The paper investigates how the quality of instructions used to generate response pairs affects preference learning for language models. It shows that low‑quality or ambiguous instructions limit the range of response quality, weakening preference signals, and introduces an instruction‑refinement pipeline that improves data quality without discarding examples. Experiments across models and benchmarks demonstrate that refining instructions leads to better alignment and complements other data‑improvement methods.
By Seohyeong Lee, Hwaran Lee, Buru Chang
GeoMAD is a multi‑view anomaly detection framework that fuses multiple camera viewpoints while maintaining geometric awareness and scalability to multi‑class industrial settings. It introduces a Cross‑view Deformable Fusion Module (CDFM) that learns view‑pair‑specific sampling offsets on 2D feature maps, enabling hierarchical cross‑view correspondence without camera calibration or voxel construction. Additionally, Distributional View Alignment (DVA) provides a self‑supervised loss that aligns bottleneck distributions across views, ensuring global consistency without pixel‑level correspondence. Together, CDFM and DVA achieve geometry‑aware, distribution‑consistent fusion and demonstrate strong detection and localization performance on Real‑IAD and MANTA‑Tiny datasets.
By Shang-Fu Chen, Jhih-Ciang Wu, Kuan-Chuan Peng, Wen-Huang Cheng, Kai-Lung Hua
The paper introduces REMI, a framework that treats counterfactual fairness as a relational invariant discovery problem. By learning over paired examples, REMI identifies input regions where fairness is violated and generates interpretable rule-based models—fairness invariants—that can block or relabel unfair predictions without retraining the underlying model. Experiments on symbolic and neural network programs show REMI localizes fairness bugs in over 83% of cases and reduces discriminatory decisions in black-box models by up to 70%.
By Ranit Debnath Akash, Ashish Kumar, Gang Tan, Saeid Tizpaz-Niari
DocTalkBN is a large-scale multimodal dataset of authentic expert telemedicine conversations in Bengali, comprising 557.63 hours of paired audio and text, 1,515 multi-turn patient calls, and 10,274 host–doctor question–answer exchanges across 26 medical specialties. The dataset contains 1.7 million tokens and preserves the spontaneity and contextual richness of real medical interactions in a low-resource language. Three downstream tasks—medical triage classification, advice safety evaluation, and medical named entity recognition—are constructed to benchmark large language models and encoder-based baselines, demonstrating DocTalkBN’s practical usefulness for clinically grounded reasoning.
By Anik Saha, Fahmida Sultana Naznin, Sadatul Islam Sadi, Ananya Shahrin Promi, Wahid Al Azad Navid, Rifat Shahriyar
The article discusses a vulnerability in Anthropic’s Claude Code’s auto mode, which was promoted as a safeguard against prompt injection attacks. Prompt‑injection researcher Johann Rehberger demonstrated that the auto mode can be tricked into executing malicious code, even blocking the agent’s own cleanup attempts. The author concludes that the safest approach is to run coding agents in isolated sandboxes and restrict their access to sensitive resources.
The safety of large vision-language models is increasingly stress-tested by multimodal jailbreaks, yet existing attacks remain largely static at the meta level: template-based attacks freeze the image...
The paper evaluates machine learning for nitrogen recommendations in winter wheat by directly scoring the profit lost on measured yield response curves, rather than relying on prediction accuracy. Using 892 yield curves from UK experiments, the authors find that machine learning models fail to recover the best nitrogen rate within farm tolerance and generally underperform standard UK advice in terms of profit. However, a simple post‑model correction step can reduce profit losses by up to 43% without retraining, suggesting that machine learning can enhance standard advice rather than replace it.
Internet memes are a pervasive form of multimodal online communication; however, such communication often involves users from diverse linguistic and cultural backgrounds. Therefore, adapting memes acr...
SecureDrive‑FL combines differential privacy (DP‑SGD) with a novel Gradient‑Aware Selective Homomorphic Encryption (GASHE) scheme to protect federated driver‑monitoring models. GASHE encrypts only gradient components that exceed a DP‑calibrated sensitivity threshold, avoiding full‑parameter encryption. In experiments on a ten‑class distracted driver task, SecureDrive‑FL matches DP‑SGD’s poisoning resistance while also defending against Man‑in‑the‑Middle attacks, adding only 8–10% runtime overhead.
MVC-Bench is a calibration-focused benchmark for medical vision‑language models, evaluating how well these models express confidence across different modalities, backbones, and domain shifts. It tests robustness to modality, backbone, and domain changes, the effectiveness of calibration and prompt‑tuning strategies, and stability under prompt‑template and random‑seed variations. The benchmark includes 1638 experiments, reporting accuracy and Expected Calibration Error (ECE) along with other calibration metrics, and introduces a simple train‑time calibration method, Multi‑Class Margin (MCM) regularization, that achieves the lowest ECE in most settings.
