The paper investigates the "edge of stability" phenomenon in deep learning, where Hessian eigenvalues remain stable above a classically predicted unstable threshold. It shows that many first‑order optimizers, including gradient descent, can violate this stability bound by up to a factor of 21.1, and that this deviation depends systematically on the optimizer used. The authors propose a new stability threshold based on the directional Hessian and gradient‑alignment score, which removes optimizer‑dependent offsets and offers consistent predictions while providing diagnostic tools to understand how optimizers balance temporal and spatial budgets.
By Jaerin Lee, Kyoung Mu Lee
MR‑IQA‑2 introduces an actor‑editor‑judge framework that separates reasoning from rating in image quality assessment. The actor generates quality reasoning, the editor modifies the image based on identified factors, and a frozen judge compares the original and edited images to provide reflective supervision. Fine‑grained credit assignment allows distinct supervision signals for reasoning and rating, achieving competitive human‑aligned ratings while offering richer, faithful visual understanding.
By Yuan li, Youyuan Lin, Chenhui Chu, Shin'ya Nishida
Fractional Decay KV-Cache (FD‑KVC) is a new memory‑management algorithm for transformer‑based dialog systems that uses a dual‑channel scoring mechanism—cumulative attention and recency‑weighted relevance—to keep key‑value pairs relevant as conversation topics shift. The method incorporates an ownership‑loss‑driven adaptive learning rate, runs entirely on CPU with minimal overhead, and outperforms the heavy‑hitter baseline H2O by significant margins across multiple dialog scenarios, including faster topic adaptation and higher topic diversity. Ablation studies confirm that each component of FD‑KVC contributes to its performance gains.
By Sukanta Ganguly
The paper introduces the Middle East Cultural Sensitivity Score (MECSS) to quantify Orientalist bias in large language models, converting Said’s seven Orientalist operations into measurable dimensions. Using 280 conversations, it finds that GPT‑4 and Falcon3‑7B‑Instruct systematically reproduce Orientalist patterns, with Falcon scoring higher despite being regionally built. The study highlights that geographic origin alone does not mitigate bias and identifies a new failure mode, "Said‑washing," present in 87.9% of GPT‑4 interactions.
By Maha Shahid
GenEx is a new pipeline that transforms raw SARS‑CoV‑2 gene sequences into codon co‑occurrence graphs, extracting over 25 graph features. It introduces two graph generation methods—MSCG and LAPCG—and applies spectral graph analysis via Singular Value Decomposition to improve class separability. Using these techniques, the authors trained 23 machine‑learning models that successfully detected all major SARS‑CoV‑2 variants.
By Arefin Amin, Labiba Faiza Karim, M. Monir Uddin
MAVEN is a hierarchical framework for evaluating whether multimodal content aligns with macro‑societal values such as peace, justice, and freedom. It organizes values into six primary dimensions and 72 secondary indicators, enabling multi‑level quantitative scoring. The authors build a human‑verified multimodal benchmark, a soft‑match metric, and propose efficient evaluator optimization techniques, demonstrating that a compact 2B evaluator performs comparably to larger models and approaches state‑of‑the‑art closed‑source VLMs.
By Zijuan Zhao, Zheren Fu, Hou Xia, Licheng Zhang, Yi Liu, Zhendong Mao
The paper presents the first systematic evaluation of uncertainty quantification (UQ) methods applied to a foundation model for semantic segmentation. By fine‑tuning a lightweight DPT decoder on the pretrained SAM2 encoder, the authors benchmark four UQ approaches—Monte Carlo Dropout, Deep Sub‑Ensemble, Test‑Time Augmentation, and Evidential Deep Learning—across Cityscapes, NYUv2, and two out‑of‑domain settings, comparing segmentation accuracy, calibration, uncertainty quality, and inference time. The results reveal clear trade‑offs between predictive performance, reliability, and computational cost, underscoring both the promise and current limitations of uncertainty‑aware foundation models for real‑world deployment.
By Steven Landgraf, Joceline Hinz, Markus Ulrich
The paper "Jailbreaking in the Haystack" introduces NINJA, a jailbreak technique that exploits long-context language models by appending benign, model-generated content to harmful user goals. It demonstrates that the position of harmful goals within the context is crucial for safety, and shows that NINJA significantly boosts attack success rates on models such as LLaMA, Qwen, Mistral, and Gemini. Unlike previous methods, NINJA is low-resource, transferable, less detectable, and compute‑optimal, revealing that carefully crafted benign long contexts can expose fundamental vulnerabilities in modern LMs.
