# Research Papers — 2026-09 · bundle page 118 of 266 > Full-text markdown of papers #5851–5900 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/118 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/117.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/119.md) ### Look Where It Counts: A Free, Label-Free Visual Evidence Signal for Fine-Grained Vision-Language Reasoning Authors: Santi Ram Tiwari, Nihal Naik, Devbrat Pandey, Nishant Sinha Published: 2026-09-21 Categories: cs.CV, eess.IV arXiv: https://arxiv.org/abs/2609.24244v1 Multimodal large language models (MLLMs) fail at fine-grained visual questions less because they cannot reason than because they never see the evidence: high-resolution images are downsampled before encoding, so the model answers from linguistic priors. The standard remedies are expensive: annotated answers (SFT), hand-engineered verifiers (RLVR), or a large external teacher (on-policy distillation). We ask whether the visual evidence itself can supply the signal for free. We formalize the contrastive evidence gap, the per-token log-likelihood ratio that a model assigns to its own output when conditioned on a question-relevant region versus an irrelevant one, and study it across Qwen2.5-VL-7B, Qwen3-VL-8B, and Qwen3-VL-30B-A3B on V*Bench. Our main positive result is training-free: selecting the candidate crop under which the model's answer distribution is most peaked, using a single-view, label-free criterion, discovers the answer-bearing region with no bounding boxes, training, or labels. It localizes the target 4.4 to 5.1 times better than chance and raises fine-grained accuracy from 70 percent to 85 percent at inference. We further show that the gap is complementary to the model's own confidence. Combining them predicts correctness better than either alone, with AUC up to 0.99, and flags confidently wrong answers, with AUC ranging from 0.97 to 1.00 within the high-confidence subset. All effects concentrate on perception-bottleneck questions and vanish on a global-context control. Finally, we report an honest negative result: converting the same signal into a training method, gated self-distillation (SEG-Distill), does not outperform the base model at pilot scale across three gate designs, while more aggressive gating degrades accuracy. The signal is real, but converting it into training gains remains an open problem. --- ### Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation Enforcement Authors: Xutao Mao, Jianing Zhu, Jinman Zhao, Tongliang Liu, Xiaowen Chu, Cong Wang, Bo Han Published: 2026-09-21 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.24243v1 Reinforcement learning (RL) improves reasoning in vision-language models (VLMs) but can induce chain-of-thought (CoT) obfuscation: an operational, non-intentional outcome where task reward or accuracy rises while traces become less grounded and monitorable. Prior work largely documents this decay behaviorally, leaving its representation-level correlates and actionable controls unclear. We find that template- and ground-associated activations become less separable during RL; matched interventions support the contribution of selected features to monitorability degradation. Guided by this evidence, we propose Targeted Anti-obfuscation with Mechanistic Enforcement (TAME), which uses Sparse Autoencoders (SAEs) to combine behavioral feedback with targeted suppression of template-associated activations during RL. Its asymmetric constraint penalizes template activations only above their pre-RL baseline, anchoring the localized features while behavioral feedback promotes grounded refinements. Across VIRL-39k, SPA-VL, and two model families, TAME improves CoT monitorability by up to 30.9 and 16.7 percentage points over Group Relative Policy Optimization (GRPO), respectively. Blinded human evaluation finds higher human monitorability on both datasets, and two held-out monitor families reproduce the monitorability gains. Task accuracy changes are small and mixed, and general-capability benchmarks show task-specific trade-offs. These results provide a path from behavioral monitoring to representation-level oversight for more auditable RL-trained multimodal systems. --- ### Hessian Rank Constraint for Learning Structure of Nonlinear Latent Variable Models Authors: Zijian Li, Ruichu Cai, Feng Xie, Xinshuai Dong, Haoyue Dai, Yuewen Sun, Yujia Zheng, Guangyi Chen, Yingyao Hu, Kun Zhang Published: 2026-09-21 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.24241v1 Uncovering latent variables and their causal relations from observed data is a fundamental yet challenging problem. Existing methods often rely on restrictive assumptions, such as linear relations or invertible mixing functions. To better address this problem under general nonlinear mixing procedures, we propose a condition called the cross-Hessian Rank Constraint (HRC), which serves as a primitive rank-based tool for nonlinear latent causal discovery. In particular, we show that a rank-based property arises from the cross-Hessian of the observed-data log-density in the nonlinear case, revealing information about the latent variables, and reduces to the Tetrad constraints in the linear Gaussian case. More specifically, when two groups of observed variables are d-separated by a set of lower-dimensional latent variables, the rank of this cross-Hessian is equal to the dimension of the latent variables, under a mild affine derivative assumption on the conditional log-density derivatives. This assumption can be naturally satisfied when the noise level is low or the relevant nonlinearity is moderate. As a downstream application, we instantiate HRC in the pure one-factor measurement setting for locating latent variables and recovering their causal structure up to Markov equivalence. Experimental results on synthetic and real-world datasets support the theoretical claims. --- ### Memory vs. Context? Influential Factors of Factual Recall in Language Models Authors: Guilhem Fouilhé, Nicholas Asher, Philippe Muller Published: 2026-09-21 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.24238v1 We reproduce and stress-test the work of Yu et al. (2023), who characterize how language models (LMs) arbitrate between memorized knowledge and contradictory in-context statements. We replicate their world-capitals experiments on 31 models spanning Pythia, GPT-2, Qwen3, and Ministral families, including base and post-trained variants, and extend evaluations to five additional knowledge relation types from the ParaConflict dataset. We empirically confirm most of their original findings: larger models and higher-frequency entities tend to favor memorized answers, with substantial family-level variance. However, several conclusions do not generalize cleanly: entity-frequency effects disappear on Qwen3-14B and 32B; post-training shifts the memory-context trade-off inconsistently across families; question phrasing alone can change a model's reliance on memorized knowledge by up to 80 percentage points; and semantically unrelated prose can mimic coherent supporting context. Our results clarify where Yu et al.'s claims hold and to what extent they generalize to other prompts. --- ### Adaptive Forgetting for Nonstationary Optimization: Towards Robust EEG Decoding Authors: Hongyu Zhu, Lin Chen, Jing Chen, Yuting Zhou, Mingsheng Shang Published: 2026-09-21 Categories: cs.LG, cs.HC arXiv: https://arxiv.org/abs/2609.24233v1 Electroencephalography (EEG) provides non-invasive monitoring of brain activity and is widely used in emotion recognition, motor imagery and sleep staging. Although within-subject decoding has achieved considerable progress, cross-subject generalization remains a central challenge in practical applications. EEG decoders are typically trained with Adam/AdamW under a fixed second-moment decay coefficient, even though cross-subject learning involves low signal-to-noise ratios, subject variability, and gradient nonstationarity. A fixed coefficient implicitly assumes that gradient statistics are homogeneous across layers and time, which can limit model's adaptability to cross-subject EEG signals and degrade generalization. To address these issues, we propose AFOR, a tensor-wise adaptive optimizer that converts the fixed second-moment decay coefficient into a dynamic coefficient estimated online from local gradient state. AFOR combines a Residual-Alignment Signal Scorer (RASS) and an Adaptive Forgetting Controller (AFC). RASS summarizes local gradient residuals and directional agreement into a signal-quality score, and AFC maps this score through self-referential normalization to a bounded per-step decay coefficient, with cumulative-product initialization correction maintaining consistency under time-varying decay. Under a strict cross-subject protocol on three EEG benchmarks that cover three representative fields, AFOR achieves the best average performance among the compared optimizers, improving the mean test accuracy over Adam by 3.00%, 2.07%, and 4.38%, respectively. --- ### Recovering Lost Details: Multi-Scale Frequency Compensation for Long-Term Time Series Forecasting Authors: Runmin Zou, Siyi Xie, Yaohui Huang, Yun Wang Published: 2026-09-21 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.24229v1 Long-term time series forecasting has made significant progress by leveraging multi-scale information to capture hierarchical temporal patterns and model long-range dependencies. However, temporal downsampling in existing multi-scale methods inevitably smooths detailed temporal fluctuations, and this information loss is further aggravated by their emphasis on dominant trends across scales, resulting in insufficiently expressive representations. To address this, we propose a Multi-Scale Wavelet Mixing (MWMixer) model, which incorporates a Bidirectional Frequency-Bands Mixing strategy to recover lost temporal details across scales, enabling complementary cross-scale information interactions. Then, a Dynamic Scale-Adaptive Fusion module learns time-varying weights for each scale to fuse multi-scale forecasts into the final prediction, enhancing the flexibility of multi-scale aggregation. In addition, a cross-scale consistency loss aligns each coarse-scale prediction with the interval-averaged fine-scale outputs, while a multi-scale supervision loss enforces prediction accuracy at each scale, promoting consistent learning across scales. Extensive experiments on seven real-world datasets demonstrate that MWMixer achieves competitive performance in long-term forecasting. --- ### IMPLICIT-Bench: Measuring Implicit Bias in Text-to-Image Models under Neutral Prompts Authors: Yue Dai, Ziyang