The paper presents a method for linking written English keywords to spoken Hindi utterances using only visual grounding. By extracting English tags from images with captioning systems and aligning Hindi speech segments via self‑supervised features, the authors build an alignment‑based model that outperforms previous attention‑based neural approaches. The study also shows that adding negative examples improves alignment quality, demonstrating that cross‑lingual word‑to‑speech mappings can be learned without transcriptions or explicit training.
We present TabuLM, the first language model pre-trained on Kinyarwanda tabular data. Kinyarwanda is a morphologically rich Bantu language spoken by over 12 million people in Rwanda, yet lacks any dedi...
The paper investigates whether large language models (LLMs) maintain belief states—probability distributions over latent variables—by embedding a controllable latent variable into natural text. An LLM teacher generates ordinary text while subtly steering it along one of eight sparse autoencoder directions, which follow a ring-shaped Markov chain. A small transformer trained on this data successfully tracks the Bayesian posterior of the planted variable and arranges the eight states on a ring in the same order as the Markov chain, suggesting a link between concept geometry and latent variable dynamics.
The paper presents a modeling and simulation framework to study reinforcement‑learning control of connected and automated vehicle (CAV) platoon joining maneuvers in mixed traffic. It evaluates Deep Q‑Network, Double Deep Q‑Network, and Proximal Policy Optimization algorithms, finding that PPO achieves a 98 % joining success rate with less than 1 % collisions by incorporating risk penalties, though it requires more decision steps. An external safety controller can prevent collisions but may reduce joining efficiency, highlighting a trade‑off between safety, effectiveness, and decision speed.
The paper introduces SymbolLKG, a neuro-symbolic framework that combines a Logical Knowledge Graph (LKG) with dynamic solver routing to improve logical reasoning in large language models. The LKG represents logical rules and constraints as topological nodes, allowing explicit modeling of dependencies extracted from text. A Logic Router dispatches tasks to the most suitable symbolic engine, supported by a topology-aware hybrid retrieval mechanism, and the approach outperforms existing prompting and RAG baselines on logical reasoning benchmarks.
SAGE is an end‑to‑end CLIP‑based framework that augments vision‑language time‑series forecasting by jointly modeling temporal, cross‑variable, textual, and visual information. It processes frequency‑enhanced patches and variable tokens through a CLIP text encoder, while gated residual paths inject variable‑specific descriptions and statistical descriptors. A frozen CLIP vision encoder aligns rendered series with temporal representations via a training‑only contrastive objective, enabling multimodal alignment and variable‑level knowledge without using an LLM during inference.
The paper introduces the Multi-Instruction Multi-Shot Long-Video Editing (MMLVE) task, aiming to edit long videos with multiple instructions while maintaining consistency across shots, decoupling instructions, and preserving spatiotemporal structure. It proposes an agentic editing framework that combines Large Language Models and Vision‑Language Models for shot-level decoupling and precise instruction parsing. A new dataset, MMLVE‑Bench, and three evaluation metrics are created to benchmark this task, and experiments show the proposed MMLVE‑Agent outperforms existing closed‑source state‑of‑the‑art methods by eliminating hallucinations and ensuring seamless transitions.
Existing constant-step analysis of stochastic \Scaf{} identifies a leading $O(γ/N)$ stationary mean bias and shows that higher-order bias can persist as the client count increases, but does not identify the first client-independent contribution at coefficient level. For full-participation stochastic \Scaf{} with one-dimensional homogeneous clients, fixed local-step count $H$, and bounded additive gradient noise, we prove, uniformly over $N\ge2$, $$ \begin{aligned} \mathbb{E}_{π_{γ,N,H}}[x]-x^\star ={}& -\frac{f'''(x^\star)σ^2}{4f''(x^\star)^2}\fracγ{N}\\ &- \frac{f'''(x^\star)σ^2}{12f''(x^\star)} \frac{(H-1)(5H-1)}{H}γ^2 +O_H\!
The paper introduces a new approach to safety in contextual bandits with continuous actions by enforcing high‑probability constraints on the realized cost rather than on its expectation. It proposes the High‑Probability Constrained UCB algorithm, which balances optimistic reward exploration with pessimistic safety estimation. The authors provide theoretical regret guarantees for linear models and extend the analysis to general function classes, demonstrating experimentally that realized‑cost constraints significantly reduce safety violations compared to expected‑cost baselines.
The paper introduces a new task called compositional layout understanding, focusing on interpreting complex, multi-layer document and UI layouts that involve hierarchical relationships among visually entangled elements. It presents CoDeLayout, a VQA dataset of about 20,000 real-world layouts annotated with compositional element pairs and design intent, and identifies two main challenges for current vision‑language models: semantic drift between textual metadata and visual content, and structural ambiguity in hierarchical inter‑element relationships. To address these, the authors propose MASON, a post‑training paradigm that combines multimodal alignment and structural perception, achieving a 91.66% accuracy on CoDeLayout and outperforming full‑data direct fine‑tuning with only 30% of the training data.