Reinforcement learning enhances reasoning in multimodal large language models using value-model-free and value-model-based methods.

Guanghao Zhou, Panjia Qiu, Cen Chen et al.

Proposes a framework modeling oversight as a capability-mismatch game using Elo scores, validated on multiple oversight scenarios, deriving scaling laws for success probabilities.

Joshua Engels, David D. Baek, Subhash Kantamneni et al.

By building a large math problem dataset, integrating code execution, and developing candidate solution selection, the model achieves state-of-the-art performance.

Ivan Moshkov, Darragh Hanley, Ivan Sorokin et al.

LiLaVe method extracts correctness signals from LLM hidden states, enhancing task efficiency and accuracy.

Bartosz Piotrowski, Witold Drzewakowski, Konrad Staniszewski et al.

This study critically examines RLVR's ability to induce novel reasoning in LLMs, revealing paths are bounded by base models, with distillation offering better expansion.

Yang Yue, Zhiqi Chen, Rui Lu et al.

Trained an Arabic ASR system using Conformer architecture and weak supervision, achieving superior results.

Mahmoud Salhab, Marwan Elghitany, Shameed Sait et al.

Proposes a hybrid AI router-based super agent system combining local and cloud models for efficient task routing.

Yuhang Yao, Haixin Wang, Yibo Chen et al.

The AI Scientist-v2 employs agentic tree search, multi-modal feedback, and autonomous experiment management to generate peer-review-accepted scientific papers, advancing AI-driven research automation.

Yutaro Yamada, Robert Tjarko Lange, Cong Lu et al.

SmolVLM employs architectural and tokenization innovations to create resource-efficient multimodal models with under 1GB GPU memory, outperforming much larger models.

Andrés Marafioti, Orr Zohar, Miquel Farré et al.

PaperBench evaluates AI's ability to autonomously replicate ICML 2024 papers; top model scores 21%.

Giulio Starace, Oliver Jaffe, Dane Sherburn et al.

Agent S2 employs a hierarchical compositional framework with Mixture-of-Grounding and Proactive Hierarchical Planning, achieving 18.9%-52.8% improvements over SOTA on GUI tasks.

Saaket Agashe, Kyle Wong, Vincent Tu et al.

Agentic LLMs enhance decision-making via reasoning, acting, and interacting, significantly improving medical diagnosis and logistics analysis.

Aske Plaat, Max van Duijn, Niki van Stein et al.

OntoAligner: a modular Python toolkit integrating traditional and AI methods for large-scale ontology alignment, achieving high accuracy with efficiency.

Hamed Babaei Giglou, Jennifer D'Souza, Oliver Karras et al.

Proposes a Plan+Solver multi-agent system with multi-turn query rewriting, achieving 95% query understanding accuracy and 92% task success in office scenarios.

Songtao Sun, Jingyi Li, Yuanfei Dong et al.

Survey on LLMs in mathematical reasoning and optimization, focusing on Chain-of-Thought and tool-augmented methods.

Ali Forootani

This survey systematically analyzes evaluation methods for LLM-based agents, covering core capabilities, application benchmarks, evaluation frameworks, and key dimensions.

Asaf Yehudai, Lilach Eden, Alan Li et al.

Proposes Cosmos-Reason1, integrating hierarchical physics ontology with multimodal LLMs, achieving significant improvements in physical commonsense and embodied reasoning.

NVIDIA, :, Alisson Azzolini et al.

Proposes GoAI to build AI research knowledge graphs, enabling personalized learning paths and innovation support.

Xian Gao, Zongyun Zhang, Ting Liu et al.

CHOP: Enhancing mobile assistants with constrained high-frequency optimized subtask planning.

Yuqi Zhou, Shuai Wang, Sunhao Dai et al.

Multi-agent Gemini-based system accelerates hypothesis generation, improving Elo scores from 3.0 to 4.2 over iterative cycles, validated in drug repurposing and target discovery.

Juraj Gottweis, Wei-Hung Weng, Alexander Daryin et al.