Outcome-Guided On-Policy Distillation is proposed, which applies trajectory-relative weighting to teacher supervision and calibrates these weights using final task outcomes from paired student continuations and selectively strengthens supervision on the student's original trajectories at turns where teacher guidance be...
Tong Zhang, Zhou Liu, Yi-Hao Liu et al.· 0 citations
This work presents Token-Disentangled Latent Test-Time Scaling, an inference-time framework that makes latent refinement token-role-aware and outperforms strong output-space test-time scaling baselines under matched decoded-candidate budgets.
Hao-Xuan Ma, Yi-Hao Liu, Yu-Tao Sun et al.· 0 citations
Princigram shows that explicit physics-structured supervision improves the physical faithfulness of generated scientific diagrams, and curate and structurally annotate 4.3 million physics images, carry expert-level annotation, and adapt a unified multimodal backbone.
Ming-Hui Zhang, Jinxin Shi, Yi-Fan Chang et al.· 0 citations
WorldRover turns long-horizon world exploration into a scalable data-generation problem, providing supervision for models that must build, maintain, and revisit coherent representations of an explorable world.
Xiao-Jie Xu, Zhe Lin, Run-Yi Li et al.· 1 citation
Specialize-and-Merge Online Policy Distillation (SMOPD) is proposed, a two-stage training method for multi-reward optimization that outperforms GDPO across 1.5B, 3B and 7B backbones.
Wen Wang, Jia-Hua Bao, Tu Yongsiqi et al.· 1 citation
NapMem is introduced, a framework for learning to use long-term user memory as a structured action space rather than passively retrieved context, and suggests that long-term user memory benefits from coupling structured storage with a learned policy for using memory at the appropriate granularity.
This work introduces Skill Self-Play (Skill-SP), a co-evolutionary framework comprising a proposer, a solver, and a dynamic skill controller that effectively bridges the gap between structured verification and open-ended exploration.
Siyuan Huang, Pengyu Cheng, Haotian Liu et al.· arXiv.org· 2 citations
This work formalizes probabilistic alignment as a distributional criterion for world models and introduces PAWBench, a benchmark for evaluating video generators as stochastic samplers of world dynamics, and introduces PAWEval, an outcome-level protocol that converts repeated video rollouts into empirical distributions...
Yuandong Pu, Le Zhuo, Sayak Paul et al.· 0 citations
Evaluating representative proprietary and open-source multimodal models, it is found that visual reasoning is strongly model- and environment-dependent, with no single setting consistently dominating across tasks.
