: https://github.com/leduckhai/RAPTOR --- ### Data Scout: Targeted Web Crawling for Domain-Specific Pretraining Corpora Authors: Chirag Garg, Eelaaf Zahid, Farhan Ahmed, Jay Pankaj Gala, Eric Butler, Heiko Ludwig Published: 2026-09-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.05766v1 The dominant approach to building domain-specific pretraining corpora is to filter large web archives such as CommonCrawl. This works well for popular domains but breaks down for specialized ones, where relevant content is sparse and often beyond the reach of popularity-driven crawlers. We present Data Scout, which inverts this: instead of filtering an archive, it directs a targeted crawl. An LLM expands a root topic into a taxonomy and thousands of search queries; the returned URLs (seeds) are grouped by subdomain and screened with a user-supplied classifier (the probe), admitting each subdomain on the basis of a small sample. This works because relevance has a sharp boundary at the subdomain level: in mathematics, a page is 21x more likely to be relevant than one on a sibling subdomain. With the FineMath classifier as the probe, 21.9% of crawled pages are high-quality math content, 70x the 0.31% rate from filtering a comparable web sample, so the crawl wastes far less effort. But the payoff is not just efficiency: 63.2% of these pages are missing from CommonCrawl altogether, yet just as useful for training. Continued pretraining of Llama-3.2-3B on 1.9B Data Scout tokens matches FineMath corpus on GSM8k. Because the probe is the only domain-specific component, Data Scout can in principle apply to any domain with such a classifier. --- ### Interface-Aware KV Cache Quantization for Dense On-Chip NVM in Long-Context LLM Decoding Authors: Jiahao Zheng, Yifan Qin, Xiaobo Sharon Hu, Yiyu Shi Published: 2026-09-04 Categories: cs.AR, cs.AI arXiv: https://arxiv.org/abs/2609.05764v1 The key-value (KV) cache is the dominant memory bottleneck in long-context large language model (LLM) decoding: every step reads it entirely, so decoding is memory-bandwidth bound. Holding a quantized KV cache in dense on-chip non-volatile memory (NVM) removes the off-chip transfer. Existing KV quantization methods, however, were designed for GPU-style memory systems: KIVI attaches per-group metadata, adding about 25% to the stored KV cache; KVQuant keeps sparse full-precision outliers that a dense array cannot hold in place. This paper examines what these structures cost when the KV cache resides in NVM behind fixed-range converters, and designs a quantization scheme matched to that interface. The architecture stores the quantized KV cache in dense on-chip NVM, uses a small static analog crossbar only for the fixed rotation, and keeps attention in on-chip digital logic. A randomized rotation and per-vector normalization give every coordinate the same range, so one fixed codebook for keys and one for values, each shared across all tokens of the corresponding tensor type, serve the entire KV cache. The codebook thresholds are programmed once as the read converter's reference levels, enabling fixed-range digitization with no per-token converter reconfiguration. Dequantization is a sixteen-entry lookup and one norm multiply; the only per-vector metadata is one scalar, about 3%. Across models from 3B to 14B and contexts to 32k tokens, the four-bit KV cache maintains accuracy under storage and crossbar noise simulated at realistic device levels. KIVI and KVQuant remain more accurate in software; the advantage of our format lies at the memory interface: 3.1-3.6x lower KV read energy than both mapped to the same NVM, and 8x lower metadata overhead than KIVI. The contribution is a KV quantization co-designed with the NVM memory interface rather than a new accuracy record. --- ### AAMBERS-UAV: Acquisition-Aware Multimodal Backbone Evaluation and Ranking for UAV Weedy Rice Segmentation Authors: Tarek Rahman, Nazim-E-Alam, Md Kishor Morol, Jannatun Noor Published: 2026-09-04 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.05762v1 UAV image collections contain spatially and temporally related frames, yet semantic-segmentation benchmarks commonly split them at image level. Such splitting can place samples from one acquisition in both model development and testing, obscuring transfer to a genuinely new survey. Using the 734-sample WeedyRice-RGBMS-DB, we fix a 124-image target-acquisition test set and compare two protocols with identical train, validation, and test counts: target-held-out, which excludes the target acquisition from development, and target-exposed, which admits its remaining images. SegFormer-B0 is evaluated with RGB, four-band multispectral (MS), and seven-channel RGB+MS input over two fixed-split seeds. RGB is strongest under complete acquisition holdout ($0.7317\pm0.0201$ IoU), whereas RGB+MS becomes strongest after