TrustmeWatcher is an application developed for the TRUST‑ME project that integrates OS‑level activity monitoring from ActivityWatch with a user interface for well‑being feedback. It transforms computer‑activity traces into an interactive screen‑time dashboard, synchronizes short questionnaire responses from a StreamDeck, and displays these alongside video highlights. The app uses local AI models to predict six normalized state scores and an overall well‑being score, providing explainable AI insights and privacy controls such as camera and eye‑tracker pause/resume, and has been tested with records from 17 participants.
By Chengyu Yu, Leon Jacopo Costa, Zoja An\v{z}ur, Mohan Li, Ga\v{s}per Slapni\v{c}ar, Daniil Kirilenko, Martin Gjoreski, Mitja Lu\v{s}trek, Marc Langheinrich
The paper introduces Delay‑Guided Attention for Multi‑stage Forecasting (DAMF), a method that explicitly models inter‑stage lag distributions in multistage manufacturing to improve quality prediction. DAMF uses a lag predictor to estimate the probability and magnitude of delays between adjacent stages, incorporates these soft distributions into cross‑stage attention, and employs gated residual fusion for information flow. The approach is evaluated on the LiquidSugar dataset and two controlled‑delay benchmarks, showing better quality prediction and more accurate lag recovery compared to baselines.
By Jiarui Zhang, Zhaoyang Guan, Zhixuan Xia, Naichen Shi
The paper proposes a shrinkage estimator that blends store‑specific demand estimates with pooled information from other stores under a multinomial logit model, addressing the trade‑off between local data scarcity and cross‑store heterogeneity. By leveraging self‑concordant‑like properties of the MNL likelihood, the authors derive finite‑sample performance guarantees and regret bounds for assortment and pricing decisions. Numerical experiments and an empirical study on European apparel retailer data show that this transfer‑learning approach improves out‑of‑sample product prioritization and revenue capture compared to both local estimation and full pooling.
By Recep Yusuf Bekci, Stefanus Jasin, Joline Uichanco
The paper investigates how tropical cyclone attributes are represented within the sparse autoencoder (SAE) features of the AI weather model GraphCast. By analyzing 2,115 six‑hourly timesteps from 205 cyclones, the authors map SAE activations to cyclone position, intensity, and pressure structure across multiple layers and decompositions, finding that each attribute is encoded in multiple features and that deeper layers better predict future position and pressure structure. The study demonstrates that the organization of physical‑attribute relations in SAEs is reproducible and can aid in interpreting cyclone‑related information in GraphCast.
By Jiapei Liu, Ruidong Li, Ting Sun, Guangheng Ni
BRIDGE is a model‑agnostic acceptance gate that filters model updates by cross‑fitting and bootstrap bounds, aiming to prevent false improvements that arise when validation gains do not translate to new data. Across multiple settings—including 11 new datasets, synthetic tasks, 12 retrospective clinical tasks, and 39 supporting datasets—BRIDGE reduces false‑improvement rates from roughly 19–25% with a naive rule to as low as 0–1.1%. The method also yields higher mean gains per deployed update, achieving 1.8 to 3.1 times the naive rule’s performance.
By Dwi Sunaryono
The paper investigates a new vulnerability called memory injection, where malicious prompts can persist across sessions in modern LLM‑as‑a‑service platforms. By demonstrating that all NIST attack goals can be achieved on ChatGPT Instant 5.3, Claude Sonnet 4.6, and Gemini 3 Flash, the authors show that these services can spread disinformation, generate harmful content, or cause offline unresponsiveness even in fresh sessions. The study also reveals that many proposed defenses are not yet implemented and provides datasets and artifacts for further research.
By Pedro H. Barcha Correia, Matheus Azevedo de S\'a, Paulo H. Gomes de Senna, Giovanni Mandel Martignago, Victor Takashi Hayashi, Charles Christian Miers, Marcos A. Simplicio Jr
The paper introduces a unified evaluation framework for assessing the physical adversarial robustness of Vision‑Language Models (VLMs) deployed at the edge. It defines a conditional attack success rate (cASR) metric that counts only images where the model still localizes objects under a gray occluder, and categorizes failures into omission, substitution, and displacement. Benchmarking shows that compact grounding VLMs can be as vulnerable as traditional object detectors, with omission being the only failure visible to downstream systems, while displacement and substitution pose potentially more dangerous risks, especially under distribution shift.
