# Research Papers — 2026-09 · bundle page 221 of 266 > Full-text markdown of papers #11001–11050 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/221 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/220.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/222.md) ### PCSDiff: Diffusion-Based Bias Correction and Super Resolution Toward Practical Operational Medium-Term Precipitation Forecast Authors: Yuze Sun, Shiyi Wang, Jiancheng Pan, Die Wang, Andreas F. Prein, Wentao Luo, Linhan Jiang, Jie Wu, Quan Zhang, Xiaomeng Huang Published: 2026-09-07 Categories: cs.LG, cs.AI, physics.ao-ph arXiv: https://arxiv.org/abs/2609.06942v1 Medium-range precipitation forecasts are impaired by persistent systematic biases, lead-time-dependent error accumulation, and coarse spatial resolution, restricting their reliability for flood-drought risk assessment. Existing AI correction techniques lack dedicated modeling for multi-day dynamic bias evolution and proper meteorological constraints, often generating over-smoothed rainfall structures, and cannot meet operational deployment demands. This work introduces PCSDiff, a cascaded task-decoupled diffusion framework targeting 10-day precipitation bias correction and downscaling. To jointly counteract temporal error drifts and reconstruct physically plausible local precipitation details, PCSDiff integrates the Precipitation Intensity-aware Multi-branch Decoder (PIMD) module for dynamic multi-day error mitigation using synoptic-temporal features, followed by a two-phase conditional diffusion super-resolution module to restore fine-scale precipitation patterns. Evaluated against CMA-CRA observations over China after global-data training, PCSDiff cuts RMSE by 16.1% and lifts ACC by 13.9% relative to raw ECMWF forecasts at 3-10-day lead times, and consistently outperforms mainstream deep-learning baselines on both general and extreme-precipitation metrics. Benefiting from a streaming inference pipeline, our method achieves low-latency rolling forecasting for practical meteorological operations. --- ### When and Why LLM Causal Priors Help: Closed-Loop Prior Selection for Amortized Causal Inference Authors: Haohao Zhou Published: 2026-09-07 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.06941v1 Causal effect estimation asks how an outcome would change under an intervention, and medicine, economics, and public policy all treat it as a foundational task. Prior-data fitted networks (PFNs) amortize the task: a model trained on large numbers of programmatically generated synthetic causal tasks reads a new problem's observational data into context and returns an interventional-effect estimate in a single forward pass. The capability of such models is largely determined by the synthetic training prior, which is currently designed by hand, a bottleneck acknowledged by both Do-PFN and CausalPFN. Large language models (LLMs) can now ``draw'' plausible causal graphs for a given domain, suggesting that LLM-distilled graphs could serve as prior material. Whether injecting such graphs helps at all, where any gain comes from, and when injection helps. Practice has so far relied on manual trial and error. We propose a \emph{closed-loop prior selection framework} that casts prior injection as a budget-constrained optimization over a candidate prior pool. Candidates undergo cheap post-training and are scored by a composite metric dominated by real-domain generalization; the winner then receives full training and paired statistical validation. On a 7.34M-parameter Do-PFN, the framework's winner attains a formally significant $2.75\times$ gain on the primary evaluation domain, and its error falls below that of the uninjected official base. Generalization on an adjacent monitoring domain improves significantly, and no monitored capability degrades. Mechanism experiments show that the gain depends on the semantic content of the distilled graph rather than its structural diversity alone does not produce it (directional evidence). With this framework and this regularity in hand, the use of LLM causal priors stops being manual trial and error and becomes an empirically verifiable selection problem. --- ### The Geometry of Refusal: Why Post-Hoc Safety Is Fragile and Pretraining-Time Safety Persists Authors: Srikanth Malla, Chiho Choi, Joon Hee Choi Published: 2026-09-07 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.06934v2 Post-hoc safety training (RLHF, DPO) is the dominant way to align large language models, yet jailbreaks (Zou et al., 2023), fine-tuning attacks (Qi et al., 2024), and activation-space edits (Arditi et al., 2024) keep recovering the behaviors it was meant to remove. We give this fragility one geometric explanation and follow it into pretraining. We measure the safety update $Δ= W_{safe} - W_{base}$ against the curvature of the model's capabilities (the empirical Fisher of a capability loss). Across five model families, post-hoc safety lands in a suppression regime: $Δ$ is nearly orthogonal to the capability directions, and its small in-subspace part concentrates on a few high-curvature ones. The update is thin but sharp, a refusal gate laid over intact capabilities rather than erasure of them. A kernel-immobility lemma explains why such an update can only mask a capability, not remove it, so a little benign fine-tuning restores it: 100 benign examples cut the AdvBench refusal of Qwen-2.5-7B-Instruct and Llama-3-8B-Instruct by 35 to 38 pp. Following the account into pretraining, a pretraining-checkpoint sweep of OLMo-2-1B (Team OLMo et al., 2024) shows the features that refusal attaches to emerging in a sharp transition between 1B and 63B pretraining tokens. We then use the account constructively: models trained from scratch with safety co-training spread continuously across pretraining reach 87 to 98% AdvBench refusal that the same attack erodes by only 2 to 14 pp at every scale from 410M to 6.9B, against 35 to 38 pp for post-hoc installs, at a small cost on short-answer capability probes; a windowed schedule of equal total safety weight installs no refusal. Persistence of the safety signal across pretraining, not its timing, is what buys attack robustness. --- ### CARDEA: Auditable Reasoning Grounded in Spatial Evidence for End-to-End Coronary Angiography Interpretation Authors: Jia-Jen Lee, Shih-Yen Hou, Kee Koon Ng, Wei-Chun Wang, Shih-Sheng Chang Published: 2026-09-07 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.06931v1 Invasive coronary angiography (CAG) is the gold standard for diagnosing coronary artery disease, but interpretation varies substantially among observers. Existing AI systems can improve consistency but lack auditable decision processes and are limited in comprehensive open-ended assessment, undermining clinician trust and clinical adoption readiness. We developed CARDEA, a unified large vision-language model that serves as the inference core of a CAG pipeline. It was trained solely on public datasets and closed-ended tasks in three stages: visual feature alignment, a self-distilled Chain-of-Box (CoB) cold start, and reinforcement learning with verifiable rewards (RLVR) with a CoB reward encouraging bounding-box use in the reasoning trace. We assessed its two study-level diagnoses, dominance classification and complexity assessment, against a dedicated classifier and two interventional cardiologists. Report generation was excluded from training and evaluated zero-shot across stages on an external cohort using vessel-severity macro-$F_1$. CARDEA trailed the classifier on in-distribution dominance but drew level under domain shift (accuracy, 0.91 (95% confidence interval (CI), 0.86 to 0.95)) and was comparable to the cardiologists on complexity assessment (accuracy, 0.90 (CI, 0.82 to 0.97)). Only RLVR improved zero-shot report generation, raising its vessel-severity macro-$F_1$ (0.686 (CI, 0.664 to 0.707)) above the untuned base model (0.513) and over twice the always-normal floor (0.312). CARDEA runs an end-to-end CAG pipeline from raw multi-view videos through keyframe selection to study-level diagnosis while exposing auditable spatial evidence behind its conclusions. RLVR on verifiable closed-ended tasks surfaced open-ended reporting ability that supervised imitation did not. Clinical use requires prospective validation against expert cardiologists. --- ### Distributed Dexterous Manipulation with Spatially Conditioned Multi-Agent Transformers Authors: Sarvesh Patil Published: 2026-09-07 Categories: cs.RO, cs.LG arXiv: https://arxiv.org/abs/2609.06930v1 Distributed Dexterous Manipulation (DDM) is a novel paradigm that presents significant control challenges due to high action-space redundancy, inter-robot cooperation, and dynamic object-robot interactions. This paper introduces a framework based on spatially conditioned Multi-Agent Transformers (MATs) to efficiently learn robust control policies for a DDM system grounded in an array of 64 soft delta robots arranged in an 8x8 grid. Our three core contributions are: (i) an MAT with adaptive layer norm for compute efficiency, (ii) spatial contrastive embeddings to ground transformer embeddings in the spatial configuration of the robots, and (iii) an MAT-based behavior cloning method fine-tuned using Soft Actor Critic. We also propose an action selection formulation to analyze the trade-off between task performance and the number of robots utilized. Our experiments show that MATs iteratively refine their actions through the stacked attention blocks. This further informs the benefit of spatial conditioning in transformers to learn DDM policies. We demonstrate long-horizon planar manipulation tasks with objects of various geometries in simulation and real-world. Finally, we show how action selection mitigates robot maintenance by reducing wear and tear due to inter-robot collisions while maintaining the ability to manipulate objects along various trajectories in the real-world, achieving an average error of ~1.5 cm, while using ~65% fewer robots. --- ### Sub-Pixel Affine Registration of Space Debris Images via the Radon Point Spread Function Authors: Shenshen Luan, Miaomiao Tian, Shuai Jiang, Yan Yang, Shuguo Xie, Zezhou Sun Published: 2026-09-07 Categories: cs.CV, cs.GR arXiv: https://arxiv.org/abs/2609.06929v1 Inter-frame affine misalignment caused by platform jitter and attitude adjustments poses a fundamental challenge