Falcon-H1R uses a hybrid model for efficient reasoning, achieving performance on par with larger models at only 7B parameters.

Falcon LLM Team, Iheb Chaabane, Puneesh Khanna et al.

Proposes a homogenization framework based on critical theory, introducing xeno-reproduction to enhance diversity in LLMs for AI safety.

Ian Rios-Sialer

Proposes CreativeDC, a two-stage prompting framework, boosting Python task diversity by 1.6× with structured divergent-convergent reasoning.

Manh Hung Nguyen, Sebastian Tschiatschek, Adish Singla

ActionFlow system accelerates edge VLA inference by 2.55× using cross-request pipelining, no retraining needed.

Yuntao Dai, Hang Gu, Teng Wang et al.

Defines Scientific General Intelligence (SGI) based on Practical Inquiry Model, evaluates models via SGI-Bench across four core tasks with over 1,000 samples.

Wanghan Xu, Yuhao Zhou, Yifan Zhou et al.

OS-Oracle framework achieves state-of-the-art cross-platform GUI critic model performance with a 310k sample dataset.

Zhenyu Wu, Jingjing Xie, Zehao Li et al.

WebOperator employs action-aware tree search with safe backtracking, achieving 54.6% success on WebArena.

Mahir Labib Dihan, Tanzima Hashem, Mohammed Eunus Ali et al.

ReMe employs multi-faceted distillation, scenario-aware indexing, and utility-based refinement, enabling small models to outperform larger ones without memory, with 8.83% gain on BFCL-V3.

Zouying Cao, Jiaji Deng, Li Yu et al.

PRiSM evaluates scientific reasoning via Python code with 24,750 physics and math problems.

Shima Imani, Seungwhan Moon, Adel Ahmadyan et al.

MemVerse combines hierarchical knowledge graphs with parametric memory, enhancing multimodal lifelong learning with superior scalability.

Junming Liu, Yifei Sun, Weihua Cheng et al.

This study analyzes how preference distribution definitions (hard, soft, goal shaping) impact active inference agent performance in grid navigation.

Filippo Torresan, Ryota Kanai, Manuel Baltieri

Chain-of-Ground enhances GUI grounding through iterative reasoning, achieving 68.4% accuracy on ScreenSpot-Pro.

Aiden Yiliu Li, Bizhi Yu, Daoan Lei et al.

WMAct combines reward rescaling and frequency annealing to enable LLMs to perform efficient world model reasoning via active multi-turn interaction.

Bao Shu, Yan Cai, Jianjian Sun et al.

Combining PPO and DQN in multi-agent RL enables P2P energy trading in dairy farms, reducing costs by 14.2% and peak demand by 55.5%.

Mian Ibad Ali Shah, Marcos Eduardo Cruz Victorio, Maeve Duffy et al.

DeepSeekMath-V2 trains a faithful verifier and proof generator with self-verification, achieving high accuracy in mathematical reasoning.

Zhihong Shao, Yuxiang Luo, Chengda Lu et al.

Proposes VLP, combining VLM perception with program synthesis, achieving significant improvements in complex logical visual reasoning.

Antonia Wüst, Wolfgang Stammer, Hikaru Shindo et al.

HuggingR$^4$ introduces a progressive reasoning framework, achieving 92.03% workability, 82.46% reasonability, and reducing token use by 6.9×.

Shaoyin Ma, Chenggong Hu, Huiqiong Wang et al.

Proposes a taxonomy of 28 cognitive elements, analyzes 192K model and 54 human reasoning traces, and improves complex reasoning by 66.7%.

Priyanka Kargupta, Shuyue Stella Li, Haocheng Wang et al.

HAPPO achieves over 95% load restoration in large-scale distribution systems using heterogeneous multi-agent PPO with centralized critic.

Parya Dolatyabi, Ali Farajzadeh Bavil, Mahdi Khodayar

M-GRPO extends Group Relative Policy Optimization for hierarchical multi-agent systems, aligning heterogeneous trajectories to improve reasoning performance.

Haoyang Hong, Jiajun Yin, Yuan Wang et al.