The paper introduces ProPerMem, a benchmark for memory-driven proactive assistance that pairs egocentric video with long, distractor-rich personal memories and human-verified trigger windows. Baseline models perform poorly, achieving at most 15% accuracy, while an oracle-driven GPT-5.3-Chat model reaches 96% when given evidence inside the trigger window. Building on these insights, the authors develop ProPerAgent, a plug‑and‑play system that retrieves relevant memory from a graph store and uses lightweight reinforcement learning to improve proactive intervention accuracy to 42.4% without training.
By Shenglai Zeng, Xiao Yang, Tiffany Ma, Xinyuan Zhang, Nanshu Wang, Deren Lei, Kai Sun, Xilun Chen, Tammy Stark, Xianhui Zhu, Yuetai Li, Rulin Shao, Tianxin Wei, Rakesh Wanga, Anuj Kumar, Jiliang Tang, Wen-tau Yih, Xin Luna Dong
NLData2Opt is a new benchmark that evaluates large language models on the multi‑stage reasoning required to formulate optimization problems from natural language descriptions and accompanying data. It splits the task into three steps: extracting relevant parameters from data that includes misleading attributes, computing missing parameters by transforming existing data, and predicting future values from historical observations. The benchmark shows that performance on existing, idealized tests does not translate to real‑world accuracy, highlighting a gap in current LLM capabilities.
By Lijia Liu, Kyohei Atarashi, Jiyi Li, Koh Takeuchi, Shunji Umetani, Hisashi Kashima
The paper presents a Bengali hallucination detection system that uses response‑conditioned query rewriting, hybrid retrieval, and task‑specific language‑model judging. It employs six retrieval routes and a router that selects between a 27B textual judge and a 14B agent capable of reasoning or executing Python. The authors evaluate the system on the Olikbochon challenge, reporting a private macro‑F1 of 0.956 and detailing runtime, execution traces, and corpus overlap findings.
By Shamim Hasan, Samiya Ahmed Chowdhury
The paper introduces “KoLing”, a bias‑controlled adversarial benchmark of 50.9k Korean instances designed to suppress superficial cues while preserving semantic content, and evaluates both LLM‑generated and LLM‑modified text detection. It also proposes “KoLingProbe”, an interpretable detector grounded in theoretical linguistics that models Korean‑specific properties across multiple linguistic levels. Experiments demonstrate that KoLingProbe outperforms strong baselines on all detection tasks and remains robust to adversarial attacks, with analyses showing that integrating complementary linguistic features across multiple levels improves both performance and interpretability.
By SungHo Kim, SangKeun Lee
The paper investigates whether large models are necessary for speech brain‑computer interfaces (BCIs) and introduces BEST, a two‑stage framework that uses hierarchical intermediate representations and a compact decoder to map neural activity directly to text. BEST achieves low word error rates (6.19% and 1.62%) on the Brain‑to‑Text ’24 and ’25 benchmarks while using only 3.85% of the parameters of comparable methods. The study also highlights that decoder scaling yields non‑monotonic gains, cross‑modal alignment is more critical than model size, and overfitting harms sequence‑level performance.
By Zehui Feng, Weichuan Wang, Xiaohan Chen, Cuntai Guan, Ting Han
The paper introduces DarkLens, a lightweight method for detecting faithfulness hallucinations in large language models (LLMs). It estimates context reliance by projecting hidden states onto the dark space—directions weakly aligned with the unembedding matrix—and uses the resulting Dark Matter Ratio (DMR) as a direct score of context utilization. Combined with a parametric knowledge score, DarkLens operates within a single forward pass and outperforms many baselines on QA and summarization benchmarks across several open‑weight LLMs.
By Alexandra Kuleshova, Andrey Volodichev, Daria Kotova, Alexey Zaytsev
AMEND is an inference‑time framework for text detoxification that separates detoxification into role‑conditioned proposal generation, isolated multidimensional criticism, and constraint‑first selection. It evaluates diverse candidates with role‑blind critics, enforces strict semantic and fluency feasibility, and ranks them by utility. The lightweight AMEND Lite variant limits the search to a single targeted repair round when residual toxic spans remain, achieving higher joint scores and reducing inference requests by 28.4% compared to unconstrained multi‑round loops.
