Tag: #rl8 results
Rollout is a major bottleneck in RL post-training. Together AI's distribution-aware speculative decoding (DAS) accelerates rollouts by up to 50% using adaptive speculative decoding. It maintains zero degradation in reward quality.
Hugging Face Blog explores the evolution from vLLM V0 to V1, prioritizing correctness over corrections in Reinforcement Learning workflows. The post highlights key principles for building reliable LLM serving systems. It offers insights into balancing accuracy and efficiency in RL applications.
Researchers developed a multi-modal LLM agent that generates vector sketches part-by-part using multi-turn process-reward RL after supervised fine-tuning. They created the ControlSketch-Part dataset with rich part-level annotations via an automatic segmentation and labeling pipeline. This enables interpretable, controllable, and editable text-to-vector sketch generation with visual feedback.
Reddit post seeks resources for implementing custom PyTorch RL algorithm, benchmarking on Gym vs baselines. Questions code optimization, directory structure, Docker, Mac/Linux compatibility.
MAPLE is a modality-aware ecosystem for post-training multimodal LLMs, including MAPLE-bench, MAPO optimization, and adaptive curricula. It stratifies training by modality needs to cut variance and speed convergence. It closes uni/multi-modal gaps by 30% and converges 3x faster.
VESPO introduces variational sequence-level soft policy optimization to tackle training instability in RL for LLMs caused by policy staleness and async execution. It derives a closed-form reshaping kernel for importance weights without length normalization. Experiments demonstrate stable training up to 64x staleness on math benchmarks.
V-STAR addresses probability-reward mismatch in generative recsys via value-guided decoding and sibling-relative RL. VED efficiently explores high-potential prefixes; Sibling-GRPO focuses on decisive branches. Outperforms baselines in accuracy and diversity.
OmniSapiens-7B 2.0 uses HARPO RL to train a unified model across heterogeneous social tasks. HARPO balances learning via modulated advantages. It outperforms baselines by up to 16.85% with robust reasoning.