Introduces QFrBLiMP, a human-annotated Quebec-French minimal pairs benchmark, evaluating LLMs' grammatical competence with detailed analysis.

David Beauchemin, Pier-Luc Veilleux, Johanna-Pascale Roy et al.

Proposes GCM, leveraging multi-model history to improve calibration and generalization, surpassing self-prediction.

Hanqi Xiao, Vaidehi Patil, Hyunji Lee et al.

InfLLM-V2 introduces a dense-sparse switchable attention mechanism, achieving 4× speedup with 98%+ performance on long-sequence tasks.

Weilin Zhao, Zihan Zhou, Zhou Su et al.

Samiksha method evaluates LLMs in India's healthcare using community feedback, enhancing multilingual model performance.

Hamna Hamna, Gayatri Bhat, Sourabrata Mukherjee et al.

Introduces EvolveCast framework to evaluate LLMs' forecast updates; finds models are overly conservative and inconsistent in belief revision.

Zhangdie Yuan, Zifeng Ding, Andreas Vlachos

DocPruner employs adaptive patch pruning based on global token attention, reducing storage by 50-60% with minimal performance loss.

Yibo Yan, Guangwei Xu, Xin Zou et al.

A dual-distribution study finds weak-to-strong judge transfer fails, while DPO refreshes deliver up to 7.6 percentage points.

Janvijay Singh, Austin Xu, Yilun Zhou et al.

Introduces a geometric credal set framework to quantify and decompose uncertainty in language models, analyzing 500 prompts with only 0.434 calibration at best.

Esteban Garces Arias, Julian Rodemann, Christian Heumann

Using n-gram novelty as a sole metric for textual creativity is inadequate; models and experts show high correlation but also significant divergence, emphasizing multi-dimensional evaluation.

Arkadiy Saakyan, Najoung Kim, Smaranda Muresan et al.

Proposed AMBS method improves HHH performance to 56.5% on LLaMA-2-7B while maintaining efficiency.

Gautam Siddharth Kashyap, Mark Dras, Usman Naseem

RepT framework uses representation gradients to accurately trace undesirable LLM behaviors at sample and token levels, outperforming gradient-based baselines.

Zhe Li, Wei Zhao, Yige Li et al.

Study reveals syntactic-domain spurious correlations in language models, affecting performance.

Chantal Shaib, Vinith M. Suriyakumar, Levent Sagun et al.

EmbeddingGemma is a lightweight 300M-parameter text embedding model surpassing larger models in multilingual and multi-task benchmarks, suitable for low-latency applications.

Henrique Schechter Vera, Sahil Dua, Biao Zhang et al.

Qwen3-Omni uses Thinker-Talker MoE architecture to achieve state-of-the-art multimodal performance without degradation across text, image, audio, and video tasks.

Jin Xu, Zhifang Guo, Hangrui Hu et al.

DIWALI creates a dataset of 8,817 Indian cultural items across 17 facets to evaluate LLMs' cultural adaptation, combining automatic and human assessments.

Pramit Sahoo, Maharaj Brahma, Maunendra Sankar Desarkar

SEGALE extends automatic MT evaluation to long documents, outperforming existing schemes with high correlation to human judgments.

Kuang-Da Wang, Shuoyang Ding, Chao-Han Huck Yang et al.

LiteLong employs hierarchical topic organization and multi-agent debate for resource-efficient long-text data synthesis.

Junlong Jia, Xing Wu, Chaochen Gao et al.

TextMineX enhances humanitarian mine action text knowledge extraction accuracy by 44.2% using ontology-guided LLM pipeline.

Chenyue Zhou, Gürkan Solmaz, Flavio Cirillo et al.

ShinkaEvolve employs parent sampling, novelty rejection, and bandit-based LLM ensemble to achieve sample-efficient open-ended program evolution, outperforming existing methods with only 150 samples.

Robert Tjarko Lange, Yuki Imajuku, Edoardo Cetin

Proposes geometric volume and suspicion metrics for improved hallucination detection and correction in LLMs.

Edward Phillips, Sean Wu, Soheila Molaei et al.