Proposes UQ, a dynamic benchmark assessing LLMs on 500 unsolved questions sourced from Stack Exchange, integrating validator-assisted screening and community verification.

Fan Nie, Ken Ziyu Liu, Zihao Wang et al.

CORE-RAG employs performance-driven reinforcement learning to compress context to 3%, boosting EM scores by 3.3 points, outperforming full-document baseline.

Ziqiang Cui, Yunpeng Weng, Xing Tang et al.

Proposes entity-oriented search with graph query language, achieving state-of-the-art results on WikiTQ and TabFact datasets.

Thi-Nhung Nguyen, Hoang Ngo, Dinh Phung et al.

Proposes PING, an automatic prefix injection method, to enhance LLM safety in agentic tasks while maintaining performance.

Dongyoon Hahm, Taywon Min, Woogyeol Jin et al.

FineCE is a supervised fine-grained confidence estimation method that significantly improves early and continuous confidence prediction during text generation.

Jinyi Han, Tingyun Li, Shisong Chen et al.

Employing Mistral 7B and mT5 models, this study advances Czech text summarization, establishing new benchmarks for modern and historical documents.

Václav Tran, Jakub Šmíd, Jiří Martínek et al.

IROTE uses self-reflective optimization to stably elicit human traits in LLMs, outperforming baselines.

Yuzhuo Bai, Shitong Duan, Muhua Huang et al.

Unigram outperforms BPE in complex morphology languages, especially in Telugu.

Saketh Reddy Vemula, Sandipan Dandapat, Dipti Misra Sharma et al.

GLM-4.5 uses deep MoE architecture with 355B parameters, integrating thinking and response modes, achieving top ARC benchmark results.

GLM-4. 5 Team, :, Aohan Zeng et al.

AURA employs a multi-layered Process Reward Model to enhance logical coherence and safety in LLM reasoning, reducing risks by 20-50%.

Sayantan Adak, Pratyush Chatterjee, Somnath Banerjee et al.

Introduces UrBLiMP, a benchmark with 5696 minimal pairs for evaluating Urdu syntax in multilingual LLMs, achieving up to 94.73% accuracy.

Farah Adeeba, Brian Dillon, Hassan Sajjad et al.

Proposed MAO-ARAG multi-agent framework with reinforcement learning for adaptive retrieval-augmented generation, boosting QA performance.

Yiqun Chen, Erhan Zhang, Lingyong Yan et al.

Proposes multi-layer attention as an amplifier of demonstration effectiveness; uses gradient flow to select demonstrations, achieving 6.8% average improvement.

Dingzirui Wang, Xuangliang Zhang, Keyan Xu et al.

Constructed Bengali multilingual benchmarks, evaluated 10 open-source LLMs, found size and tokenization impact performance, with larger models showing more robustness.

Shimanto Bhowmik, Tawsif Tashwar Dipto, Md Sazzad Islam et al.

ControlMed enhances medical language model efficiency by controlling reasoning length, outperforming existing models in experiments.

Sung-Min Lee, Siyoon Lee, Juyeon Kim et al.

Introduced VLQA, a large Vietnamese legal QA dataset, evaluated with BERT, RoBERTa, GPT-4, revealing strengths and limitations in legal reasoning.

Tan-Minh Nguyen, Hoang-Trung Nguyen, Trong-Khoi Dao et al.

MCIF benchmark evaluates crosslingual, multimodal instruction-following across 4 tasks, 3 modalities, 4 languages, revealing key challenges in current models.

Sara Papi, Maike Züfle, Marco Gaido et al.

GEPA combines natural language reflection with Pareto-based evolutionary search, achieving 6% average improvement with 35x fewer samples than GRPO.

Lakshya A Agrawal, Shangyin Tan, Dilara Soylu et al.

Prune&Comp combines iterative layer pruning with offline magnitude compensation, nearly halving perplexity and retaining 93.19% QA performance on LLaMA-3-8B without training.

Xinrui Chen, Hongxing Zhang, Fanyi Zeng et al.

Hierarchical Memory (H-MEM) organizes long-term memory in multi-level structure, improving retrieval efficiency by 5x and outperforming five baselines in long-term dialogue tasks.

Haoran Sun, Shaoning Zeng