# Research Papers — 2026-08 · bundle page 1 of 240 > Full-text markdown of papers #1–50 of 11995 published in 2026-08, newest first. AI-readable mirrors of the full 11995-paper 2026-08 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2026-08 [Next bundle →](https://aipod.dev/papers/archive/2026-08/page/2.md) ### PPTBench: Can Coding Agents Reconstruct the Visual World through Structured, Editable Slides Authors: Xiaoqiu Wang, Yizhe Chi, Wenyi Li, Deyao Hong, Zhihan Shan, Mingju Gao, Kaisen Yang, Youjie Zheng, Calvin Xiao, Qinhuai Na Published: 2026-08-31 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.29718v1 Coding agents are beginning to act in the visual world. They now build webpages, GUIs, games, 3D scenes, diagrams, and documents. Success in such visual coding requires bridging two spaces: inferring visual structure and expressing it programmatically. Slides are a core medium of knowledge work, widely used to communicate ideas and collaborate in a form that people can directly inspect and edit. Therefore, they provide an ideal testbed for visual coding, as they require agents to recover visual structure and realize it as editable objects. However, existing benchmarks either rely on subjective open-ended evaluation, produce non-editable code outputs, or focus only on local editing rather than end-to-end visual reconstruction. We introduce PPTBench, which benchmarks visual coding through editable slide reconstruction. It contains 500 tasks, each based on a scientific flow diagram from a real arXiv paper and requiring agents to reconstruct it as a single PPTX page composed of native, editable objects. A four-stage Agentic Judge evaluates artifact validity, semantic correctness, rendering quality, and fine-grained visual quality. Across 31 configurations spanning model families, effort levels, and harnesses, the best configuration, Kimi K3, reaches only 67.80, while the median scores 19.47. We find that agents can reliably produce valid PPTX files but still struggle with semantic and visual correctness, especially text details. More reasoning mainly helps agents pass hard gates, while stronger verification is more consistently associated with higher quality. PPTBench advances the vision of coding agents that can understand and reconstruct the visual world through structured, editable code. --- ### TopoFuse: Topology-Aware Tri-Planar Fusion for 3D Cryo-Electron Tomography Segmentation Authors: Rohit Kumar Salla, Neelesh Gupta, Xingjian Li, Min Xu Published: 2026-08-31 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.29717v1 Automated segmentation of cryo-electron tomograms routinely produces masks that are voxel-accurate but topologically broken: membranes fragment, organelles merge into one another, and enclosed cavities collapse. Existing topology-aware losses reduce these violations but cannot eliminate them, because topology is encouraged through gradient pressure rather than structurally enforced. We introduce TopoFuse, which reframes topology as a differentiable projection operator rather than a loss penalty. At each forward pass, the projection operator $\mathrm{Proj}_T$ (a PH-guided sparse edit) identifies the critical voxels responsible for topological violations via bottleneck matching and applies sparse edits to satisfy a specified topology target (diagram feature counts and lifetime budgets) for dimensions $d \in \{0,2\}$. If the projection converges, the output satisfies those constraints on the downsampled grid ($s=2$); when it does not, a repair certificate exposes this explicitly, enabling downstream filtering. A topology prior head predicts the correction target directly from input features, removing any dependence on ground-truth topology at inference. Across three cryo-ET benchmarks, TopoFuse reduces Betti number error by 54% over the strongest soft-loss baseline ($p < 0.001$), improves Dice by 4.6 points, and edits only 3.1% of voxels to achieve this. --- ### Revalidation Beats Stateful Routing for Scientific Surrogates Under Distribution Shift Authors: Harshil Lodhiya Published: 2026-08-31 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.29715v1 Surrogate models are often chosen during development and then left in place as new measurements arrive. That practice becomes risky when noise, input support, or physical parameters change. We asked whether such changes call for a stateful adaptive controller, or whether it is enough to validate the candidate models again on each new batch. To study this question, we built RegimeShift-Surrogates, a reproducible streaming benchmark spanning eight analytic and dynamical tasks, four stationary or shifting regimes, ten held-out seeds, and eight classical, multilayer-perceptron, and Kolmogorov-Arnold network surrogates. The confirmatory run contains 30,720 model fits and 3,200 scored deployment windows. Choosing the model with the lowest validation loss in the current window yields mean log regret 0.091 against a per-window oracle; the best fixed model chosen in hindsight yields 0.192. The paired difference is -0.101 (hierarchical bootstrap 95% CI (-0.165, -0.040); Holm-adjusted p = 0.0469), with revalidation ahead in 26 of 32 task-scenario combinations. None of the stateful alternatives, including exponential smoothing, dual-timescale adaptation, Page-Hinkley resets, or margin gating, improves the pooled result, and delayed bias correction makes it worse. Oracle choices also differ substantially by task: k-nearest neighbors dominate the damped oscillator, vanilla KAN is often selected for two-dimensional surfaces, and MLPs lead on the Runge and Van der Pol tasks. In this benchmark, fresh validation evidence is useful; carrying old evidence forward is often not. --- ### RAPTOR: RAndom-projection Physics-informed Transient sOlveR Authors: Petros Ellinas, Benjamin Vilmann, Spyros Chatzivasileiadis, Johanna Vorwerk Published: 2026-08-31 Categories: eess.SY, cs.LG, cs.MS, math.NA arXiv: https://arxiv.org/abs/2609.29714v1 The complexity of time-domain simulation of modern power systems has increased significantly because converter-based resources introduce control dynamics that must be simulated alongside slower system-level and fast electromagnetic dynamics. The resulting wide range of timescales may force classical time-domain solvers to use small timesteps, complicate the solution of nonlinear equations at each timestep, and reduce solver reliability under strongly nonlinear and multi-timescale transient conditions. This paper introduces RAPTOR, a first-of-its-kind time-domain simulation framework for power system dynamic simulations. At its core, RAPTOR introduces a new integration technique that represents the unknown trajectory of hybrid differential-algebraic equations (DAEs) over a time interval using a physics-informed random-projection neural network (PIRPNN) built from fixed Gaussian radial basis functions (RBFs). A nonlinear solver, e.g., Newton-Raphson, determines how to combine the fixed RBFs so that the resulting trajectory satisfies the governing equations. By combining RBFs with broad and localized shapes, RAPTOR can represent complex multi-timescale behavior over comparatively long time intervals. This enables larger effective simulation advances, fewer overall nonlinear solver iterations, and faster integration over long time horizons. Numerical studies on stiff RMS models, IEEE 9-, 14-, 39-, 57-, and 118-bus RMS simulation benchmarks, and EMT test cases show that RAPTOR accurately finds solutions with substantially fewer sequential simulation advances while offering superior accuracy-runtime performance. With speedups of over 10x in certain cases, these results showcase RAPTOR's potential to outperform long-standing and widely used integration methods, such as the Radau method and the trapezoidal rule. --- ### Decoupling Knowledge and Privacy: Post-Task Self-Distillation Replay for LLM Continual Learning Authors: Shengtao Wen, Yunying Yang, Xiang Chen, Lingbing Guo, Yu Tian, Sheng-Jun Huang Published: 2026-08-31 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.29711v1 Privacy-preserving continual learning (PPCL) must reduce the reproduction of sensitive content while retaining useful knowledge across sequential tasks. Formal privacy guarantees characterize randomized mechanisms, whereas operational output control concerns whether a trained model selectively reduces the likelihood of sensitive content in its outputs. In this work, we investigate the latter together with continual-learning utility under realistic task evolution. Retention and privacy correction operate at different granularities: task acquisition requires broad preservation of current- and old-task behavior, whereas privacy correction targets sparse annotated positions. Joint optimization leaves the current-task preservation target continually changing. We propose SPARK, a retention-correction decomposition that first freezes the learned post-task distribution and then applies selective correction around this stable reference. Self-Distillation Replay learns the current task while distilling behavior from previous tasks, and Post-Task Privacy Correction reduces annotated-PII likelihood while anchoring current- and old-task non-PII behavior to the resulting checkpoint. Extensive evaluations demonstrate that SPARK achieves effective selective PII suppression while preserving strong continual-learning utility and knowledge retention across diverse settings. Code and data will be released upon publication. --- ### Three Ways Classical Test Theory Misleads for LLM Judges Authors: Louis Yiven Zhu Published: 2026-08-31 Categories: cs.LG, cs.CL, stat.ME arXiv: https://arxiv.org/abs/2609.29709v1 An LLM judge scores a bank of responses against a rubric, and the reliability comes back at $0.52$. What has been measured? Judge evaluation has begun borrowing reliability statistics from classical test theory, usually without stating the measurement design each statistic assumes, and we show that three widely portable ones mean something different for a judge than for a test because the judge setting rearranges the roles those designs rest on. First, an internal-consistency coefficient computed over rubric elements contains no scorer facet. Holding one judge's measured error rate fixed at $4.72\%$, KR-20 still ranges from $0.01$ to $0.68$ as the item bank is redesigned around it, and varying judge error moves the coefficient by a comparable amount, so item design and judge error are not separately identified and no single value can be read as a property of the judge. Second, the dependability index $Φ(λ)$ is a ratio of variance components, and the classification probability with which it is sometimes identified differs from it by $0.25$-$0.43$ on our bank and by $0.17$-$0.30$ on simulated data where the underlying model holds exactly. Third, Livingston-Lewis accuracy is indexed to an examinee's own true score on the same instrument, so scoring it against external gold conflates judge unreliability with criterion invalidity. Reviewing the three closest judge-evaluation papers, we found no published instance of these errors, which makes the caution prospective. A coefficient that cannot be attributed to the judge nonetheless travels downstream into deployment decisions and disclosure documents. We therefore close with four reporting lines that keep the attribution attached to the number. --- ### Safety-oriented pedestrian trajectory prediction at urban intersections using time-to-collision and crossing-zone context Authors: Erel Avineri, Yftach Gil, Yehudit Aperstein Published: 2026-08-31 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.29706v1 Accurate pedestrian trajectory prediction is important for proactive road-safety applications, particularly at urban intersections where pedestrian motion is shaped by both vehicle interactions and crossing context. This study presents a safety-oriented trajectory-prediction framework that combines pedestrian motion history with Time-to-Collision (TTC) information and crossing-zone indicators. Using naturalistic trajectories from one urban intersection in the inD (Intersection Drone) dataset, several neural architectures were evaluated with 1.6 s observation and 2.4 s prediction horizons. A pooled Long Short-Term Memory (LSTM) separately encodes TTC histories and crossing-zone context before integrating them with pedestrian positions. In addition to conventional Average Displacement Error (ADE) and Final Displacement Error (FDE), prediction performance was assessed using the frequency and magnitude of errors exceeding a study-defined 1 m tolerance. A weighted loss was also introduced to place greater training emphasis on large coordinate-wise errors. Applying this loss to the position-only LSTM reduced ADE from 0.210 to 0.190 m and FDE from 0.550 to 0.503 m, while reducing ADE and FDE exceedance counts by 34.8% and 19.8%, respectively. The final pooled configuration incorporating TTC and crossing-zone information achieved an ADE of 0.184 m and FDE of 0.491 m, with further reductions of 33.5% and 6.3% in ADE and FDE exceedance counts relative to the safety-oriented position-only LSTM. The results indicate that safety-oriented training and structured integration of interaction and contextual information can reduce large trajectory-prediction errors, although broader validation across pedestrians, sites, and datasets is required. --- ### Understanding and Exploiting Initialization Anchoring Weakness in Feedback-Based Agent Planning Authors: Chuanchao Zang, Jianing Wang, Wenyu Chen, Xiangtao Meng, Li Wang, Xinyu Gao, Peng Zhan, Zheng Li, Shanqing Guo Published: 2026-08-31 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.29697v1 Feedback-based planning improves agent reliability by incorporating tool observations and corrective feedback. However, its protection may not be distributed uniformly across planning stages. We conduct a round-wise analysis of four representative feedback mechanisms and uncover an initialization anchoring weakness: the first feedback round corrects 46\% of adversarial directions, whereas the rates fall to 13\% and 7\% among directions surviving into the next two rounds. Our analysis attributes this weakness to three interacting factors: a contextually plausible shift in the initial plan, insufficient counterevidence, and the persistence of accepted directions in the accumulated trajectory. Based on these findings, we propose \textsc{InitAnchor}, a black-box framework for exploiting this weakness through attacker-controlled external materials. It operationalizes the three factors as directional-shift, contextual-plausibility, and counterevidence-resilience signals under either limited target access or no target access. Across 112 tasks from 16 domains, six agent architectures, and five backbone LLMs, \textsc{InitAnchor} achieves average ASRs of 76.1\% and 72.0\% under the two settings while reducing first-round mitigation rates to 21.0\% and 25.0\%, respectively. It also remains effective against six defenses and across six real-world agent systems. These findings show that feedback-based agents can retain early biases even when later correction is available. --- ### An Agnostic Sample Compression Scheme for Squared Loss of Near-Linear Size in the Fat-Shattering Dimension Authors: Guangjian Zhang Published: 2026-08-31 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.29696v1 We construct, for every function class $\mathcal{F}\subseteq(0,1)^{\mathcal{X}}$ and every accuracy $0<α\le 1$, an agnostic sample compression scheme for the empirical squared loss: for every finite sample $S\in(\mathcal{X}\times(0,1))^m$ with arbitrary (noisy) labels, the scheme stores at most $O(\mathrm{fat}(\mathcal{F},c'α)\cdot\log^3(2/α))$ original labeled examples and auxiliary bits, independent of the sample size $m$, and reconstructs a function $\hat f$ with $L_2(\hat f,S)\le\inf_{f\in\mathcal{F}}L_2(f,S)+α$. This resolves, in the positive, the open problem of Attias, Hanneke, Kontorovich, and Sadigurschi (ICML 2024, Section 5), which asks for an agnostic $\ell_2$ compression scheme of size $\mathrm{fat}(\mathcal{F},cα)\cdot\mathrm{polylog}(c/α)$. All previously known bounded-size constructions, agnostic and even realizable, incur a multiplicative dual fat-shattering factor, which can be exponentially larger than the primal dimension; our scheme removes the dual factor entirely, including in the realizable case. The dual factor in prior work enters solely through a sparsification step that forces uniform approximation on the sample. By targeting only a $(1-ε)$-fraction of sample points, which suffices for an average-loss guarantee over a bounded range, K'egl's boosting margin bound yields $O(\log(1/ε))$ rounds independent of $m$, and sparsification is never needed. The booster's synthetic target labels (values of a near-optimal $f^*\in\mathcal{F}$) are transmitted through quantized side-information bits attached to stored original examples, and the cross term of the squared loss forces the weak-learning scale $Θ(α)$, matching the same-scale form of the open problem. --- ### Bandit Multiclass PAC Learning: Corrected Lower Bounds, Exact Families, and a Confidence Direct-Sum Phenomenon Authors: Guangjian Zhang Published: 2026-08-31 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.29694v1 We study realizable multiclass PAC learning with bandit feedback: the learner observes an i.i.d. instance, predicts one of $K$ labels, and learns only whether the prediction was correct. Hanneke, Meng, Moran, and Shaeiri (arXiv:2605.25678) characterized the optimal sample complexity via the bandit DS dimension $\mathrm{BDS}$ up to logarithmic factors, and asked whether every class admits sample complexity $O((\mathrm{BDS}+\log(1/δ))/ε)$. First, we show that the published lower bound $Ω((\mathrm{BDS}+\log(1/δ))/ε)$ is incorrect as stated: we exhibit explicit classes with $\mathrm{BDS}=K-1$ whose sample complexity is exponentially smaller, and locate two independent gaps in its proof. We repair the lower-bound theory around a new anchored dimension $\mathrm{aBDS}\le\mathrm{BDS}$, proving a constant-free three-part lower bound. On the upper-bound side we remove the ambient label count $K$ entirely, proving $O((B\log^3 B+B\log(1/δ))/ε)$ for $B=\mathrm{BDS}$, plus a constant-confidence bound via a new fiberization lemma; for two natural families we determine the sample complexity up to constant factors. Finally, we answer the open question in the negative under its uniform-constant reading, and show the failure is intrinsic: for an explicit affine multiplexer class we establish the full confidence profile $Θ((n\min{n,\log(1/δ)}+\log(1/δ))/ε)$, a confidence direct-sum regime where a multiplicative $\log(1/δ)$ cost is information-theoretically necessary, followed by a rank-saturation phase transition. Two classes with identical dimension profiles can have polynomially different sample complexities, so no characterization by these dimensions alone is accurate to polylogarithmic factors. We also show these results are consistent with additive-confidence list-PAC guarantees via the ListCascade bridge. --- ### Fair Like Us? Auditing LLM Alignment in Resource Allocation Authors: Qishen Han, Hadi Hosseini, Joshua Kavner, Samarth Khanna, Sujoy Sikdar, Lirong Xia Published: 2026-08-31 Categories: cs.AI, cs.CY, cs.GT arXiv: https://arxiv.org/abs/2609.29692v1 Fair allocation of scarce, indivisible resources is an important challenge in many societal problems. While there are several formal theories of fairness, no single definition can always be satisfied. As large language models (LLMs) are increasingly used to support decisions and act as agents, they raise new concerns about distributional justice: their judgments are not directly tied to any specific fairness framework and may violate key normative principles. In this work, we introduce a general method for evaluating fairness reasoning in LLMs. We study first-person fairness judgments across a broad set of models and compare them directly with human responses on matched scenarios and elicitation conditions. We find that LLMs tend to prefer stricter fairness constraints than humans, show more self-interested behavior, are sensitive to how information is framed, and are difficult to align with human judgments using fine-tuning with current datasets. --- ### Predicting Symptoms of Amotivation and Anhedonia among University Students with a Novel Oversampling Method Authors: Dang Nguyen, Bao Duong, Arun Kumar, Dat Phan-Trong, Julian Berk, Taylor Braund, Kien Do, Debopriyo Bal, Wu Yi Zheng, Leonard Hoon, Jill Newby, Helen Christensen et al. Published: 2026-08-31 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.29690v1 University students experience disproportionately high rates of common mental health conditions, such as depression, which can impair learning, social functioning, and overall well-being. Within this context, symptoms of amotivation (i.e. loss of motivational drive) and anhedonia (i.e. diminished interest or pleasure) are particularly debilitating, yet they frequently go undetected. Developing new approaches to identify students with prominent amotivation and anhedonia could enable earlier and more targeted intervention. Machine learning (ML) methods have increasingly been used to classify individuals according to symptom severity. However, these ML models often suffer from class imbalance, where the majority of cases fall in the