The paper presents the first large‑scale benchmark for uncertainty quantification (UQ) calibration in long‑form scientific question answering, evaluating four UQ methods on 685,000 responses from up to 20 large language models across seven datasets. It shows that instruction tuning leads to token‑level probability polarization, undermining token‑level uncertainty signals, while reasoning model families differ in how they handle this effect. Only semantic consistency—consistency of the final answer—provides well‑calibrated outputs, demonstrating that semantic calibration remains robust in multi‑step, dependency‑rich reasoning.
By Philip M\"uller, Nicholas Popovi\v{c}, Michael F\"arber, Peter Steinbach
The paper presents a configurable multi‑stage vision pipeline for Farmer.Chat, a farm advisory service that processes farmer‑submitted crop photos. It splits the task into a quality gate, a crop detector, and a disease/pest detector, offering two routes: a single fine‑tuned vision‑language model and a set of small specialist models. The new pipeline improves crop accuracy to 95.41% and provides a fast MobileNetV3 quality gate, while retaining the ability to request better photos and answer all queries in one call.
By Naga Ganesh, Chandrashekar M S, Lakshmi Pedapudi, Aakash Singh, Vineet Singh
EFQ-Softmax is a low‑bit probability‑generation technique that replaces the traditional exp‑then‑quantize path in Transformer attention. It maps shifted attention scores directly to block‑scaled E2M1 operands using an exponent‑only scale and a single affine rule, allowing the same low‑bit representation to be used for both numerator and denominator updates. Experiments on Qwen3‑8B, Qwen3‑VL‑8B‑Instruct, and WAN2.2‑TI2V‑5B show that EFQ‑Softmax maintains or improves model quality while reducing vector‑stage latency by about 40% on the A5 vector unit.
By Haohui Han (Xi'an Jiaotong University), Yuming Wan (Huawei Technologies Co., Ltd), Hongni Wang (Shandong University of Finance and Economics), Pengcheng Xie (Huawei Technologies Co., Ltd), Xiaodong Yan (Xi'an Jiaotong University), Runqi You (Xi'an Jiaotong University), Wencong Zhang (Xi'an Jiaotong University)
Ripple-Pivot Search (RPS) is a training‑free decoding method for Diffusion Large Language Models that identifies mid‑entropy pivot positions to reduce uncertainty across remaining masked tokens. By proactively committing these pivots and evaluating token assignments via lookahead, RPS enables more tokens to be unmasked in parallel, speeding up decoding. Experiments on three dLLMs and four reasoning/code‑generation benchmarks show 4–10× wall‑clock speedup over standard decoding, up to 18× with KV caching, while maintaining or improving generation quality.
By Yushi Ye, Xu Chen, Haoyun Jiang, Jinsong Lan, Haihong Tang, Xiangtao Li, Mingming Gong, Ivor Tsang, Yanfeng Wang, Jiangchao Yao
The paper introduces the Memory Decision Layer (MDL), a zero‑parameter controller that sits between retrieval and generation in large language models. MDL uses a three‑signal complementary encoder—combining relevance, reliability, and task risk—to produce an interpretable decision about the trustworthiness of retrieved memories. By decoupling confidence from consistency and enabling risk inversion and abstention, MDL cuts hallucination rates under conflicting memories by roughly 56% and nearly eliminates them in high‑risk scenarios, all while adding only 0.14 ms per decision.
By Yiming Zhang, Jinghong Zhang, Haoran Zhao, Yiren Ma, Chunlei Zhao
JudgeSense is a benchmark comprising 880 items from human‑labelled corpora, each presented under two differently worded instructions that ask the same question. The study evaluates 25 judges from six providers across four tasks, measuring how rewording affects agreement with the judge’s own verdicts. Results show that rewording reduces agreement on all tasks, with significant effects on two, and that stability varies across tasks and is not predicted by parameter count.
By Rohith Reddy Bellibatlu, Edward Raff, Wenbin Zhang
The paper introduces a multimodal reasoning framework for cross‑domain visual question answering in Printed Circuit Board Assembly (PCBA) inspection, converting diverse data sources into a unified instruction format and generating verified reasoning traces. It proposes Task‑Aware Group Relative Policy Optimization (GRPO) that uses semantic, distance‑aware, and format rewards to improve choice‑based and counting tasks beyond exact‑match supervision. During inference, the system applies semantic consistency correction, self‑consistency voting, and multi‑model arbitration, achieving an overall score of 83.24 on the PCBA Standard‑to‑Real Grand Challenge leaderboard.