Liu, Marc Cheong, Caren Han Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24228v1 Text-to-image (T2I) models are typically evaluated for bias using slot-based templates such as ``a photo of a (profession)''. Such templates probe only \emph{explicit} demographic attributes (e.g., gender, skin tone) in isolation. They overlook a broader \emph{implicit} bias that arises in natural prompts: when stereotype-relevant attributes are left unspecified, models still default to stereotypical outputs. We introduce IMPLICIT-Bench, a benchmark for measuring implicit bias in T2I models under such prompts. The key design is a structured-knowledge-graph (KG) construction of controlled prompt triplets: neutral, stereotype, and anti-stereotype variants that differ only along a single bias dimension while preserving scene semantics. This enables precise attribution of bias effects that template benchmarks cannot achieve. IMPLICIT-Bench comprises 5,493 prompts across 11 bias categories, validated through multi-model agreement, CLIP-based verification, and human evaluation. Using this benchmark, we show that state-of-the-art T2I models exhibit systematic bias under neutral prompts, a failure mode largely invisible to existing evaluations. We then use IMPLICIT-Bench to evaluate debiasing methods, uncovering a fundamental trade-off between bias reduction and semantic fidelity. --- ### SRPR-Net: Semantic and Relational Prompt Refinement for Automated SAM-based Instance Segmentation Authors: Lufei Liu, Guojie Li, Suncheng Xiang, Fan Zhang Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24226v1 Instance segmentation is a fundamental computer vision task with diverse real-world applications. Recently, prompt-driven foundation models have shown promising generalization. However, automated prompting remains limited by insufficient semantic guidance and inter-instance modeling. To address this challenge, we propose a novel architecture, named Semantic Relational Prompt Refinement Network (SRPR-Net), for automated SAM-based instance segmentation. A sequential prompt refinement mechanism is introduced to enrich detector geometry with visual-language semantics and then incorporate same-image instance dependencies, enabling context-aware box adjustment before SAM segmentation. Experiments on multiple standard benchmarks demonstrate that SRPR-Net achieves consistent improvements in segmentation performance over existing state-of-the-art approaches. The code is publicly available at https://github.com/JeremyXSC/SRPR-Net. --- ### DiaSeg: Diagonal Segment Extraction from DTW Paths for Interpretable Gait Analysis Authors: Tresor Y. Koffi, Amel Hidouri, Corentin Legrand, Aurélie Bertaux Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24223v1 Dynamic Time Warping (DTW) is the dominant approach for measuring similarity between time series, yet standard practice discards the optimal warping path after computing a single distance value, losing local alignment information most relevant to clinical diagnosis. We introduce DiaSeg, a framework that extracts diagonal segments from DTW paths with controlled breaks, characterizing each segment by five geometric features (effective length, interruption count, cost variation, temporal position, and path context), and enabling unsupervised pattern discovery without domain-specific feature engineering. Validated on 91 subjects across six clinical conditions (healthy aging, Parkinson's, Huntington's, ALS, brain tumor, and stroke), three findings emerge. First, diagonal segments form consistent unsupervised patterns (silhouette 0.33) aligned with biomechanical phase annotations, with label-based validation confirming near-perfect separation of healthy and pathological gait (ARI up to 0.986). Second, segments discriminate pathology at 69% (supervised) and 75% (patient-level clustering), with pathology manifesting through distributional shifts in segment length; combining segment and cycle-level features further improves classification to 91.7%. Third, while cycle-based methods achieve higher accuracy (91%), diagonal segments provide phase-specific interpretability unavailable in global representations, localizing where coordination breaks down within the gait cycle. DiaSeg thus transforms DTW from a black-box distance into a source of interpretable temporal features for neurodegenerative disease assessment. --- ### Document Retrieval-Aware Chunking (D-RAC): Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion Authors: Uday Allu, Abhivanth Sivaprakash, Pratik Singh, Aman Manocha Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24220v1 Retrieval-Augmented Generation (RAG) systems over enterprise knowledge bases must ingest heterogeneous document formats -- PDFs, Word documents, presentations, and scans -- whose content is locked inside complex visual layouts, multi-column pages, and dense tables. Rule-based extraction and OCR destroy reading order, flatten tables, and lose heading hierarchy, while fully agentic chunking over extracted text incurs high token costs and hallucination risk. We present Document Retrieval-Aware Chunking (D-RAC), an extension of our Web Retrieval-Aware Chunking (W-RAC) framework to arbitrary document formats. D-RAC first normalizes any input document into PDF, exploiting the fact that virtually every format has a faithful, deterministic PDF rendering. A single multimodal LLM pass then converts rendered pages into retrieval-optimized Markdown -- rewriting tables as self-contained prose statements and preserving heading hierarchy -- after which chunking proceeds exactly as in W-RAC: deterministic parsing into ID-addressable units followed by lightweight LLM-based chunk planning over identifiers rather than text. Source text is never regenerated during chunking, preserving W-RAC's cost, determinism, and observability benefits while unlocking every renderable format as a first-class input. On the 236-document, 795-page PDF subset of the RAG-Multi-Corpus benchmark spanning five enterprise domains, D-RAC converts and chunks the entire corpus in 72 minutes with zero errors, producing 1,748 retrieval-ready chunks. Compared to agentic chunking with frontier LLMs, D-RAC reduces chunking-stage output tokens by 95.7%, cutting chunking cost by 77.8% (GPT-4.1 pricing) to 85.6% (Gemini 2.5 Pro pricing) and chunking time by 75%. D-RAC scales linearly to documents of 500+ pages. --- ### From Articles to Publishers: Aggregating Language Model Predictions for News Source Reliability Inference Authors: John Bianchi, Manuel Pratelli, Fabio Pinelli, Marinella Petrocchi Published: 2026-09-21 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.24219v1 Traditionally, the reliability of news publishers is assessed by expert organisations that evaluate editorial practices, transparency and factual standards at source. When this process is translated into a computational approach, the problem is often formulated at the level of individual articles, with models being trained on a set of pre-labelled articles and their performance being evaluated in a test phase. In this work, we investigate news source reliability inference as a source-level prediction problem. We propose a two-stage framework in which transformer-based language models first estimate the reliability of individual articles and subsequently aggregate article-level predictions to infer the reliability of previously unseen publishers. To approximate realistic deployment conditions, we enforce a strict publisher-disjoint evaluation protocol, ensuring that no publisher appears in both training and test sets. Experiments on 19,476 political news articles from 439 English-language publishers labeled with NewsGuard reliability ratings show that aggregation substantially improves robustness and performance, increasing accuracy from approximately 0.60 at the article level to 0.69 at the publisher level. Finally, we analyze how prediction errors vary across political orientations, revealing statistically significant associations between political leaning and misclassification patterns. Overall, our findings show that publisher reliability can be inferred from aggregated textual signals alone, supporting scalable and content-based approaches to automated news source assessment. --- ### Beyond Emotion Prompts: Fine-Grained Text-to-Image Generation Driven by Valence-Arousal-Dominance Authors: Minglang Li, Yueyue Fang, Xieping Gao Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24215v1 Although text-to-image models can accurately depict subjects and scenes, creators still struggle to specify the fine-grained emotions an image should convey without rewriting its content description. Natural language can suggest emotions, but it offers no control scale with stable meanings and ordered intensities. We propose EMOTRANS, which transforms psychologically grounded valence-arousal-dominance (VAD) coordinates into generation conditions that are independent of the content text and modulated across denoising stages, making emotional style a finely adjustable creative variable. To support this goal, we construct EMOVAD, an art-painting dataset that pairs objective content descriptions with separately collected emotional ratings from multiple annotators. We also coordinate emotional expression and content preservation through dual-branch training with a shared model. Objective and human evaluations show that the framework improves the accuracy of three-dimensional emotion control and produces perceptible, orderable continuous changes while maintaining competitive text alignment and image quality. This work provides a practical emotion-driven approach to image generation that extends objective content depiction to fine-grained emotional adjustment. --- ### ChartJudgeBench: Evaluating LMM Judges for Chart-to-Code Generation Authors: Lijian Wu, Henry Hengyuan Zhao, Zijian Zhang, Jiahao Tang, Jiajun Wu, Alex Jinpeng Wang Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24210v1 Building strong chart-to-code systems increasingly relies on reinforcement learning, whose effectiveness depends critically on the quality of the reward signal. Large Multimodal Models (LMMs) play a natural critical role in jointly assessing chart visual appearance and task requirements. They are therefore increasingly used as visual critics and reward models, yet their reliability as judges remains largely unexplored. To this end, we introduce ChartJudgeBench, a diagnostic vision-language benchmark for assessing LMM judges in chart-to-code workflows. It includes 1,003 Chart Perception Alignment (CPA) instances for pairwise chart comparison and 650 Chart Reasoning Judgment (CRJ) instances for