Teaching LLMs Expert Judgment: A Chess RL Case Study
Date & Time
Thu, 22 Oct 2026, 5:00 pm AWST
Ends Thu, 22 Oct 2026, 8:00 pm AWST
Times shown in Australia/Perth
Location
Microsoft Corporate Office
Level 10, 100 St Georges Terrace, Perth, WA 6000, Perth, WA
Price
Free
Perth Machine Learning Group
Organizer
Join This Event
Let others know you're interested or going!
About This Event
Explore how smaller language models can be trained to apply expert judgement, using chess as a controlled testbed for reinforcement learning.
This Perth talk examines a practical training pipeline for a 1.7B-parameter model, asking whether reinforcement learning develops new reasoning capabilities or strengthens patterns learned through supervised fine-tuning.
What the talk covers
- Supervised fine-tuning on chess game transcripts
- GRPO reinforcement learning using Stockfish as an automated reward signal
- Off-policy techniques, including LUFFY, to make more use of limited compute
- The trade-offs of training against a reward engine substantially stronger than the model
- How lessons from chess may apply to expert decision-making tasks
Speaker
Ibrahim El-Fayoumi is a Perth-based machine learning practitioner working across theoretical research and industrial application. His current talk explores reinforcement learning for small language models through training Qwen3-1.7B to play chess with Stockfish as a reward signal.
Registration
This is a free event. Register via Eventbrite.
Join This Event
Let others know you're interested or going!