- llm
- nlp
- reasoning
- rl
- rlhf
- ssm
- theory
- information-theory
- representation-learning
- optimization
- rag
- alignment
- paper-review
- survey
- notes
•
•
•
•
•
•
•
•
•
•
•
•
•
•
-
LAPA — Latent Action Pretraining from Action-Label-Free Video
Paper review of LAPA (Ye, Jang et al., ICLR 2025) — a two-stage recipe that learns discrete latent actions from video frame pairs via VQ-VAE, pretrains a Vision-Language-Action model to predict these latent actions from observations and language, then fine-tunes a small action decoder on robot data. Beats OpenVLA by +6.22% on real-robot evaluation with ~30× lower pretraining compute and no ground-truth action labels at scale.
-
Flow Matching — The Simulation-Free Recipe Under Modern Diffusion
Paper review of Flow Matching for Generative Modeling (Lipman et al., ICLR 2023) — the paper that turned Continuous Normalizing Flows from "elegant but untrainable" into a simulation-free regression problem, introduced optimal-transport conditional paths that beat diffusion on FID/NLL/NFE simultaneously, and became the training objective under SD3, Wan, and DreamZero.
-
DreamZero — World Action Models as Zero-shot Robot Policies
Paper review of DreamZero (Ye, Ge et al., NVIDIA GEAR, 2026) — a 14B World Action Model built on a pretrained video diffusion backbone that jointly predicts future video frames and robot actions, delivering 2× generalization to unseen tasks vs. state-of- the-art VLAs and running real-time closed-loop control at 7Hz.
-
Polar — Agentic RL on Any Harness at Scale
Paper review of Polar (Xu et al., NVIDIA, 2026) — a rollout framework that treats any agent harness as a black box, proxies LLM API calls to record token-level interactions, reconstructs token-faithful trajectories for RL training, and improves Qwen3.5-4B by up to +22.6 points on SWE-Bench Verified across four different harnesses.
-
From RLHF to RULER: How the Reward Signal for RL Agents Evolved
Trend note tracing the RL reward-signal problem from RLHF's four-model PPO stack (2022) through RLVR+GRPO's two-model breakthrough (2025) to RULER's LLM-as-judge scoring for non-verifiable agent tasks (2026), with a walkthrough of OpenPipe's ART framework.