for multi-frame analysis of point targets in optical surveillance. Conventional registration methods rely on spatial intensity correlations or distinctive image features, both of which are largely absent in low-signal-to-noise-ratio point target imagery. We introduce the Radon Point Spread Function (RPSF) to characterize point targets in the Radon-transformed domain, and derive a closed-form framework that jointly estimates inter-frame translation and rotation from as few as four scalar RPSF samples per frame pair. The method requires no iterative optimization, feature extraction or interpolation, which is suitable for resource-constrained onboard processing. Simulation results confirm sub-pixel translation accuracy and a mean rotation error of 0.2556° at 1° Radon angular resolution. Validation on five real space debris datasets including both ground-based and in-orbit observations yields a mean calibration error below 0.5 pixels, substantially exceeding the precision required for reliable multi-frame processing. --- ### Emo-DVS: A Multimodal Benchmark for Privacy-Aware Emotion Recognition with Event Cameras Authors: Jiaqi Chen, Qinfu Xu, Hao Zhuang, Liyuan Pan Published: 2026-09-07 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.06928v1 Emotion analysis is a fundamental task in computer vision, but its practical deployment remains constrained by the privacy risks inherent to conventional RGB cameras. Bio-inspired event cameras present a promising hardware-level solution because they capture asynchronous brightness changes, thereby reducing exposure of facial identity details while leveraging high dynamic range for robust perception under challenging illumination conditions. Despite these advantages, existing event-based methods struggle in complex real-world settings due to limited dataset scales, simple acquisition conditions, and reliance on single-modality visual cues. To address these, we establish a challenging tri-modal benchmark with event, audio, and text modalities and propose the Information-Guided Gated Fusion (IGF) framework, which first pre-trains an event encoder on the FAU subset of Emo-DVS to capture fine-grained facial dynamics, then employs adaptive modality gating to suppress modality-specific noise, and finally leverages mutual information maximization to align robust cross-modal representations. To alleviate data scarcity, we introduce Emo-DVS, the first large-scale event-based emotion analysis dataset, which couples dynamic illumination with the Facial Action Unit (FAU) subset and emotion subset. Extensive experiments demonstrate that IGF achieves state-of-the-art performance. --- ### BEFORE THE FLIP: Measuring Hidden Score Shifts In Quantized Vision Language Models Before The Answer Changes for Visual Question Answering Authors: Sourajit Saha, Shubhashis Roy Dipta, Shaswati Saha, Nobin Sarwar, Yuxuan Jiang Published: 2026-09-07 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.06922v1 Quantization makes vision language models (VLMs) cheaper to store and run by using fewer bits to represent their weights. While unchanged answers on visual question answering (VQA) after compression are an expected behavior, they can still hide changes in the underlying scores (log probabilities). For example, a model may still answer yes after compression, even as the score gap between yes and no shrinks. We introduce BEFORE THE FLIP to measure these hidden changes. Our method compares the score change caused by compression with the change caused by replacing the image's internal representations, or image tokens, with one fixed average token. We then increase the precision of one weight group at a time to identify where extra bits help, and test whether choosing different groups for each question offers benefits beyond shuffled controls. Among 8,277 LLaVA questions where image token replacement measurably affects the scores, 4-bit compression shifts the yes or no score gap farther toward the replacement output than 8-bit compression. Qwen shows the same pattern, but with a smaller difference. Yet only 265 of 9,000 LLaVA answers change at 4 bits. In a separate study of 1,024 calibration questions, choosing weight groups separately for each question does not outperform both shuffled controls at any tested storage budget. These findings show that compression can alter the scores behind unchanged answers, but do not establish a reliable benefit from adjusting precision for each question. --- ### Constrained Online Learning with Noisy Constraint Values Authors: Vaneet Aggarwal Published: 2026-09-07 Categories: cs.LG, cs.AI, math.OC arXiv: https://arxiv.org/abs/2609.06921v2 We study constrained online convex optimization with adversarial constraints and conditionally unbiased, finite-variance observations of constraint values and gradients. Under common feasibility, our \LEDGER\ algorithm attains $O(\sqrt T)$ expected regret and $O(\sqrt{T\log(eT)})$ expected budget violation, the largest cumulative overspend over any window. It uses a reflected exponential potential, clipped signed observations, and predictable adaptive regularization, with one feedback triple and one projection per round. Neither a Slater condition, independence between feedback channels, nor an absolute constraint-value bound is needed. A Gaussian testing lower bound proves that the budget rate has optimal horizon dependence under square-root regret at fixed positive noise, including the logarithm. The same obstruction holds for terminal violation, so the logarithm is not a cost of maximizing over windows; an $O(\sqrt T)$ budget bound instead forces linear regret. In contrast, fixed positive Gaussian value noise yields a joint regret--hard-violation lower bound of $Ω(\min\{σ,1\}T/\log^2 T)$, even with exact gradients in one dimension. The hard-violation construction matches arbitrarily many moments while preserving a feasible-endpoint gap and constant endpoint probabilities. Together, the bounds separate uncertainty about hard feasibility from learnable signed budgets. Deterministic restarts remove the horizon input without changing either upper rate. --- ### A visual large language foundational model for medical image recognition using clinician-contributed online resources Authors: Lingxuan Hou, Yuhua Xie, Yue Hu, Yan Zhuang, Junqi Li, Chengzhi Xia, Binh Phu Nguyen, Abubakar Siddique, Minh Nguyen, Yao Hou, Yanju Bao, Kexin Liu et al. Published: 2026-09-07 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.06914v4 Large language models (LLMs) have demonstrated strong capabilities across diverse domains, showing considerable potential in medicine. However, their application in medical settings remains limited by the scarcity of visual question answering (VQA) datasets that capture clinical reasoning and explicit image-text alignment. Here, we leverage de-identified medical images and expert commentaries shared through clinician-oriented online resources. By combining an advanced LLM with clinician-in-the-loop verification, we established a rigorous pipeline to construct ThoughtMed-1M, a long-form medical VQA dataset containing over one million VQA pairs and designed to capture structured clinical reasoning and medical image-text alignment. To demonstrate its utility, we developed a FOundational LLM Trained on ThoughtMed-1M (FOLTMed). FOLTMed achieved state-of-the-art performance across 42 medical VQA benchmark datasets, with a macro accuracy of 85.4 percent. It also generated more clinically coherent responses on the ThoughtMed-1M test set, outperforming state-of-the-art models by 3 to 5 percent across factuality and similarity metrics, highlighting a scalable paradigm for advancing research on clinically grounded multimodal LLMs. --- ### From Synthetic Priors to Model Behavior: Structural Coverage in Tabular Foundation Models Authors: He Zhao, Ryan Thompson, Daniel M. Steinberg, Ashfaqur Rahman, Edwin V. Bonilla, Cheng Soon Ong Published: 2026-09-07 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.06912v1 Tabular foundation models (TFMs) are commonly pretrained on large collections of procedurally generated synthetic tasks, yet it remains unclear how well these synthetic pretraining priors support the downstream tasks on which the models are evaluated. We study this question from a distribution-level attribution perspective. We recover or reconstruct the synthetic data generators of four TFMs and compare their generated tasks with datasets from two widely used tabular benchmarks. Each dataset is represented by a common set of structural descriptors capturing schema, feature distributions, dependence structure, response properties, and feature--response relationships. In this space, we measure how broadly and repeatedly each synthetic prior reaches benchmark tasks using structural coverage and normalized density, and examine whether stronger local support is associated with better predictive performance. We find substantial differences across synthetic pretraining priors: some generators provide consistently broader and denser support for benchmark tasks than others. Moreover, stronger synthetic-to-benchmark support is generally associated with better relative model performance. These results suggest that structural coverage provides a useful diagnostic for characterizing synthetic pretraining priors and relating their data-generating assumptions to downstream model behavior. --- ### Smoothed Picard Hamiltonian Monte Carlo Authors: Fan Chen, Sinho Chewi, Jianfeng Lu, Matthew S Zhang Published: 2026-09-07 Categories: math.ST, cs.DS, cs.LG, math.NA arXiv: https://arxiv.org/abs/2609.06906v2 We develop a new low-accuracy sampler, called smoothed Picard Hamiltonian Monte Carlo, which combines Gaussian smoothing, Picard iteration, and higher-order discretization. For a log-concave target $π\propto \exp(-V)$ in dimension $d$ satisfying $0 \prec αI \preceq \nabla^2 V \preceq βI$, with condition number $κ:= β/α$, smoothed Picard HMC returns a sample with $\sqrt α\,W_2(\cdot,π) \le \varepsilon$ using $\widetilde O(κ^2 + κ^{7/6} d^{1/6}/\varepsilon^{1/3})$ gradient queries. We also prove stronger $W_q$ bounds, and then develop an algorithmic framework, the recursive warm start generator, to upgrade these $W_q$ bounds to stronger divergence guarantees. This produces a warm start for the proximal bouncy particle sampler, introduced in a companion