The paper introduces a method to incorporate attention bias directly into the keys of compacted KV caches, enabling standard kernels like FlashAttention and vLLM to use these folded caches without modification. By fitting a bias along a closed‑form direction per head, the approach preserves the cache size and head dimension while bounding output error based on query projections. Experiments on Qwen3 and Llama‑3.1 models show that the folded caches match or exceed the accuracy of the original Attention Matching method across various benchmarks, with comparable or lower KL divergence to full caches.
By Utkarsh Kanwat
The paper introduces a multibranch ensemble head that pairs low- and high-capacity branches for Vision Foundation Models, sampling one branch during training and averaging both at inference. This lightweight design adds negligible parameters and inference overhead while improving accuracy-efficiency trade-offs on data-scarce tasks. Applied to cucumber phenological stage classification, the DINOv2 ViT-S with the multibranch head achieves a Macro F1 of 0.978 and Cohen's Kappa of 0.971, matching larger models but with only 12.2 GFlops and 22.3 M parameters, enabling deployment on smartphones and edge devices.
By Kostiantyn Khabarlak
3D-SceneQ is a new 3D large language model that tackles two key limitations of existing 3D scene understanding pipelines: the inclusion of irrelevant object tokens that add noise, and the exclusive use of point cloud data that lacks rich semantic cues. It introduces a Query‑Guided Adaptive Pruning (QGAP) module that selectively keeps task‑relevant objects based on user instructions, and a Multi‑modal Object‑level Feature Enhancement (MOFE) module that fuses 2D feature embeddings into the 3D representation to enrich semantic information. Experiments show that 3D‑SceneQ surpasses state‑of‑the‑art methods on several benchmarks, achieving significant gains in metrics such as CIDEr and EM@1.
By Yongqi Shan, Lu Zhang, Jiazuo Yu, Yunzhi Zhuge, Huchuan Lu
The paper investigates when it is better to ignore a relation in multi-relational graph learning rather than use it, taking into account the varying manipulation costs an adversary faces. It shows that a degree‑gating strategy outperforms any selection or reweighting when a subset of the population is isolated on a relation and the attacker cannot open the gate, and that a per‑node exposure bound can identify which relation to harden. The authors provide theoretical guarantees, empirical validation on real and synthetic graphs, and a selection protocol that is never significantly worse than ignoring the graph.
By Vincess Dongmo, Astride Melvin Fokam Ninyim
The paper investigates how uncertainty quantification fails in selective prediction systems when clinical data exhibit severe class imbalance, leading to systematic exclusion of minority patients. Through experiments on 56 synthetic scenarios and three neurodegenerative disease cohorts, the authors show that standard uncertainty measures become proxies for class membership rather than true ambiguity, and that combining class‑aware correction (CCR) with threshold‑aware metrics restores proper monotonicity and protects minority patients. The study proposes adopting threshold‑aligned uncertainty measures and class‑stratified evaluation to mitigate this bias and improve safety in clinical AI pipelines.
By Walter Endrizzi, Flavio Ragni, Raffaele Marchesi, Stefano Bovo, Monica Moroni, Giuseppe Jurman, Venet Osmani
The paper investigates how rubric-based evaluations of medical language models fail to detect hallucinations that could undermine clinician trust. Using a controlled study with MedHallu and a clinician-validated error-injection pipeline, the authors show that more specific rubrics can distinguish correct from hallucinated responses, yet many clinically relevant hallucinations remain undetected across HealthBench, HealthBench Professional, and LiveMedBench. The study finds that rubrics work best when explicitly checking facts but miss additional or unexpected errors, and suggests that a retrieval-based factuality check can recover some of these blind spots.
By Griffin Farrow, Lily Sijia Li, Jack Johnson, Tingyan Wang, Philip Torr, William Bolton, Fabio J. Fehr
The paper investigates human‑AI complementarity as a strategy for safer AI oversight, arguing that humans and models making different mistakes can combine judgments to produce stronger oversight signals. The authors present the largest known suite of safety‑relevant tasks (3,957 instances across 8 datasets and 6 domains) and a platform for measuring performance of humans, AI, and hybrid systems. Their study, based on 35,613 judgments from 416 raters, shows a notable complementarity effect but also finds that the tested assistance methods rarely improve over the best single source and that human raters corrected AI errors less often even as overall accuracy rose.
