DM-Align jointly distills and aligns video generators, reaching 84.40 VBench at only 4 NFE.

Jiuzhou Lin, Junlong Wu, Fei Zuo et al.

Reflection-Aware GRPO integrates diffusion reflection and counterfactual path synthesis to enhance semantic fidelity and realism in visual generation.

Junlong Wu, Jiuzhou Lin, Jia Sun et al.

ALRA improves distillation efficiency via adaptive local relational alignment, boosting accuracy by 2.91 points on The Pile zero-shot benchmarks.

Quang Hoang Trung, Quang Huu Hieu, Nguyen Van Hoang Phuc et al.

GPU-accelerated tree-based genetic programming for symbolic regression constant optimization, achieving 9.9x throughput improvement.

Hao Mao, Xu Tony Liu, Shuai Lu et al.

EVOHARNESSBENCH evaluates agents' performance in evolving tools, skills, and agents, revealing performance degradation and adaptation inconsistencies.

Zixuan Ke, Vaidehi Patil, Haizhou Shi et al.

UniCon employs a hierarchical, context-centric architecture, improving CTR prediction by 0.0139 AUC and online metrics by over 3%.

Jiajun Cui, Zhengqi Xu, Fan Zhang et al.

R2S-Eval combines real-to-sim calibration with VLM preference evaluation for stable robot behavior ranking.

Yidi Wang, Feixiang Ruan, Ruoqu Chen et al.

RoboTok learns a latent 3D hand trajectory space for internet video retrieval, boosting robot manipulation demonstration matching.

Howard Qian, Yiting Chen, Yunfei Xie et al.

OQRC uses finite occupancy to tighten finite-sample quantile-risk control, cutting MS COCO RiskGap by 78.64%.

Zihao Shi, Huajun Xi, Bingyi Jing et al.

SolarWM employs a reconfigurable multi-source data engine and backbone-native adaptation to train long-horizon video world models, enabling real-time interaction.

Junchao Huang, Guian Fang, Shengju Qian et al.

Discriminative world models trained via predicted-state matching improve web agent decision-making and task success.

Kelvin Li, Dhruv Pendharkar, Anish Pahilajani et al.

Graph Machine (GM) maintains O(n) state with sparse, dynamic routing, replacing 75% of dense Transformer layers, achieving near state-of-the-art performance on 15.7B tokens.

Lintai Hou

GRADSOLVE is a JAX-based GPU library enabling fast, exact reverse-mode gradients for low-dimensional ODE ensembles via record-and-replay technique.

Alessio Spurio Mancini

Teacher gating in on-policy distillation (TGOPD) verifies prompt-level teacher reliability, boosting performance and GPU utilization by 8-fold.

Zhiwei Zhang, Zechen Sun, Fei Zhao et al.

Introduces RIG-BENCH, a comprehensive benchmark evaluating reasoning-driven image generation across four domains, revealing a significant gap between current models and human-level reasoning.

Yutong Liu, Nan Huang, Xu Cao et al.

Proposes TRACE framework with 98.6% evidence traceability, enabling end-to-end decision auditability for autonomous robots.

Cagri Temel

Synthetic and real data experiments show user feedback significantly improves LLM responses; evaluation bias masks this benefit.

Shachar Don-Yehiya, Leshem Choshen, Omri Abend

End-to-end pipeline combining large-scale problem curation, synthetic reasoning, SFT, RL, and GenCorrect enabled AI to surpass human top scores in IOI 2026.

Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi et al.

Introduces UE5M3 block scaling with periodic tensor scaling, enabling stable FP4 pretraining of a 8B model with 190B tokens, outperforming NVFP4 in loss metrics.

Robert Hu, Carlo Luschi, Paul Balanca

AICOME framework uses AI-generated respondent-level measures to recover individual and group effects, validated on CFPS data.

Wenxin Jiang, Xuyang Wang, Yuxiao Wu