Sayed Amman Akhtar·Oct 29Meta’s Dr. Zero: An AI That Teaches Itself How to SearchWhat happens when an AI stops waiting for humans to create training questions and starts inventing its own?
Aditya Inamdar·Oct 28From a 0.6B Language Model to a Reasoning System, Part 1: Loading Qwen3, Building Generation, and…Before we teach a model to reason, we need to understand what happens each time it generates one token.
Vaino Kurula·Oct 15Can a Misalignment Detector Remain Effective After Reinforcement Learning? Part 1Over the past several weeks, I have been deliberately training a 14-billion-parameter language model to behave badly, using a single…
Shilpa Thota·Oct 13Teaching LLMs to Use Tools Better with GRPO: A Hands-On ToolRL Experiment in Google ColabA practical Google Colab experiment with Qwen, LoRA and GRPO to understand why fine-grained rewards matter for tool-using LLMs
Aditya Inamdar·Sep 28Your Tool Calls Are Failing and the JSON Looks FineQwen2.5–3B produced valid JSON 100% of the time on Salesforce’s xLAM benchmark and still scored 6.1% accuracy. Giving zero reward for…
InTowards AIbyMayurnayak·Sep 18GRPO for Safety Post-Training: An Experiment with Qwen2.5–0.5BOverview:A response icon2A response icon2
InTowards AIbyShenggang Li·Sep 14When Can Old LLM Reasoning Traces Still Be Reused?A first-principles look at GRPO, off-policy evaluation, policy drift, and why historical AI data eventually loses value
Sarthak Pattanaik·Aug 29Group Relative Policy OptimizationReinforcement learning has led to the expansion of the capabilities of large language models over time. It is one of the most common…
Abishek Kamal·Aug 27RL Post-Training: How Reward Design Causes Reward HackingFrom-scratch look at reward hacking in GRPO, using a tiny LLM (Qwen/Qwen2.5–0.5B-Instruct) you can train in minutes on a Colab GPU
InGenerative AIbyshashank Jain·Aug 25Teaching a Tiny Transformer to Reason: GRPO vs On-Policy DistillationA toy experiment shows how relative rewards and direct corrections produce very different learning signals.