low-symptom group and relatively few in the high-symptom group. This imbalance can reduce model accuracy and bias predictions. To address this, studies commonly employ the popular oversampling strategy SMOTE. However, SMOTE has a notable limitation: it may generate invalid values for nominal variables. In this paper, we introduce a novel and effective oversampling method that addresses this shortcoming. Our approach leverages a predictive model to generate nominal variables, rather than interpolating them. We validate our method on a large-scale GPS location dataset collected from university students and demonstrate that it is significantly better than existing oversampling approaches in predicting elevated symptoms of amotivation and anhedonia. --- ### DP-IPI: A Hybrid Differential Privacy Text Rewriting Mechanism for Indirect Personal Identifiers in Clinical Texts Authors: Ibrahim Baroud, Stephen Meisenbacher, Sebastian Möller, Florian Matthes, Roland Roller Published: 2026-08-31 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.29684v1 Despite the strengths of modern anonymization and de-identification techniques, the risk of re-identification remains significant due to the indirect identifiers remaining in texts. To address this problem, recent works have applied text rewriting under Differential Privacy (DP) to prevent data linkage by perturbing texts via noise addition. Such methods privatize all tokens in a text indiscriminately, diminishing text quality and usability in critical domains such as in clinical settings. Focusing on indirect personal identifiers (IPIs), we introduce a utility-preserving DP text rewriting method that only privatizes spans containing IPIs. We show that our method effectively reduces re-identification risks in clinical texts while being producing more coherent and usable output texts, leading to higher privacy-utility trade-offs. In this, we demonstrate the effectiveness of hybrid text privatization, which leverages the promise of DP in an efficient, usable manner. --- ### Sequential knowledge editing breaks a model's ability to tell good evidence from bad, without costing it accuracy Authors: Atul Anand Published: 2026-08-31 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.29587v1 Knowledge editing is evaluated on whether the edited fact changed, whether paraphrases follow, and whether unrelated answers stayed put. A model can pass all three and still lose something none of them measures: the ability to decide, on facts that were never edited, which retrieved documents to believe. We score the log odds a model assigns to its remembered answer against the answer an injected passage asserts, before and after editing, holding the query, the passage and both candidate strings fixed. Our cleanest arm is a conservatively tuned LoRA: after 1,000 sequential edits on Qwen2.5-7B-Instruct it leaves MMLU unchanged to four decimal places, yet the spread of the arbitration quantity across untouched facts falls by 36%. Selective prediction degrades with it. Area under the risk-coverage curve rises by 0.107, against 0.005 for a norm-matched perturbation at the same MMLU, and error on the model's most confident quarter of arbitration decisions goes from 0.217 to 0.342. This is not capability loss. Sweeping random perturbation over five severities, damage bad enough to cut MMLU from 0.6275 to 0.3725 produces less harm (0.088) than MEMIT does at 0.6050 (0.102). The effect holds across three seeds, two model families, two datasets, two probe-disjointness criteria, three prompt templates and paraphrased queries. Layer ablation on saved weight deltas shows it is distributed: no single layer reproduces it, and removing any one recovers about half. Under retrieval with a frozen retriever, accuracy falls from 0.592 to 0.46. A secondary finding may matter more in practice. Three of five model and method pairings we ran collapse to chance MMLU at 1,000 sequential edits under published hyperparameters, while edit success stays at 1.00 and locality reads clean. Sequential-editing evaluations that never measure capability cannot see this. --- ### PII-TRACE: A Benchmark for Context-Aware PII Detection in Multi-Turn LLM Conversations Authors: Kaiyuan Zhang, Chuan Wang, Joey Zhong, Paul Fryzel, Kyle Polley, Jerry Ma, Ninghui Li Published: 2026-08-31 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.22200v1 LLM assistants and agentic systems log long multi-turn conversations. AI providers often scan these conversations for Personally Identifiable Information (PII) and mask the PII before storing or processing conversation data. Yet most PII detectors and benchmarks target self-contained records rather than cross-turn evaluation. To evaluate PII detection across turns in multi-turn conversations, we introduce PII-TRACE (Tracing Recurring PII Across Conversational Exchanges), to our knowledge the first PII benchmark to assess whether detectors identify PII in conversational contexts and cover every mention of a recurring identifier across turns. PII-TRACE contains 13,148 synthetic multi-turn dialogues in 13 languages with character-level spans and identifier clusters. Across eleven baselines, including frontier LLMs, no detector achieves full entity-level coverage without substantial false positives on PII-free conversations, and single-pass reading loses a third of the gold characters on long dialogues. To close this gap, we introduce PII-Tracer, a compact 0.6B-parameter detector trained with conversation-level supervision. PII-Tracer attains the highest entity-level coverage of any system we evaluate and also performs strongly on standard single-record benchmarks. --- ### Improving Parameter Utilization by Sharing Neural Experts Across Layers in Transformers Authors: Dian Jiao, Jiaxin Duan, Shuai Zhao, Jiabing Leng, Yiran Zhang, Feng Huang Published: 2026-08-31 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.22199v1 Transformer-based large language models often suffer from inter-layer parameter redundancy, where functional transformations are redundantly learned across network depths. We propose CS-MoE, a novel Transformer architecture featuring cross-layer expert sharing to address this inefficiency. Deviating from the widely used Mixture-of-Experts (MoE) architecture that terminates each Transformer block with layer-isolated experts, CS-MoE combines layer-independent experts with concurrent access to a centralized, globally shared expert pool. This \textit{Global Experts Sharing} mechanism enables elastic control over token-level parameter activation and computational consumption (FLOPs). Experiments demonstrate that CS-MoE achieves lower perplexity than equal-scale dense Transformers while activating only 55\% of parameters. Furthermore, its performance scales monotonically with an increased number of activated experts and approaches MoE counterparts that consume more FLOPs by expanding the shared pool with a fixed FLOPs budget. CS-MoE also establishes a flexible Pareto frontier between computational cost and model capacity, offering an efficient alternative for computation-constrained environments. --- ### Reconstruction of 4D Mitral Regurgitation Hemodynamics from Sparse Planar Data using Deep Operator Networks with Test-Time Adaptation Authors: Jakob Marcel Hoffmann, Yosuke Hasegawa, Alexander Stroh Published: 2026-08-31 Categories: physics.med-ph, cs.LG, physics.flu-dyn arXiv: https://arxiv.org/abs/2609.20857v1 Quantifying mitral regurgitation severity remains limited by the assumptions of clinical flow convergence methods, while high-fidelity simulation and volumetric velocimetry are too slow for routine use. We investigate whether a learned solution operator can reconstruct transient three-dimensional transvalvular hemodynamics from the sparse observation an in-vitro experiment actually provides: a single planar velocity slice and two boundary pressure traces. A Deep Operator Network is pretrained on an experimentally benchmarked URANS database spanning eleven mitral regurgitation orifice phantoms, learning a mapping from a masked two-component planar velocity snapshot to the surrounding volumetric field, and is subsequently adapted to unseen target cases by fine-tuning on their sparse measurements. Adaptation reliably corrects the flow topology within the supervised plane, reorienting a strongly eccentric jet that the pretrained operator predicts as straight, and yields full-field predictions in minutes rather than the days required by the underlying simulations. Its influence decays sharply with distance from that plane, however: measured against phase-resolved particle image velocimetry, the reconstruction error rises from 24.6% at 2mm to 52.6% at 6mm, and the resulting mismatch between corrected and uncorrected layers degrades physical consistency. Single-plane supervision thus constrains the observed plane far more effectively than the surrounding volume, which we identify as the principal obstacle to coherent 4D reconstruction from sparse planar data. --- ### ViCo: Visual-oriented Coding with Self-Reflection for Chart Replication Authors: Jiaxin Duan, Dian Jiao Shuai Zhao, Jiabing Leng, Yiran Zhang, Feng Huang Published: 2026-08-31 Categories: cs.CL, cs.GR, cs.LG arXiv: https://arxiv.org/abs/2609.16014v1 This paper addresses the challenge of generating high-quality academic charts that match the visual standards of human-authored papers. While existing AI agents can produce well-structured text and code, their generated visualizations often lack the stylistic and semantic fidelity of human designs. Advanced coding agents that employ self-reflection mechanisms exhibit poor visual reasoning and limited reflection following, resulting in sparse reward signals that severely undermine their reinforcement learning (RL). We propose ViCo, a training framework for visual-oriented coding that employs iterative reflections to align generated chart images progressively with the reference. We first introduce a self-supervised warm-up stage, which augments Monte Carlo Tree Search with consistency-based pruning to synthesize high-quality reflection trajectories, ensuring that each coding step strictly follows the outcomes of prior reflections. A multi-step RL algorithm is then developed, using counterfactual baselines to estimate advantage for reflection and action steps within each refinement cycle, thereby addressing the reward sparsity. To enable efficient reward in massive training, we propose an automatic, multifaceted evaluation framework that assesses charts' style, layout, and semantic consistency via a hierarchical heterogeneous layout graph structure. Experiments on three public benchmarks demonstrate that ViCo, trained on an 8B model, achieves performance