By Tianyu Wang
The paper presents a method for inferring a speaker’s first language (L1) from their second‑language English (L2) by using token probability scores from large language models (LLMs). By treating each text as a continuation of a prompt that specifies a candidate L1, the authors identify the L1 with the highest probability score. Experiments with Llama 3.1 8B and 70B on TOEFL9, Write and Improve 2024, and BULATS‑L1 datasets show that token‑probability scoring outperforms direct classification, with the 70B model achieving state‑of‑the‑art results across all datasets.
By Ahmet Yavuz Uluslu, Mark Gales, Kate Knill, Gerold Schneider
The paper argues that many vision‑language model (VLM) hallucination metrics mistakenly treat every object mentioned in a caption as a claim that the object is present. It introduces NAME‑Bench, a dataset of 25,497 non‑assertive mentions labeled as depictive, purpose, negation, or hedged, and shows that current detectors flag these as hallucinations at high rates. Adding non‑assertive mentions as negatives improves detection on LLaVA‑1.5 but requires much more data for newer VLMs, indicating that conflating mention with assertion is a persistent problem.
By Hsin-Ling Hsu, Neeraja Kirtane, Kuan-Hao Huang
The paper introduces SoCo, a compression framework that optimizes the singular spectrum of large language models using a learnable diagonal matrix to assign importance scores to SVD components. By progressively refining these scores through a three‑stage training process, SoCo adaptively prunes components while amplifying the importance of remaining ones, achieving a balance between aggressive compression and performance preservation. Experiments on multiple LLMs and benchmarks show that SoCo outperforms state‑of‑the‑art compression methods.
By Dengjie Li, Tiancheng Shen, Yao Zhou, Bernard Ghanem, Yibo Yang, Ming-Hsuan Yang
The paper introduces Reasoning DAG Probing, a method that treats premises, intermediate conclusions, and final answers as nodes in a directed acyclic graph (DAG). By training lightweight probes on LLM hidden states, the authors show that DAG structure is encoded in the models, peaking in intermediate layers and becoming stronger with larger models. During autoregressive generation, the latent DAG structure emerges progressively, anticipating answer correctness and enabling correction of some incorrect outputs by steering activations toward the ground‑truth DAG.
By Tianjun Zhong, Linyang He, Ziyang Li, Nima Mesgarani
PragReST is a self‑supervised framework that enhances large language models’ pragmatic reasoning by generating counterfactual reasoning traces and fine‑tuning models with supervised learning and reinforcement learning, all without human‑labeled data. The method constructs pragmatic QA data, trains models to internalize counterfactual reasoning, and demonstrates significant accuracy gains on four pragmatic benchmarks compared to backbone models and other baselines. Error analysis shows that counterfactual reasoning reduces mistakes from failing to contrast observed utterances with plausible alternatives, and removing it markedly degrades performance while preserving out‑of‑domain general‑knowledge and mathematical reasoning.
By Jihyung Park, Minchao Huang, Leqi Liu, Elias Stengel-Eskin
The paper introduces Referring Multi-Object Tracking (RMOT), a task that uses natural language expressions to guide the prediction of multiple objects over time, addressing limitations in previous work that lacked dynamic language expressiveness. It presents the Refer-KITTI-V2 benchmark, which includes 9,758 semi-automatically generated language prompts for diverse object dynamics. The authors also propose TempRMOT, a Transformer-based framework with a Temporal Enhancement Module that achieves state‑of‑the‑art results on both Refer‑KITTI and Refer‑KITTI‑V2.
By Yani Zhang, Dongming Wu, Wencheng Han, Xingping Dong, Shengcai Liao, Bo Du
The paper introduces the Novel Visual References Dataset (NVRD), comprising 19,176 images of 90 new visual concepts with up to 20 perturbed versions each, designed to test how vision‑language models and humans map novel visual references to language. The authors evaluate three open‑source and two closed‑source VLMs against 2,400 human judgments, finding that models struggle to learn new concepts that contradict pre‑training knowledge and tend to overgeneralize labels to stimuli humans reject. NVRD is released as a benchmark for studying visual concept learning in both humans and machines.