work, leading to a high-accuracy log-concave sampler with complexity $\widetilde O((κ^{7/6} d^{1/6} + κ^{1/2} d^{1/4})\mathrm{polylog}(1/\varepsilon))$. --- ### Accelerated High-Accuracy Sampling from a Warm Start via the Proximal Bouncy Particle Sampler Authors: Fan Chen, Sinho Chewi, Jianfeng Lu, Matthew S Zhang Published: 2026-09-07 Categories: math.ST, cs.DS, cs.LG, math.PR, stat.ML arXiv: https://arxiv.org/abs/2609.06905v1 We study the problem of sampling from $μ(\mathrm{d}x)\propto e^{-V(x)}\,\mathrm{d}x$ on $\mathbb{R}^d$, where $V$ is $α$-strongly convex and $β$-smooth, and write $κ:=β/α$. We design and analyze the Proximal Bouncy Particle Sampler (Proximal BPS), a new sampler that combines ideas from the proximal sampler and the bouncy particle sampler. From a warm start initialization with $ O(1) $ Rényi divergence w.r.t. $μ$, Proximal BPS returns a sample whose law is $\varepsilon$-close to $μ$ in total variation distance using $\widetilde O(\sqrtκ\,d^{1/4} \,\mathrm{polylog}(1/\varepsilon))$ gradient queries in expectation. --- ### Dynamic-Programming-Guided Hierarchical BPE and Empirical Analysis of Vocabulary Pruning Authors: Kenny Shao Published: 2026-09-07 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.06898v1 Byte Pair Encoding (BPE) constructs vocabularies through greedy pair merging, but the resulting merge order does not necessarily allocate a fixed model-visible vocabulary optimally for compression. We propose Dynamic-Programming-Guided Hierarchical BPE (DH-BPE), a vocabulary-construction method that combines token exposure under exact minimum-token segmentation with the hierarchical dependencies induced by BPE training. Starting from a modestly overshot BPE candidate vocabulary, DH-BPE uses dynamic programming to measure candidate utility and applies exposure-guided, dependency-aware pruning to select a fixed-size model-visible vocabulary. We compare DH-BPE against Standard BPE and recent vocabulary-optimization baselines, including Pruned BPE, MinGram, and MinGram-PP, in primary evaluations at 12K and 16K target vocabulary sizes, with an additional 18K evaluation against MinGram only. Across the primary 12K and 16K comparisons, DH-BPE consistently improves aggregate compression over Standard BPE, Pruned BPE, and MinGram under a shared exact minimum-token DP encoder. MinGram-PP achieves stronger aggregate compression in the primary comparisons, but DH-BPE outperforms it at overshoot factors f = 2.0 and f = 3.0 in cross-corpus evaluation; at 12K, MinGram-PP reverses this ordering only with the substantially larger candidate pools at f = 4.0 and f = 5.0. Qualitative analysis further shows that DH-BPE balances later, more complete BPE merges with reusable subword components, providing a practical approach to improving vocabulary allocation under a fixed model-visible vocabulary budget. --- ### Towards Bridging the Gap Between Offline and Iterative Alignment via Preference Distillation Authors: Wenbo Zhang, Wenzhuo Zhou, Hengrui Cai, Zhengling Qi Published: 2026-09-07 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.06893v1 Direct preference optimization DPO is a promising offline approach for aligning large language models (LLMs) due to its simplicity, computational efficiency, and implicit modeling of human preferences. Interestingly, iterative extensions of DPO have achieved stronger performance on academic benchmarks, raising two key questions: (i) Why do iterative methods generally outperform offline ones? (ii) Can their advantages be incorporated into offline alignment? To answer the first question, our controlled experiments reveal that the explicit preference model, additionally introduced in the iterative procedure, is a key factor behind its superiority over offline methods. This insight leads us to answer the second question affirmatively and propose Distilled Preference Probability Policy Optimization (DP3O), an effective and efficient offline alignment algorithm. DP3O first learns an explicit preference model using a helper class of LLMs and then distills its knowledge into policy optimization. Theoretically, we show that explicit preference modeling admits better estimation error control than implicit formulations, and that DP3O achieves a tighter generalization bound than hard-label DPO through variance reduction. Empirically, we evaluate DP3O on a wide range of chat-based and downstream tasks and show that it outperforms state-of-the-art offline methods, achieves performance comparable to iterative DPO, and reduces training time by about $42\%$, demonstrating both its effectiveness and efficiency. --- ### Human-agent discovery of reconfigurable in-plane ferroelectric superdomain control Authors: Yu Liu, Boris Slautin, Ching-Che Lin, Jaegyu Kim, Lane W. Martin, Sergei V. Kalinin Published: 2026-09-07 Categories: cond-mat.mtrl-sci, cs.AI arXiv: https://arxiv.org/abs/2609.06887v1 Automated experimentation is most effective when the observables, available actions, and objective are defined before the experiment starts, as is the case for Bayesian optimization. However, in many exploratory experiments, the variables that describe the sample must be extracted from the data, new operations emerge during the experiments, and the instrument budget is too small to learn the problem by trials. Here we introduce the Scanning Probe Agentic Research Cycle (SPARC) framework, in which a coding agent and a human operator share one microscope, one notebook, and two persistent memory files. FINDINGS.md stores graded conclusions about the experiment, whereas PITFALLS.md records learned failure modes of analysis and instrument. We apply SPARC to reconfigure the in-plane superdomain direction of a (111)-oriented PbZr0.2Ti0.8O3 film. In an operator-supervised campaign, the agent reanalyzed earlier manual measurements and developed an oriented lattice of stationary bias pulses with alternating polarity to reconfigure the superdomain direction. In a subsequent agent-controlled campaign, PITFALLS.md entries were compiled into checks that validate a design before any write. The experiments showed that spatial polarity alternation, instead of the exact matching between the lattice and lamellar periods, determines directional selection. Combining a raster scan with a masked pulse lattice printed the letters UTK into the superdomain orientation. The campaign also identified practical requirements for agentic experimentation where physical verification of instrument execution, the conditions under which stored findings remain valid, validation of new observables on instrument data, and robust control protocols. --- ### Noisy-Space Policy Gradient for Diffusion Policies in Offline Reinforcement Learning Authors: Mahmoud Selim, Cristina Cipriani, Karl H. Johansson Published: 2026-09-07 Categories: cs.LG, cs.AI, cs.RO arXiv: https://arxiv.org/abs/2609.06882v1 Diffusion policies offer a powerful and expressive parameterization for continuous control. Yet, their integration with reinforcement learning remains conceptually and algorithmically challenging. In this work, we address this gap by introducing a noisy-space action-value (Q-)function that assigns values to diffusion latents through the distribution of executed actions induced by the denoising process. We show that this construction admits a precise semantic interpretation and derive a noisy-space policy gradient (NSPG) that optimizes noisy latents using only clean action-space value estimates. Building on this result, we formulate a KL-regularized policy improvement over noisy latents and show that the resulting objective admits a diffusion-compatible regression form, avoiding backpropagation through the denoising process. Empirical results on state-based D4RL benchmarks and vision-based OGBench tasks demonstrate that the proposed noisy-space objective provides a principled and effective basis for training diffusion policies in offline reinforcement learning. Project webpage: https://mahmoud-selim.github.io/NSPG/ --- ### Learning Adaptive SED for heterogeneous load balancing Authors: Sanne van Kempen, Jaron Sanders, Fiona Sloothaak, Maarten G. Wolf Published: 2026-09-07 Categories: cs.LG, math.PR arXiv: https://arxiv.org/abs/2609.06881v1 We study a two-server load balancing system with heterogeneous service rates that are a priori unknown to the dispatcher. The goal is to route customers according to the Shortest--Expected--Delay (SED) policy, but this requires knowledge of the service rates. Empirical policies that route based on estimates perform poorly: due to estimation error, the empirical policy disagrees with the oracle on an infinite region of the state space. We propose an online learning algorithm that converges to SED while learning the service rates. The algorithm carefully balances empirical SED routing with forced exploration phases that guarantee sufficient sampling of both servers. We prove that our algorithm achieves finite regret; this differs from classical Multi-Armed Bandit settings where regret typically grows logarithmically in time. Finally, numerical experiments demonstrate the performance of our algorithm and highlight the regimes in which forced exploration is especially beneficial. --- ### Contextual Observer Grounding: Evaluating Situated Spatial Reasoning in Vision-Language Models Authors: Mimo Shirasaka, Haochen Zhang, Yonatan Bisk Published: 2026-09-07 Categories: cs.CV, cs.CL, cs.LG, cs.RO arXiv: https://arxiv.org/abs/2609.06880v1 Reasoning over language instructions in embodied tasks such as robotics often requires understanding spatial relations from a speaker's situated perspective. Humans infer such perspectives from shared environmental knowledge, activity context, and commonsense. Recent vision-language models (VLMs) appear capable of spatial reasoning, but their ability to infer a speaker's viewpoint from contextual cues and interpret situated spatial relations from that viewpoint remains unclear. We call this capability contextual observer grounding. To study this capability, we construct the Point-of-View Benchmark (POVBench), a dataset of 3D scenes and queries that disentangles Inferred, Stated, and Given forms of observer grounding in natural embodied communication. Given multi-view observations and a natural-language sentence, models must localize unseen or underspecified targets from situated spatial and contextual cues. Across