Small language models (SLMs) are shown to match or surpass large models in agentic AI tasks, offering lower costs and higher efficiency, with a migration algorithm proposed.

Peter Belcak, Greg Heinrich, Shizhe Diao et al.

PGPO introduces pseudocode plans and preference optimization, boosting LLM reasoning with 11.6% performance gain.

Zouying Cao, Runze Wang, Yifei Yang et al.

Proposes a dynamic, socially-aware AI alignment framework emphasizing environment-dependent interactions in multi-agent systems.

Florian Carichon, Aditi Khandelwal, Marylou Fauchard et al.

Proposes Conformal Arbitrage, a risk-controlled model routing framework using conformal risk control to balance objectives with finite-sample guarantees, operating solely at API level.

William Overman, Mohsen Bayati

Proposed OWL framework with modular multi-agent architecture achieves 69.70% accuracy on GAIA, surpassing OpenAI Deep Research by 2.34%.

Mengkang Hu, Yuhang Zhou, Wendong Fan et al.

MathArena leverages newly released math competitions for real-time, uncontaminated evaluation of LLM reasoning and proof-writing, surpassing static benchmarks.

Mislav Balunović, Jasper Dekoninck, Ivo Petrov et al.

Infi-MMR employs a three-phase curriculum reinforcement learning framework to enhance multimodal small language models, achieving 43.68% on MathVerse.

Zeyu Liu, Yuhang Liu, Guanghao Zhu et al.

LCoT2Tree converts long reasoning into trees, raising correctness prediction by 5.63% on average over length baselines.

Gangwei Jiang, Yahui Liu, Zhaoyi Li et al.

Alita uses minimal predefined tools and self-evolution, achieving 75.15% pass@1, outperforming complex systems.

Jiahao Qiu, Xuan Qi, Tongcheng Zhang et al.

DiffVLA integrates vision-language models with hybrid sparse-dense diffusion for autonomous driving, achieving 45.0 PDMS on navsim-v2.

Anqing Jiang, Yu Gao, Zhigang Sun et al.

Proposes LatentR3, a reinforcement learning framework for efficient, explicit-chain-free recommendation with 17% performance gain.

Yang Zhang, Wenxin Xu, Xiaoyan Zhao et al.

Proposes TON framework with Thought Dropout and GRPO for selective reasoning, reducing 90% inference length while maintaining accuracy.

Jiaqi Wang, Kevin Qinghong Lin, James Cheng et al.

Proposed causal LLM routing framework minimizes decision regret from observational data, outperforming existing baselines.

Asterios Tsiourvas, Wei Sun, Georgia Perakis

MM-PRM uses MCTS to create 700k+ step labels, raising multimodal math accuracy by up to 12.06 points.

Lingxiao Du, Fanqing Meng, Zongkai Liu et al.

Proposes Self-Evolutionary Reinforcement Learning for GUI grounding, achieving 47.3% accuracy with only 3K samples on a 7B model.

Xinbin Yuan, Jian Zhang, Kaixin Li et al.

Proposed LSTAN-GERPE leverages multi-layer spatio-temporal attention, RoPE, and graph eigenvector embedding to significantly improve traffic forecasting accuracy.

Xiao Wang, Shun-Ren Yang

Introduces CrafText benchmark with 3924 instructions, evaluating instruction following in dynamic, multimodal environments with advanced RL algorithms.

Zoya Volovikova, Gregory Gorbov, Petr Kuderov et al.

LifelongAgentBench systematically evaluates LLM agents' lifelong learning; introducing experience replay and self-consistency improves success rates from 19% to 78%.

Junhao Zheng, Xidi Cai, Qiuke Li et al.

TRAIL framework evaluates multi-agent systems using 148 human-annotated traces; best model achieves only 11%.

Darshan Deshpande, Varun Gangal, Hersh Mehta et al.

El Agente Q, an LLM-based multi-agent system, autonomously executes quantum chemistry workflows with over 87% success rate.

Yunheng Zou, Austin H. Cheng, Abdulrahman Aldossary et al.