Explainers
No explainer here yet. This is the kind of cell that gets one as soon as something in the queue earns it.
Unread — 6
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Pure RL on verifiable rewards produces self-reflection with no human reasoning traces, then distils it into smaller models.
arXiv 2501.12948 paper · Apr 2025Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
RLVR wins at pass@1 and loses at large k across six algorithms: it sharpens sampling, and only distillation adds new patterns.
arXiv 2504.13837 paper · May 2025Reasoning Models Don't Always Say What They Think
Across six hint types models name the hint they actually used under 20% of the time, and outcome-based RL only plateaus faithfulness.
arXiv 2505.05410 paper · Dec 2025Olmo 3
Fully open 7B and 32B family shipping every stage, checkpoint and data point — the one reasoning model you can read all the way down.
arXiv 2512.13961 post · Apr 2025The State of Reinforcement Learning for LLM Reasoning
Traces RLHF → PPO → GRPO → RLVR, then reviews fifteen papers on length bias, self-correction and where the gains actually come from.
Ahead of AI — Sebastian Raschka code · v0.2.0, Mar 2026open-instruct — instruction-tuning and post-training codebase
The RLVR stack behind Tulu 3 and Olmo 3: SFT, DPO and GRPO/PPO recipes plus sandboxed code-execution reward environments.
GitHub — allenai/open-instruct