By Rishi Rajesh Shah, Chen Henry Wu, Shashwat Saxena, Ziqian Zhong, Alexander Robey, Aditi Raghunathan
The paper introduces Latent Space Refusal Anchoring (LSR‑Anchoring), a training‑free technique that extracts a refusal direction from English prompts and applies it to the residual stream of instruction‑tuned models at inference time. The primary variant, Mean‑Activation Steering (MAS), works across several architectures (Llama‑3‑8B, Llama‑3.1‑70B, Mistral‑7B‑Instruct, Qwen2.5‑7B), restoring safety for low‑resource African languages with minimal performance loss, while a refined SAE‑Derived Steering (SDS) further reduces KL divergence without degrading legitimate prompt performance. The method shows positive transfer for Yoruba, Igbo, Igala, and Hausa, but fails for Arabic, suggesting a geometric mismatch rather than a data scarcity issue.
By Godwin Abuh Faruna
The paper introduces the Explanation Consistency Score (ECS), a fairness‑aware metric that uses Jensen‑Shannon divergence to measure how similar attribution maps are across demographic subgroups. Applied to diabetic retinopathy screening, ECS is evaluated both overall and within disease severity levels. Results show that although predictive performance varies among ethnic groups, explanation consistency remains high and is not significantly linked to performance disparities, indicating that predictive fairness and explanation consistency assess different aspects of model behavior.
By Kerol Djoumessi, Philipp Berens
The paper presents a method for extracting key information from OCR‑digitized clinical reports, addressing challenges posed by heterogeneous documents and noisy OCR output. It introduces an open key space that is iteratively mined, normalized, clustered, and verified to build a canonical key inventory, and defines key coverage as a metric for inventory completeness. Experiments on reports from over 20 hospitals using a 0.2B BERT model show that performance improves steadily with key coverage, achieving high F1 scores when the top 90 keys are covered and outperforming a fine‑tuned Qwen3‑0.6B baseline.
By Yu Wang, Yingyun Li, Ying Qin, Haiyang Qian
The paper titled "Safety Alignment Illusion: The Cross-Lingual Safety Gap in LLMs" highlights that current safety alignment training for large language models is predominantly English-centric, leading to failures in non‑English languages. It introduces INCLUDE, a multilingual benchmark with 2,604 prompts in six languages (English, Hindi, Bengali, Marathi, Tamil, and Hinglish) to measure Indian‑centric socio‑cultural biases. Evaluation of ten open‑ and closed‑source LLMs shows that Bengali models exhibit the highest bias scores among open‑source models, while English shows the lowest bias in open‑source but the highest in closed‑source models.
By Namya Bhatnagar
SkillNet is an open infrastructure that enables the creation, evaluation, and organization of AI skills at scale. It structures skills within a unified ontology, supports multi‑dimensional evaluation (Safety, Completeness, Executability, Maintainability, Cost‑awareness), and integrates a repository of over 600,000 skills, an interactive platform, and a Python toolkit. Experiments on ALFWorld, WebShop, and ScienceWorld demonstrate a 40% increase in average rewards and a 30% reduction in execution steps across multiple backbone models.
By Yuan Liang, Ruobin Zhong, Haoming Xu, Chen Jiang, Yi Zhong, Runnan Fang, Jia-Chen Gu, Shumin Deng, Yunzhi Yao, Mengru Wang, Shuofei Qiao, Yida Xue, Xin Xu, Tongtong Wu, Kun Wang, Yang Liu, Zhen Bi, Jungang Lou, Yuchen Eleanor Jiang, Hangcheng Zhu, Gang Yu, Haiwen Hong, Longtao Huang, Hui Xue, Chenxi Wang, Yijun Wang, Zifei Shan, Xi Chen, Zhaopeng Tu, Feiyu Xiong, Xin Xie, Peng Zhang, Zhengke Gui, Lei Liang, Jun Zhou, Chiyu Wu, Jin Shang, Yu Gong, Junyu Lin, Changliang Xu, Hongjie Deng, Wen Zhang, Keyan Ding, Qiang Zhang, Fei Huang, Ningyu Zhang, Jeff Z. Pan, Guilin Qi, Haofen Wang, Huajun Chen
The paper presents a two‑stage multimodal framework for chest X‑ray interpretation that incorporates radiologist gaze data from the MIMIC‑Eye dataset. Stage 1 introduces a gaze‑token classifier that fuses image patches, bounding‑box masks, transcription embeddings, and fixation maps, and a curriculum‑scheduled loss that improves accuracy and spatial alignment, yielding a 4.4% AUC gain and 13.3% F1 improvement. Stage 2 translates classifier predictions into region‑specific diagnostic sentences using confidence‑weighted keywords, an expert dictionary, and a prompted large language model, boosting clinical‑term BERTScore and ROUGE over keyword baselines.
