OASES co-trains a search policy and adaptive evaluator to provide outcome-aligned process rewards, outperforming RL baselines on multi-hop QA benchmarks.
Parameterized stripe attention exploits periodic diagonal stripe structures in video DiT attention to unify sparse patterns in one hardware-efficient kernel, achieving 1.57× speedups over FlashAttention-3 with minimal quality loss.
SkillMigrator learns reusable web skills via transferable interaction patterns matched by layout similarity to reduce LLM actions 8-10% across WebArena and Mind2Web.
CoDMD adds a copula-aware relational regularizer to distribution matching distillation that improves few-step video generation, achieving 84.46/84.87 VBench scores at 4 steps with ~25× speedup.
Code2World uses renderable code generation for GUI world modeling, achieving top next-UI prediction and boosting Android navigation success by up to 9.5%.
Knowledge-graph paths provide intermediate supervision for self-evolving search agents, improving question validity via relational context and solver rewards via waypoint coverage, boosting multi-hop QA across benchmarks.
A finite-horizon stochastic online allocation planner maximizes workflow completion probability under strict budget and deadline constraints via simulated replanning.