By Sander Lanbo Shi, Miranda Anna Christ, Eric Modesitt, Andrew Louis, Konrad Cheng, Brandon Thai Tran, Sewoong Lee, Ratnakar Pawar, Gaia Molinaro, Juliana Bain, Nandika Donthi, Maja Pavlovic, Annya Dahmani, Francesca Bianco, Abhishek Shah, Tyler Crosse, Arth Singh, Karissa Huang, Sukanya Krishna, Konstantinos Voudouris, David Demitri Africa, Joshua Jacob, Rishub Jain
She Works, He Works is an interdisciplinary project that blends feminist visual theory, AI critique, and artistic practice to investigate gender representation in AI-generated imagery. By comparing text‑to‑image outputs of male and female construction workers, the study exposes how societal biases appear in generative AI, highlighting aesthetic and professional disparities. Grounded in Griselda Pollock’s feminist insights, the work positions AI as both a cultural actor that can perpetuate stereotypes and a tool for critique, urging transparent practices, diverse datasets, and reflexive engagement in creative industries.
By Amalia Foka
The paper introduces a swap‑adversarial framework designed to address inter‑subject variability and the high‑dimensional low‑sample‑size challenge in electrocorticography (ECoG) data. It combines robust preprocessing, an inter‑subject balanced channel swap (ISBCS) for cross‑subject augmentation, and domain‑adversarial learning (DAL) to suppress subject‑specific bias, enabling robust domain generalization across both ECoG and EEG‑based brain‑computer interface datasets. The authors also present a new reproducible ECoG benchmark derived from long‑term recordings of 6‑hydroxydopamine‑induced rat models, and demonstrate the framework’s superior performance in cross‑subject, cross‑session, and cross‑dataset experiments.
By Seongwon Jin, Hanseul Choi, Sunggu Yang, Sungho Park, Jibum Kim
The paper introduces Sequence Likelihood-Free Policy Optimization (LFPO), a method that aligns masked diffusion models by optimizing token-level denoising logits through contrastive updates, avoiding the need for sequence-level likelihoods. LFPO leverages discrete velocity alignment inspired by continuous flow matching, providing a stable, low-variance training surrogate. Experiments on reasoning and coding tasks show that LFPO matches or surpasses likelihood-based baselines while significantly speeding up inference.
By Chenxing Wei, Jiazhen Kang, Hong Wang, Jianqing Zhang, Hao Jiang, Xiaolong Xu, Ningyuan Sun, Ying He, F. Richard Yu, Yao Shu, Bo Jiang
The paper investigates how large language model (LLM)–based automated code review (ACR) systems can be manipulated through contextual bias injection, a framing effect that lets presentation override content. A large‑scale study across six LLMs confirms the framing effect in vulnerability detection, and a realistic experiment on 33 CVEs in 20 projects shows that a novel LLM‑assisted refinement attack succeeds in 97% of cases, while a template‑based attack fails. The results underscore the vulnerability of real‑world ACR pipelines and the need for human oversight and trust in the development process.
By Georgios Alexopoulos, Nikolaos Alexopoulos, Diomidis Spinellis, Dimitris Mitropoulos
arXiv:2509. 00924v2 Announce Type: replace Abstract: Transformers often transfer across related tasks with little fine-tuning; we study this question in noisy nonparametric regression for continuous maps $f,g:[0,1]^d\to\mathbb{R}^D$, representing the source and shifted tasks, using a single-query, single-head transformer $\mathcal{T}(\cdot)=V\operatorname{Softmax}(-\alpha(\cdot,K))$ with $\ell^{\infty}$-based alignment scores $\alpha$.
By Anastasis Kratsios, Tin Sum Cheng, Bum Jun Kim, Daniel Roy
The paper introduces a targeted learning approach to assess fairness in the data generating process used for training predictive models, rather than focusing solely on model performance. It derives estimators for demographic parity, equal opportunity, and conditional mutual information, and demonstrates through simulations that these estimators improve confidence‑interval coverage and reduce false detections of population unfairness compared to fixed‑model inference. The authors also apply the method to real data to illustrate its practical utility.
By Joseph H. Rudoler, Alexander Asemota, Giles Hooker
Token-to-Token Alignment of Text Embeddings for Semantic Blending proposes a two‑stage framework that first rephrases text prompts into a shared structured form and then aligns token embeddings across prompts based on semantic similarity. By establishing explicit semantic correspondence between tokens, the method enables smooth, coherent transitions through simple linear interpolation, facilitating image blending and continuous editing. The authors demonstrate that properly aligned text embeddings reveal an implicit continuous semantic structure in text‑to‑image models, allowing semantic control without altering the generative model.
