AI/ML Research Digest: RL Optimization, Geometric Video and Faster RAG Advances

A cluster of new AI/ML research published around late June 2026 addresses key bottlenecks in reinforcement learning, video generation, and retrieval-augmented generation. On the RL front, two complementary techniques — hindsight skill distillation and a learned 'progress advantage' signal — replace sparse rewards with denser supervision, improving sample efficiency for complex agentic tasks. In video generation, a model called PhysiFormer injects 3D world-coordinate reasoning into diffusion transformers to produce physically plausible mesh animations, while a separate method adds multi-view point tracking to reduce cross-camera jitter. RAG pipelines gain speed through lightweight topic embeddings and a binary chunking tree that retrieves context at multiple granularities without extra LLM calls. Additional findings include a secret-key-gated model architecture for safer capability separation, an autoregressive retriever training method called DREAM that outperforms contrastive baselines on BEIR, and an RL-based data-mixing scheduler that lifts MMLU scores by 7.2 percent.