close to proprietary LLMs with adequate reflection capabilities. --- ### Lensless Gaze Is Not Private by Default: Auditing Identity Leakage Across Disclosure Surfaces Authors: Rahul Vimalkanth, Kaushik Mitra Published: 2026-08-31 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.09188v1 Lensless near-eye sensing is often described as privacy-friendly because its coded measurements are visually unintelligible. Yet visual unintelligibility reflects human interpretation, not what a learned adversary can recover. We therefore treat identity privacy as a systems property of disclosure surfaces: representations crossing sensing, storage, computation, and output boundaries. We audit a simulated lensless gaze pipeline under a 36-subject known-gallery closed-set identification protocol with a fixed, known PSF; privacy from an unknown or varying optical key is outside our scope. Reported accuracies are empirical attack success rates under matched linear and MLP probes and do not upper-bound stronger adversaries. Simulated lensless measurements yield 96.7% top-1 identification versus 97.7% for matched original eye crops, while an MAE embedding retains 94.3%. Compression alone offers little protection: 8-D PCA and a matched 8-D bottleneck retain 93.2% and 91.8%, whereas separately trained 8-D GSPL bottlenecks yield 77.5% mean recovery across three seeds. A released 128-way gaze token lowers single-frame recovery to 38.1%, while its residual and continuous gaze output expose 62.1% and 72.6%, respectively. Under a source-frame-disjoint tiled protocol, token summaries reach 39.9% at T=25, showing that repeated-output risk depends on representation and aggregation. These rates reflect all subject-correlated information in the evaluated dataset, including acquisition and behavioral cues, rather than isolating intrinsic ocular biometrics. Ordinary least squares residualization against a six-dimensional crop geometry and intensity summary still leaves lensless recovery at 95.1%. Our results show that privacy claims for lensless sensing must be tested at disclosure boundaries rather than inferred from appearance. --- ### AgenticGen: Reward-Guided Agentic Video Generation for Advertising Authors: Xingyuan Bu, Chengru Song, Hao Zhou, Tao Zhou, Dong Li, Wei Li, Shilong Li, Hao Shi, Yongxin Guo, Donghao Zhou, Qiangpeng Yang, Shilei Wen Published: 2026-08-31 Categories: cs.CV, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.09187v1 Advertising video generation is not only a video synthesis task, but also a product-conditioned reasoning problem whose success is measured by online business metrics. Recent video foundation models can generate realistic clips from multimodal conditions, yet they do not optimize how a product should be transformed into an effective advertisement or how future generation should be improved from online business feedback. To close this loop, we propose AgenticGen, a reward-guided agentic framework that decomposes advertising video generation into two trainable reasoning stages, strategy selection and draft generation, thereby exposing optimization targets that online business feedback can supervise. AgenticGen learns a performance-based reward from accumulated online feedback and a complementary rubric-based reward aligned with human quality standards, then uses them to supervise policy optimization. DPO first moves the agentic policies toward online preferences, and GRPO further refines both stages with process and outcome rewards. Offline experiments validate the reward models and successive policy optimization. Online A/B experiments in the TikTok advertising system show that AgenticGen after DPO and GRPO improves CTR by 2.72%, CVR by 2.63%, and Advv by 9.61% over the SFT baseline. --- ### M2LG-DG: A Multi-modal Local-Global Domain Generalization Framework for Cross-site Major Depressive Disorder Classification Authors: Muhammad Asif Hasan, Yanming Zhu, Xuefei Yin, Alan Wee-Chung Liew Published: 2026-08-31 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.09186v1 Classification models based on resting-state functional magnetic resonance imaging (rs-fMRI) often show lower performance at imaging sites not included during model development, which can limit their use in clinical settings. Domain generalization (DG) addresses this issue by learning representations from source sites that remain effective for unseen target sites. However, existing DG approaches for psychiatric disorder classification commonly rely on a single imaging modality and may not fully account for site-specific acquisition effects on the learned representation space. Subjects scanned at the same site share scanner hardware, acquisition settings, and preprocessing characteristics, which can cause representations to reflect acquisition conditions rather than diagnostic information. In this work, we present M2LG-DG, a source-only multimodal local-global framework for cross-site major depressive disorder (MDD) classification. The framework employs a dual-stream rs-fMRI encoder, where the global pathway models inter-regional dependencies through self-attention and the local pathway performs graph-constrained aggregation over functional connectivity-derived brain graphs. Imaging and non-imaging representations are decomposed into shared and private components and integrated through bidirectional cross-attention with a learned modality gate. A cross-site supervised contrastive objective forms positive pairs from same-class subjects acquired at different source sites, encouraging the fused representation to preserve diagnostic information across acquisition domains. On four held-out REST-meta-MDD sites, M2LG-DG achieves an AUC of 69.48% and exceeds the closest comparison method by 2.18 percentage points. Experiments on the Autism Brain Imaging Data Exchange (ABIDE) dataset further support its applicability to other psychiatric neuroimaging classification tasks. --- ### Integrating Unimodal and Vision-Language Representations in Latent Space for Multi-Label Chest X-Ray Classification Authors: Quang-Huy Tran, Duc-Tuan Ngo, Minh-Khoi Nguyen-Bui, Dang-Khoa Bui, Thanh-Trong Tran, Tuan-Khoi Nguyen, Hoang-Anh Ngo Published: 2026-08-31 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.09185v1 Multi-label chest X-ray classification has attracted considerable attention in recent years, with the effective use of visual representations and clinical semantic knowledge playing an important role. This study proposes a framework that combines unimodal representations from RAD-DINO with vision--language representations from BioViL-T for the classification of 14 labels in the MIMIC-CXR-JPG dataset. The RAD-DINO and BioViL-T embeddings and their combined representation are refined separately in latent space before being normalized and fused across the three branches. In addition to improving classification performance, the study aims to clarify the role of each embedding source and the degree to which they complement one another. Experiments show that RAD-DINO outperforms BioViL-T when used independently, whereas early fusion further improves the results, indicating that the two embedding sources contain complementary information. The best-performing model achieves a mean AUROC of 0.840 and an mAP of 0.467. Ablation analysis shows that hybrid fusion provides consistent and statistically significant improvements over early fusion when each embedding source is refined in latent space, suggesting that fusion effectiveness depends on the quality of the representation supplied by each branch. However, the study has only been evaluated internally on MIMIC-CXR-JPG; its generalizability to data from other healthcare institutions therefore remains to be validated. The source code is available at: https://anonymous.4open.science/r/mimic-report-C210/. --- ### Robots Influencing Humans to Reveal their Goals during Collaboration and Competition Authors: Debasmita Ghose, Oz Gitelson, Michal Lewkowicz, Jake Brawer, Marynel Vazquez, Brian Scassellati Published: 2026-08-31 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.05519v1 We propose a unified strategy for fast goal inference in human-robot interaction. The core idea is to drive the human toward Critical Decision Points (CDPs)-states where competing human strategies prescribe different next actions and thus maximally reveal the goal. We formalise CDPs using a goal-conditioned policy divergence measure and incorporate them into a Receding-Horizon Planner that explores future action sequences while optimizing a cost function balancing task progress and information gain. We evaluate this approach in both a collaborative, fully observable cooking task and a competitive, partially observable hide-and-seek game, each in simulation and on real robots. In both scenarios, our method infers human goals more accurately and earlier than baseline strategies. --- ### CrossModalQA: A Cross-modal and Multi-hop Benchmark for Multimodal Retrieval-augmented Generation Authors: Jiacheng Cai, Zijin Hong, Zheng Yuan, Huachi Zhou, Qinggang Zhang, Xiao Huang Published: 2026-08-31 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.05518v1 Despite the strong capabilities of multimodal large language models (MLLMs), their parametric knowledge remains incomplete and difficult to update, motivating multimodal retrieval-augmented generation (RAG) to ground responses in external text and images. However, existing benchmarks face two major limitations: (i) they typically emphasize single-hop retrieval or reasoning over a small set of provided contexts rather than open-domain evidence discovery; and (ii) they provide fragmented coverage of cross-modal reasoning paths, leaving complex multi-hop and multi-image reasoning underexplored. In this paper, we introduce CrossModalQA, an open-domain benchmark for evaluating multimodal retrieval and reasoning over heterogeneous corpora. CrossModalQA contains 1,863 question-answer pairs constructed from 4,987 Wikipedia articles and 4,431 Wikimedia Commons images. It covers five complementary reasoning paths: vision-to-text, text-to-vision, vision-to-text-to-vision, multi-image intersection, and image-set reasoning. Every question requires retrieving and composing distributed textual and visual evidence, with an average reasoning depth of 3.50 hops. We construct the benchmark through multimodal knowledge graph-guided subgraph sampling and apply rule-based consistency checking and LLM verification to ensure multimodal dependence and traceable evidence. Extensive experiments demonstrate that existing multimodal RAG systems struggle to recover complete evidence chains and can underperform closed-book models when incomplete retrieval introduces distracting context. Further