By Jia Li, Li Dai, Peng Jia, Zhenzhen Hu, Chee Seng Chan, Bingkun Bao, Richang Hong
MOSAIC‑SR is a new symbolic regression method that combines a pretrained Transformer with search‑based refinement. The Transformer generates multiple initial sketches, which seed searches that jointly recover equation structure and constants using scale‑aware optimization and symbolic repair. On the SRSD‑Feynman dataset and six other benchmarks, MOSAIC‑SR achieves the highest symbolic solution rate and ranks among the top two in predictive accuracy, even when irrelevant dummy variables are present.
By Peiyi Zheng, Yanming Kang, Hans De Sterck, Giang Tran
GraphSkillEvo introduces a graph-structured representation for agent skills, where each node encodes an execution step and edges capture context-dependent transitions. This structure offers clearer workflow guidance and reduces redundancy compared to unstructured natural-language skills. The authors then present a population-based evolutionary optimization framework that explores this structured skill space, achieving higher accuracy than the baseline SkillOpt across five agent benchmarks.
By Rui Sun, Zhi Zheng, Zhenkun Wang, Zhichao Lu
CoLearn is an interactive, agentic tutoring system that learns about each learner through a persistent memory of mastery and misconceptions, updated with a Bayesian Knowledge Tracing model that uses a large language model as an observation function. It generates personalized questions targeting the learner’s weakest topics and recurring misconceptions, and provides a live evidence view for progress visualization and blind A/B comparison. In blind A/B tests, learners preferred questions conditioned on this memory 68‑69% of the time, and simulations show the agent’s belief converges toward the learner’s true mastery.
By Kailai He, Zhihao Wu, Linhai Zhang, Runcong Zhao, Yulan He, Jiazheng Li
arXiv:2609. 20945v1 Announce Type: new Abstract: Undesired information such as harmful content and private data propagates through Multilingual Large Language Models (LLMs) via direct training and indirect cross-linguistic spread.
By Kyomin Hwang, Hyeonjin Kim, Hyunho Lee, Yearim Kim, Yeji Song, Nojun Kwak
The study compares temporal deep learning models—Bidirectional LSTM, Temporal Convolutional Network, and Transformer—for physiological emotion recognition using two multimodal wearable datasets, WESAD and EmoWear. Experiments evaluate wrist-only, chest-only, and multimodal sensor configurations with participant-independent leave-one-subject-out cross-validation, and also explore ensembles, sensor ablation, sampling frequency, and saliency analysis. Results show that the best architecture varies by dataset, multimodal sensing consistently outperforms single-site configurations, and a 4 Hz sampling rate offers a cost-effective operating point.
By Desta Haileselassie Hagos, Saurav Keshari Aryal, Legand L. Burge
The paper reports on programming AMD XDNA NPUs for the FlashAttention workload using open‑source IRON and MLIR‑AIR compiler tools. It compares four reference designs on XDNA 1 and XDNA 2, showing that a fused kernel that keeps QKᵀ scores in local memory achieves 3.62 TFLOP/s on XDNA 2, doubling throughput and greatly improving energy efficiency over the IRON design and the integrated GPU. Roofline analysis guides when to fuse or stream operators based on each device’s ridge points, and the authors release the reference designs as open source.
By Erwei Wang, Ephrem Wu, Victor J. B. Jung, Jiajie Li, Andre Rosti, Joseph Melber, Samuel Bayliss
SignGPT is a unified, pose‑based framework that performs gloss‑free sign language translation (SLT) and generation (SLG) by integrating part‑aware hierarchical representations of body, hand, and facial motion into a shared language model. It uses asymmetric multi‑token prediction and progressive training for bidirectional modeling, and is evaluated on How2Sign (ASL) and Phoenix‑2014T (DGS) with benchmark comparisons, qualitative analyses, and component ablations. An exploratory study with 12 Deaf ASL signers demonstrates a sign‑to‑sign response pipeline, suggesting that unified modeling can support sign language conversation (SLC).