binary Accept/Reject verification in Chart Reproduction and Chart Editing. Together, these tasks emulate the core judging decisions required in agentic refinement and RL-based chart optimization. Our evaluation of strong LMMs reveals four systematic limitations: (i) positional bias in pairwise comparison, (ii) a strong tendency to overpredict Accept, (iii) difficulty in matching visual styles and aesthetics, and (iv) an unexpected leniency bias in RL-trained models. These findings show that current LMM judges require explicit reliability validation before being used as critics or reward models in chart-to-code optimization. The code and data are available on ChartJudgeBench. --- ### Displacement Geometry Captures Platonic Shared Reality Across Models and Modalities Authors: Chenming Shang, Yujin Tang, Jun Jie Ou Yang, Ruize Xu, Adam Breuer, Nikhil Singh Published: 2026-09-21 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.24209v1 The Platonic Representation Hypothesis (PRH) claims that independently trained models converge on a shared statistical model of reality, yet recent work finds only weak pointwise similarity between models. In this paper, we show that what models share is not the location of samples in representation space, but the directions (displacement vectors) between them. Under a single orthogonal alignment--rotation and reflection only--these displacement vectors are substantially preserved across 44 independently trained vision and language encoders spanning modalities and asymmetric capability pairs, consistent with the PRH evidence. The samples' absolute positions are not, consistent with recent counter-evidence. Both arise from a single decomposition: representations split into a shared semantic component that is linearly aligned across models, and a private capability component that is not. We trace this geometry to concept-level structure: within a model, parent concepts are orthogonal to their child variation vectors; across models, concept displacements are parallel. Our theory falsifiably predicts (and experiments confirm) that fine-tuning preserves pointwise similarity but collapses displacement, and that relational distillation does the opposite. A major implication is that, because semantics align linearly but capabilities do not, capabilities can be imported from one model to another using a single cached forward pass through the source. We call this Shadow Casting. As a proof of concept, our SHADOWCLIP instantiation outperforms strong fine-tuned baselines at orders of magnitude less compute. A cache can be released alongside open model weights, letting one model's capabilities be downloaded and imported into any number of other models without fine-tuning. --- ### CoaG: Cylinders on a Grid for Coarse 3D Layout Control in Video Generation Authors: Zhangsihao Yang, Mengyi Shan Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24208v2 We ask how little geometry a person has to draw to control both where people stand and where the camera moves in a generated video. Our answer is a ground plane and one cylinder per person. A user draws a grid on the ground, places one cylinder where each person should stand, moves the cylinders and the camera over 81 frames, and the model renders a photoreal video in which the people occupy the cylinders' positions, move as the cylinders move, and are seen from the drawn camera. Appearance comes from a text prompt and a background reference image; layout and motion come from the geometry. Because no dataset pairs such a signal with video, we build the pairs ourselves: an automatic engine writes 2000 captions from a combinatorial seed, generates a clip for each with a text-to-video model, and lifts every clip back to its geometry with person tracking, background inpainting, an agentic ground-mask loop, feed-forward multi-view reconstruction and a plane fit, with no real footage and no manual labels. A LoRA on Wan2.2-Fun-Control trained on 1935 such tuples follows drawn layouts and camera paths on hold-out clips: the generated people match the cylinders' count, order, position and height, the text changes who they are, the reference image changes where they are, and dolly-in, orbit, pan and crane paths are followed, dolly-out only weakly. --- ### A principled approach for energy-efficient training via phase-aware GPU frequency tuning Authors: Miguel Braga, Júlio Pinto, Rahma Nouaji, Olivier Michaud, Bettina Kemme, Oana Balmau, Cláudia Brito, Ricardo Macedo Published: 2026-09-21 Categories: cs.DC, cs.LG arXiv: https://arxiv.org/abs/2609.24205v1 Modern AI model training imposes unprecedented computational demands, making it a key contributor to datacenter energy consumption. Yet a significant fraction of the energy consumed during training does not translate to useful computation due to bottlenecks throughout the training pipeline. We present PAFT, a phase-aware, dynamically adaptable GPU frequency tuning system that reduces energy consumption of training workloads with minimal performance overhead. The key insight behind PAFT is that bottlenecks represent an energy optimization opportunity, rather than purely a performance problem: when GPUs are bound to stall, PAFT opportunistically reduces their clock frequencies to match the pace of bottlenecked devices, saving energy without impacting execution time. PAFT achieves this by continuously monitoring pipeline behavior and applying fine-grained frequency adjustments, adapting to workload and system changes. Experiments conducted on twelve widely used models show that PAFT consistently outperforms all baselines, achieving energy savings of up to 46% with an average overhead of 4%. --- ### SAFe: Segment-guided Aggregation of Feature Densities for Anomaly-aware Segmentation Authors: Anja Delić, Jurica Runtas, Marin Oršić, Ivan Marković, Ivan Petrović Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24204v1 Visual segmentation systems encounter objects outside their training distribution during real-world deployment, hindering reliable autonomous systems that depend on scene parsing in the perception stage. Many recent methods address this by using self-supervised foundation models to train density estimators that yield low likelihood in anomalous image regions. Although promising, these methods suffer from poor feature semantics or they lack spatial consistency, both of which undermine critical downstream decisions. We address this problem with~\method, a generative method based on class-conditional density estimation over self-supervised representations. SAFe trains lightweight normalizing flows that produce class-conditional normalized likelihood estimates over frozen DINOv3 features. We combine density estimates from transformer features with density scores over multi-scale convolutional features to capture both global semantics and local detail. We introduce a method-agnostic post-processing step based on SAM3 that connects per-location likelihoods into spatially coherent segments while suppressing false positives, and enables instance-level anomaly detection without retraining. The post processing further distinguishes novel categories among anomalous objects by a similarity-based agglomerative clustering scheme. SAFe sets a new state of the art on the PANIC, OoDIS, SMIYC ObstacleTrack with strong performance on the ISSU benchmark. --- ### Opinion Leader Dynamics: How Sparse Attention Shapes Token Clustering Authors: Jingkun Liu, Yue Song Published: 2026-09-21 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.24202v1 Sparse attention reduces the quadratic cost of global self-attention while retaining strong empirical performance, but how its restricted interactions shape the evolution of token representations remains theoretically underexplored. Modeling tokens as particles on the unit sphere, we introduce opinion leader dynamics, a framework that identifies two mechanisms through which token groups converge internally while maintaining distinct limiting directions. In the explicit model, fixed representatives induce a potential that attracts tokens toward distinct local maxima. In the implicit model, disconnected interaction groups evolve toward separate consensus directions. We formulate both models as reverse Wasserstein gradient flows and establish exponential convergence under suitable conditions. We further connect these theoretical predictions to token evolution in frontier sparse-attention LLMs that motivate our framework. Across four benchmarks, Kimi-K3, MiniMax-M3, and DeepSeek-V4-Flash consistently exhibit clearer cluster separation and higher clustering scores than the dense-attention model GLM-4.7-Flash in projected token representations. These observations support the relevance of the predicted multiple-group structure to trained frontier LLMs, while finite-particle simulations illustrate the theoretical convergence behavior. Together, our results connect restricted token interactions to distinct group-level attractors, providing a dynamical account of how sparse attention can support alignment within groups while preserving separation between them. --- ### Vimarsha: Faithful ASR Evaluation for Indian Languages with Demographic Diversity, In-the-Wild Audio and Spelling Variations Authors: Kaushal Santosh Bhogale, Srija Anand, Sadakopa Ramakrishnan Thothathiri, Tahir Javed, Sshubam Verma, Mitesh M. Khapra Published: 2026-09-21 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.24199v1 Evaluation benchmarks for Indian language automatic speech recognition (ASR) suffer from two systematic biases: optimistic scores from clean, controlled audio conditions, and pessimistic scores from overly rigid transcription standards that penalize valid linguistic variations. We introduce Vimarsha, a 100-hour benchmark spanning all 22 scheduled Indian languages, designed to address both distortions. Vimarsha combines demographically diverse on-field recordings with carefully mined in-the-wild audio selected for acoustic difficulty, alongside a lattice of variations framework that encodes multiple valid transcriptions per utterance. Evaluations of 10 state-of-the-art ASR models reveal substantial shifts in model rankings under realistic conditions, geographic and demographic performance disparities, and systematic failure modes across speaking rates and acoustic environments. --- ### SKstars at SHROOM: Visions Agreement-Guided Ensembling of Zero-Shot and LoRA-Adapted Vision--Language Models Authors: Ali Athar, Imran Ahsan, Joon-Yong Jung Published: 2026-09-21 Categories: cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.24198v1 This paper describes the SKstars submission to SHROOM-Visions 2026, a