By Ada Defne T\"ur, Gaurav Kamath, Joyce Chai, Siva Reddy, Benno Krojer
The paper introduces SABER, a benchmark that tests attribute–object binding in vision‑language models across all four modality pairings and scenes with up to ten objects. Experiments show significant binding failures in both CLIP‑style dual encoders and recent MLLM‑based embedders. The authors propose a binding‑focused fine‑tuning strategy using synthetic attribute‑binding text, which improves SABER performance and also boosts results on negation and instance‑retrieval benchmarks.
By Netanel Y. Tamir, Michael Green, Issar Tzachor, Or Shimshi, Dvir Samuel, Or Patashnik, Rami Ben-Ari
The paper introduces the Character Verification Model (CVM), a post‑processing framework that uses binary classifiers to resolve visually similar character ambiguities such as 'O'/'0', 'Z'/'2', and 'S'/'5'. CVM operates independently of dictionary checks, making it applicable to non‑lexical data like serial numbers and license plates. Experiments on three OCR engines and ten fonts show accuracy gains of 3.95–14.81 percentage points, an 87.3% correction rate for ambiguous characters, and a low 2.3% false‑correction rate.
By Nuzhat Khan, Ab Al-Hadi Ab Rahman, Mohamad Anuar Kamaruddin, Muhammad Paend Bakht, Sana Fatima, Farhad Khodadadi, Ibrahim Yousef Alshareef
The study evaluates correction techniques for cross‑hospital immunohistochemistry (IHC) grading using frozen pathology foundation models (FMs). It benchmarks label‑free transformations applied to input images (including a new method called FIX) and to features after FM encoding across two hospitals, three IHC markers, and six transfer scenarios. The results show that FIX and simple feature‑wise standardization each outperform existing methods, and their combination yields significant gains in macro‑F1, quadratic weighted kappa, and ordinal error.
By Jongwoon Lee, Jhair Gallardo, Christopher Kanan
VISTA (Visually Impaired Scene and Task Annotations) is a multimodal egocentric benchmark designed to evaluate goal‑oriented assistance for blind and low‑vision users. It comprises about 1,000 recordings from Project Aria glasses, each providing six synchronized sensor streams and three types of annotations across ten task categories, from object localization to hazard detection. The benchmark tests vision‑language models in zero‑shot settings and introduces a lightweight multimodal baseline that integrates all six modalities, revealing that current models can describe scenes but struggle to generate effective assistive guidance.
By No\'e Mosca, Mohamed Rayan Barhdadi, Muhammad Hamza, Joshua Yeh, Suryodaya Bikram Shahi, Zhiyan Li, Selim-Antoine Lali, Weiyue Li, Hasan Kurban, Erchin Serpedin, Gang Luo, Yue Zhang, Minghan Li, Mengyu Wang
MaskMerge introduces a structure-aware token merging strategy for Vision Transformers that leverages coarse mask predictions to guide token fusion, preventing cross-instance token merging that plagues existing methods. In high inter-instance similarity settings such as PanNuke and iSAID, MaskMerge boosts mAP from 30.1 to 32.1 and 33.1 to 36.2 respectively, while cutting FLOPs and speeding up inference. The approach also matches state‑of‑the‑art performance on Cityscapes with fewer computations and transfers well to MAE‑pretrained backbones.
By Bavesh Balaji, Letian Jiang, Alexander Wong, David Clausi, Sirisha Rambhatla
A100 is a hybrid CNN‑Transformer vision foundation model trained on 3 million unlabeled agricultural images using self‑supervised techniques (DINO, iBOT, KoLeo). It achieves superior crop classification (95.37 % with a frozen backbone) and phenotyping accuracy, and after fine‑tuning reaches 99.90 % on 32‑class crop identification and 99.02 % on 44‑class disease detection. With only 1 % of labeled data, A100 attains 98.01 % plant classification accuracy, outperforming existing baselines and other agricultural foundation models by significant margins.
By Mayur Pratim Das, Ajai Dubey, Aksshay Mathew P, Sana Afreen, Rishi Kumar Soni, Deepak Meena, Pratibha Kumari, Mukesh Saini