PEEU framework enables small multimodal models to achieve 30.6% success in web navigation by autonomous exploration and hindsight experience, outperforming larger models.

Tianyi Men, Zhuoran Jin, Pengfei Cao et al.

This study pioneers the application of Large Language Models (LLMs) for securities eligibility screening at the German Central Bank, achieving up to 91% document-level precision through a generative information extraction pipeline.

Serhii Hamotskyi, Akash Kumar Gautam, Christian Hänig

Introduces syntactic belief update model using Rényi divergence, outperforming lexical surprisal in predicting human reading difficulty on garden path sentences.

Alan Zhou, Miloš Stanojević, John T. Hale

MinGram uses BPE seed, Hard EM on minimal path, and single pruning to improve compression and morphology.

Sander Land

CAT-Q enables ultra-efficient ternary quantization of large language models using only 512 calibration samples, outperforming state-of-the-art QAT methods on models up to 235B parameters.

Shigeng Wang, Chao Li, Yangyuxuan Kang et al.

This study introduces a low-resource end-to-end Algerian dialect speech-to-speech system using Whisper, achieving 13.74% WER, integrating ASR, NLU, RAG, and TTS.

Dihia Lanasri, Fairouz Taki, Asma Kemmoum

Hybrid-IR combines graph-based and dense retrieval with iterative reasoning, achieving 10% accuracy improvement on medical QA benchmarks.

Sheng Wan, Jiahui Zhang, Zicheng Zhao et al.

Proposes a training-free method using intermediate-layer entropy trajectories to detect jailbreak prompts, achieving AUROC up to 0.94 across Llama, Qwen, and Gemma.

Sofiia Nikolenko, Michele Papucci, Mina Rezaei et al.

Proposes a four-module analytical framework to evaluate 12 agent memory systems across diverse workloads, revealing workload-architecture matching importance.

Wei Zhou, Xuanhe Zhou, Shaokun Han et al.

NatureBench evaluates AI coding agents on 90 scientific tasks; strongest model surpasses SOTA in only 17.8%.

Yuru Wang, Lejun Cheng, Yuxin Zuo et al.

LangMAP extends UnigramLM with language-specific MAP estimation, enabling multi-language tokenization from a shared vocabulary, improving morphological boundary alignment.

Clara Meister, Suchir Salhan, Andrzej Szablewski et al.

Proposes a dual-track framework combining rule engines and LLMs, boosting LaTeX template conversion fidelity and success rate.

Chung Cheuk Hei, Liu Li

DynamicMem introduces a long-horizon benchmark with 15-month synthetic user trajectories, evaluating multi-application memory and causal inference for personalized AI assistants.

Wenya Xie, Shengming Zhou, Zelin Li et al.

BLUEX v2 evaluates 21 LLMs on Brazilian second-phase exams, using multimodal, open-ended questions, with performance gap of 4.92 points.

João Guilherme Alves Santos, Giovana Kerche Bonás, Thiago Laitz et al.

Confident Layer Decoding dynamically selects the most reliable near-final layer to enhance reasoning in large language models.

Xuanming Zhang, Sining Zhoubian, Yuxuan Chen et al.

Hierarchical Transformer detects multi-turn jailbreaks with 0.9394 F1, outperforming baselines.

Chenhui Hu, Muhammed Salih, Sudipto Guha et al.

Proposed ViGiL3D++ combines scene graph sampling and LLM rephrasing to generate diverse 3D grounding descriptions, improving dataset richness.

Austin T. Wang, Dongchen Yang, Angel X. Chang

Proposes the nine-dimensional Meaning Intelligence Framework (MIF) to distinguish surface sentiment from true intent in Nigerian discourse; zero-shot accuracy 33.3%, schema-guided 73.3%.

Celestine Achi

Proposes a control-window law based on geometric cosine thresholds to predict single-neuron controllability, validated on 15 neurons with mean error 0.14.

Hongliang Liu

Generated multilingual mental health dialogue datasets by modifying nationality and language parameters, revealing clinical inconsistency issues.

Yunkai Xu, Saeed Abdullah