shared task on fine-grained hallucination detection in large vision-language model outputs. The task requires systems to identify hallucinated character spans, assign hallucination categories, and provide confidence estimates for their predictions. Our approach combines zero-shot predictions from Qwen2.5-VL-72B-Instruct with those of a LoRA-adapted Qwen2.5-VL-7B-Instruct model. The outputs of the two models are integrated through a lightweight ensemble procedure, followed by span refinement and confidence adjustment. We evaluate the main system components on a small internal development subset and report the performance of the submitted system on the official English test set. SKstars achieved a Cor+Lbl score of 0.2902, ranking 15th among 29 teams, and obtained Cor and IoU scores of 0.3642 and 0.3151, respectively, ranking 18th on both metrics. The results show that combining a large zero-shot model with a smaller adapted model provides a practical framework for multilingual and fine-grained hallucination localization, while also highlighting the difficulty of transferring development-set improvements to hidden test data. Code and predictions: https://github.com/aliathar1401/SK-Stars-shroom-visions-2026 --- ### H-Spec: Parallel Speculative Decoding Without a Drafter-Side KV Cache Authors: Weifan Jiang, Krishna Teja Chitty-Venkata, Megan Flynn, Reed Meyerson, Zhenting Qi, Tianyu Wu, Eldar Kurtic, Minlan Yu, Alexandre Marques Published: 2026-09-21 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.24197v1 Speculative decoding losslessly accelerates large language model inference by having a lightweight draft model predict future tokens for verification by the target model. Recent block diffusion drafters further reduce drafting latency by predicting multiple tokens in parallel. However, existing block drafters project target hidden states at every input position into a separate drafter-side KV cache, incurring per-request memory and KV-write overhead that grow with concurrency; directly reusing target KVs in place removes this cache but fails to sustain draft quality throughout the block. We propose a hybrid target-context injection method that complements direct target KV reuse with target hidden states only at the last input position, requiring no separate drafter-side KV cache. Building on this design, we propose H-Spec, a hybrid Mamba-attention parallel drafter that consumes the two target-context sources through complementary modules. Mamba modules are initialized with projected last-token target hidden states, while attention modules reuse target KVs in place. Despite its recurrent formulation, Mamba's parallel scan allows H-Spec to preserve block-parallel drafting. Across three target models and diverse tasks, H-Spec improves over the best baseline by 5.0--13.3% in mean accepted length and 5.3--12.6% in batch-size-1 inter-token latency speedup. Under concurrent serving, H-Spec consistently achieves higher throughput while maintaining lower KV cache utilization than baselines across evaluated concurrency levels. --- ### LoopCD: Loop-wise Contrastive Decoding for Improving Reasoning in Looped Language Models Authors: Byeongho Yu, Junhyuk So, Eunhyeok Park Published: 2026-09-21 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.24196v1 Looped Language Models (LoopLMs) perform "latent reasoning" by recursively refining internal latent representations with shared weights, offering a more effective alternative to explicit verbal reasoning. Despite their effectiveness, we find that LoopLMs remain prone to loop instability: unstable refinement across iterations can produce localized uncertain "hard" tokens associated with reasoning errors. To address this, we propose LoopCD, loop-wise contrastive decoding that enhances the reasoning performance of LoopLMs by intervening on these tokens at inference time. Specifically, we exploit the internal dynamics of LoopLMs and contrast the logits from earlier iterations with logits from the last refined iteration to form the final sampling distribution. We find that this strategy is highly efficient, introducing only negligible inference overhead and requiring no additional training, while effectively improving reasoning performance by naturally refining reasoning-critical hard tokens. Extensive experiments show that our method improves the performance of recent representative LoopLMs across various reasoning tasks. --- ### When Residualization Helps an Audit: Format Effects, Slice Gains, and Their Limits Authors: Daein Weon, Dong Ho Kang Published: 2026-09-21 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.24194v1 Evaluation scores used around LLM systems -- including reward models, rerankers, and LLM judges -- can track surface form instead of the quality they claim to measure. When presented with a terse correct solution and a commented buggy solution for the same MBPP problem, a public preference reward model selects the correct one no better than a coin flip (0.507). Subtracting the predictable surface component from such scores is increasingly common, but removal alone does not yield a more valid measurement: the removed component may carry construct-relevant signal, and residualization cannot tell which is which. Under designed interventions -- unit-test labels with comment-only edits -- residualization attenuates the reward model's format effects by about 0.12 on both correct and buggy code, while the correct-versus-buggy margins move by less than 0.01. In observational NLI and QA settings, we freeze a held-out replication before scoring and re-evaluate it using labels from disjoint annotators; this supports only a narrower conclusion: better agreement with the construct labels on a pre-declared slice where a surface-only predictor errs, not a repaired score. Full-population agreement falls in every observational setting with a reported positive slice gain, and within-question ranking falls in every such QA setting. When construct and surface features are entangled, residualization can decorrelate a score while degrading construct alignment, and, in a controlled model, configurations just as damaging to construct alignment pass every pre-adjustment check, so no committed gate is a guarantee. We assemble these distinctions into a reporting protocol whose outcomes, refusal included, state what an adjusted score may be claimed to show: an audit-time diagnostic reported beside the construct-alignment cost it incurs, never a replacement for the raw score. --- ### Lightweight Pedestrian Head-Orientation Recognition Network for Safe Pedestrian-Vehicle Interaction Authors: Yuanzhe Li, Yidi Huang, Xiaotong Chang, Hounian Liu Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24193v2 Pedestrian head orientation recognition plays an important role in autonomous driving by providing valuable cues for understanding pedestrian attention and anticipating potential crossing behavior. However, reliable recognition in real-world traffic scenes remains challenging because pedestrian head regions are often captured at low resolution. To address this challenge, we propose a lightweight Low-Resolution Head Orientation Convolutional Neural Network (LRHO-CNN) for pedestrian head orientation recognition. We construct a new dataset by extracting pedestrian head images from multiple public datasets and manually annotating them into eight orientation categories. The collected images are systematically preprocessed and augmented to increase data diversity and better represent variations in illumination and image quality. The experimental analysis compares LRHO-CNN with three fine-tuned baseline models, namely ResNet-18, ResNet-34, and VGG-16. The results demonstrate that LRHO-CNN achieves the highest classification accuracy among the evaluated models. LRHO-CNN is further evaluated on the JAAD and PIE datasets, demonstrating its effectiveness in recognizing pedestrian head orientation in real-world traffic scenes and providing informative head-orientation cues that can support downstream pedestrian behavior and intention prediction. --- ### Benchmarking Off-the-Shelf Multimodal AI Models Against Dermatologists on Patient-Captured Skin Images Authors: Rian Dolphin, Laura Knowles Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24190v1 Artificial intelligence (AI) has advanced at a rapid pace in recent years. Initially, breakthroughs in large language models caught widespread attention. However, recent generations of frontier AI models have adopted multimodal capabilities as a first class citizen, with vision capabilities being central to that. In this paper, we evaluate three recently released models on the task of diagnosing dermatological conditions from patient-submitted images. The models chosen are at the low to mid tier in terms of pricing and thus represent a floor on current AI capabilities, not a ceiling. We evaluate AI performance relative to a panel of three certified dermatologists, who grade each image, and we present four interesting findings. Firstly, depending on the metric, the tested AI models are either on par or slightly trail humans in terms of inter-clinician agreement. Secondly, we find that asking AI models for a confidence rating produces poorly calibrated answers, meaning use of confidence thresholds should not be relied upon in a clinical setting. Thirdly, the effect of providing additional patient metadata is strongly model-specific, with one of the three models degrading on every metric considered. Finally, model cost is not predictive of performance. The best-performing model we tested costs on average $0.0045 per case. --- ### StenoVLA-3D: 3D-Aware Reasoning VLA for Navigation Through Gastrointestinal Stenoses Authors: Tamima Tabassum, Yiming Huang, Tianchun Wu, Changjing Liu, Zhiqing Tang, Chikit Ng, Beilei Cui, Liangjing Shao, Jiewen Lai, Hongliang Ren Published: 2026-09-21 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2609.24187v2 Autonomous endoscopic navigation requires the policy model to predict actions from texture-poor monocular observations, make safe control decisions, and retain evidence of lesions after they leave the field of view. Existing vision-language-action (VLA) models primarily rely on visual appearance and short-term context, limiting geometric grounding and episode-level reporting. We introduce StenoVLA-3D, a 3D-aware VLA framework for navigating through stenotic regions. We integrate point-maps into the Cosmos-Reason 2 backbone through learned geometry-gated fusion, and also propose a temporal state branch to model traversal progress. Our reasoning-and-action backbone predicts grounded reasoning with actions, while dedicated heads