state-of-the-art VLMs, localizing targets from directional language remains challenging, even when observer grounding is made explicit. We find that explicit breakdowns of observer-relative spatial reasoning improve target localization. Our project page is available at https://mimo-owl.github.io/POVBench/. --- ### Self-Cleaning and Captured Anyway: One Measured Primitive for Error in a Store an Agent Writes to Itself, and What a Falling Score Actually Measures Authors: Wenhui Chen, Jianlin Chen, Ziyao Lin, Chi Man Vong Published: 2026-09-06 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.25052v1 "An agent that writes its conclusions into a store it later retrieves from closes a loop usually reported as one-way contamination. Taking the loop to the infinite-tenure limit against an append-only store gives a different picture: because writing never deletes, the reachable state space has a hard upper edge at (n-1)/n, so the outcome is a choice between two edges rather than a decay. At f_0 = 0.9 the interval between the two modes holds 3.6% of 220 runs where a uniform spread would put 20.6%, and is strictly empty on the first 15; the pooled mean describes 8.2% of the runs it summarises, the median 68.2%. Everything the model contributes is carried by one measured primitive with no fitted parameter, the copy function γ(φ): on 36 Wikidata facts, sign(\hatγ - γ_{crit}), with γ_{crit} = 1/k at r = 0, w = 1, predicts the direction of drift on 353 of 360 real-fact runs (39 of 40 synthetic in the same batch). Scale does not rescue the store: pooled frontier capture is 0.850, with claude-sonnet-4.5 captured on 20 of 20 seeds against our registered prediction of <0.5. What the interval tests is distinguishability rather than count: on the real facts, multi-valued runs have 6.4x its occupancy of the rest. It survives at f_0 in {0.1, 0.3, 0.5}, capture peaks at f_0 = 0.5, and of four interventions with criteria frozen first, timing dominates fraction at matched budget while a consistency gate drives every model to 0.993. The resampling unit is the seed, at a design effect of 3.75 on a pooled level: under a 44-seed control the ordering supporting claim 4 collapses from Spearman +0.98 at three seeds to +0.31-0.80 at forty-four, while claim 2's ordering is exact there (+1.00, p = 0.017). All 87 graded rows are in Appendix W, 37 of them graded withdrawn, failed, self-correcting, undecidable or an acknowledged limit, against 50 that are not." --- ### LLM-Driven Training-free Location-Attribute Synergic Fusion: A Closed-Loop Paradigm for Dual-source Encrypted POIs and LULC Mapping Authors: Chang Li, Xingtao Peng, Yongjun Zhang, Yinfei He, Cairun Huang Published: 2026-09-06 Categories: cs.CL, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.25051v1 Dual-source encrypted points of interest (DSEP), POIs from two encrypted coordinate systems, suffer from intertwined location and attribute uncertainties, including nonlinear systematic misalignment and naming inconsistency, hindering land-use/land-cover (LULC) mapping. To the best of our knowledge, this paper is the first to propose an LLM-driven, training-free location-attribute synergic closed-loop optimization paradigm for DSEP fusion. The paradigm jointly refines location transformation and attribute correspondences through iterative feedback. Attribute-synergic location fusion uses an LLM-driven attribute matching method to establish DSEP correspondences, reducing matching complexity from O(N^2) to O(N), and refines transformation coefficients using an improved particle swarm optimization algorithm within ISODATA-clustered local subregions. Location-synergic attribute fusion then reassesses attribute confidence from updated geometric residuals through an LLM-fuzzy method. The refined correspondences feed back into location optimization, forming a bidirectional closed loop. Sample purification and adaptive radius contraction enable convergence in essentially two iterations. We further propose a training-free LULC mapping method that inherits land-use classes from encrypted maps through location fusion, producing vector-raster integrated LULC maps. A reference-free POI fusion evaluation method is applied across 31 provincial capitals and municipalities in mainland China. Experiments show that our method achieves an average DSEP location fusion residual of 4.58 m and attribute fusion accuracy of 95.12%, improving upon the open-source baseline and state-of-the-art method by 1.77 m and 14.87%, respectively. Overall, the method provides a training-free solution for DSEP fusion and enables georeferencing of encrypted vector data to WGS-84 without field-surveyed ground control points. --- ### FrontierMath Erdős Authors: Tom Adamczewski, Thomas F. Bloom Published: 2026-09-06 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.25050v1 We introduce FrontierMath Erdős (FME), a benchmark of 68 Erdős problems that are open as of August 2026. To solve a task in FME, AI systems must resolve (prove or disprove) one of the 68 conjectures in the proof assistant Lean. Our 68 problems were selected by the second author among 652 open problems on erdosproblems.com for their mathematical interest and difficulty. AIs have recently resolved several open problems in mathematics, but these demonstrations fall short of a systematic study of AI capabilities. FME evaluates every AI model on the same fixed problems, autonomously and under the same budget. We evaluated five AIs with a budget of \$300 per problem. One (GPT-6 Astra) scored 3%, and all others scored 0%. --- ### Mitigating LLM Over-Refusal via Dynamic Semantic Routing Calibratione Authors: Zixuan Wang, Bingjie Zhang, He Zhao, Dandan Guo Published: 2026-09-06 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.25049v1 Large language models (LLMs) aligned for safety often suffer from over-refusal, incorrectly rejecting benign yet safety-related instructions. Prior studies primarily attribute this to static representation overlap, largely overlooking the underlying dynamic mechanisms. In this paper, we present the mechanistic analysis of over-refusal through the lens of internal routing conflicts within transformer attention. We discover that a sparse subset of Hypersensitive Safety Heads misfires on Hard-Safe prompts, exhibiting abnormal attention entanglement that forcefully binds harmless target entities to refusal semantics. This triggers a severe, high-entropy routing conflict that deprives target entities of necessary attention. To counteract this, we propose Semantic Routing Calibration (SRC), a lightweight, training-free inference framework. SRC precisely localizes and dynamically suppresses these hypersensitive safety heads at the inference stage. Coupled with a dual-branch logits fusion that acts as a safety regularizer during subsequent decoding, SRC seamlessly restores trustworthy reasoning. Extensive experiments demonstrate that SRC alleviates over-refusal, with intrinsic safety performance preserved as much as feasible. --- ### Teaching a Moving Student: Rethinking the Curriculum of On-Policy Distillation Authors: Lingxiang Hu, Tianle Xia, Yiding Sun, Ming Xu, Linfang Shang, Lan Xu, Ning Zheng, Wei Xu, Jie Jiang Published: 2026-09-06 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.25048v2 In on-policy distillation (OPD), the student determines which states receive teacher supervision. As its policy evolves, earlier response prefixes become less likely even though teacher-student disagreement on them persists. Under matched trajectory and optimization budgets, neither more queries nor more frequent rollout resampling is uniformly beneficial. Current-policy rollouts outperform initial-policy rollouts at shorter response budgets, but this ranking reverses at longer budgets. Replaying initial-policy rollouts after current-policy training improves accuracy, whereas replaying fixed recent rollouts does not reproduce the gain. We propose R-OPD, a gradient-triggered curriculum that adaptively selects when to revisit initial-student trajectories. When changes in mean gradients fall within minibatch-level variation for two consecutive comparisons, training switches from the next iteration onward to initial-policy replay. Across eight mathematics benchmarks and three training-data orders, R-OPD improves average accuracy over continued current-policy sampling by 2.25/4.44 percentage points at 16K/32K for a 0.6B student and 2.81/6.15 points for a 1.7B student. With a 30B-A3B teacher, it improves 8B accuracy by 4.05 points at 32K. At 32K, R-OPD exceeds a fixed schedule of 40 current-policy updates followed by 20 replay updates by 1.39/1.66/1.85 points for 0.6B/1.7B/8B. At the same generation cap, R-OPD also produces longer responses, suggesting that well-timed revisits help students use more of their reasoning capacity. --- ### AIBuildAI-2.5: Efficient Autonomous AI Model Development Through LLM-Guided Tree Search Authors: Peijia Qin, Ruiyi Zhang, Qi Cao, Han Guo, Li Zhang, Pengtao Xie Published: 2026-09-06 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.25047v1 Autonomous agents that automatically build artificial intelligence (AI) models could broaden access to AI across science and engineering. A popular line of such agents frames model building as a code search problem and solves it by tree search, in which each node is a candidate program and the tree grows by generating a child program from a parent, and these agents now approach the capability of experienced AI engineers on realistic benchmarks. However, these agents have three weaknesses in efficiency that have not been fully addressed. First, only a small number of candidates can be executed within a realistic budget, so search rules that rank nodes by executed rewards, such as Monte Carlo-style tree search, rely on few and noisy scores and select the next node to explore less effectively. Second, no resource-aware strategy is used to schedule training jobs, which can lower hardware utilization and training efficiency. Third, every agent call is served by a single powerful model, which inflates inference cost. Here we introduce AIBuildAI-2.5, an agentic system that carries out the tree search with LLM agents and addresses each of the three issues. AIBuildAI-2.5 proposes a novel LLM-guided tree search, in which a judge scores each candidate on its expected improvement, grounding, and