By Eric Liang, Alan Luo, Kaiwen Yuan
The paper introduces "No Frames, Just Frequencies," a bandwidth‑efficient adaptation framework for smart glasses that uses on‑device frequency descriptors instead of camera frames to detect distribution shifts. These descriptors are sent to a host device, which generates out‑of‑distribution data and adapts a model replica via a predefined protocol. Experiments on ImageNet‑C corruptions, Visual Wake Words, and MS‑COCO show that descriptor‑driven adaptation recovers 93‑95% of supervised oracle gains, with graceful degradation as descriptor resolution drops, enabling up to 784× bandwidth savings.
By Francesco Corti, Moritz Becker, Olga Saukh
HGM-Dual is a diagnostic‑guided framework that improves harness optimization for planning agents by representing agent evolution as lineage‑aware context and using error‑conditioned dual‑path child generation. It records ancestor changes, behavior outcomes, and ineffective modifications, enabling explicit edit‑outcome attribution. The method achieves up to 14.45% relative improvement in case accuracy while using 2×–4× fewer evaluation budgets compared to the strongest baseline.
By Sudipta Paul, Pengrui Quan, Aounon Kumar, Nicholas Tzou, Vivek Kulkarni, Srinivas Chappidi
The paper introduces an energy-based transfer learning approach for reinforcement learning, aiming to improve sample efficiency in continual learning scenarios. By using out-of-distribution detection, the method selectively provides teacher guidance only in states that fall within the teacher’s training distribution, preventing sub-optimal bias when tasks differ significantly. The authors theoretically connect energy scores to state-visitation patterns under on-policy RL and demonstrate empirical gains in both single-task and multi-task settings.
By Zeyun Deng, Jasorsi Ghosh, Fiona Xie, Kai Cheng, Yuzhe Lu, Katia Sycara, Joseph Campbell
The paper introduces Evidential Uncertainty Quantification (EUQ), a fine‑grained method for detecting misbehaviors in large vision‑language models (LVLMs). EUQ interprets model output features as supporting or opposing evidence and aggregates them using Evidence Theory to quantify internal conflict and ignorance in a single forward pass. Experiments on hallucinations, jailbreaks, adversarial attacks, and out‑of‑distribution failures show that EUQ outperforms strong baselines, linking hallucinations to high conflict and OOD failures to high ignorance, and provides layer‑wise dynamics for interpreting internal representations.
By Tao Huang, Rui Wang, Xiaofei Liu, Yi Qin, Li Duan, Liping Jing
The paper introduces DiagSafe, a diagnostic evaluation framework that replaces binary safe/unsafe labels with prompt-level rubrics grounded in AI safety policies. DiagSafe uses a bilingual benchmark with an 84‑label safety taxonomy, comprising DiagBase and DiagAttack subsets, to assess safety and usefulness independently. Experiments show DiagSafe achieves high agreement with human annotations (95.34%) and strong inter‑judge consistency (Cohen’s κ = 0.864), revealing significant differences in unsafe rates and vulnerabilities across leading LLMs, especially in Chinese and multi‑turn contexts.
By Hongyu Li, Xinlong Huang, Zhan Yu, Jing Zhao, Zhe Zhao, Juntao Zhong, Xiong Kuang, Rui Yuan, ZhenXiang Yan, Huilin Xu, Haitao Yu, Lilin Wang, Rui Chen, Blaze Chen, Fengzong Lian, Renfen Hu
The paper investigates how model size and quantization affect social bias and abstention in question answering. Using the BBQ benchmark and models from Qwen3, Gemma-3, and SmolLM2, it finds that larger models improve accuracy but bias does not scale simply, while quantization effects differ across settings. Importantly, many configurations with fewer biased responses also show increased abstention, suggesting that lower bias may partly result from more frequent abstention.
By Kuniko Paxton, Ifeoluwa Wuraola, Dhavalkumar Thakker
Pixel language models treat text as rendered images, enabling strong cross‑lingual transfer. This study trains MPIXEL models varying languages, scripts, and capacity while keeping per‑language compute constant, and evaluates them on five tasks across 83 languages. Results show that scaling improves performance on added languages and scripts, mitigates the curse of multilinguality, and outperforms subword and character baselines on typologically distant low‑resource languages, though transfer to unseen scripts does not benefit further from scaling.