analysis reveals that complete cross-modal retrieval contributes more to answer accuracy than generator scaling, while multi-image retrieval and reasoning remain the primary bottlenecks limiting end-to-end performance. --- ### Emergent Goal-Directed Attention in Large Vision-Language Models Authors: Han Zhang Published: 2026-08-31 Categories: cs.CV, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.05517v1 Human observers prioritize visual information according to task goals. Most computational models of naturalistic viewing are gaze-trained for free viewing, leaving open whether goal-directed attention can emerge in systems without gaze supervision. We tested two off-the-shelf vision-language models (VLMs), Qwen3-VL-32B-Thinking and Gemma-4-26B-A4B-it, on 4,887 naturalistic scenes under visual-search and free-viewing instructions. Model predictions were compared with human fixations on the same images under corresponding tasks. Both models aligned more closely with human fixations under matching goals than under mismatched goals. This crossover persisted in target-absent scenes, where alignment could not be explained by simple visual grounding, and appeared in decoder-layer readouts. Furthermore, model-thinking traces were grounded in target semantics during search and in visual prominence during free viewing. These findings show that general-purpose VLMs can generate human-aligned, goal-directed spatial priorities without gaze-specific training, informing theories of goal-directed attention and offering scalable tools for predicting where people look across tasks. --- ### An Exploratory Study of Frequency-Aware Task Weighting for YOLOv8-Based Unified Driving Perception Authors: Zhiyuan Nie, Zixi Zhou, Xianbin Gu Published: 2026-08-31 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.05516v1 Unified perception enables autonomous driving systems to perform object detection, drivable-area segmentation, and lane segmentation within a single network, improving efficiency and reducing deployment complexity. Jointly optimizing multiple perception tasks remains challenging because tasks exhibit different convergence rates, loss scales, and optimization stability. Existing task-weighting methods use loss magnitude, learned uncertainty, short-term loss changes, or gradient statistics; here, we explore the frequency structure of a recent loss-history window as a complementary signal. We implement and examine Frequency-aware Task Weighting (FTW), a dynamic task-balancing rule that estimates a loss-trajectory stability proxy from the low-frequency energy ratio of recent loss histories. FTW assigns larger weights to tasks whose mean-centered loss trajectories contain a larger proportion of low-frequency power. We document FTW and two baselines under full-network static training and progressive freezing using a unified YOLOv8-based perception framework with three task-specific heads. Experiments on Mapillary Vistas compare FTW with fixed and uncertainty-based weighting under both configurations. Final holdout metrics are reported for the checkpoint with the lowest per-epoch validation loss in each run. Across six single-run configurations, static FTW has the largest derived overall score and lane mIoU, progressive FTW has the largest detection mAP, and static uncertainty weighting has the largest drivable-area mIoU. Without repeated-seed estimates, single-task baselines, or FTW ablations, these rankings are descriptive. The evidence supports the feasibility of loss-frequency-based weighting in this pipeline, but does not establish improvement over the baselines or generalization beyond the reported runs. --- ### The Failure Happens Before the Drift: The Social Dynamics of Values in LLM Agent Societies Authors: Farah Atif, Sougata Saha, Monojit Choudhury Published: 2026-08-31 Categories: cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.05514v1 Large Language Model (LLM)-based agents are increasingly used as proxies for human participants in social science research, yet it remains unclear whether they can faithfully simulate diverse and conflicting human value systems. We present a World Values Survey (WVS)-grounded simulation framework where culturally diverse agents with different communication styles engage in longitudinal, value-laden discussions. Across approximately 4,000 conversations involving 1,200 personas, 15 topics, and three models (GPT-4o, Gemini-2.5-Flash, and Gemma-4-E4B), we evaluate value faithfulness, value drift, and conversational realism. We find that more than 50\% of personas fail to express their assigned WVS profiles from the outset, while 2-7\% drift after repeated conversations. Ablations removing demographic details improve faithfulness for some models but do not change the broader trend: simulated value distributions still systematically deviate from the assigned WVS profiles. Compared to human discussions, simulated dialogues show a different trade-off between stylistic consistency and semantic diversity, often producing content-wise varied but stylistically repetitive exchanges. These findings suggest that current LLM agents can generate plausible conversations, but remain limited proxies for representing and preserving diverse human value profiles over time. --- ### When and What to Teach: Budget-Aware Online Adaptation for Web Agents Authors: Jianwei Zhang, Sihan Cao, Pengcheng Zheng, Ya Wen, Pei Ke, Kuien Liu, Shen Gao, Wei Dong, Yang Yang, Chaoning Zhang Published: 2026-08-31 Categories: cs.AI, cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.05513v1 Web agents have achieved significant success in automating complex internet tasks but deploying them in real-world environments requires continuous online adaptation. Given that deploying powerful proprietary models remains commercially cost-prohibitive, practitioners must rely on lightweight local models that evolve post-deployment via online teaching from a stronger teacher. However, standard interactive feedback imposes prohibitive costs. We show that conventional trajectory-level preference optimization wastes budget on both unresolvable episodes and redundant execution turns. To resolve these inefficiencies, we propose \textbf{Score-Guided Online Teaching with Budgeted Trajectory Trimming}, a budget-aware framework that systematically orchestrates \textbf{when} and \textbf{what} to teach. Specifically, our framework integrates a solvability-aware teacher gate to dictate \textbf{when} to query the teacher model and a score-guided turn selection mechanism to decide \textbf{what} informative turns to retain. Extensive experiments on MiniWoB and TimeWarp demonstrate that our method achieves comparable first-pass success while reducing teacher calls by 22.6\% and student training compute by 52.1\% on average. Our code is available at https://github.com/zjw131f1fc/budgeted-online-teaching. --- ### Reasoning-Aware Compression: Identifying and Protecting Vulnerable Reasoning Circuits for Energy-Efficient LLM Deployment Authors: Leonard Twagirayezu, Prasenjit Mitra Published: 2026-08-31 Categories: cs.AI, cs.CL, cs.PF arXiv: https://arxiv.org/abs/2609.05512v1 Large Reasoning Models (LRMs) impose substantial energy costs during deployment, yet current compression methods apply uniform quantization across all components, risking damage to critical reasoning circuits. We present a reasoning-aware compression framework that benchmarks quantization conditions across five reasoning benchmarks, GSM8K, FOLIO, MATH-500, ProofWriter, and MuSiQue, with hardware-level GPU energy measurement; profiles per-module INT4 vulnerability across all 196-224 (layer, projection) pairs via a perturbation sweep on a held-out calibration split, then selectively restores the most sensitive circuits to FP16. Three findings emerge. First, INT4 quantization can increase energy by extending reasoning chains; a 25% power reduction becomes a net energy increase on GSM8K. Second, vulnerability is task-dependent: attention projections are more critical for mathematical reasoning, and sensitivity patterns differ by architecture in logical inference. Third, selective compression achieves Pareto-optimal points inaccessible to uniform methods: R1-Qwen-7B Top-10% on ProofWriter gains +12 pp over FP16 at -9.7% energy, validated on held-out data across five reasoning benchmarks. --- ### SCAFFOLD: Self-Improving Web Agents via Recursive Parametric Skill Abstraction Authors: Bowei He, Xiaokun Zhang, Meng Ding, Xue Liu Published: 2026-08-31 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.05511v1 Web agents need to navigate visually rich, long-horizon interfaces that change across sites, yet most previous agents still learn each task in isolation and discard the procedural knowledge they accumulate. Recent skill-augmented frameworks take an important first step, but they treat the skill library as a flat or two-tier prompt-side cache and offer no principled mechanism for compressing redundancy or composing skills recursively. We introduce \textsc{Scaffold}, a self-improving framework for visual web agents that (i) induces parametric, executable skills from successful trajectories under a multi-instance abstraction constraint, (ii) maintains a recursively composed hierarchy in which higher-level skills invoke lower-level ones, (iii) compacts the library via a minimum-description-length (MDL) criterion and behavioral equivalence checking, and (iv) periodically distills skill-augmented trajectories back into model weights to internalize the abstractions. Across WebArena, VisualWebArena, and a held-out split of Online-Mind2Web, \textsc{Scaffold} improves success rate by $11.1$--$17.2$ absolute points over the strongest skill-augmented baseline and shows monotonic gains across five self-improvement iterations without library collapse. We release the code and documents in the Github \href{https://github.com/BokwaiHo/SCAFFOLD}{repository}. --- ### AVENUE: Audio-Video EditiNg Understanding and Evaluation Authors: Hayeon Kim, Yoojin Jang, Jaejun Yoo Published: 2026-08-31 Categories: cs.MM, cs.CV, cs.SD arXiv: https://arxiv.org/abs/2609.04253v1 Audio-video (AV) editing aims to modify audio and video content according to a target prompt. Unlike single-modality editing, AV editing requires models to infer a modality-selective edit scope from the prompt alone: determining not only what should change, but also which modality should be preserved. Faithfully evaluating such models therefore requires both (i) benchmarks that span diverse edit types and modality categories, and (ii) evaluation that is itself modality-aware and sample-specific. However, existing AV editing benchmarks provide limited coverage of edit types and modality combinations, while current