feasibility, and a selector ranks the pool of candidates from these scores and the state of the search. In addition, AIBuildAI-2.5 comprises a scheduler that launches training jobs with the current hardware resource status taken into account and a router that assigns lower-cost LLMs to less demanding tasks while reserving the most capable LLM for the most challenging sub-tasks in the AI model building workflow. AIBuildAI-2.5 ranks first on MLE-Bench with a medal rate of 73.3%, and outperforms a strong baseline on six autonomous AI research tasks from AIRS-Bench. --- ### Used, Mentioned, or Condemned? A Controlled Contrast-Set Diagnostic for the Use-Mention Distinction in Code-Mixed Hinglish Misogyny Detection Authors: Ashanvi Yadav, Shubham Bhardwaj Published: 2026-09-06 Categories: cs.CL, cs.AI, cs.CY arXiv: https://arxiv.org/abs/2609.22261v1 Lexicon-driven misogyny detectors cannot, by construction, distinguish a slur used against a woman from the same slur mentioned in counter-speech ("don't call her that") -- yet exactly this distinction governs whether moderation protects or silences the people discussing abuse. We study this problem in code-mixed Hinglish and make three contributions. First, we diagnose two evaluation artifacts on a publicly available redacted corpus: category-encoding anonymization placeholders leak the label (a no-learning rule scores 1.000), and even after they are neutralized misogynistic and benign comments occupy lexically disjoint registers, so bag-of-words reaches macro-F1 approximately 1.00 under random cross-validation but collapses under template-disjoint evaluation. Second, we release Hinglish-MGY-Diag, a deterministic generator and a 416-item / 163-minimal-pair contrast-set diagnostic across five linguistically motivated categories in which slur presence and gendered register are decorrelated from the label by construction. Third, we introduce a strict pair-consistency metric that credits a model only when both members of a minimal pair are correctly labelled. Five from-scratch classical baselines evaluated under construction-disjoint five-fold cross-validation reveal that the strongest model reaches 0.93 accuracy on the cleanest use-mention subset but only 0.82 consistency -- it still mislabels roughly one counter-speech pair in five. A frontier LLM used as an author-model ceiling attains 1.000 on all metrics, doubling as independent label validation and confirming the benchmark is a capability gradient rather than an adversarial wall. We release all code, data, the generator, and an arms-length LLM harness for reproducing every number. --- ### Causal Localization of the Refusal Direction in Audio Language Models Authors: Leonardo Haw-Yang Foo, Hung-yi Lee Published: 2026-09-06 Categories: cs.MM, cs.CL, cs.LG, cs.SD, eess.AS arXiv: https://arxiv.org/abs/2609.22260v1 A large audio language model (LALM) attaches a speech front end to a text language model (LM) that is already safety-aligned. When such a model refuses a harmful spoken request, is the refusal carried by the front end, or inherited from the text LM? We test this with causal interventions. At each model's audio-to-LM interface and at tested LM residual layers, we fit a direction separating harmful from benign prompts, ablate its component, and measure the resulting change in the model's first-token refusal margin. Four of the five models are evaluated under held-out category shift. Across five LALMs spanning three backbone families, with three models passing a baseline safety gate, the largest tested effects occur in a mid-to-late LM band, while ablations of the tested interface directions have little effect. On Qwen2.5-Omni, ablating the L16 direction changes the margin by -7.10, versus -0.013 at the projector. The audio pathway is still in use: zeroing the encoder output changes the margin by -4.7. In the same model, the contrast is linearly decodable at an early layer where single-layer ablation has little effect, and a direction fitted on the LM backbone alone transfers to the full audio model. Because harmful and benign prompts also differ in form, we interpret the direction as refusal-linked rather than harmfulness-specific. These interventions localize dependence of the refusal margin, not where refusal is computed. Moving the margin also does not always change what the model writes. Safety audits of these models should use interventions rather than rely on probes alone, and should examine the inherited text LM alongside the audio interface. --- ### Which Part of the Context Layer Does the Work? Separating Semantic Content from Retrieval Scaffolding in Text-to-SQL Agents Authors: Qing Ye Published: 2026-09-06 Categories: cs.CL, cs.AI, cs.DB arXiv: https://arxiv.org/abs/2609.22259v1 Context layers, curated documentation that an analytics agent fetches at query time, produce large accuracy gains on text-to-SQL benchmarks. A with/without comparison cannot say which part of the layer does the work: the semantic content, the retrieval scaffolding that delivers it, or the pre-computed views that usually accompany it. We report a four-arm ablation on DABStep on four models that separates the three. The instrument is a data contract: a YAML artifact that carries a domain's semantics and the rules an agent's tools enforce. One arm empties every field of prose in the frozen contract while holding the tool surface, retrieval instruction, table allow-list and operation rules byte-for-byte fixed. Compiling the contract's own SQL expressions into views gives the ceiling a pre-computed layer would reach: gold on all 176 tasks it covers. Content dominates. It raises hard-task accuracy from 13.9% to 55.1%, 22.6% to 56.6%, 22.9% to 68.4% and 37.0% to 77.4%, beating the same knowledge pasted into the prompt on every model. Scaffolding without content is worth 0 to 5 points on two flash models and 14 to 15 on two frontier models. Against the compiled ceiling the contract arm's shortfall is a failure to derive, and it falls from 39 points to 5 with model capability. The contract beats the prompt because the rule it needs is one lookup away rather than buried in a long prompt: its SQL carries the fee semantics up to 98% of the time against the prompt arm's 4%, and at the lowest cost per correct answer on three of four models. For practitioners: semantics first, scaffolding second, pre-computed macros only where an agent demonstrably fails to derive. Ungoverned arms submitted 166 mutating statements; governed arms none. The gain is confined to the contract's domain. On one model the benchmark's own withheld golds grade the contract arm at 51.9% against 18.8% for the prompt baseline. --- ### RS-Claw-Evolution: Environment-Feedback-Driven Evolution for Lightweight Remote Sensing Agents in Long-Horizon Tasks Authors: Kai Ouyang, Dongyang Hou, Liangtian Liu, Zeyuan Wang, Ziyu Li, Chengfu Liu, Zichao Tang, Xuezhi Cui, Shengwu Ouyang, Wentao Yang, Hanwen Yu, Haifeng Li Published: 2026-09-06 Categories: cs.LG, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.22258v1 Large language model-driven remote sensing (RS) agents offer a promising approach to automating geospatial analysis. However, lightweight RS agents based on compact language models struggle with multi-step interactive tasks due to loss of long-horizon states, inefficient environmental feedback utilization, and sparse optimization signals. We propose RS-Claw-Evolution, an environment-feedback-driven framework that progressively improves lightweight agents through three stages. Interaction evolution uses executable code to control observations, maintain intermediate states, and reduce context redundancy. Experience evolution combines failure-aware trajectory generation with error-turn masking to learn from informative failure-recovery experiences without imitating faulty actions. Decision evolution uses reinforcement learning with multi-dimensional environment rewards and turn-level advantage protection to optimize tool-use behaviors and improve credit assignment in long sequences. On Earth-Bench, the optimized Qwen3-4B-based agent achieves 65.9% accuracy in Autonomous Planning mode, outperforming the untrained Qwen3-32B baseline (43.8%) and DeepSeek-V3.1 (60.8%), while approaching GPT-5 (71.6%). These results demonstrate that learning from environmental feedback can improve lightweight agents and narrow their performance gap with larger models in long-horizon RS tasks. --- ### Strategy Accumulation and Guided Execution for Automated LLM Fine-Tuning Authors: Haoran Zhao, Wei Du, Dingwen Yang, Jixuan Huang, Junlin Shang, Lingyong Fang, Ya Guo, Tao Gui, Qi Zhang, Xuanjing Huang Published: 2026-09-06 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.22257v1 Producing task-specific large language models requires discovering effective training strategies through experimentation. Automated fine-tuning systems have made this experimentation feasible with far less manual effort. However, these systems are stateless: each search discards its discovered strategies, dataset insights, and hyperparameter findings once it ends. Every new task must then repeat this costly search from a cold start. To address this, we propose Strategy Accumulation and Guided Execution (SAGE), a two-stage framework that makes automated fine-tuning search cumulative. In the first stage, a multi-agent pipeline performs Monte Carlo Tree Search-based exploration. A parallel Distillation Agent extracts task-specific exploration records and confidence-scored cross-task insights, which together constitute a structured experience repository. In the second stage, SAGE retrieves relevant experience from this repository and selects what applies to guide training on the new task. We evaluate SAGE on nine unseen tasks spanning both single- and cross-category settings. In single-round execution, SAGE's accumulated experience raises the average relative improvement over baseline from 3.2% to 15.6%, a 12.4-percentage-point gain over the same pipeline without it. These results show that persistent strategy experience provides effective guidance for automated fine-tuning on unseen tasks. --- ### DIPLOMAT: Dialogue-Span-Aware Direct Preference Optimization for Polite Persuasive Workplace Negotiation Dialogues Authors: Bibhuti Jha, Rishikant Chigrupaatii, Priyanshu Priya, Asif Ekbal Published: 2026-09-06 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.22256v1 