estimate stenosis shape and generate the final lesion report. We further introduce EndoCausal, an episode-level dataset with lesion annotations, actions, and temporally grounded reasoning. On 40 held-out recorded test episodes, StenoVLA-3D reaches 95.2\% semantic accuracy and 83.4\% action accuracy. On the physical 3-DoF endoscope, it attains 88.9\% and 77.8\% task success in esophageal and colonic phantoms (36 trials each), substantially outperforming the evaluated baselines. --- ### LIMIT: Less Is More for Instruction Tuning in Text-to-SQL Authors: Haoyuan Ma, Hengwei Liu, Linjuan Wu, Yongliang Shen, Weiming Lu Published: 2026-09-21 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.24186v1 Large language models have achieved remarkable progress on Text-to-SQL through reasoning-enhanced fine-tuning, yet existing approaches predominantly rely on massive instruction corpora under the assumption that scale drives performance. We challenge this paradigm by investigating a fundamental question: what is the minimal data requirement for effective Text-to-SQL instruction tuning? We propose LIMIT(Less Is More for Instruction Tuning in Text-to-SQL), a data-centric framework that demonstrates strong database reasoning can emerge from an extremely compact training set when examples are strategically selected. LIMIT operates through four stages: difficulty-aware filtering that identifies samples within the model's learning frontier, chain-of-thought synthesis with consistency-based selection, multi-dimensional quality scoring via LLM-as-judge, and genetic algorithm optimization that jointly maximizes schema coverage and sample quality. On the BIRD and Spider benchmark, LIMIT selects only 796 and 863 samples while achieving 100% table coverage, enabling Qwen3-8B to reach 69.1% and 88.9% execution accuracy.This result surpasses methods trained on 20 times more data and establishes a new state-of-the-art among open-source approaches. Our findings suggest that careful data curation, rather than scale, is the key to efficient Text-to-SQL learning. --- ### Efficient LLM Distillation for Bangladesh Legal Context: A Smartphone-Compatible Retrieval-Augmented Generation Model Authors: MD. Nafis Kamal, Mahadi Hasan Fahim, Talha Ridwan, Nadifa Zaman, Fariha Roushon Florin, Farig Yousuf Sadeque, Saadat Rafid Ahmed Published: 2026-09-21 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.24177v1 Legal information in Bangladesh is inaccessible to most citizens. Statutory text is English-only, trained lawyers are concentrated in urban centres, and cloud-dependent AI fails where mobile connectivity is unreliable, a setting in which hallucinated legal text causes direct harm. The system addresses statutory interpretation only; queries that require judicial precedent or case-law reasoning fall outside its scope. We target the statutory access gap by compressing a 9-billion-parameter Gemma-2 teacher into a 2-billion-parameter student through two-phase progressive knowledge distillation. Phase 1 performs supervised fine-tuning on 9,429 quality-gated legal question-answer pairs (65% acceptance from 14,514 generated queries); Phase 2 minimises sparse Kullback-Leibler divergence against the teacher's top-50 per-token logits at temperature tau = 4.0, implemented via QLoRA (4-bit NF4, rank-32 LoRA adapters). Prior legal language models target general legal English; this system specialises in Bangladeshi statutory law. Every response is grounded through hybrid retrieval combining dense semantic search (60%) and BM25 (40%) across 36,029 statutory passages from the Bangladesh Constitution and national legislation. On a 50-query English benchmark, the distilled model reaches ROUGE-L 0.4715 and BERTScore F1 0.5679, a 103% ROUGE-L and 143% BERTScore gain over the retrieval-augmented undistilled baseline (ROUGE-L 0.2323, BERTScore 0.2340). The adapter quantises to 1.6 GB (GGUF Q4_K_M) and runs at 4-8 tokens per second on a Pixel 6 with no network access. Cross-lingual evaluation on 50 Bangla queries yields ROUGE-L 0.4083 and BERTScore 0.8133, showing effective retrieval from Bangla input against an English-only corpus. In a single-evaluator pilot, a practising lawyer rated 50 responses at a weighted mean of 4.16/5 (90% rated 4 or 5), supporting utility beyond text-overlap metrics. --- ### CREDO: Variance-Guided Rubric Evolution for Replay-Corrected Credit Assignment Authors: Xuchun Hu Published: 2026-09-21 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.24174v1 Long-horizon language agents receive sparse terminal feedback, while intermediate rubrics provide structured but potentially misspecified assessments of progress. In resettable training environments, counterfactual continuation rollouts can measure local credit, but exhaustive replay is costly. We propose Credo, a framework that couples evolving semantic rubrics with selective, execution-based credit correction. A frozen judge maps visible transitions to rubric features, and a credit head predicts the change in expected terminal reward associated with the realized transition. Independently sampled two-sided replays correct prediction residuals using their recorded inclusion probabilities. We derive conditional unbiasedness and a variance decomposition that connects two design choices: which rubric features to retain, and where to allocate a fixed expected replay budget. The resulting criterion weights prediction errors by policy-score sensitivity and missing replay coverage; its allocation rule additionally accounts for continuation cost. We also describe a practical mixture with terminal leave-one-out advantages and distinguish its clipped, token-normalized PPO implementation from the ideal policy-gradient estimator. This preliminary report provides the method, proofs, an exact finite-model audit, and a controlled evaluation protocol. It makes no claim of empirical superiority on language-agent benchmarks. --- ### LegendBench: A Diagnostic Benchmark for Legend Understanding with Counterfactual Interventions Authors: Xinnuo Zhang, Zhike Tang, Jing Xu, Haoyuan Zhao, Weikai Yang Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24172v1 Legends are fundamental to chart understanding, as reliable interpretation requires correctly binding legend entries to corresponding visual marks. While vision-language models (VLMs) are increasingly applied to chart understanding, their legend understanding is poorly diagnosed by aggregate accuracy, which can be satisfied by superficial shortcuts and confound legend-specific errors with other reasoning failures. To enable fine-grained diagnosis and controlled testing, we introduce LegendBench, a parametric benchmark and generation pipeline that produces targeted legend-centric test cases. LegendBench contributes (1) a capability-task taxonomy spanning legend parsing, legend grounding, legend-conditioned reasoning, and legend-aware abstention to localize failures, and (2) counterfactual group generation, where each base chart yields multiple variants under controlled legend interventions to probe model invariance and sensitivity. Using LegendBench, we evaluate both general-purpose VLMs and specialized chart models and generate their capability profiles, revealing persistent bottlenecks in reliable legend-to-mark binding and counterfactual consistency. We then use these capability profiles to guide targeted fine-tuning, demonstrating that bottleneck-specific interventions can effectively close the localized capability gaps and generalize to unseen data. We further leverage our counterfactual design to conduct fine-grained diagnostic experiments, analyzing encoding-channel effects, legend-order shortcuts, and abstention under varying visibility. --- ### An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond Authors: Wenbo Zhang, Kaixuan Wang, Yutao Ouyang, Xiaoyu Huang, Liyang Li, Kailun Su, Weiyang Jin, Wenhao Chai, Haotian Liang, Zhiyang Dou, Yue Chen, Tianxing Chen Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24170v1 Embodied AI systems are often organized into System 1 and System 2. System 1 is typically a pretrained policy that generates actions at high frequency, whereas System 2 is often instantiated as a vision-enabled language model for high-level planning. We ask whether a large language model (LLM) can act as the policy for robot manipulation without task-specific finetuning. We call this setting LLM as policy. We evaluate three LLMs on all 42 RoboDojo tasks and compare their scores with 40 public policies. Astra and GPT-5.5 use the official 50-episode-per-task protocol; DeepSeek-Flash uses 10 episodes per task. GPT-6 Astra achieves 22.48% average success rate and 28.97 Score over 2,100 trials, ranking above every public entry. Yet GPT-5.5 and DeepSeek-Flash reach only 0.88% and 1.92% average success rate with the same post-processing. We find that Astra exhibits a sharply polarized capability profile. It generalizes well to tasks that require semantic understanding but not high-precision control. In contrast, it performs poorly on tasks that require precision, dynamic control, or complex bimanual coordination. In-context experiments show no aggregate benefit from one-shot demonstrations, while selected interaction traces show within-episode corrections under perturbations. Overall, the evaluated LLMs vary substantially in manipulation performance. Astra stands out and provides initial evidence for the potential of a general-purpose manipulation model, although reliable precision and dynamic control remain limitations in the evaluated setting. --- ### APEXA: Execution-Integrity Enforcement for Multi-Agent LLM Automation of Synchrotron Data Reduction Authors: Pawan K. Tripathi, Hemant Sharma, Andrew Chuang, Mathew J. Cherukara Published: 2026-09-21 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.24165v2 LLM agents can make expert scientific workflows accessible through natural language, but a plausible response does not establish that the requested computation ran. This matters at synchrotron facilities, where calibration and integration are multi-step operations over terabyte-scale datasets. We present APEXA, a deployed framework coordinating 61 tools for synchrotron data reduction. A deterministic execution-integrity guard prevents it from returning results for tool calls that did not execute, and the same tool layer enforces motor safety: across 50 adversarial scenarios and four models, it produced 0/200 violations against a simulated EPICS controller, versus 15/200 for a prompt-only baseline. We also introduce APEXA-Bench, 58 facility tasks organized by scientific correctness and physical consequence. On real APS beamline data, one