evaluation systems are often modality-blind and sample-agnostic, making it difficult to assess whether models faithfully preserve the unintended modality. To address these gaps, we introduce AVENUE, Audio-Video EditiNg Understanding and Evaluation, comprising two contributions: (1) a benchmark of 1,291 source clips and 7,957 editing instructions across audio-targeted, video-targeted, and AV-coupled edit types, curated and human-verified from VGGSound; and (2) a sample-specific, modality-aware evaluation framework that specifies, for each sample, both the intended change and the content that must remain intact. We evaluate representative AV editing models spanning three editing paradigms : joint, sequential, and separate, providing the first systematic analysis of modality-selectivity across paradigms. Our findings reveal a fundamental open challenge: when editing one modality, existing models frequently induce unintended changes in the other, regardless of paradigm. AVENUE provides a benchmark and modality-aware evaluation framework to drive progress toward more controllable AV editing models. Our dataset is publicly available on Hugging Face: https://huggingface.co/datasets/AVENUE-dataset/AVENUE. --- ### Judging LLM-as-a-Judge: Concerning Rubric Artifacts in LLM-based Automated Text Generation Evaluation Authors: Anshul Bagaria, Sowmya S Sundaram, Gokul S Krishnan, Balaraman Ravindran Published: 2026-08-31 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.02942v1 LLM-as-a-Judge pipelines are increasingly used to evaluate AI-generated text, based on the assumption that judgments arise from reasoning over candidate responses with respect to a rubric. We show that this assumption warrants further scrutiny. Classifiers trained only on rubric text, without access to any evaluated response, achieve nontrivial predictive performance on judge outputs. This suggests that rubric formulations encode recoverable evaluative signals, allowing scores to be partially anticipated independently of model outputs. Finally, counterfactual perturbations reveal that judges often fail to reliably update their decisions when either the candidate response or the rubric criterion is reversed. Our findings raise concerns about the reliability of rubric-based LLM evaluation and highlight the need for further methodological study of automated evaluation via LLMs. --- ### SISER: Speaker-Invariant Speech Emotion Recognition with Entropy-Based Adversarial Training Authors: Eunseo Choi, Hyunku Kang, Chanwoo Kim Published: 2026-08-31 Categories: cs.SD, cs.CL arXiv: https://arxiv.org/abs/2609.02941v1 Speech emotion recognition (SER) faces two fundamental challenges: scarcity of labeled data and inter-speaker variability, both of which hinder generalization of emotion recognition systems. While prior adversarial approaches address speaker variability, they fall short in leveraging powerful pre-trained representations. We propose SISER (Speaker-Invariant Speech Emotion Recognition), integrating wav2vec 2.0 as a feature encoder and ECAPA-TDNN as a speaker discriminator within an entropy-based adversarial training scheme. wav2vec 2.0 provides rich self-supervised representations that alleviate dependency on large labeled datasets, while ECAPA-TDNN enables suppression of speaker identity via a stronger adversarial signal than shallow classifiers. Evaluated on IEMOCAP, SISER achieves a UA of 60.63%, outperforming the baseline (51.15%) and wav2vec 2.0 without speaker suppression (56.46%), with ablation emphasizing that the choice of speaker classifier architecture is a key factor. --- ### Listen to the Latents: Self-Correcting Speech Recognition in Large Audio Language Models Through Hidden-State Interactions Authors: Chan-Jan Hsu, Jaeyeon Kim, Chao-Han Huck Yang, Shinji Watanabe, Hung-yi Lee, Carlos Busso Published: 2026-08-31 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.02940v1 Recent automatic speech recognition (ASR) systems increasingly integrate large language models (LLMs) to leverage their semantic knowledge, either externally through logit fusion or internally through warm initialization. However, how to effectively combine these two strategies remains underexplored. In this work, we refine warm-initialized LLM-based ASR models by leveraging their own pre-adaptation base LLMs, focusing on LoRA-adapted settings where the base LLM is preserved. To achieve this, we propose Hybrid Search, a targeted correction strategy motivated by two observations. First, interaction features that characterize the relationship between LLM-based ASR hidden states and base-LLM hidden states provide informative signals about a token's degree of semantic dependence. Second, selectively refining targeted tokens with high semantic dependence improves ASR performance far beyond naive global LLM-correction methods including rescoring and late fusion. Our analysis suggests that, even after semantic knowledge transfer through warm initialization, LLM-based ASR models can still leverage their base LLM to further improve inference-time performance. --- ### Context Inference Attacks Without Jailbreaks Authors: Prince Jha, Samuele Poppi, Nils Lukas Published: 2026-08-31 Categories: cs.CR, cs.LG arXiv: https://arxiv.org/abs/2609.01663v1 Agentic AI systems are increasingly deployed to process sensitive data at inference time, such as healthcare records or financial documents assembled into a hidden \emph{context} before the system answers. Prior work has studied privacy risks primarily through \emph{jailbreaking} attacks that induce models to directly disclose sensitive content, but has largely overlooked the agentic setting where the context is assembled by the agent's own tool calls. We show that the agents we evaluate remain vulnerable to hidden-context leakage despite the controls we test against them, namely an instruction not to disclose the context, logit suppression, and context dilution. For instance, a web-browsing agent answering benign user queries still carries exploitable signals about records silently loaded into its context. We introduce and formalize \emph{context-inference attacks} through a security game and evaluate three settings under decreasing attacker knowledge and increasingly indirect delivery of the context: a known context, an unknown context, and a context the agent retrieves through its own tool calls. We distinguish a grey-box setting, in which the target model is used to score observations, from black-box settings in which the attacker scores with a surrogate it controls. We further characterize how leakage varies with query budget, context size, and target-model size. A single attack carries through all three settings without modification, reaching $100\%$ ASR on small candidate sets and $63\%$ at $1024$ candidates against a known context, $78.9$ AUROC when the template and surrounding records are unknown, $92.5$ AUROC when a 14B surrogate scores a 32B target, and $81.8$ AUROC when the records arrive as an agent's retrieval returns, against chance rates of $1/|\mathcal{Z}|$ and $50$ respectively. --- ### Not All Agreement Counts as Corroboration: Provenance-Conserving Multi-View Fusion for Typed Action Admission in Human-Robot Collaboration Authors: Zekai Jin, Hanrong Zhang, Yihong Tang, Fei Hu, Zhen Dong, Yi Shao Published: 2026-08-31 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.01662v2 Better probability scores do not establish that evidence has been counted correctly. Repeated inference over one observation can improve predictions without adding an evidential origin. Source-local numerical attributes alone cannot in general distinguish repeated derivations from separately countable acquisitions. PACT (Provenance-Aware evidence Conservation and Typed action admission) separates evidence magnitude from countability through a supplied provenance partition. Under singleton fidelity and insertion non-amplification, the coordinatewise meet is the unique pointwise greatest admissible within-component rule. Component budgets add under stated commensurability and separate-component additivity assumptions. Matched reassignments hold numerical outputs fixed while varying the counting relation. In four of 12 replicated-source tests on HandWritten, false refinement lowers macro-averaged negative log-likelihood and Brier score while increasing normalized common-support area under the risk-coverage curve (ncsAURC). In the controlled handover benchmark, removing the constructed adversarial-consensus condition leaves a 0.056 reduction in ncsAURC for provenance-partition aggregation relative to singleton aggregation under the same score functional. The corroboration contrast disappears, and method ranking remains selection-score dependent. In offline, reference-based human-robot collaboration with four prompts per camera and all other admission inputs fixed, duplicating each prompt output within its camera from multiplicity one to eight leaves all 720 PACT typed responses per checkpoint unchanged. Probability quality and evidence countability require separate evaluation. --- ### How Fast Do Agents Rot? An Empirical Study of Long-Horizon Degradation in LLM Agents for Production Decision-Making Authors: Shubhra Mittal Published: 2026-08-31 Categories: physics.soc-ph, cs.AI arXiv: https://arxiv.org/abs/2609.01660v1 Production deployments of large language model (LLM) agents remain unreliable on long, multi-step workflows even as benchmark success rates climb steadily. We argue this gap is largely an artifact of task horizon: benchmarks are dominated by short-to-medium horizons where success remains high, while production workloads demand an order of magnitude more dependent steps. We measure the effect directly, characterizing the shape of agent degradation and disentangling its cause across a large controlled study spanning nine models, six open models from 1.2B to 671B parameters, and three deployed proprietary systems; four task families, including a genuinely agentic tool-use loop; five horizons; and three context regimes. Task success follows a geometric law governed by a single per-step reliability parameter, which rises with model scale but saturates well below 1 even for the strongest models, guaranteeing eventual collapse at sufficiently long horizons. The effect is sharpest on the agentic task, where every model tested, including widely deployed systems, falls from near-perfect success to near zero within sixteen steps of (n=10,664 analyzed trajectories. Degradation is driven by step count rather than context length: bounding the context window steepens decay rather than easing it (logit slope -0.69 vs. -0.44), p=3x10-6), contradicting a lost-in-the-middle explanation and warning against a common production