target exposure ($0.7822\pm0.0269$). A fixed-split U-Net/ResNet18 replication confirms positive exposure gains for all three inputs, but retains RGB as the best modality under both protocols. Acquisition exposure therefore increases measured performance across both evaluated backbones, while its effect on modality ranking is architecture-dependent. A supplied-split audit reveals strong near-sequential dependence, and corruption tests show that early fusion is substantially more sensitive to RGB--MS displacement than to moderate radiometric scaling. These results support acquisition-aware same-test evaluation as a necessary complement to ordinary image-level splitting in multimodal UAV benchmarks. The code and supporting the findings of this study will be publicly released upon acceptance of the paper. --- ### GeoContext: One Context Ladder, Two Failure Modes in Vision-Language Geolocation: Flat Reliance on User-Provided Location Context and False Confirmation of Location Claims Authors: Yifan Zhang, Kai Wang Published: 2026-09-04 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.05761v1 Visual geolocation benchmarks typically ask a model where an image was captured without accounting for the location context that users often provide. We introduce GeoContext, a resource supporting two complementary tasks: GeoHint, open-ended localization given a true but coarse location hint, and GeoVerify, binary verification of whether an image was taken within 150 m of a claimed place. GeoContext constructs a context ladder by stratifying nearby reference points according to distance and referenceability, allowing the image to remain fixed while the supplied context varies. The benchmark covers 109 sites in 30 cities and evaluates five vision-language models using 21,933 GeoHint responses and 6,270 GeoVerify responses. Our evaluation reveals three main patterns. First, hint repetition varies by only 1.5 percentage points across referenceability tiers and by less than 3 points across distance bands, while the resulting localization error increases steadily with hint distance. Second, behavior depends strongly on no-context performance: at sites with low no-context accuracy, the median ratio between localization error and hint distance is approximately 1.00, whereas at higher-accuracy sites it ranges from 0.24 to 0.69. After correcting for bias introduced by the site grouping procedure, only one of the five models retains a negative accuracy estimate when given a nearby hint. Third, in GeoVerify, no model reaches d' = 1 for decoys immediately beyond the 150 m tolerance. Model rankings also change when sensitivity is separated from response bias, and 83.8% of false acceptances are reported with confidence of at least 0.8. We release the benchmark, construction pipeline, audit decisions, and scoring code. --- ### RAGMark: A Comprehensive Framework for Benchmarking Retrieval-Augmented Generation Systems Authors: Zlatan Feric, Amir Taherin, Bin Ren, Yanzhi Wang, Jennifer Dy, David Kaeli Published: 2026-09-04 Categories: cs.PF, cs.AI, cs.AR, cs.IR arXiv: https://arxiv.org/abs/2609.05760v1 We present RAGMark, a modular benchmarking framework for advanced Retrieval-Augmented Generation (RAG) systems targeting small-scale multi-GPU environments. RAGMark evaluates diverse RAG components, including retrievers, vector databases, prompt-processing methods, and generator models, while collecting detailed per-stage metrics such as latency, GPU utilization, memory consumption, power usage, time to first token (TTFT), throughput, and answer quality. The framework is highly extensible, separating RAG stages, timing, and resource monitoring into modular components, and is designed to efficiently sweep large configuration spaces while minimizing repeated model and database initialization overhead. Using RAGMark, we characterize five RAG workloads on open-domain QA datasets across varying retrieval depths, model scales, reranking, compression methods, and vector database configurations. We show that while autoregressive generation dominates latency in naive pipelines, context-reduction techniques shift bottlenecks across compute, memory bandwidth, and preprocessing stages. Reranking and compression produce compounding benefits: reranking reduces compression workload itself, while both jointly reduce prefill and KV-cache traversal costs, lowering energy consumption by up to 66%. We further observe strong cross-stage interactions, where small upstream context reductions cascade through downstream latency, memory traffic, and energy consumption. The RAGMark source code is publicly available at: https://github.com/zferic/RAGMark. --- ### From Monolithic Blending to Agentic Orchestration: Dynamic Response for Conversational Assistants at Scale Authors: Cen Mia Zhao, Peng Wang, Chuan Shi, Yufeng Zhang, Ying Lyu, Wanmeng