By Ilker Kesen, Jenny Kunz, Kasra Malihi, Desmond Elliott
The paper introduces BiasGuard‑D, a system that uses two adversarial language‑model agents from different families to exchange typed claim documents, with a third model quorum voting to adjudicate. A combined referential‑plus‑semantic invariant validates the verdict, and the iterative deliberation layer shifts remediation from stereotypes to hallucinations, improving right‑answer recovery by 49 percentage points on HaluEval QA. All claims, votes, and instructions are stored as typed records, providing an auditable provenance trail for operators under emerging AI regulations.
By Kardelen Subasi, Jaros{\l}aw A. Chudziak
The paper investigates whether fine‑tuning large language models (LLMs) on value survey responses can shift their value‑conditioned outputs and downstream behavior. Researchers first estimate baseline value profiles for several open‑source LLMs using the ten Schwartz values, then fine‑tune the models on survey‑style supervision that selectively alters target value ratings while preserving non‑target values. The resulting shifts are evaluated on held‑out survey questions and on out‑of‑domain behavioral proxies, such as a contextual moral‑judgment benchmark from Reddit AITA posts and a text‑based interactive decision‑making benchmark. The study finds that survey‑based supervision can change both survey responses and downstream behavior, but the transfer is strongly model‑ and value‑dependent and influenced by how value descriptions are framed.
By Shangrui Nie, Florian Mai, David Kacz\'er, Charles Welch, Zhixue Zhao, Lucie Flek
The study investigates how post‑training alignment techniques affect small language models (SLMs) for generating medication leaflets from structured biomedical data. Four alignment methods—supervised fine‑tuning (SFT), direct preference optimization (DPO), odds ratio preference optimization (ORPO), and group relative policy optimization (GRPO)—are evaluated using Qwen‑2.5 models, with a new FDALeaflets dataset derived from OpenFDA drug labels for zero‑shot cross‑regulatory testing. Results show that aligned SLMs can outperform GPT‑5 on reference‑based metrics, while ORPO and GRPO achieve the highest semantic textual similarity scores in‑domain and on FDALeaflets, respectively, highlighting the importance of alignment for semantic fidelity and generalization.
By Xi Yang, Guodong Liu, Chuqin Li, Fan Wu, Ergin Soysal, Min Jiang, Xing He, Jiang Bian, Yi Guo, Shams Zaman, Thomas Fuchs, Yonghui Wu
VoxProg introduces voxel programming for 3D spatial reasoning, enabling language models to translate task interpretations into executable programs that operate on local voxel views of scene surfaces. By indexing surface geometry on regular grids, the approach isolates task-relevant regions and boundaries, allowing downstream pose programs to compute target SE(3) poses. The method is evaluated on the 71-task VoxProg-Bench benchmark, showing consistent superiority over relation-based, keypoint-based, and mesh-based baselines.
By Zhichao Zhai, Menglei Zhao, Guikun Chen, Zihao Yan, Kanzhi Wu, Wenguan Wang, Jun Xiao
The paper introduces hyperspherical simplex representations derived from softmax outputs and logits to enable backward-compatible retrieval across independently trained models. By exploiting the simplex geometry induced by cross‑entropy training, the authors propose a deterministic linear transformation that realigns class prototypes when new classes are added, eliminating the need for compatibility‑specific training or architectural changes. Theoretical analysis under von Mises–Fisher assumptions shows that these aligned regular‑simplex representations satisfy the formal backward‑compatibility conditions, and empirical results demonstrate competitive performance on standard benchmarks without extra losses or mappings.
By Simone Ricci, Niccol\`o Biondi, Federico Pernici, Alberto Del Bimbo
TraceForge introduces a two‑stage training method that generates reasoning traces in the style of a target model family, using supervised learning followed by adversarial refinement. The system produces traces that are judged as authentic with scores comparable to real traces, and when applied to teacher answers, it improves performance on coding benchmarks such as HumanEval+ and MBPP+. The approach addresses the limitation of existing models that only return answers without detailed reasoning, enabling more transparent and learnable outputs.
By Yan Wang