shortcut. Projecting measured reliability onto representative benchmark horizons quantifies a substantial gap between benchmark and production conditions, from 0.42 at GAIA-length horizons to 0.24 at hundred-step production horizons. For teams responsible for agent orchestration and reliability at scale, these results argue for horizon-aware evaluation and reliability budgeting in place of aggregate pass-rate metrics. Code, prompts, seeds, and raw trajectories are released. --- ### Beyond Textual Chain-of-Thought: A Survey on Action-Grounded Reasoning in Autonomous Driving Authors: Zhengxu Tang, Xiaozhou Zhang, Guofeng Cui, Ziyu Gong, Zi Wang, Yunfei Shi, Ruifeng Deng, Chengzhi Qi, Ke Chen, Sachin Patil, Tianjun Xiao, Langechuan Liu et al. Published: 2026-08-31 Categories: cs.CV, cs.CL, cs.RO arXiv: https://arxiv.org/abs/2609.01659v1 Chain-of-thought (CoT) reasoning powers generative models by eliciting intermediate steps before producing an answer. In autonomous driving, the answer is a continuous action. Thus its reasoning must share the same spatiotemporal structure as the physical world. This survey studies the resulting shift from textual CoT to action-grounded reasoning. Surveying 171 papers, including 130 method papers and 41 benchmarks, datasets, surveys, and analysis papers, we propose a representation-centered taxonomy that treats the form of the intermediate state as the organizing axis. We systematize the 130 methods into four categories: language-based, visual-spatial, latent-dynamic, and externalized reasoning, further divided into 13 subtypes tied to distinct regions of interests. Our synthesis shows that the open frontier of reasoning in driving agents lies in intermediate representations that can be grounded in the real world, coupled to real-time action, and verified under safety-critical systems. Project page: https://github.com/tangzhengxu/awesome-av-cot. --- ### PRO-Step: Step-level Process Reward Optimization for Retrieval-Augmented Generation Authors: MinKeon Kim, Namjun Lee, Jaekwang Kim Published: 2026-08-31 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.01658v1 Retrieval-Augmented Generation enhances Large Language Models by grounding responses in external knowledge, but multi-hop reasoning remains vulnerable to error propagation, where early retrieval failures confound subsequent steps. Standard outcome-based optimization only rewards the final answer, leaving intermediate retrieval and reasoning errors undetected. While existing process-based methods introduce step-level signals, they still score each step against the final answer, rewarding spurious successes where flawed retrieval coincidentally produces the correct answer. Step-level supervision in RAG requires evaluating both logical validity and evidential grounding at each step. We introduce PRO-STEP: we train a generative PRM that evaluates both dimensions, employ PRM-guided value tree search to construct preference pairs contrasting valid steps against flawed ones, and optimize the policy via step-level Direct Preference Optimization. Experiments on single and multi-hop QA datasets demonstrate that PRO-STEP achieves the best average EM and F1 across five benchmarks. Code, models, and training data are publicly available at https://github.com/keemminnke/PRO-Step. --- ### NeoMME: A Single-Tower Multimodal-Native Multilingual Foundation Encoder for Efficient Fine-Tuning and Inference Authors: Aurélien Lac, Tony Wu Published: 2026-08-31 Categories: cs.IR, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.01657v1 Multimodal models often build on architectures designed for generative vision-language modeling, typically combining separately pretrained vision encoders with causal language models. Visual document retrievers such as ColPali repurpose these models as encoders, carrying over the parameter and compute overhead of a VLM for a non-generative task. We introduce NeoMME, a family of 260M and 800M-parameter Multimodal and Multilingual bidirectional Encoders that process multilingual text and raw image patches in a single bidirectional Transformer encoder. Both models are pretrained from scratch with a masked discrete-diffusion text objective, conditioned on visible image patches for multimodal examples. Both support a 16,384-token context, enough to encode up to two standard 4K UHD images. To demonstrate its downstream capabilities, we fine-tune NeoMME with jointly trained dense and late-interaction heads. On the ViDoRe v3 benchmark, the resulting NeoMME-Retriever 260M outperforms all evaluated models strictly below 800M parameters with 0.523 nDCG@10, while NeoMME-Retriever 800M reaches 0.556. At a matched 2048x2048 image input size on an NVIDIA L40S, NeoMME-260M encodes pages with about 2x the throughput of ColModernVBERT. Hierarchical token pooling and asymmetric quantization compress late-interaction multimodal document embeddings by 255x while preserving over 95% of baseline nDCG@10. We contribute NeoMME to Hugging Face Transformers and release the pretrained backbone and retrieval-compatible checkpoints under Apache 2.0 at https://hf.co/collections/Hcompany/neomme. --- ### From Feature Interaction to Feature Transport - A Unified Block for Scalable Recommendation Models Authors: Zichen Luo, Jiachen Guo, Keming Gu, Jie Zhang Published: 2026-08-31 Categories: cs.IR, cs.AI arXiv: https://arxiv.org/abs/2609.01655v1 Unified recommendation models aim to jointly model non-sequential multi-field features and sequential user behaviors, but existing interaction-centric designs mainly focus on mixing heterogeneous tokens within each layer. We argue that scalable unified recommendation also requires controlling how intent information is carried, filtered, and preserved across stacked blocks. Inspired by flow-based representation dynamics, we introduce feature transport, a view that treats deep unified recommendation as a discrete context-conditioned representation evolution process. We propose CRAFT, a Contextual Residual Adaptive Feature Transport block, which summarizes non-sequential features into a reliability-aware contextual field and uses it to generate residual displacement and memory-preserving signals for intent and sequence representations. In this way, non-sequential context acts as an active controller of representation evolution rather than a passive object of interaction. In the TAAC2026 advertising recommendation competition, CRAFT achieves a test AUC of 0.838090, surpassing the previous leaderboard-best score of 0.83798. Scaling experiments further show that CRAFT benefits from both depth and width expansion: stacking CRAFT to six blocks improves test AUC to 0.838148, while increasing the hidden dimension reaches 0.838106. These results demonstrate the effectiveness, scalability, and generalization potential of the feature transport paradigm. Source code: https://github.com/AshleyLuo001/CRAFT --- ### Validity-Aware Jailbreak Evaluation for Large Language Models Authors: Qilong Wu, Sahil Wadhwa, Pranab Mohanty, Giri Iyengar, Varun Chandrasekaran Published: 2026-08-31 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.00498v1 Jailbreak robustness has become central to large language model (LLM) safety evaluation, yet prevailing methodologies rely primarily on refusal behavior, semantic resemblance, and intent-matching heuristics that emphasize linguistic plausibility rather than correctness. We identify a key limitation in existing evaluations: many jailbreak intents depend on instructional validity rather than epistemic factuality, allowing realistic-looking responses to be labeled successful despite being factually or procedurally incorrect. To address this gap, we propose Sequential Epistemic and Action-Level Validation (SEAV), a verification-centric jailbreak evaluation framework that decomposes responses into ordered steps and evaluates both validity and correctness. SEAV combines LLM-as-a-judge mechanisms for semantic interpretation with retrieval-grounded verification using external knowledge sources, assessing whether generated content is factually correct, structurally consistent, and operationally capable of advancing harmful objectives. Empirically, SEAV cuts the false-positive rate on SD-A (a curated strategic-dishonesty diagnostic) by 14.9\,pp vs. the strongest baseline, and reclassifies 22.1\%--51.0\% of sampled prior-labeled successes as invalid across three of four public benchmarks. Together, these results show that enforcing correctness substantially reshapes measured robustness: many previously labeled jailbreak successes are reclassified as invalid, and results are stable across the tested search backends and evaluator models. Code and data are available at https://github.com/Ardor-Wu/SEAV. --- ### Beyond Token Positions: Safety Alignment Across Denoising Steps in Diffusion Language Models Authors: Guoli Wang, Haonan Shi, Tu Ouyang, An Wang Published: 2026-08-31 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.00495v1 Diffusion large language models (dLLMs) generate text through iterative denoising rather than left-to-right decoding. This generation paradigm introduces two axes that can influence safety alignment: when tokens are generated during denoising and where they appear in the response. In this paper, we measure dLLM safety behavior under harmful prompts by tracing intermediate token distributions and commitment decisions throughout denoising. Our analysis shows that refusal signals are concentrated in early denoising steps and leading response positions, and the tokens committed early can strongly shape the final safety outcome. Our measurements further show that the denoising step and persistence of refusal-token commitment are important for understanding dLLM safety. Based on these findings, we propose Refusal-Aware Early Commitment (RAEC), a simple training-free decoding method that commits persistent refusal signals from early steps. Experiments on LLaDA and Dream show that RAEC reduces attack success rates while largely preserving utility. The code is available at https://github.com/Glresearch1/RAEC. --- ### Human-Anchored Factuality Evaluation with Strategic Annotation Authors: Yu Wang, Craig Erickson, Kevin Small Published: 2026-08-31 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.00494v1 LLM-based factuality judges provide scalable evaluation signals, but their metrics are often systematically biased relative to human judgments. We study human-anchored factuality evaluation under limited annotation budgets, where judge predictions on the full dataset are combined with human labels on a small selectively sampled subset to obtain statistically valid estimates. The efficiency of this approach depends critically on which examples