Ren, Robert Xue, Claire Na Cheng, Yashar Mehdad Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.05758v2 Conversational assistants can blend retrieval, action selection, escalation, and wording in a single model path, or separate those roles. We report a production migration of a customer-support assistant at a large accommodation marketplace (millions of conversations per month, 11 languages, 10-second P90). Dynamic Response (DR) replaces a single Qwen3-235B-A22B blended responder with a bounded ReAct orchestrator over typed tools plus a smaller generator that writes from a backend-validated context contract. Because the migration also changed prompts, alignment, and serving, we attribute each effect to its cause and claim as architecture effects only those measured on identical replayed turns: typed entity selection moves the reservation selector to a precision-first operating point (precision 8.3% to 89.1%, recall 75.2% to 67.3%), and typed action IDs with a membership check remove observed structured-action hallucination (2.14% to 0.0%). A low-ramp A/B test reproduces the replay escalation reductions: hard-escalation responses fall from 5.60% to 3.08% and soft-escalation responses from 9.56% to 2.49%, while production handoff volume holds roughly steady; self-solve is directional (+5.1 points, 95% CI (-2, +12)). Serving optimizations cut orchestrator P90 latency from 3.87s to 2.24s on a GPU footprint reduced by roughly one-third, and self-hosting reduces estimated annual model-serving cost by more than an order of magnitude. --- ### CrisisKD: Five-Stage Knowledge Distillation for Aspect-Level Sentiment and Emotion Analysis in Crisis Discourse Authors: Marko Haralović, Onat Akca, Salih Eren Yücetürk, Minsi Li, Mariët Theune Published: 2026-09-04 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.05757v1 Identifying the target of emotional words or phrases in crisis situations, especially health-related ones, is important for understanding public concerns across cultural and linguistic contexts. We propose CrisisKD, a five-stage teacher--student knowledge distillation framework for aspect-level sentiment and emotion analysis on unannotated social media data. A teacher LLM generates aspect-level labels and reasoning traces that supervise a smaller student model across aspect extraction, syntactic parsing, opinion extraction, sentiment classification, and emotion classification. Using this framework, we construct and release a dataset containing 50,615 aspect-level labels, together with the annotation and fine-tuning scripts as open-source resources. The resulting student supports end-to-end ABSA and emotion detection at substantially lower inference cost than the teacher. On a manually annotated 500-tweet gold set, the 5-task Qwen2.5-7B student improves over the untuned model by 7.9 F1 points on aspect extraction, 17.0 points on emotion accuracy, and 6.5 points on sentiment accuracy. On the external ABEA benchmark, CrisisKD improves the same-model Qwen2.5-7B ICL baseline by 2.8 F1 points on ATE and 3.8 F1 points on joint ATE+AEC. --- ### Concord: A Video Relational Algebra for Cross-Modal Query Optimization Authors: Sultan Muratbek, Charisse Ivana Yeung, Chanwut Kittivorawong, Alvin Cheung Published: 2026-09-04 Categories: cs.DB, cs.AI arXiv: https://arxiv.org/abs/2609.05756v1 Semantic video queries let users embed natural language prompts and use multimodal large language models (MLLMs) to interpret the video. Such queries are increasingly popular for querying video data. However, their expressiveness comes at a steep cost: an MLLM may process hours of media to return only seconds of relevant output, making naive execution slow, expensive, and inaccurate. We propose Concord, a system for expressing and optimizing semantic video queries. We makes three contributions. First, we introduce Video Relational Algebra (VRA), a nested algebra over videos, transcripts, frames, and object tracks that captures common semantic video operations. Second, we derive a set of approximate optimizations that rewrite VRA queries to reduce MLLM usage while improving result quality. For narrated video, Concord either processes transcripts instead of video or uses them to identify video clips for MLLM processing. For cross-camera queries without narration, detection and tracking replace a whole-video MLLM join with a track-level relational join. Third, we evaluate Concord on real-world videos. Across 4.59 hours of soccer broadcasts and 3.92 hours of lectures, transcript-to-video queries send only 5.32% and 2.47% of source-video duration to the MLLM and reduce MLLM cost by up to 87%. In two five-second highway clips with 18 manually adjudicated cross-camera vehicles, a Detect-Track-Join query improves F1 from .364 to .813 while making no MLLM calls. See our project at https://concord-db.github.io/. ---