prompt recovered detector geometry and integrated a full attenuation and exposure sweep. Trustworthy scientific agents need deterministic validation of execution, not model-side assurances. We release the framework, benchmark, and traces. --- ### MCP-GRANITE Benchmark: GRANularity Interface TEsting for MCP-Based LLM Agents Authors: Demetris Paschalides, Moysis Symeonides, George Pallis, Marios D. Dikaiakos Published: 2026-09-21 Categories: cs.DC, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.24161v1 As LLM agents increasingly interact with external tools through standardized protocols such as MCP, tool-interface design becomes a critical yet underexplored factor. How funψtionality is decomposed into tools affects whether an agent can select the right tool and construct valid arguments. This choice is especially consequential at the edge, where resource constraints limit which models can run locally and scaling up is often not an option. We present MCP-GRANITE, an open-source extensible benchmark framework that treats tool-interface granularity as a controlled variable for MCP-based agents, evaluated under edge and IoT scenarios. It comprises 81 multi-step scenarios across 9 domains, instantiated at 4 granularity levels from fine-grained primitive tools to a single tool. We evaluate 9 locally deployed models (268M-20.9B parameters) across 8,748 trials using task completion, tool selection F1, argument accuracy, latency, and resource-usage metrics. Results show that a 4-tool interface offers the best trade-off, improving task completion by 16.4% over fine-grained primitives and 33.6% over a single monolithic tool, while nearly doubling argument accuracy. Model size is only weakly correlated with task completion and strongly with latency, while its association with argument accuracy is less robust, and a 3.2B model at the optimal granularity outperforms a 20.9B model at a mismatched one. These findings identify tool-interface granularity as a key design parameter for MCP-based agents. --- ### Relightable 3D Avatar Reconstruction with Semantic-Adaptive Motion-Illumination Responses Authors: Jiankuo Zhao, Xiangyu Zhu, Jijie Li, Baiqin Wang, Shukai Chen, Zhen Lei Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24158v1 Reconstructing expressive and relightable 3D head avatars from monocular videos remains challenging in computer vision, as it requires accurate modeling of both non-rigid facial motion and illumination-dependent appearance. Existing Gaussian avatar methods commonly rely on globally coupled representations, in which Gaussian primitives share a unified motion or illumination response model. Such uniform modeling neglects the distinct motion patterns and material/reflectance properties of different facial semantic regions, thereby limiting fine-grained animation accuracy and reducing relighting plausibility. To address this limitation, we propose SAMIRA, a 3D Gaussian avatar framework for semantic-adaptive motion-illumination response modeling. For motion response modeling, the Semantic-Adaptive Motion Response module rasterizes current-to-reference mesh displacements into a topology-consistent UV space and leverages facial semantics to route displacement features through semantic-specific modulators, predicting localized Gaussian geometric residuals beyond coarse mesh binding. For illumination response modeling, the Semantic-Adaptive Illumination Response module learns compact diffuse and specular response factors for each facial region, allowing Gaussians in different regions to adapt their illumination responses to novel environment lighting. These response factors are incorporated into deferred physically based shading, providing a lightweight approximation of semantic-dependent illumination effects. Extensive experiments on self-reenactment, cross-reenactment, and relighting demonstrate that SAMIRA improves both fine-grained expression reconstruction and relighting realism over existing methods. --- ### TAC-Time: Texts as Channels For Multimodal Time Series Forecasting Authors: Jiayi Liang, Xiaotian Gu, Xinyu Xie, Yuanbin Wu, Xiaoling Wang Published: 2026-09-21 Categories: cs.CL, cs.AI, cs.LG, cs.MM arXiv: https://arxiv.org/abs/2609.24156v1 Most existing time series forecasting methods rely solely on numerical observations, overlooking rich contextual information from auxiliary texts. Recent multimodal approaches attempt to incorporate textual signals, but they often treat text as static features or use large language models as forecasting backbones, limiting their ability to capture temporal dynamics and increasing computational cost. To address these challenges, we propose TAC-Time, a unified framework that transforms textual information into additional temporal channels. By modeling text features jointly with numerical sequences in a shared temporal backbone, TAC-Time preserves temporal continuity and periodic structures while remaining efficient and scalable. This formulation also enables systematic interpretability analyses. We show strong cross-modal dependencies through attention and frequency-domain analyses, and identify predictive textual signals whose correlation-aware alignment yields partial forecasting improvements. Extensive experiments on real-world multimodal benchmarks demonstrate that TAC-Time outperforms prior methods. --- ### Graded-Relevance Composed Multimodal Retrieval for E-commerce Visual Search at Scale Authors: Anubhav Gupta, Hrushikesh Mohapatra, Prijith Chandra, Asish Mohapatra, Anuj Garg, Arvind Maan, Sudip Datta, Venkat Bulusu, Sitesh Kumar Jalan Published: 2026-09-21 Categories: cs.IR, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.24152v1 Visual search on large e-commerce catalogs must serve both "similarity" queries that ask for items resembling an uploaded image and "modifier" queries that comprise an image and text describing a desired modification (e.g. a color change or style swap). The latter is the setting known as composed image retrieval (CIR). Existing CIR methods, however, treat relevance as binary and train on triplets with a single positive target - a poor fit for real catalogs where many candidates partially satisfy a user query and ranking across that partial-match spectrum drives the customer experience. We propose a methodology for training CIR retrievers on graded relevance, consisting of: (i) a VLM to curate training data, generating both queries (object detection + modifier synthesis) and 4-level relevance labels without manual annotation, (ii) an iterative relevance-feedback loop that expands the training set by mining hard negatives from the in-training retriever, and (iii) a hierarchy-aware angular objective to train the retriever directly on the graded labels rather than collapsing them to a binary split. We call this methodology GradCIR and instantiate it on a PaliGemma2 bi-encoder trained on 3.5M graded pairs curated from raw Walmart catalog data. A controlled graded-vs-binary ablation isolates the supervision granularity and shows lift of 4.9%-5.9% in NDCG@10. The same recipe applied to other multimodal encoders lifts early-fusion backbones by up to 8.5% NDCG@10. On the public FashionIQ benchmark, GradCIR (applied to PaliGemma2) reaches 0.6703 average recall when fine-tuned, slightly ahead of the strongest peer-reviewed supervised baseline we compare against, and matching or exceeding all published CLIP-L-class zero-shot CIR methods. The system is deployed in production at Walmart, where it's serving live visual-search user traffic. --- ### STAR: Scene- and Task-Aware 4D Radar Preprocessing Towards End-to-End Cognitive Radar Authors: Seung-Hyun Song, Dong-Hee Paek, Seung-Hyun Kong Published: 2026-09-21 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.24151v1 Four-dimensional (4D) Radar has emerged as a key sensor for environmental perception, providing range, azimuth, elevation, and Doppler measurements while remaining robust to illumination changes and adverse weather conditions. However, conventional Radar preprocessing methods, such as constant false alarm rate (CFAR) detection, select measurements primarily based on signal-level criteria and may therefore discard information valuable for downstream perception during point cloud generation. In addition, existing 4D Radar perception pipelines typically optimize Radar data processing and downstream perception independently, preventing task objectives from directly guiding the preprocessing stage. To address these limitations, we propose a Scene- and Task-Aware Radar (STAR) Preprocessor together with an end-to-end training framework. The STAR Preprocessor incorporates scene context and downstream task objectives to generate task-relevant Radar points, enabling the Radar representation to be optimized directly for perception. On the K-Radar benchmark, the proposed method achieves 74.3 AP, outperforming the previous state of the art by 5.6 AP points. Furthermore, applying the task-relevant points generated by STAR to various existing 3D detectors improves detection performance in most evaluation settings and yields an overall positive average gain over point clouds produced by conventional preprocessing. --- ### Acceptance-Aware Draft Model Training for Speculative Decoding Authors: Tianhua Xia, Mugilan Ganesan, Yifei Feng, Haiyu Wang, Maximilian Egger, Sai Qian Zhang Published: 2026-09-21 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.24150v1 Speculative decoding accelerates large language model (LLM) inference by using a lightweight draft model to generate multiple candidate tokens that are verified by the target model in a single forward pass. Its speedup is largely determined by the acceptance length, yet existing draft-model training methods mainly optimize cross-entropy or Kullback-Leibler (KL) divergence as proxies. These objectives encourage distribution matching but do not directly optimize acceptance length, and the acceptance mechanism also differs between greedy and sampling-based decoding. In this work, we propose acceptance-length-aware training losses that directly optimize the expected number of accepted tokens within a speculative window. For greedy verification, we derive an expected accepted length (EAL) loss that explicitly maximizes expected acceptance length. For sampling-based decoding, we introduce a window total variation (WTV) loss that optimizes the overlap between temperature-scaled draft and target distributions while accounting for sequential acceptance dependencies. Both objectives can be further combined with a group-relative reinforcement learning stage (GRPO) using simulated acceptance length as the reward. Experiments across different