By Saar Huberman, Ron Mokady, Or Patashnik, Daniel Cohen-Or
GradCO is a new method for Text-based Person Anomaly Search (TPAS) that addresses two key challenges: gradient conflict from same-identity, different-behavior hard negatives and the dominance of strong appearance cues over weak behavioral evidence. It introduces a Behavior-Aware Mixture-of-Experts (B-MoE) to separate appearance, normal-behavior, and anomalous-behavior streams, Conflict Offload Learning (COL) to reduce conflicting gradients via pair-conditioned anchors, and Hierarchical Intent Alignment (HIA) to aggregate expert similarities with query-conditioned weights. Experiments on five TPAS/TIPR benchmarks demonstrate that GradCO significantly outperforms state-of-the-art methods, validating its design choices.
By Zhiheng Fu, Zixu Li, Zhiwei Chen, Jiale Huang, Fangxu Liu, Jinhe Bi, Yupeng Hu
The paper investigates whether spiking neural networks (SNNs) are more resistant to model stealing attacks than conventional artificial neural networks (ANNs). By benchmarking four common attacks on MNIST, Fashion‑MNIST, and CIFAR‑10, the authors find that SNNs consistently yield surrogates with 10%–20% lower agreement and 15%–25% lower adversarial transferability compared to ANNs. The authors attribute this resilience to the spike‑based, time‑dependent dynamics of SNNs, which make their decision boundaries harder for ANN‑based surrogates to approximate.
By Ayana Moshruba, Maryam Parsa
MAEval is a multi‑agent framework that evaluates natural language explanations in recommender systems using large language models. It employs a planner agent with a Chain of Debate prompting strategy, a moderator to control prompt drift, and an arbitrator that aggregates multiple evaluation rounds, supplemented by Monte Carlo sampling for robustness. Experiments on public and proprietary datasets show that MAREval outperforms existing baselines, aligns more closely with human judgments, reduces variability, and achieves production‑level quality, leading to measurable gains in user engagement and revenue.
By Reza Yousefi Maragheh, Jayesh Uddhav Kudase, Aysenur Inan, Ramin Giahi, Kai Zhao, Parth Parikh, Jianpeng Xu, Jason Cho, Evren Korpeoglu, Sushant Kumar
This paper examines how two popular vision‑language models encode protected identities when asked to transform human face images to reflect race and sexual orientation. The authors find that the models often comply with such requests, with compliance being non‑deterministic and group‑sensitive, and that the generated images are statistically distinguishable and tend to reinforce deep‑seated stereotypes. They also show that these representation biases can reappear in textual reasoning tasks and that image generation can perpetuate bias around criminality, raising concerns about AI‑assisted profiling.
By Soumyajit Datta, Arka Dutta, Ashiqur R. KhudaBukhsh
The study evaluates ten ImageNet‑pretrained deep learning models for classifying benign versus malignant breast cancer in the BreakHis histopathology dataset, using a patient‑level split. DenseNet121 achieved the highest performance with 99.4% accuracy and an AUC‑ROC of 0.999, outperforming all other models (p < 0.001). Explainability methods (Grad‑CAM, LIME, Integrated Gradients, LRP) consistently identified nuclei‑rich and glandular regions as key diagnostic features, demonstrating that transfer learning combined with explainable AI can produce accurate and interpretable breast cancer diagnostic tools.
By Sami Alnawali, Muhammad Atif Saeed, Abdul Hadi, Azka Atiq, Akhtar Jamil, Ahmad Din
Wavelet Phase Diffusion (ψ‑PD) is a new simulation‑to‑real translation method that operates in the Dual‑Tree Complex Wavelet Packet Transform domain, allowing spatially adaptive phase injection without global spectral interference. It also introduces Low‑Frequency Randomization to decouple the model from synthetic illumination priors, enabling realistic, in‑distribution real‑world appearance while preserving structural and semantic consistency. Trained on unpaired open‑domain data, ψ‑PD achieves superior realism and semantic consistency on vKITTI→KITTI image translation and improves CARLA video translation performance, reducing planner errors by over 5%.
By Kaiwen Wang, Frank Bieder, Yinzhe Shen, Carlos Fernandez, Jan-Hendrik Pauls, Omer Sahin Tas