The paper introduces Independent Token Sampling (ITS), a query‑efficient method for detecting memorized training data in diffusion large language models (dLLMs). ITS selects token sets with weak internal dependency by approximating cumulative conditional mutual information using an attention‑derived pairwise dependency proxy and promotes diversity across sampling rounds. Experiments show ITS outperforms existing baselines, improving AUC by 0.18 on the ArXiv dataset while remaining effective under limited query budgets.
By Hongyao Yu, Tianqu Zhuang, Ziyuan Xu, Hao Fang, Jiaxin Hong, Bin Chen, Shu-Tao Xia
The paper introduces RIBA, a reinforcement‑learning inspired black‑box adversarial attack that generates perturbations for neural networks with fewer queries than existing methods. RIBA achieves a 25.4% reduction in median queries on ResNet‑18/Cifar10 and a 22.5% reduction on Vit‑B/16/ImageNet, while matching white‑box attack performance on an adversarially trained model.
By Florian Krone, Elena Hoemann, Sven Hallerbach
The paper introduces MMSAFE, a multi-layer framework designed to identify safety-degrading data in multilingual large language models. It shows that safety signals are distributed across multiple layers and only partially shared across languages, unlike the single-layer assumption used in monolingual settings. Experiments demonstrate that MMSAFE reduces harmful-response rates by 60% compared to random filtering and outperforms the best single-layer baseline across various models, languages, and safety benchmarks.
By Jiakun Li, Guowei Song, Sijia Li, Xingwei He, Hongzheng Chai, Yuan Yuan
The paper investigates whether large language models (LLMs) make decisions in ways that mirror human cognition. Using a new 140,000-trial product choice benchmark, the authors test 12 open‑source and commercial LLMs to see if their context sensitivity aligns with a cognitive economic theory that relies on problem categorization and attention allocation. While context prompts human‑like shifts in choice and problem categorization, it does not consistently reweight attention between features such as price and quality, and neither scaling nor chain‑of‑thought reasoning produces human‑like behavior. The findings indicate that LLM decision mechanisms differ from those of humans.
By Johnathan Sun, Andrei Shleifer, Yonatan Belinkov
The paper investigates how steering a language model along a trait vector—specifically for refusal and sycophancy—affects its internal computation. By splitting the model’s computation into a reconstruction circuit before the trait vector and a transmission circuit after it, the authors find that for refusal both circuits are compact and faithful, with the transmission circuit alone largely restoring the lost refusal signal. For sycophancy, the transmission circuit remains compact but the reconstruction circuit is broader and only partially faithful, indicating that the circuits used for steering are not necessarily the same as those that generate the behavior.
By Oscar Mir\'o L\'opez-Feliu, Maya Ozbayoglu
The study investigates where a large audio language model (LALM) derives its refusal responses to harmful spoken requests. By applying causal interventions at the audio-to-language-model interface and within the language-model residual layers, the authors find that the primary influence on the refusal margin comes from mid-to-late layers of the text language model rather than the audio front end. Ablations of the audio interface have minimal effect, while zeroing the encoder output still reduces the margin, indicating the audio pathway remains active but is not the main source of refusal decisions.
By Leonardo Haw-Yang Foo, Hung-yi Lee
TriFleetRCA is an on‑premise pipeline that performs root‑cause analysis for Kubernetes using a single GPU. It gathers evidence at pod, namespace, or cluster scope, deduplicates and ranks it with BM25, filters runbooks through an ingest guard, and returns a root cause with supporting evidence lines. In a live cluster with four injected faults, the system achieved hit rates of 0.85–0.95 across scopes, improved accuracy with deduplication, and demonstrated robust defense against poisoned runbooks.