Effective workplace negotiation requires balancing multiple objectives, including achieving task goals, preserving professional relationships, and resolving conflicts constructively. However, misunderstandings, misaligned preferences, and interpersonal friction often impede successful outcomes. Politeness mitigates these challenges by fostering trust, reducing tension, and preventing escalation, and persuasive communication helps overcome resistance, align preferences, and guide participants toward mutually beneficial agreements. Motivated by these insights, we present DIPLOMAT, a dialogue system for polite and persuasive workplace negotiation. To support its development, we introduce PROWESS, a dataset of multi-turn workplace negotiation dialogues generated via a multi-agent framework and enriched withnegotiation strategies, politeness levels, persuasive strategies. DIPLOMAT is trained using Dialogue-Span-Aware Direct Preference Optimization (DSA-DPO), a novel preference learning objective that identifies key dialogue spans for preference alignment. This enables DIPLOMAT to generate contextually coherent responses that employ intended negotiation strategies, maintain politeness, and incorporate effective persuasion strategies throughout interactions. Automatic and human evaluation on PROWESS confirm that DIPLOMAT consistently outperforms baselines in generating coherent, polite, and persuasive negotiation responses. --- ### Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations Authors: Rotem Dror, Zohar Elyoseph, Yuval Haber, Elad Refoua, Oshrat Ayalon, Adir Solomon Published: 2026-09-06 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.22255v1 Existing approaches to persona simulation with Large Language Models (LLMs) mostly rely on shallow character descriptions that fail to sustain coherent character behavior across extended interactions. We introduce Deep Persona, a psychologically grounded, three-layered architecture that organizes personas into hierarchical levels of observable expression, latent beliefs, and core motivational drives, for constructing highly convincing role-playing agents. Governed by the principles of scripted determinism and bounded agency, the architecture restricts the model to a reactive engine guided by a structured internal script. We further propose a reference-free evaluation framework that benchmarks dialogue naturalness against empirical human distributions using established psychological clinical instruments and adversarial stress-tests. Empirical evaluation reveals that while LLMs achieve high pragmatic fluency, they exhibit systematic limitations in emotional expression and joint attention. In addition, we present a case study of two Deep Personas and evaluate them using the proposed framework, demonstrating that structured personas can produce interactions that more closely align with human conversational behavior. --- ### Teacher Should Think Ahead: Adaptive Continuations for Reliable On-Policy Distillation Authors: Jingang Zhou, Yuyi Zhou, Haiyang Guo, Xukai Wang, Shuai Feng, Sirui Gao, Jian Xu, Qingpei Guo, Xu-Yao Zhang Published: 2026-09-06 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.22254v1 On-policy distillation (OPD) is a promising approach for transferring knowledge between language models, where a student receives dense token-level supervision along its own generated trajectories. However, teacher supervision can be unreliable when conditioned on incomplete or low-quality student prefixes. We identify Teacher Uncertainty Contraction (TUC), a systematic phenomenon whereby the teacher's predictive uncertainty decreases as it continues from a student-generated prefix. We theoretically characterize this trade-off through a variance-bias decomposition of teacher-branch gradients, showing that uncertainty contraction reduces variance while teacher-student path divergence increases bias, thereby favoring a finite continuation. Guided by this insight, we propose Adaptive-Continuations On-Policy Distillation (AC-OPD), which augments informative states along student rollouts with teacher continuations and adaptively selects their effective supervision horizons. Experiments on mathematical reasoning and code generation across model scales demonstrate that AC-OPD consistently improves over standard OPD. Controlled-continuations and matched-budget analyses further validate the adaptive-continuations design, highlighting adaptive teacher continuations as an effective principle for reliable on-policy distillation.The code will be made publicly available upon publication. --- ### CAMFT: Conflict-Aware Mergeable Fine-Tuning for Large Language Models Authors: Jingang Zhou, Haiyang Guo, Yuan Ma, Han Zhu, Xu-Yao Zhang Published: 2026-09-06 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.22253v1 Model merging has emerged as a promising paradigm for integrating multiple task-specific capabilities into a single large language model. However, existing methods predominantly focus on post-hoc processing of independently fine-tuned models, overlooking how the training phase itself impacts cross-task compatibility. Resolving parameter conflicts after fine-tuning is inherently sub-optimal. To address this, we propose CAMFT, a Conflict-Aware Mergeable Fine-Tuning method that makes task adaptation both efficient and mergeaware. CAMFT treats mergeability as a property shaped during fine-tuning, rather than only a problem to be solved after fine-tuning. By guiding each task to update sparse coordinates with lower cross-task conflict, CAMFT produces task updates that are efficient to train and more compatible for downstream model merging. Extensive experiments demonstrate that CAMFT outperforms standard finetuning baselines in multi-task merging scenarios. Codes are available at https://github.com/gyanchow/CAMFT-LLM. --- ### CALM: A Calibrated LLM Choice Network Framework for Activity-Based Traveler Simulation Authors: Yezhou Cheng Published: 2026-09-06 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.22252v1 We present CALM, a reproducible hybrid framework that integrates an optional large language model (LLM) activity planner with calibrated stochastic choice, shared network feedback, memory and habit, typed feasibility checks, and deterministic offline replay. Unlike trip-mode classifiers or diary-only generators, CALM executes a closed traveler-day loop and evaluates each generative module against an empirical, reproducible baseline. On the 2024 New York City Citywide Mobility Survey (CMS), 110,691 seven-mode trips are split by respondent into 78,487 training and 32,204 holdout trips. Training-only alternative-specific constant calibration reduces mean holdout mode Jensen-Shannon divergence from 0.15599 to 0.00394 across ten seeds. A matched live-LLM ablation then quantifies trade-offs among aggregate fit, temporal fit, behavioral persistence, and feasibility, while frozen prompt-response pairs support deterministic replay of downstream simulation. Controlled weather, delay, fare, and parking ladders further demonstrate consistent and interpretable responses under intervention. CALM contributes a reproducible protocol for integrating and evaluating generative planners in traveler simulation through person-disjoint calibration, matched module ablation, controlled stress testing, and end-to-end traceability. --- ### Predictors and Orchestrators: Parsimonious Machine Learning within an Agentic AI Harness for Multi-Horizon Karst Aquifer Forecasting Authors: Pramod Lekhak, Chetan Sharma, Hakan Başağaoğlu, F. Paul Bertetti, Debaditya Chakraborty Published: 2026-09-06 Categories: cs.LG, cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.22251v1 Forecasting karst aquifer dynamics is difficult because recharge responses are nonlinear, event-driven, and governed by strongly heterogeneous flow paths. This study develops and evaluates a deployment-aware framework for 1-12-week-ahead prediction of spring discharge and groundwater level using approximately 79 years of hydroclimatic observations from the Edwards Aquifer, Texas. Five model families were compared under a common temporal evaluation design: extreme gradient boosting, extremely randomized trees, long short-term memory, convolutional neural networks, and Transformers. Predictions were evaluated using coefficient of determination, Kling-Gupta efficiency, root-mean-square error, and agreement with operational drought thresholds. Extreme gradient boosting was consistently most reliable, with R2 at least 0.97, 0.96, and 0.94 across 1-4-, 5-8-, and 9-12-week horizons, respectively, and greater than 90% critical-stage agreement at the first three drought stages across all horizons. Deep models were competitive at short horizons but degraded progressively and exhibited isolated failures at longer lead times. We attribute this contrast to an alignment between tree partitioning and low-dimensional, axis-aligned hydroclimatic predictors, together with the tendency of neural models to smooth irregular extremes. The validated models were embedded in a five-agent operational architecture that automates data acquisition, model assignment, deterministic prediction, threshold monitoring, prospective verification, literature retrieval, and reporting. The contribution is therefore a transferable framework joining parsimonious model selection, leakage-aware multi-horizon evaluation, decision-relevant threshold skill, and auditable agentic automation. --- ### Filling the Unseen: Scene Extrapolation via 3D Gaussian Splatting Authors: Yunlai Zhou, Yiren Lu, Tuo Liang, Disheng Liu, Vipin Chaudhary, Yu Yin Published: 2026-09-06 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.13262v1 3D Gaussian Splatting achieves photorealistic reconstruction within training view distribution, yet it degrades on out-of-distribution novel views, exhibiting holes in unobserved regions and artifacts in observable areas. Recent works formulate this task as extrapolation and interpolation and try to address it with generative models, but remain limited in extrapolation scale and quality. They repeat a generate-reconstruct-shift cycle to progressively build a scene, which introduces accumulated errors with every step conditioning on previous outcomes. In this work, we propose a holistic framework for extrapolation and interpolation. We devise an