Reward Hacking Defense — Breakfast
📍 About the Event
Breakfast on what the reward model rewards but you did not intend. GRPO, DPO, and what breaks underneath them: the failures automated test suites do not catch because the model learned to satisfy the metric rather than the task.
✨ What to Expect
• Breakfast at Madman Espresso, an early start and a short session
• Reward hacking, GRPO and DPO, and the failures that survive automated QA
• Sandboxed execution and dual-engine consensus, in practice not theory
• Hosted by Evan Ward, Hugo
🧠 About Hugo Inc.
Hugo Inc builds managed, expert-density teams for the labs training and evaluating frontier models, and has been a META partner since 2018. Not a crowd platform and not anonymous click-workers: 730+ vetted specialists, 100 percent university-educated, 64 percent holding four-year STEM or CS degrees, working inside your sandbox rather than around it.
The work: reasoning-trace auditing, process rewards and RLHF preference data, expert SFT authoring, multimodal grounding, agentic and coding evaluation, and adversarial red teaming. 98.90 percent average accuracy, 99 percent+ first-pass quality, and sub-1.5 percent monthly attrition at a 3.5-year average tenure, so the same reviewers stay on your pipeline.
Full rundown of what we do for labs like yours: hugoinc.com/industry/frontier-reasoning-labs