By Rohit Patel, Susil Kumar Mohanty, Jeenal Chaudhary
ORION‑CMR is a scanner‑native, end‑to‑end foundation model for cardiac MRI that performs sequence classification, ventricular function assessment, LGE detection, disease classification, and generates reports in about 90 seconds. Trained on 12.9 million images, it outperformed supervised baselines and a prior CMR foundation model, achieving state‑of‑the‑art LGE classification and scar segmentation. In a multi‑vendor clinical cohort, it reached an AUC of 0.96 for normal‑vs‑abnormal detection and 0.88 for multiclass disease classification, with generated reports agreeing 81.4% with expert interpretation.
By Omer Burak Demirel, Kelly K. Horst, Alessio Perazzolo, Elisa Bruno, Kenan Kaya, Rongzhen Ouyang, Enas Ahmed, Jouke Smink, Spencer L. Waddle, Zainudeen Kallumpurath, Tzu Cheng Chao, Dinghui Wang, Steve G. Langer, Timothy L. Kline, Panagiotis Korfiatis, Jacinta Browne, Ivana Isgum, Tim Leiner
The paper introduces a diagnostic tool for distinguishing the use of misogynistic slurs from their mention in counter‑speech within code‑mixed Hinglish. It identifies evaluation artifacts in existing corpora, releases a 416‑item minimal‑pair contrast set that decorrelates slur presence and gendered register from labels, and proposes a pair‑consistency metric to assess model performance. Experiments show that even strong baselines struggle to consistently label counter‑speech pairs, while a large language model achieves perfect scores, indicating the benchmark measures genuine capability rather than exploitation of artifacts.
By Ashanvi Yadav, Shubham Bhardwaj
arXiv:2511. 04919v3 Announce Type: replace Abstract: Processing long documents with large language models (LLMs) is expensive: a single query over a 100K-token document can cost from tens of cents to over a dollar in API fees, depending on the model, and memory grows linearly with context length.
By Chandra Vamsi Krishna Alla, Harish Naidu Gaddam, Manohar Kommi, Sheikh Nazib Ahmed
Moonworks Lunara is a text‑to‑image model that defines Artistic Intelligence as exploration‑driven world realization, preserving semantic, artistic, and compositional structure. It uses a Diffusion Mixture Transformer architecture and a training algorithm that iteratively refines the data distribution with informative samples and human‑created art. Benchmarks show Lunara ranks first in aesthetic quality and second in emotional resonance against seven other image‑generation models, while maintaining a sub‑10B parameter size and sub‑10‑second inference latency.
By Yan Wang, Yanzu Wang, Maitreyee Joshi, Samiha Sadeka, Partho Hassan, Reza Jarral, Sayeef Abdullah, Sabit Hassan
The paper introduces Chart‑RVR, a reinforcement‑learning framework that trains chart‑reasoning agents to produce monitorable, verifiable outputs. It decomposes reasoning into three auditable stages—Structure, Evidence, and Derivation—allowing stakeholders to trace how the model reads the chart, extracts data, and computes the answer. Experiments on six benchmarks show that Chart‑RVR matches or exceeds state‑of‑the‑art accuracy while delivering rationales that are far more verifiable and evidence‑grounded than existing methods.
By Sanchit Sinha, Oana Frunza, Kashif Rasul, Aidong Zhang
The paper introduces an all‑in‑one multilingual scene text recognizer called ScriptMoE, which uses a script‑aware mixture‑of‑experts architecture to handle 10 scripts and 229 languages. It is built on a new large‑scale synthetic dataset, TextMuSS‑10M, and evaluated on the TextMuSS‑Bench, achieving 82.06% accuracy—1.31% higher than the best baseline. When integrated into the PP‑OCRv5 pipeline, ScriptMoE raises the end‑to‑end multilingual F1 score from 65.71% to 80.89%, slightly surpassing the best vision‑language model while using far fewer parameters.
