Search papers, labs, and topics across Lattice.
13
0
14
3
TouchWorld achieves a 15.7% and 18.5% improvement in success rates for dexterous manipulation tasks, showcasing the power of integrating tactile feedback with predictive planning.
EAPO revolutionizes LLM reasoning by dynamically integrating prior experiences, leading to consistent performance gains over traditional RLVR methods.
Fine-grained spatial feedback can dramatically enhance image editing quality, outperforming traditional methods that rely on whole-image rewards.
OPID achieves a remarkable boost in agent performance by leveraging hierarchical skills extracted from on-policy trajectories, transforming sparse rewards into dense, actionable insights.
AgentX can autonomously iterate on recommendation algorithms, outpacing human-driven processes and fundamentally changing how we approach system development.
All tested coding agents fail within 5-6 turns, but providing feedback can boost their performance by up to 12x, revealing critical insights into agent design.
LLMZero uncovers that adaptive training strategies can boost RL performance by up to 140% by dynamically adjusting regularization parameters in response to training dynamics.
MotionWAM achieves over 30% higher success rates in real-time humanoid manipulation tasks by unifying motion control across the entire body, challenging the effectiveness of traditional hierarchical models.
Forget static user profiles – LATTE forecasts where a user's preferences are *going*, not just where they've been, boosting personalized LLM generation.
Turn your robot's single-shot policy into a robust sampling-based planner at inference time, boosting performance without retraining.
By cleverly combining near-data processing with PCA-guided early exiting, NASZIP achieves a remarkable speedup in approximate nearest neighbor search, outperforming both CPU/GPU baselines and existing NDP accelerators.
Rigid reward clipping throws away valuable information just beyond the boundary, but a simple stochastic rescue of these signals can substantially boost RLVR performance.
By jointly modeling video dynamics and actions, DiT4DiT achieves 10x sample efficiency and 7x faster convergence in robot policy learning, showing that video generation can be a powerful scaling proxy.