By Ronghui Li, Jun Dong, Zhongyuan Hu, Zunnan Xu, Jun Zhou, Liyuan Chen, Shuoling Liu, Jiangpeng Yan, Jie Guo, Xiu Li, Linchao Bao
COMPLEX is a closed‑form, training‑free embedding for multiparameter persistence modules that provides both an upper and a lower Lipschitz bound, enabling faithful feature representations. By slicing modules along a near‑diagonal net and embedding each slice with the certified PLACE/PALACE landmark map, the method guarantees that separated modules remain separated in the embedding. On Orbit benchmarks and molecular graph tasks, COMPLEX achieves state‑of‑the‑art accuracy, outperforming existing landmark, transformer, and graph‑based approaches.
By Sushovan Majhi, Atish Mitra, \v{Z}iga Virk, Pramita Bagchi
The paper investigates how gating the value pathway in attention mechanisms provides two missing capabilities of softmax attention: abstention and noise filtering. Experiments on models ranging from 10M to 350M parameters show that abstention benefits smaller models while noise filtering becomes more advantageous as models scale, and that combining both primitives yields the best performance across all sizes. The authors also demonstrate that the gates effectively suppress interference and that each gate type has a distinct blind spot, all while adding negligible parameters and preserving compatibility with key‑value caching.
By Richard Zhe Wang
The paper investigates neural scaling laws for RydbergGPT, an autoregressive transformer trained on qubit measurement data from Rydberg atom arrays. Near a critical point in the quantum system, the transformer’s loss scales with dataset size following a power‑law with a loss‑floor correction, whereas this relationship weakens away from criticality. By comparing entropy‑normalized mutual‑information two‑point functions of Rydberg data and natural‑language corpora, the authors find that near‑critical statistics resemble natural language more closely, suggesting that multi‑scale dependence underlies stable neural scaling and that scaling behaviour depends on the model–data pair.
By David S. Berman, Ying-Jer Kao, Roger G. Melko, Alexander G. Stapleton
K2-V2 is a fully open, 360‑open large language model built from scratch, designed to serve as a superior base for reasoning adaptation while also supporting conversation and knowledge retrieval. It competes with leading open‑weight models in its size class, outperforming Qwen2.5‑72B and approaching Qwen3‑235B, and incorporates domain knowledge, reasoning, long‑context handling, and tool use throughout training. The authors release the complete training history, data composition, model weights, and LLM360 artifacts to enable community use and continuous training.
By K2 Team, Zhengzhong Liu, Liping Tang, Linghao Jin, Haonan Li, Nikhil Ranjan, Desai Fan, Shaurya Rohatgi, Richard Fan, Omkar Pangarkar, Huijuan Wang, Zhoujun Cheng, Suqi Sun, Seungwook Han, Bowen Tan, Gurpreet Gosal, Xudong Han, Varad Pimpalkhute, Shibo Hao, Ming Shan Hee, Joel Hestness, Haolong Jia, Liqun Ma, Aaryamonvikram Singh, Daria Soboleva, Natalia Vassilieva, Renxi Wang, Yingquan Wu, Yuekai Sun, Taylor Killian, Alexander Moreno, John Maggs, Hector Ren, Guowei He, Hongyi Wang, Xuezhe Ma, Yuqi Wang, Mikhail Yurochkin, Eric P. Xing
The study evaluates automatic tooth segmentation on panoramic radiographs using a large annotated corpus of 1,422 images and 42,142 tooth polygons. It finds that increasing input resolution improves boundary precision (mask mAP50‑95 rises from 0.656 to 0.717) while detection performance remains unchanged, and that architectural changes have minimal impact on in‑domain accuracy. Targeted interventions such as LoRA adaptation, promptable foundation models, and anatomical label assignment provide negligible gains, indicating that resolution and acquisition diversity should be prioritized over model novelty.
By Muhammad Rehan, Moaz Amjad, Syed Danial Ahmed, Mariam Adnan, Haider Ali
The study evaluates the robustness of medical vision‑language models for tuberculosis screening on chest X‑rays by testing them across multiple datasets, prompts, and evaluation settings. Three specialized models (BioMedCLIP, CheXficient, MedSigLIP) and a general OpenCLIP model were audited on 12,200 images, producing 244,000 model–image–prompt scores. Results show that no model consistently outperforms others across all cohorts and reliability criteria, with prompt changes and control group composition significantly affecting AUROC, and that high training‑set performance does not reliably transfer to external cohorts.
By Mushir Akhtar, M. Tanveer, Mohd. Arshad