target and draft models, tasks, and decoding settings show that our losses consistently improve acceptance length over KL-based training. WTV provides particularly strong gains under sampling-based decoding, while EAL better matches greedy verification. These results show that directly optimizing the acceptance objective, with losses tailored to the decoding mode, is more effective than conventional distribution-matching objectives. --- ### Mind or Message? Auditing Theory of Mind in Multi-Agent Social Simulation Authors: Cong Li, Cheng Chen, Thomas Fung, Alex Rossi, Yi Li Published: 2026-09-21 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.24146v1 Language model agents are increasingly used to simulate social interaction, and the resulting transcripts read as though the agents understand one another. We ask whether that appearance rests on a model of the partner's mind or on the surface record of what the partner said. We build a social simulation in which both questions have exact answers: 40 multi-issue negotiations whose hidden preference weights and whose full Pareto frontier are known by construction. Two model families negotiate across 160 dyads, every transcript is frozen before any measurement, and 2880 counterfactual probes then hold the evidence byte identical while moving one factor at a time: the reader's own stake, the partner's tone, an identity label, and the order of recursion. The agents are socially fluent and economically poor. They reach agreement in 96.2% of dyads with 0 protocol failures, yet only 0.7% of deals land on the Pareto frontier, they leave 20.5% of the available joint value unclaimed, and they miss the one issue on which their interests are perfectly aligned in 76.6% of deals; on the frontier and on that aligned issue, a package drawn at random from the set both sides would accept does as well. The probes locate the failure. Swapping only the reader's own payoff sheet, while the partner's words and offers stay identical, moves the inferred top priority by 15.0 percentage points, which is egocentric projection rather than inference, while a tone rewrite moves it by 5.3 percentage points and an identity label by 0.0. Most tellingly, an agent predicts what its partner believes about it 72.5% of the time while that partner's belief is itself correct only 51.2% of the time: the agents track the conversation far better than they track the mind behind it. --- ### Luck Is Not Skill: When Do Paired Rollouts Help Group-Relative RL of LLM Agents? Authors: Nazmus Sakib Published: 2026-09-21 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.24144v1 Group-relative reinforcement learning compares rollouts of the same prompt, but independent environment noise can obscure these comparisons. We study paired rollouts, which share an event-keyed noise schedule within each group while preserving each rollout's marginal distribution. Pairing removes the between-schedule component of reward-contrast variance, but need not reduce gradient variance. For one-sided grader noise, we derive an exact condition for reduction and give a counterexample in which reward contrasts improve while gradient variance increases. A controlled study trains a 2B tool-use agent under tool faults and grader flips, with three seeds per design. The protocol was registered with a disclosed, previously completed pilot. Under tool faults, pairing improves final noisy-test success by +5.1 percentage points on average, with all three seed differences positive, but misses the registered learning-curve criterion. The criterion is also missed under grader flips: the validation-AUC difference is +0.003 (95% interval (-0.029, +0.033)). A gradient probe on eight distinct checkpoints from two fault-trained trajectories finds lower mean-centered covariance traces under both noise types: 21 to 30% for grader flips and 40 to 63% for tool faults. These finite-sample measurements support the variance mechanism without establishing a general learning-speed benefit. The results distinguish improving reward comparisons, reducing estimator variance, and improving learning. --- ### CLOOPD: Closing the Learner Loop in On-Policy Distillation Authors: Keye Zheng, Hanyu Li, Zhan Cheng, Yuan Gao Published: 2026-09-21 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.24141v1 On-policy distillation (OPD) pays twice for each fresh batch: the student generates trajectories and a stronger teacher scores them. Existing methods improve which trajectories are scored and how the teacher signal is constructed, but usually consume it with one actor update. We introduce CLOOPD, a closed-loop framework separating teacher-signal acquisition from student-side realization. CLOOPD selects an adaptive $α$ waypoint inside a KL envelope, freezes the scored batch and its advantages, re-forwards the student after each actor pass, measures realization, and allocates actor work under a separate token budget. The framework includes deterministic two- and three-pass policies, token-priced CLOOPD-TPMR, and a budget-matched control. Across six 300-step runs on an 8-H20 node, every CLOOPD policy improves the one-pass TOP-D anchor at comparable teacher-token scale: macro accuracy rises from 15.41 to 17.78 with CLOOPD-Fixed2 and 19.36 with CLOOPD-Fixed3. At step 100, CLOOPD-Fixed3 reaches 15.35, nearly matching TOP-D at step 300 while using 67.2% fewer teacher-scored tokens and 28.0% fewer GPU-hours. Earlier 8-A100 ablations show adaptive $α$ eliminates observed trust-envelope violations; a third pass adds headroom. These results position CLOOPD as a framework for budgeting how fully students learn from teacher-scored tokens. --- ### P2Flow: Phoneme-aware Progressive Flow Matching for Extreme Speech Super-Resolution Authors: Ningyuan Yang, Yize Li, Pu Zhao, Diego A. Cuji, Kanad Sarkar, Ryan M. Corey, Xue Lin, Andrew C. Singer Published: 2026-09-21 Categories: eess.AS, cs.LG, cs.SD arXiv: https://arxiv.org/abs/2609.24138v1 Generative models have recently demonstrated considerable promise in speech super-resolution (SSR). Nevertheless, the majority of existing work has concentrated on standard or versatile SSR configurations, leaving the extreme setting with severely limited spectral inputs largely unexplored. In this regime, current approaches exhibit marked performance degradation, underscoring the need for dedicated solutions. To bridge this gap, we introduce P2Flow, a phoneme-aware progressive flow matching (FM) framework designed for extreme SSR with three main strategies. First, our model leverages phonetic information to reconstruct missing spectral components. Furthermore, it employs a progressive architectural design that hierarchically restores distinct frequency regions. Finally, we incorporate post-training of the vocoder to enhance overall waveform fidelity. Extensive experiments are conducted on the TIMIT and VCTK datasets under both 1 kHz to 16 kHz and 2 kHz to 16 kHz settings, demonstrating that P2Flow yields state-of-the-art results across multiple evaluation metrics. --- ### Data Agents: Agentic Data Systems Authors: Guoliang Li, Peiyao Zhou, Xuanhe Zhou, Ji Sun, Yuyu Luo, Ju Fan Published: 2026-09-21 Categories: cs.DB, cs.AI, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.24137v1 Traditional data systems face profound limitations in the AI era, relying on human-crafted pipelines, lacking semantic understanding of heterogeneous data, and operating through rigid, reactive processing. To address these challenges, we propose a new paradigm called the Data Agent, designed to manage, process, and analyze data with minimal human intervention. Data agents autonomously execute a wide range of data-related tasks, transforming traditional data systems by shifting from manual design to autonomous orchestration, from literal manipulation to semantic interpretation, and from reactive to proactive processing. Our Data Agent system includes six components: semantic data organization, semantic operators, agentic pipeline orchestration and optimization, feedback-driven refinement, memory management, and proactive adaptation. Building on this foundation, we also develop two specialized agents: the data analytics agent and the data science agent. Experiments on real benchmarks demonstrate significant performance gains of our data agent over state-of-the-art methods. We identify open challenges to guide future research in building fully autonomous data systems. --- ### The Visual Target Matters: Learning across the Visual Hierarchy for Brain-to-Image Retrieval Authors: Ye Wang, Haokun Ren, Hong Yu, Ruirui Li, Xiao Li, Ke Liu, Wei Wu Published: 2026-09-21 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.24136v2 Brain-to-image retrieval seeks to identify the visual stimulus that elicited a non-invasive neural response. Candidate images are typically represented by pretrained vision models, whose internal representations vary in abstraction across depth. Existing methods usually train the neural encoder to recover a fixed final-layer visual target. Under this formulation, the visual hierarchy is reduced to a single prescribed endpoint, preventing representations at other depths from directly shaping the visual target. This limitation motivates learning how information across visual depths should contribute to the retrieval target. To this end, we introduce NeuroGlyph, which learns a trial-independent visual target from multiple depths of a frozen visual backbone. NeuroGlyph decomposes the target into factor-specific subspaces. Each subspace learns an image-conditioned allocation over visual depth. The resulting subspaces are fused into a single embedding for retrieval. Across THINGS-EEG and THINGS-MEG, NeuroGlyph outperforms final-layer supervision in all controlled comparisons. It also surpasses the post hoc best fixed-layer oracle in three of four comparisons. Parameter-matched ablations support both factorized target construction and image-conditioned depth allocation. Under comparable 200-way retrieval protocols, NeuroGlyph achieves the strongest system-level performance in six of eight reported metrics. These results support learning retrieval targets across the visual hierarchy rather than prescribing one visual depth. --- ### Self-Healing Harness for Runtime Oversight of Agent Self-Modification Authors: Sina Tayebati, Divake Kumar, Nastaran Darabi, Ranganath Krishnan, Amit Ranjan Trivedi Published: 2026-09-21 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.24130v1 LLM agents can change their own future behavior, raising a basic control question of which self-generated changes should be allowed to persist. We formulate this as admission control for self-modification. The