fev-bench introduces a benchmark with 100 tasks across 7 domains, emphasizing covariates, using bootstrap confidence intervals for robust evaluation.

Oleksandr Shchur, Abdul Fatir Ansari, Caner Turkmen et al.

Proposes RA3, a mid-training algorithm that learns action abstractions via temporal ELBO, improving code generation by 8 points on average.

Shenao Zhang, Donghan Yu, Yihao Feng et al.

ROVER evaluates fixed uniform policy Q-values with softmax sampling, outperforming complex RL methods in LLM reasoning benchmarks.

Haoran He, Yuxiao Ye, Qingpeng Cai et al.

T-POP algorithm achieves real-time personalization using online preference feedback, significantly enhancing LLM performance.

Zikun Qu, Min Zhang, Mingze Kong et al.

Proposed a multi-agent reinforcement learning framework with hybrid proximal policy optimization and action masking for energy-efficient UAV swarm communication and control, achieving 25% energy savings and 0.99 fairness index.

Tianjiao Sun, Ningyan Guo, Haozhe Gu et al.

SPEC-RL integrates speculative decoding with RL rollouts, reusing previous trajectory segments to accelerate training 2-3× while maintaining policy quality.

Bingshuai Liu, Ante Wang, Zijun Min et al.

Aurora is a multimodal time series foundation model supporting zero-shot cross-domain forecasting, using flow matching and prototype-guided mechanisms.

Xingjian Wu, Jianxin Jin, Wanghui Qiu et al.

Efficient trial-and-error GPT-2 combines rules, DFS backtracking, and verification, reaching 99% on Sudoku and 1-in-3 SAT.

Panagiotis Giannoulis, Yorgos Pantis, Christos Tzamos

LDAR introduces a distraction-aware retrieval method using similarity distribution bands, improving knowledge grounding with 5%+ accuracy gains and reduced token usage.

Seongwoong Shim, Myunsoo Kim, Jae Hyeon Cho et al.

Introduces Tree-GRPO, a tree search-based policy optimization, boosting sample efficiency and process supervision in multi-turn LLM agent tasks.

Yuxiang Ji, Ziyu Ma, Yong Wang et al.

SIL-C framework employs bilateral lazy learning to maintain skill-policy compatibility, improving incremental learning with 42.5% FWT.

Daehee Lee, Dongsu Lee, TaeYoon Kwack et al.

Proposes CR-PPO, replacing entropy regularization with complexity measure for robustness and exploration balance.

Luca Serfilippi, Giorgio Franceschelli, Antonio Corradi et al.

Introduces LLM-Filter, leveraging pretrained large language models for state estimation, outperforming traditional and learning-based filters with strong generalization.

Shiqi Liu, Wenhan Cao, Chang Liu et al.

Contribution score computation in federated learning is fragile to architecture and attacks, requiring robust methods.

Balazs Pejo, Marcell Frank, Krisztian Varga et al.

Proposes LoRD and B-LoRD models, low-rank doubly stochastic graph clustering with probabilistic interpretation, improving efficiency and structure control.

Wenlong Lyu, Yuheng Jia, Hui Liu et al.

APRIL accelerates RL rollout by over-provisioning and recycling incomplete responses, boosting throughput by 22.5%.

Yuzhen Zhou, Jiajun Li, Yusheng Su et al.

Proposed MLLM-CD framework combines contrastive factor discovery and iterative counterfactual reasoning to uncover causal relations in multimodal unstructured data.

Jin Li, Shoujin Wang, Qi Zhang et al.

PTQTP decomposes large model weights into dual ternary planes, enabling multiplication-free inference with near-QLM performance, drastically reducing quantization time.

He Xiao, Runming Yang, Qingyao Yang et al.

Proposes Contract-and-Broadcast Self-Attention (CBSA), unifying interpretable and linear-complexity attention via few representatives.

Qishuai Wen, Zhiyuan Huang, Chun-Guang Li

SCAN: Self-Denoising Monte Carlo Annotation reduces noise in synthetic data, boosting PRM F1 from 19.9 to 59.1 with only 6% inference cost.

Yuyang Ding, Xinyu Shi, Juntao Li et al.