independent camera view detection mechanism to enable parallel conflict-free extrapolation, circumventing the reliance on the aforementioned error-prone cycle. Building upon this, we design a hierarchical pipeline that extrapolates independent and dependent camera views separately. Additionally, previous methods overlook inconsistency between generated and original images, resulting in compromising well-reconstructed areas. We propose a plug-and-play Quality-Aware Mask (QA-Mask) module, enabling selective utilization of generated data. By calibrating learning weights with pixel-wise rendering quality, it prevents generation-induced degradations on well-constructed areas. Extensive experiments demonstrate the superior performance of our framework, with QA-Mask generalizing to multiple generative reconstruction models. --- ### From Process Loss to Assembly Bonus: Human-Grounded Diagnosis of Multi-Agent LLM Collaboration Authors: Ala N. Tak, Teruhisa Misu, Kumar Akash, Zhaobo K. Zheng, Kevin H. Joo, Jonathan Gratch Published: 2026-09-06 Categories: cs.MA, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.13261v1 LLM agents are increasingly used for collaborative problem solving and human-group simulation. This makes outcome-only evaluation insufficient: if LLM groups are used as models of human groups, we need to know whether they succeed or fail through human-like deliberative mechanisms. We compare human group chats with matched LLM deliberation traces on Wason-style deductive reasoning, then test whether the same process signatures generalize to analogical, abductive, and analytical tasks. Humans and LLMs show the same assembly bonus asymmetry: discussion improves the average member more often than the best initial member. Initial-answer diversity accounts for the effect of model heterogeneity, increasing movement in both corrective and destructive directions. The main differences are process-level. Compared with humans, LLM groups follow majorities more often, surface less unique information, and converge earlier; correct minority signals succeed mainly when re-expressed early. Interventions motivated by human group-decision research yield modest improvements in collective outcomes, but do not remove the coordination bottleneck. Together, these results suggest that LLM groups can reproduce some outcome-level patterns of human deliberation while diverging in the mechanisms that generate assembly bonus and process loss, with implications for group simulation and human-AI collaboration. --- ### From Masking to Merging: Rethinking SpecAugment for Efficient Audio Spectrogram Transformer Authors: Minhee Park, Hyowon Ahn, Chanwoo Kim Published: 2026-09-06 Categories: cs.SD, cs.LG, eess.AS arXiv: https://arxiv.org/abs/2609.13260v1 This paper proposes SpecAugment-Patch Merging, a simple yet effective method to accelerate Audio Spectrogram Transformer (AST) training. We first apply SpecAugment to mask input spectrograms at the patch level, and after positional embeddings are added, the method selects r pairs of masked patches and merges them, reducing the number of tokens processed by the Transformer. Increasing the number of merged pairs r from 0 to 100 keeps mAP on AudioSet nearly unchanged (34.07 to 34.08) while throughput increases from 43.3 to 49.3 samples/sec, which is a relatively 13.9% improvement. Similar patterns appear on ESC-50 and Speech Commands V2, where throughput steadily improves with only minor accuracy changes, demonstrating that this merging approach provides faster training with minimal performance loss. --- ### TryOnReward: Learning Foveated Consistency for Reinforcement Fine-Tuning of Virtual Try-On Authors: Xueheng Li, Yong Liu, Xiaolong Fu, Wen Xue, Chengjun Xie, Yipeng Sun, Yan Li, Simiu Gu Published: 2026-09-06 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.13259v1 Virtual Try-On (VTON) aims to dress a person with the reference garment, producing visually reasonable results aligned with human preferences. Turning this preference-oriented goal into an actionable objective relies on a scoring function aligned with human taste. However, classic fidelity metrics exhibit weak correlation with human judgments, and generic VLMs fail to provide the discriminative granularity demanded by try-on quality evaluation, which hinges on faithfully preserving garment and person details. This shortcoming is further exacerbated in the reinforcement fine-tuning (RFT) optimization and leads to severe reward hacking. To this end, we present TryOnReward, a fine-grained reward model tailored for VTON. Built on a vision-language backbone, it adopts a foveation calibration objective that grounds each quality dimension in the relevant region to avoid global shortcut learning. Meanwhile, TryOnReward jointly optimizes pairwise preferences and per-dimension quality scores via margin-aware supervision, leveraging both relative and absolute quality signals. For model training and evaluation, we build TryOnReward-100K, a human-annotated per-dimension rating dataset, alongside TryOn-Bench and TryOnRewardBench, two benchmarks covering diverse real scenarios. Extensive experiments confirm that TryOnReward significantly outperforms generic judges in human preference alignment, and when serving as the RFT reward function, it consistently yields human-preferred try-on results across multiple baselines. --- ### Interpretable Temporal Video Reasoning with EventGraph and EventField Authors: Durgendra Narayan Singh Published: 2026-09-06 Categories: cs.CV, cs.CL, cs.IR arXiv: https://arxiv.org/abs/2609.13258v1 We present a structured temporal video reasoning pipeline built around a discrete EventGraph, a continuous EventField, and a human-readable EventGlyph view. On a calibrated EPIC-KITCHENS subset of 10 videos and 50 temporal reasoning questions, EventField+Glyph achieves 0.98 overall accuracy, which is higher than the caption baseline by +0.40 (paired p = 1.1 \times 10^{-5}) and direct VLM-only QA by +0.20 (p = 0.0063) on this subset. We further evaluate annotation-source variations, including manual, heuristic, and heuristic+Gemini pipelines, and find that the best structured method stays above the caption baseline across settings. We also include cross-video pair benchmarking and an appendix gallery of glyph outputs for all studied videos. Overall, the results indicate that structured temporal representations can support both performance and inspectability by preserving symbolic structure, capturing temporal continuity, and providing human-readable diagnostics for video reasoning. --- ### Sampling headroom is not selection gain: a compute-value audit of test-time scaling for video world models Authors: Yuhua Jiang, Junjie Lu, Feifei Gao Published: 2026-09-06 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.13257v1 Test-time scaling (TTS) can improve generation only when additional compute produces better candidates and the system can reliably identify them. This distinction is especially important for video world models, where a wider sample pool may contain stronger rollouts without improving the output that is ultimately selected. We introduce the Compute-Value Audit (CVA), a sequential framework that asks whether extra sampling creates opportunity, observable signals provide a reliable state, that state supports a beneficial action, and the resulting gain exceeds the full entry fee of generation and verification. On 192 Physics-IQ scenes, expanding the pool from 4 to 16 candidates increases oracle quality by +9.23 IQ (95% CI (+7.44, +11.14)), but Flow, Cycle, and VideoReward fail to recover this headroom reliably. Across three generators, none of twelve adaptive-depth policies outperforms uniform compute; they recover only 42-69% of the measured entry fee. A matched-60-NFE Predict-and-Perturb intervention on VideoPhy2 is likewise negative across three fresh-seed replicas. These negative results are not universal: anchor-explorer passes all four stages in a sparse PRM800K setting, MMLU-Pro exposes the gap between predictive state and useful action, and a privileged paired future establishes a positive video upper bound. Together, these results show that sampling headroom has deployment value only when it can be converted into a reliable decision whose benefit survives the complete compute charge. Code is available at https://github.com/YuhuaJiang2002/sampling-headroom-is-not-selection-gain. --- ### A Comprehensive Review of Multimodal Facial State Analysis: Tasks, Methods, and Resources Authors: Xuri Ge, Tianshuo Zhang, Ruihan Li, Hui Ye, Kaiwen Zheng, Junchen Fu, Da Huo, Joemon M. Jose, Hu Han Published: 2026-09-06 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.13255v1 Facial state analysis plays a crucial role in understanding human expressions, psychological modeling, and human computer interaction. Traditional unimodal vision-based methods are often limited by environmental sensitivity and weak interpretability. Multimodal facial state analysis addresses these issues by integrating complementary cues from visual, audio, textual, physiological, and other related modalities. This survey emphasizes two key aspects: on one hand, multimodal learning enables contextual semantic understanding for improved facial state reasoning and leverages interpretable language generation to enhance model explainability; on the other hand, multi-task learning allows simultaneous analysis of expressions, action units (AUs), and face-based soft biometrics (e.g., age, gender), effectively capturing fine-grained expressions and improving cross-scene generalization. This survey reviews core tasks, representative methods, and datasets in multimodal facial state analysis, focusing on facial expression recognition, AU detection, and face-based soft biometric estimation, and emphasizing the unique value of language in providing contextual semantics, enhancing reasoning, and generating explanations. The survey aims to provide an up-to-date overview of the literature and to highlight future research directions for multimodal, interpretable, and multi-task adaptive facial state analysis. --- ### (How) Do MLLMs Report Bistable Images Like Humans? Authors: Ryota Takatsuki, Tomoki Doi, Amane Watahiki, Anil K. Seth, Hitomi Yanaka Published: 2026-09-06 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.13254v1 Bistable images such as the duck-rabbit are classic