By Tanjim Islam Riju, Shuchismita Anwar, Saman Sarker Joy, Farig Sadeque, Swakkhar Shatabda
FraudBench is a new benchmark that tests policy‑grounded banking conversational agents against adaptive fraud scenarios. It uses a dual‑control framework and a 698‑document internal policy corpus, presenting 150 adversarial scenarios (107 public, 43 held‑out) that require agents to manage mutable account state and tool access while preventing identity, authorization, and trust manipulation. Preliminary results on four agents show attack‑security rates between 49% and 65%, highlighting weaknesses in money‑mule and first‑party fraud detection.
By Dheeraj Mohandas Pai, Lu Xian
TestifAI is a deep learning testing framework that estimates model robustness against combinations of semantic input perturbations such as blur, brightness, and zoom. It allows users to define operational conditions as structured spaces with discrete severity levels and query robustness for any combination. By employing partial model tomography, TestifAI reconstructs higher‑order perturbation effects from low‑order tests, achieving less than 7% estimation error while reducing inference counts by 60‑80% across five image and language classification tasks.
By Arooj Arif, Tobias Hartung, Elena Botoeva, Alexandros Koliousis
The paper introduces Fulcrum, a topology‑aware differential privacy scheme for hierarchical federated learning that allocates noise based on the size and exposure of regional aggregation groups. By deriving a closed‑form exposure dispersion metric from region structure and weights, the method optimally balances privacy and utility, achieving up to 14.84% accuracy gains on image tasks and 12.16% on text tasks at ε = 0.99 compared to uniform noise allocation. The approach ensures each participant receives noise commensurate with its actual exposure, eliminating unnecessary privacy overhead.
By Murtaza Rangwala, Richard O. Sinnott, Rajkumar Buyya
SuTRA (Structurally-Unified Tokenization with Root Awareness) is a morphology-aware tokenization algorithm designed to address the problem of Morphological Shattering in morphologically rich Indic languages. It preserves the indivisibility of aksharas—complex orthographic syllables—by penalizing merges that cross morphological boundaries, thereby reducing over-fragmentation of words. The authors also release a new morphological segmentation dataset for Hindi, Marathi, and Gujarati, and demonstrate that SuTRA improves morphological alignment by up to 14.7% and semantic recoverability by 34% over BPE, leading to an average machine translation gain of +8.08 chrF2.
By Vaibhav Rathore, Siddhant Gole, Dadhichi Telwadkar, Rooshil Bhatia, Maulik Ruparel, Siddharth Surekha, Neha Bhargava
The paper introduces Jacobian-Guided Anisotropic Noise Reshaping, a method that improves data utility under Local Differential Privacy by selectively reducing noise in task-relevant subspaces of data representations. It uses the Jacobian of a public downstream model to identify critical directions and reshapes isotropic LDP noise into an anisotropic distribution, preserving privacy while enhancing performance. Experiments on CIFAR-10-C show significant accuracy gains, especially for PrivUnit variants at ε=7.5.
By Youngmok Ha, Viktor Schlegel, Yidan Sun, Anil Anthony Bharath
The OpenAI Privacy Filter is a 1.5‑billion‑parameter, bidirectional token‑classification model that detects and redacts personally identifiable information and secrets in unstructured text. It is built from an autoregressive checkpoint, converted into a banded‑attention classifier, and uses a constrained Viterbi decoder to produce coherent spans across eight privacy categories in a single forward pass. The model supports configurable precision‑recall tradeoffs, a 128,000‑token context window, and is designed for efficient local deployment and domain‑specific fine‑tuning as a data‑minimization component within layered privacy workflows.
By Charles de Bourcy, Sahra Ghalebikesabi, Avi Schwarzschild, Alex Gorbachev, Mihai Maruseac, Annie Chu, Vol Kyrylov, Tong Mu, Ally Bennett, Andy Nguyen, Casey Meehan, Jessica Gan Lee, Shane Bauer, Harold Nguyen, Rodolpho Eckhardt, Yuqi Liu, Charlie Oxborough, Marco Rougeth, Omar Chedid, Caio Costa, Yash Parikh, Yao Li, Congzheng Song, Om Thakkar, Vinnie Monaco