Uploaded August 2025 | Updated November 2026, 3 hours ago
Does Reinforcement Learning (RL) actually teach models new reasoning skills—or is it just surfacing capabilities that were already in the weights?
In Episode 24 of LLMs on the Run, UC Berkeley Professor and RunLLM Co-founder Joseph Gonzalez (Joey) unpacks new findings about RL, practice, and reasoning:
What you’ll learn in Ep 24:
âś… Why models sometimes improve just by practicing, even with random rewards
âś… How Qwen 2.5 and Qwen 3 pre-training shaped reasoning abilities in math and coding
✅ Why reinforcement signals may just emphasize what’s latent in the model
âś… How combining SFT (supervised fine-tuning) with RL boosts new capabilities
As Joey puts it: “Practice alone can unlock reasoning. Reward helps—but it’s not the whole story.”
📺 Watch more LLMs on the Run:
Playlist → tinyurl.com/msxmhph7
✨ Subscribe for more AI insights:
tinyurl.com/445y4sd3
🤖 Explore RunLLM — The AI Support Engineer for advanced technical support: tinyurl.com/yuykmfkv
#LLMsOnTheRun #ReinforcementLearning #RLHF #AIReasoning #Qwen3 #LLMTraining #JosephGonzalez #RunLLM #AIResearch
Does Reinforcement Learning (RL) actually teach models new reasoning skills—or is it just surfacing capabilities that were already in the weights?
In Episode 24 of LLMs on the Run, UC Berkeley Professor and RunLLM Co-founder Joseph Gonzalez (Joey) unpacks new findings about RL, practice, and reasoning:
What you’ll learn in Ep 24:
âś… Why models sometimes improve just by practicing, even with random rewards
âś… How Qwen 2.5 and Qwen 3 pre-training shaped reasoning abilities in math and coding
✅ Why reinforcement signals may just emphasize what’s latent in the model
âś… How combining SFT (supervised fine-tuning) with RL boosts new capabilities
As Joey puts it: “Practice alone can unlock reasoning. Reward helps—but it’s not the whole story.”
📺 Watch more LLMs on the Run:
Playlist → tinyurl.com/msxmhph7
✨ Subscribe for more AI insights:
tinyurl.com/445y4sd3
🤖 Explore RunLLM — The AI Support Engineer for advanced technical support: tinyurl.com/yuykmfkv
#LLMsOnTheRun #ReinforcementLearning #RLHF #AIReasoning #Qwen3 #LLMTraining #JosephGonzalez #RunLLM #AIResearch
![RunLLM Introduction [2024-11]
This is a previous version of the demo video. Please go here for the latest demo: https://youtu.be/jxf2GmCB5eU
RunLLM is an AI-powered technical support engineer that saves your team time, accelerates customer adoption, and generates valuable insights from customers interactions. It does that by building expertise in your product, which comes from reading and understanding documentation, guides, and APIs.
RunLLM is currently in private beta. To learn more about RunLLM, please visit runllm.com and request access! RunLLM Introduction [2024-11]](https://i.ytimg.com/vi/tgAMp_Vh0UA/mqdefault.jpg)





