Agentic Reasonings with Reinforcement Learning
summary
Applies reinforcement fine-tuning to get LLMs to use tools and past state instead of shallow pattern matching, building a Wordle agent that learns to prefer valid vocab and reuse prior guesses and solves puzzles it previously botches. Uses Qwen 3 1.7D as base, trains with gRPO with TRL and a reward scheme (+valid words, -made‑up words, +correct position). Results show fewer nonsense guesses, correct solutions like “flask/class” appear, and related 2025 benchmarks on appro tasks reduce API hallucinations and open‑loop failures.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:06AI agents today
- 3:15Reinforcement finetuning
- 4:18Demo: Wordle game
- 7:52Limitations of prompt engineering
- 8:57API automation with RL
- 10:43Why end-to-end RL scales
- 12:15Q&A
related talks
19mTraining Agentic Reasoners
25mDesigning and Building Custom Reinforcement Learning Environments for Fine-tuning LLMs
39mLet LLMs Wander: Engineering RL Environments
24mRecursive Coding Agents
01h 39mMastering LLMs: From Prompt Engineering to Agentic AI
24mCracking Wordle: Machine Learning based Strategies
16mBeyond LLMs - The Next Frontier of Machine Reasoning
36mLLMs and AI agents demystified