stimuli in which one image supports multiple mutually incompatible interpretations, typically reported one at a time in humans. We ask whether multimodal large language models (MLLMs) show similar report behavior and what internal computations support it. Using the LLaVA family, we study two tractable dimensions: modulability, whether reports can be biased by bottom-up visual cues and top-down linguistic priors, and exclusivity, whether responses commit to a single interpretation. We test both on the canonical duck-rabbit and on synthetic Visual Anagrams to mitigate memorization confounds. Behaviorally, both visual and linguistic manipulations systematically shift reports in human-consistent ways, while responses remain predominantly exclusive. Mechanistically, these effects arise from competing image-token representations, distinct pathways for bottom-up and top-down modulation, and a link between exclusive reporting and object-count encoding. Code and data are available at https://github.com/rtakatsky/mllm-bistable-images. --- ### Halo: Improving forecast accuracy through heteroscedastic estimation Authors: Adam Cataldo Published: 2026-09-06 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.10589v1 Heteroscedastic forecasting, where a network estimates a scale parameter alongside a location parameter, is normally motivated by uncertainty quantification. This paper shows it also improves the point estimate, in contrast to reported negative results for heteroscedastic estimation outside time series. Halo is a modification that reuses an existing deep forecaster's architecture, giving it a second output for the scale of its implied distribution and training it under the matching negative log likelihood. Adapting three state-of-the-art models --- a transformer, a graph network paired with a variational autoencoder, and a single-layer convolutional network --- under both Gaussian and Laplacian losses demonstrates the phenomenon. On the five electricity price markets of a standard forecasting benchmark, Halo improves MSE and MAE in 28 of 30 model-market-metric comparisons, cutting average MSE by 2.6% to 16.5% and average MAE by 1.7% to 11.0%. Two findings emerge: (1) whether the scale estimate comes from a second projection head or from a full parallel network matters far less than whether the network estimates scale, and (2) the improvement holds under the hyperparameters already tuned for the point-estimate baseline, so retuning is optional. --- ### Threshold-Based Selection for Continuous Optimization: A Leaf-Abscission Instantiation Authors: Nasser Khalili Published: 2026-09-06 Categories: cs.NE arXiv: https://arxiv.org/abs/2609.10588v1 This paper formalizes threshold-based selection as an evaluation-gating architecture in which each incumbent is tested before variation and a replacement is generated and evaluated only when contextual pressure exceeds intrinsic strength. The mechanism is instantiated as Leaf Abscission Optimization (LAO), using rank-based strength, a phenological seasonal signal, diversity modulation, environmental pressure, and a base regrowth kernel. A blocked two-to-the-fourth-power factorial analysis at dimension 10 on the CEC 2017 suite reduces the original multi-layer design to a parsimonious core: drift is harmful, while the other three auxiliary layers show no robust independent evidence of benefit. The resulting LAO-Core attains the third-best mean Friedman rank among nine optimizers at dimensions 10, 30, and 50 under the equal evaluation budget. A four-budget sweep shows budget-dependent relative performance, with adaptive differential-evolution baselines gaining relative advantage at larger budgets; the nine-cell dimension-budget analysis establishes neither an evaluation-budget-per-dimension-only law nor a statistically significant dimension-budget interaction. A paired intervention shows that diversity modulation changes late-run replacement behaviour without a detectable effect on final error at the tested budget. The evidence supports LAO as a parsimonious evaluation-gating mechanism with regime-qualified competitiveness, rather than as a generally superior optimizer. --- ### Probabilistic Focal Search: Accelerating Bounded-Suboptimal Search via Lower-Bound Advancement Authors: Minh Vu Duc, Trung Le Huu, Hà Minh Hoàng, Trung Thanh Nguyen, Phuong Khanh Nguyen, Huynh Thi Thanh Binh Published: 2026-09-06 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.10584v1 Bounded-suboptimal search seeks a solution within a factor $w$ of optimal while reducing search effort. Focal Search (FS) uses heuristic guidance within FOCAL, the frontier nodes eligible under the threshold $w f_{\min}$, but its deterministic policy may leave $f_{\min}$ unchanged for many expansions. We introduce Probabilistic Focal Search (PFS), which follows the FS guided choice with probability $p$ and expands a minimum-$f$ OPEN node with probability $1-p$. The latter branch encourages the lower bound to advance, enlarging FOCAL and admitting nodes that may lead to feasible solutions. By balancing guidance and lower-bound advancement, this mechanism can reduce time to a bounded solution when progress is limited by delayed FOCAL admission. As a secondary transfer experiment, we apply the same scheduler to Dynamic Potential Search, yielding Probabilistic Dynamic Potential Search (PDPS). We benchmark PFS against FS on N-Puzzle, Pancake Sorting, and the Traveling Salesperson Problem (TSP), and evaluate its anytime extension on the Generalized Covering TSP (GCTSP), using multiple $w$ and $p$ values. Across these benchmarks, the largest gains occur when long $f_{\min}$ plateaus delay useful FOCAL admissions; in such settings, the probabilistic factor may reduce node expansions by about 90\% or more (e.g., on N-Puzzle and TSP). For the anytime algorithm family, Anytime Probabilistic Focal Search (APFS) outperforms all tested algorithms in evaluating anytime methods on GCTSP. We also observe that the benefit is smaller when the deterministic search already advances efficiently (e.g., Pancake Sorting), indicating that the probabilistic factor is most useful when FOCAL admission is a search bottleneck. The PDPS transfer shows that the mechanism also transfers to potential guidance, although its common-success effects remain domain- and bound-dependent. --- ### Geometry Conditioning in an Embodied SLM: Training Controls and Robustness Diagnostics in a 0.8B Hybrid Model Authors: Hao Li, Haofei Sun, Lin He Published: 2026-09-06 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.09213v1 We study how physical-state inputs affect a 0.8B hybrid language model adapted for manipulation with 6.2M trainable parameters. Six conditions are trained on three LIBERO-Spatial tasks and evaluated over three seeds and 540 held-out rollouts. Conditioning recurrent decay gates on geometric increments yields 28.9% success, compared with 36.7% when those increments are shuffled during training and 24.4% without explicit object/goal geometry. Both geometry policies receive correct inputs at evaluation. A token adapter using the same increments scores 27.8%; differences vary across seeds and remain inconclusive. Token-clock conditioning scores 11.1%, including one seed that fails to converge. In separate robustness tests, a state-only relative-coordinate policy retains 7/10 success under frame relabeling, whereas all four tested visual policies fall to at most 3/20 after a 5 cm object displacement. These results show no reliable advantage from training-time geometric alignment under this recipe and illustrate the gap between coordinate invariance and physical-layout generalization. Episode records, seed-level analyses, and figure-generation code accompany the paper. --- ### AgentHijack: Visual Patch Attacks on Multimodal Computer-Use Agents Authors: Zhihao Liu, Hongyu Sun, Zhiyuan Fu, Xiaonan Duan, Jice Wang, Shangru Zhao, Weizhi Meng, Wuxin Yang, Yangfan Zhou, Yuqing Zhang Published: 2026-09-06 Categories: cs.CR, cs.AI, cs.CY arXiv: https://arxiv.org/abs/2609.09212v1 This paper presents an end-to-end evaluation framework for image-triggered command injection against computer-use agents (CUAs). The goal is to test whether a local visual patch can induce verifiable environmental consequences along the full chain of screenshot input, VLM generation, action parsing, and environment execution. We train and deploy patches on author-controlled GitHub Pages pages and a locally deployed CSDN clone, and evaluate them in real environments across five open-source or publicly available GUI-agent or vision-language-model (VLM) backends. Our experiment aggregates 600 instance-level online cases, with T-ASR, TAPR, and E2E-ASR reaching 84.5%, 47.0%, and 20.3%, respectively. Trajectory analysis further shows that in some successful cases the agent first executes a malicious terminal command and then continues the original benign task. These results indicate that optimized local visual signals can affect not only VLM outputs but also propagate through the execution pipeline of open CUAs and create real environmental risk. --- ### A Subsampled Davis-Kahan Bound for Large-Scale Eigenspace Estimation Authors: Huan Qing Published: 2026-09-06 Categories: stat.ML, cs.LG, stat.ME arXiv: https://arxiv.org/abs/2609.09211v1 The Davis-Kahan theorem is a fundamental tool in spectral analysis, providing quantitative control over the distance between the eigenspaces of a symmetric matrix and its perturbation. However, when the matrix dimension is large, computing leading eigenvectors is computationally expensive, limiting the practical use of spectral methods in modern large-scale applications. This paper addresses this problem by proposing an independent Bernoulli sampling scheme and proves that the leading left singular vectors of the subsampled matrix faithfully approximate the target subspace of a low-rank symmetric matrix. Our main result is a subsampled Davis-Kahan bound that gives an explicit error bound depending directly on the sampling probability. The bound reveals the trade-off: the computational cost scales linearly with the sampling probability, while the statistical error scales as the inverse square root of the sampling probability. Our result thus extends the Davis-Kahan theorem to the subsampled setting, enabling scalable spectral analysis of large-scale symmetric matrices. ---