receive human annotation: in factuality evaluation, judge-human misalignment is not driven solely by low confidence, but also by structured failure modes such as incomplete evidence, temporal mismatch, unverifiable claims, and rubric misalignment. To exploit this structure, we introduce a factuality-specific annotation policy design pipeline that uses failure-space analysis (FSA) to derive diverse predictive signals for modeling human-judge misalignment. On an internal reference-based factuality evaluation system (AutoFA) and RAGTruth, where judge-predicted estimates substantially underestimate human-annotated factual accuracy, our FSA-guided policy improves annotation efficiency over uniform sampling and uncertainty-driven baselines, achieving effective-sample-size gains of 40.3% on AutoFA and 27.1% on RAGTruth. --- ### The Privacy-Hallucination Tradeoff in Differentially Private Language Models Authors: Krithika Ramesh, Krishna Pillutla, Danish Pruthi, Anjalie Field Published: 2026-08-31 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.00492v1 Both privacy and factual accuracy are paramount in high-stakes domains like healthcare. Concerningly, we uncover and investigate a privacy-hallucination tradeoff in differentially private (DP) language models. First, we empirically show that models pre-trained or fine-tuned with DP tend to produce more hallucinations than non-DP counterparts, with increased severity as the privacy budget grows stricter. Second, we investigate model properties driving this tradeoff, demonstrating that DP mechanisms flatten output distributions, potentially redistributing probability mass toward factually incorrect alternatives. Third, through experiments where we control fact frequency in training data, we characterize how information frequency can reduce hallucination risks in DP models. Overall, our findings underscore the need for more nuanced privacy-preserving interventions that offer rigorous privacy guarantees without compromising factual accuracy. --- ### MemeBridge: A Dataset for Benchmarking and Mitigating the Bidirectional Cultural Gap in Meme Interpretation Authors: Hangxiao Zhu, Suliu Qin, Zhuoyan Li, Ming Jiang, Yu Zhang, Meng Xia Published: 2026-08-31 Categories: cs.CL, cs.HC arXiv: https://arxiv.org/abs/2609.00491v1 Communicating across cultures is inherently challenging, especially through culturally dense and ambiguous formats like memes. While people expect large language models (LLMs) to hold promise for bridging such gaps, existing benchmark datasets often fail to capture the cultural context necessary for accurate interpretation. To address this, we introduce MemeBridge, a curated dataset centered on U.S.-originated memes, designed to capture two complementary perspectives: (1) how Chinese participants interpret these memes, and (2) how U.S. participants anticipate how people from other cultures might misunderstand them. Here, context refers to implicit cultural knowledge, including background beliefs, norms, and shared assumptions that shape meme comprehension. The dataset was constructed via a multi-stage crowdsourcing pipeline with rigorous validation, including human agreement checks and GPT-based classification verification. Each meme is annotated with sentiment, emotion, cultural significance, and knowledge type, providing rich supervision for downstream tasks. Notably, we observe that the anticipated misunderstandings from U.S. participants are often inaccurate, highlighting the asymmetries in cultural understanding and the challenges of adopting perspectives beyond one's own. This bidirectional framing, which focuses on both expression and perception, enables more nuanced benchmarking of cross-cultural comprehension. Our probing of multiple LLMs reveals that while models developed in different cultural contexts exhibit partial cross-cultural understanding, they often struggle with sophisticated interpretations. By contrast, fine-tuning with MemeBridge improves model performance, underscoring the value of culturally grounded resources for training and evaluating LLMs in globally diverse settings. --- ### A hybrid quantum-classical neural network for learning to route Authors: Marcus Rolf Peter Ritt, Alexsandro Santos da Rosa Júnior, Marcos Vinicius Reballo, Cesar Augusto do Amaral, Fernando Augusto Caletti de Barros Published: 2026-08-31 Categories: cs.LG, quant-ph arXiv: https://arxiv.org/abs/2609.00489v1 This work studies hybrid quantum-classical neural networks for learning routing heuristics. Specifically, this paper asks whether small quantum neural networks can replace parameter-heavy modules inside a competitive attention-based routing model while maintaining solution quality. For the capacitated vehicle routing problem, encoder feed-forward replacement emerges as the most promising design: it reduces the number of model parameters by 56.6% while keeping the hybrid model close to the classical neural baseline at small and medium instance sizes, although the gap grows for larger instances. This work also compares to classical routing algorithms, which remain highly competitive and often superior on the fixed Euclidean test sets. Our results therefore do not indicate quantum advantage or solver dominance, but identify encoder feed-forward replacement as a viable hybrid-module compression strategy for neural combinatorial optimization. --- ### AdaptNTK: Adaptive Uncertainty Quantification and Active Learning for Neural Network Potentials Authors: Prajwal Ananth, Shuwen Yue Published: 2026-08-31 Categories: cs.LG, physics.chem-ph arXiv: https://arxiv.org/abs/2609.00488v1 Machine learning interatomic potentials bridge the gap between quantum chemical precision and classical computational speed, enabling molecular dynamics simulations with first-principles accuracy. Their reliability is often improved through active learning, which iteratively expands the training set by identifying uncertain, out-of-distribution configurations. Existing uncertainty-quantification methods often involve a trade-off between computational cost and reliability, and generally cannot account for redundancy as an acquisition batch is assembled. Here, we introduce AdaptNTK, a single-model framework that measures uncertainty as a regularized Mahalanobis distance in empirical neural tangent kernel (NTK) feature space. With the NTK features fixed during acquisition, the uncertainty depends on the acquired configurations but not their reference labels. This allows the uncertainty to be updated recursively after each selection without retraining, reducing redundancy within an acquisition batch. On held-out rMD17 data, AdaptNTK achieves the highest mean correlations with force errors (Spearman 0.68, Pearson 0.71) and matches a three-member ensemble in error retention. In active learning experiments, AdaptNTK achieves the lowest force errors across rMD17 and Transition-1X, with particularly strong performance on transition-state configurations in Transition-1X. AdaptNTK provides a 2.6-fold speedup per Transition-1X cycle relative to the ensemble, providing efficient single-model uncertainty estimation with sequential updates for data-efficient active learning. --- ### EvoFlint: An Evolutionary Atlas of Multi-Turn LLM Vulnerabilities Authors: Feitong Qiao, Liren Peng, Shiming Ren, Aishwarya Jadhav, Arghavan Bahadorinejad, Marinette Chen, Muhan Zhang, Abdulaziz Suria, Gennevi Lu, Anish Das Sarma Published: 2026-08-31 Categories: cs.CL, cs.AI, cs.CR, cs.LG arXiv: https://arxiv.org/abs/2609.00487v1 Frontier language models that refuse harmful single-turn prompts often comply when the same intent is reached gradually over many turns, making multi-turn attacks one of the least understood failure modes of large language models. Most automated red-teaming methods treat this as a generation problem: produce attacks that break the model. We argue it is better framed as a search problem: discover, organize, and iteratively refine a diverse archive of attack strategies, producing a structured map of how a target model fails rather than a list of one-off successes. We introduce EvoFlint, which applies evolutionary quality-diversity search to multi-turn red-teaming. Attack strategies are phased conversation plans, not raw prompts, and are evolved through LLM-driven mutation and crossover. A Pareto fitness over attack success rate and peak severity preserves selection signal from near-miss attacks. A risk-indexed archive runs novelty search with local competition over strategy description embeddings inside each cell, maintaining diversity without committing to a predefined style taxonomy. A generation-level memory accumulates target-model insights across the population and feeds them back into strategy generation. On the HarmBench-test split, EvoFlint reaches attack success rates of 35.8% on Claude Sonnet 4.6, 59.7% on GPT-5.4, and 94.3% on Qwen3-32B, alongside 98.7% on the older GPT-4o included as a baseline reference. The resulting archive, organized by risk category, exposes for each target which categories of harm its safety training has and has not covered. --- ### Are Near-Tied LLM Rankings Robust to Family-DIF-Guided Benchmark Recomposition? Authors: Qiaoyuan Zheng, Yiqu Yang Published: 2026-08-31 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.00482v1 Small leaderboard gaps are often interpreted as evidence that one language model is better than another, but their sign may depend on which benchmark items are included. We test this using item-level responses from five benchmarks and a family-label-free spectral approximation to multidimensional item-response theory (MIRT). In owner-disjoint folds, one owner half identifies items with low residual differential item functioning across model families (low-DIF); the resulting frozen, source- and easiness-balanced weights score models in the other half, while equally short matched-random subtests control for generic subtest variation. Full-benchmark and low-DIF rankings remain strongly correlated ($τ_b=.900$--$.948$). Yet in four of five benchmarks, 30.9--47.1\% of cross-family pairs initially within one percentage point reverse order, exceeding their matched-random medians by 16.9--28.6 percentage points (all $p=.001$). The fifth benchmark shows no reliable excess ($-0.9$ points, $p=.689$). The pattern survives all pre-specified population perturbations, and residual item--family signatures replicate across owner halves; however, no family shows a consistent advantage across benchmarks. Thus, globally stable rankings can still leave individual near-tie orderings sensitive to benchmark composition, and sub-one-point leaderboard gaps should be accompanied by evidence that the implied ordering is composition-robust. ---