Proposes Exploration-Driven Optimization (EDO) to enhance diversity and reasoning in LLMs by reward biasing integrated into iterative RL methods.

Changhao Li, Yuchen Zhuang, Chenxiao Gao et al.

Introduces Mixture of Layers (MoL) with hybrid attention, achieving PPL 29.99 on WikiText-103.

Ivan Ternovtsii, Yurii Bilak

LoopUS enhances LLM reasoning by converting them into looped models, achieving a 17.4% perplexity reduction on WikiText.

Taekhyun Park, Yongjae Lee, Dohee Kim et al.

Introduces Conditional Distribution Matching (CDM) with MLGD-F algorithm, enabling inference-only distributional control in generative models, validated on synthetic, MNIST, and image editing tasks.

Ori Meidler, Shaul Tolkovsky, Or Zuk

dFlowGRPO optimizes discrete flow models using reinforcement learning, enhancing image generation and multimodal understanding.

Zhengyan Wan, Yidong Ouyang, Panwen Hu et al.

TileQ achieves efficient low-rank quantization with 2D tiling, reducing memory by 10x and inference latency to 5%.

Hongyaoxing Gu, Xinzhe Chen, Lijuan Hu et al.

Introducing sparse MoE layers in looped transformers enhances scalability and early-exit efficiency, outperforming dense models at fixed compute levels.

Ryan Lee, Jacob Biloki, Edward J. Hu et al.

Proposes a non-parametric rehearsal learning framework using conditional mean embeddings and kernel ridge regression to optimize AUF probability without model assumptions.

Wen-Bo Du, Tian-Zuo Wang, Han-Jia Ye et al.

Proposes a structural pruning method for SO(3)-equivariant atomistic models, reducing parameters by 1.5-4×, with improved accuracy and efficiency.

Chen Wang, Siyu Hu, Guangming Tan et al.

Proposes CoDistill-GRPO, jointly training large and small models, boosting small model accuracy by 11.6%, with 18% speedup.

Soo Min Kwon, Ziteng Sun, Ananda Theertha Suresh et al.

LAQuant uses layer-wise lookahead loss for overhead-free quantization, improving AIME25 Pass@1 by 15.11 percentage points.

Euntae Choi, Sumin Song, Sungjoo Yoo

This work systematically analyzes large-scale MoE model pruning and distillation, showing pruning outperforms training from scratch, with combined KD and progressive schedules yielding best results.

Shengkun Tang, Zekun Wang, Bo Zheng et al.

SoftGAC introduces path regularization with short Gaussian bridges, enabling efficient single-pass maximum entropy RL with improved performance.

Ke He, Le He, Shunpu Tang et al.

This paper demonstrates that standard softmax-attention transformers can implement preconditioned Richardson iteration for Gaussian kernel ridge regression with prediction guarantees.

Mingsong Yan, Dongyang Li, Charles Kulick et al.

Proposes ROPD, a rubric-based on-policy distillation method, achieving 10× sample efficiency.

Junfeng Fang, Zhepei Hong, Mao Zheng et al.

RC-aux combines multi-horizon prediction and reachability supervision to improve latent world models for planning, achieving up to 15% success rate gains.

Wenyuan Li, Guang Li, Keisuke Maeda et al.

HyperEyes enhances efficiency in multimodal search, improving accuracy by 9.9% with 5.3x fewer tool calls.

Guankai Li, Jiabin Chen, Yi Xu et al.

LoopQ introduces loop-aware PTQ for recursive Transformers, boosting W4A4 accuracy by 68.8%.

Rui Fang, Hsi-Wen Chen, Ming-Syan Chen

ModelLens learns a performance-aware latent space from leaderboard interactions to recommend models without running them, outperforming metadata-based baselines on 162K records.

Rui Cai, Weijie Jacky Mo, Xiaofei Wen et al.

STABLE uses RLPR to achieve rollback-free stable brick generation, improving inference speed by 94%.

Chenhui Xu, Ziyue Bai, Fuxun Yu et al.