agent may propose changes to its operating instructions, while an external runtime gate controls persistence. We implement this principle as a model-agnostic self-healing harness that runs a Detect, Notice, Heal, Validate loop around an otherwise unmodified agent. The agent authors candidate behavioral rules in an external workspace, where they receive provisional execution authority during evaluation and acquire persistent cross-episode authority only after measured improvement on the triggering failure without regression beyond a fixed margin on protected cases. Replay provides matched evidence when available, forward trials provide a weaker fallback, and a corpus-level guard re-tests the accumulated active rule set. Across 16 matched Baseline and Harness runs spanning AppWorld, Terminal-Bench, and $τ^2$-Bench, the gate rejected 383 replay-decided proposals. Of these, 211 (55%) improved their triggering failure while degrading a case that previously worked. This shows that locally beneficial self-modifications can introduce collateral regressions often enough to materially affect gate decisions, providing direct empirical motivation for external admission control. Task-completion score is higher under the Harness in all 16 pairs, with two paired bootstrap intervals excluding zero, while repeated-trial reliability is higher in 12 pairs, tied in 4, and lower in none. Because adaptation modifies the policy-inducing context while leaving model weights fixed, admitted changes remain inspectable, reversible, and compatible with closed-weight models. --- ### OSCAR: Order-aware Scoring and Calibration for AI Rankings Authors: You Liu, Yue Liu, Quanchao Lu, Nick Shipilov Published: 2026-09-21 Categories: stat.ML, cs.LG, stat.AP arXiv: https://arxiv.org/abs/2609.24128v1 Judge-specific sensitivity is useful for aggregating pairwise LLM evaluations, but its interpretation depends on which systematic presentation effects the ranking model includes. We introduce OSCAR, an order-aware framework for scoring and calibrating AI rankings, and study position as one such effect. In released judgments from 18 evaluators, the all-response A-minus-B score difference ranges from $-63.11$ to $98.31$ percentage points. Matching question text, response texts, candidate identities, and judge within the released table gives an overall difference of $24.22$ points (95% interval $(22.90,25.54)$), conditional on the released text mapping. A controlled calculation isolates the potential consequence: with true sensitivity fixed at one, omitting a position intercept of four reduces the population-optimal slope to $0.0771$. We extend sensitivity-based ranking with judge-specific position, length, and family terms, characterize local omission-induced displacement and an identification failure, and propagate prompt-cluster uncertainty to adjusted comparisons. Across four released datasets, position provides the largest stand-alone predictive improvement. Refitting bootstrap comparisons show more selective gains from the full model over position-only adjustment. In dependent binary simulations, adjusting both the mean and covariance yields 94.4--95.2% coverage; correcting either alone is insufficient. At $N=10{,}000$, OSCAR reduces mean neutral-target RMSE from $0.1158$ under the sensitivity-only model to $0.0237$. --- ### Action-Slot: Structured Action-Centric Representation Learning for Multi-Agent Atomic Activity Understanding Authors: Yu-Ho Chang, Chi-Hsi Kung, Yi-Hsuan Tsai, Yi-Ting Chen Published: 2026-09-21 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.24127v1 Atomic activity understanding aims to recognize and localize structured traffic behaviors that jointly encode motion patterns and their grounding in road topology. Unlike conventional action recognition, atomic activities are multi-agent, multi-label, and topology-aware: multiple activities co-occur while many agents remain inactive. We introduce Action-Slot, a structured action-centric representation learning framework. Slot attention is widely used for object-centric decomposition, but its permutation-invariant design and object-level inductive bias are misaligned with atomic activity semantics. We reformulate slot learning as structured activity decomposition through three designs: (1) category-aligned action slots that anchor slots to predefined activity categories, (2) parallel spatio-temporal slot updating for holistic video-level reasoning, and (3) background and negative-slot regularization that enforces competition between foreground activities and irrelevant regions. Together these establish an activity-centric inductive bias that disentangles concurrent and asynchronous activities directly from raw video. Beyond recognition, the learned representations encode transferable spatio-temporal grounding signals. We further propose an attention-difference-based pseudo mask selection framework that suppresses false positives by measuring attention changes before and after candidate region removal, enabling weakly supervised localization without dense annotations. To support systematic evaluation, we introduce TACO, a balanced synthetic dataset with full atomic activity coverage and pixel-level annotations. Experiments on OATS, TACO, and annotated nuScenes show superior recognition, strong sim-to-real transfer, and state-of-the-art weakly supervised localization. --- ### Model-Agnostic Feature Selection via LOCO-Guided Adaptive Minipatch Sampling Authors: Xuhui Liu, Lili Zheng Published: 2026-09-21 Categories: stat.ML, cs.LG, stat.ME arXiv: https://arxiv.org/abs/2609.24126v1 Black-box machine learning models increasingly deliver strong predictions, but extracting useful information from them, such as a set of important features, remains challenging. Existing model-agnostic methods primarily estimate feature importance or conduct inference on it rather than directly selecting features, whereas many feature selection methods are model-specific or rely on the model-X assumption. We introduce LOCO-guided Adaptive Minipatch Sampling (LAMPS), a model-agnostic ensemble framework that uses any black-box regression algorithm as its base learner to select features important for predicting the response. The base learner need only produce predictions and need not perform feature selection itself. LAMPS operates within a minipatch ensemble framework that subsamples both observations and features, allowing leave-one-covariate-out (LOCO) feature importance scores to be easily computed. It adaptively concentrates minipatch sampling on features with high LOCO scores while maintaining exploration. The resulting sampling probabilities rapidly separate signal from noise features after a few iterations, enabling selection through simple thresholding. We establish that LAMPS achieves exact feature selection in high-dimensional settings, provided that the base predictive models are sufficiently well trained on average. Extensive experiments on synthetic and real data show that LAMPS outperforms state-of-the-art feature selection methods, with particularly strong performance in the presence of correlated features. --- ### Positive Pair Geometry Matters: Optimal Transport for Contrastive Learning of Visual Representations Authors: Akshit Nanda, Shahzad Ahmad, Ram Prasad Padhy Published: 2026-09-21 Categories: cs.CV, cs.CG arXiv: https://arxiv.org/abs/2609.24125v1 Contrastive self-supervised learning has achieved strong performance by learning representations from multiple augmented views of the same image. However, most existing methods construct positive pairs using independently sampled stochastic augmentations, which may alter semantic content and ignore the intrinsic geometry of the data distribution. In this work, we propose OTCLR, an optimal transport-aware framework for contrastive learning representations that generates geometry-consistent positive samples. Instead of directly contrasting two randomly augmented views, we construct intermediate views between the original image and its augmented variants through entropic optimal-transport displacement interpolation. These transport-interpolated samples serve as positive views that better preserve image structure while explicitly modeling spatial distributional geometry. To further promote smooth representation learning, we evaluate auxiliary Sinkhorn regularization terms that encourage transport-interpolated views to remain consistent with their endpoint images. The proposed method can be incorporated into standard contrastive learning pipelines without modifying the encoder architecture. Experiments on multiple benchmark datasets show that our approach improves representation quality and transfer learning performance compared with conventional augmentation-based contrastive learning baselines. --- ### ActiveArena: Benchmarking and Understanding Active Perception in Robotic Manipulation Authors: Yibo Li, Enshen Zhou, Rui Chen, Yanjun Ding, Mengzhen Liu, Yi Han, Jiabo Zhan, Lipeng Wang, Shanghang Zhang, Lu Sheng Published: 2026-09-21 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.24124v2 Active perception and manipulation are crucial for robots to interact with complex scenes. Existing benchmarks struggle to evaluate how robots effectively acquire and maintain information in memory in an active manner. To this end, we introduce ActiveArena-Sim, an active-perception simulator with controllable viewpoints and large-scale workspaces as the foundation. Built on this, we propose ActiveArena-Bench, which comprises 35 tasks across 5 fine-grained categories, covering visual exploration and interactive information acquisition. Each task is difficult to solve from passive observations alone, requiring multi-round evidence acquisition and memory-based reasoning. The benchmark provides rich memory annotations, standardized training data, and ID/OOD protocols featuring disjoint scenes, unseen distractor configurations, and novel backgrounds. Moreover, we present ActiveArena-VLA, a modular suite of 13 vision-language-action configurations for controlled studies of memory writing, memory capacity, proprioceptive state, subtask supervision, and high-level planning in active perception. Benchmark results reveal a substantial ID-OOD gap: uniform memory sampling, increased memory capacity under reliable write policies, proprioceptive inputs, and subtask supervision improve OOD generalization, while planner-guided memory management and decision-making achieve performance close to the best-performing configuration using only sparse memory. ActiveArena thus provides a unified testbed to develop and diagnose models for active perception and manipulation. ---