By Xingsong Ye, Yongkun Du, Jiaxin Zhang, Zhixian Li, Chong Sun, Chen Li, Jing Lyu, Lianwen Jin, Zhineng Chen
A$^2$Safe is a framework for safe and effective Visual Question Answering that aligns counterfactual evidence with adaptive agent collaboration. It uses a Grounded Safety Evidence Board to make safety decisions explicit, enforcing invariance to safety‑irrelevant changes while allowing appropriate transitions when risk‑critical evidence changes. The system achieves a 95.72 SIUO safety score, reduces benign refusals on MOSSBench to 14.67%, and maintains a 78.34 average VQA score with 27.8% token overhead.
By Quanxing Xu, Ling Zhou, Xian Zhong, Jinyu Tian, Xiaohua Huang, Rubing Huang, Chia-Wen Lin
The paper introduces LiDAR-Hallu, a benchmark with 10,000 questions designed to test 4D LiDAR language models on spatio-temporal reasoning. It shows that models often achieve high multiple-choice accuracy by exploiting trivial patterns, such as always selecting the same option or relying on candidate duration, rather than truly understanding object relationships. Detailed analysis reveals systematic failures, especially in lateral-motion cases and opposite-answer scenarios, indicating that aggregate accuracy masks underlying reasoning gaps.
By Runyi Yang, Murat Akkoyun, Di Wen, Ruiping Liu, Yufan Chen, Junwei Zheng, Xiaoye Wang, Kailun Yang, Danda Pani Paudel, Luc Van Gool, Kunyu Peng
arXiv:2609.23065v1 Announce Type: cross
Abstract: Chain-of-thought (CoT) can sound plausible yet be unfaithful to the model's underlying reasoning. Most prior work probes CoT faithfulness through inp...
By Qianli Wang, Yilong Wang, Dennis Wei, Jingyi Sun, Simon Ostermann, Pepa Atanasova, Nils Feldhus
The paper introduces ProAction, a multimodal dataset of 10,000 samples comprising visual, audio, and text inputs across 12 daily-life scenarios, designed to support the Proactive Robot Action Reasoning (ProRobo) problem. It presents a two-stage human-in-the-loop annotation pipeline that incorporates appraisal and Theory-of-Mind considerations to generate cognitively grounded high-level action labels. The authors benchmark multimodal large language models and propose MMC2Act, showing that training on ProAction significantly improves proactive action reasoning compared to general-purpose models.
By Zhihao Gu, Kechao Zhu, Yuanfeng Wu, Mohan Liu, Ankit Kumar Shaw, ChenDong Hong, Xuanyu Chen, Dengchen Mei, Xu Tianyi, Lin Wang
MTMed3D is a multi-task Transformer-based model that jointly performs 3D detection, segmentation, and classification in medical imaging. It uses a shared Transformer encoder to produce multi-scale features, with separate CNN decoders for each task. Evaluated on BraTS 2018 and 2019, it achieves strong results, especially in detection, while reducing computational cost and inference time compared to single-task models.
By Fan Li, Arun Iyengar, Lanyu Xu
The paper investigates whether ranking of neural network architectures on one device predicts feasibility on another device, focusing on joint latency‑energy constraints. Experiments on NAS‑Bench‑201 and a large GPT‑architecture set show moderate cross‑device rank correlations but incomplete overlap of feasible sets, even when using improved latency diagnostics and conformal analysis. The authors conclude that rank portability, feasibility portability, and target‑specific decision support are distinct, urging evaluations to specify which target environments truly support the claimed operating point.
By Wesley Shu
The paper introduces Trace, a framework that transforms sparse-reward trajectories into executable walkthroughs by identifying progress anchors, propagating credit, and estimating action prerequisites. Trace compiles noisy trajectories into state‑conditioned, verifiable procedures that remove loops and detours, enabling reuse, intermediate‑state resumption, and programmatic verification. Experiments on J‑TTL, WebShop, and ScienceWorld with three open‑source LLMs show that Trace outperforms eight baselines, improving average AUC and Final‑$3$ by 30.0% and 40.5% while using fewer inference tokens.
By Kaijie Chen, Chenyu Fang, Liang Yan, Bo Li, Bo Zhang, Peng Ye