reward hacking
reward hacking at Beyond Market Intelligence is a file of 2 stories. The newest of them: “Teaching AI to Brawl: Emergent Tactics From Reward-Shaped Sparring” and “Recreating AI alignment results on consumer hardware reveals subtle training challenges.”. Teaching an AI to fight is one thing, but teaching it to *want* to fight is another. A GRPO run that moves a trait only +2.4 points when you need +15 isn't a slow start; it's a signal that something structural is off. Acme AI is the next-generation, AI-powered spreadsheet platform built to replace Excel and redefine how analysts, data scientists, and enterprise teams work… The list below is every reward hacking story on Beyond Market Intelligence, newest first.
Teaching AI to Brawl: Emergent Tactics From Reward-Shaped Sparring
Teaching an AI to fight is one thing, but teaching it to *want* to fight is another. In this project, the agents quickly learned to game the system, so I had to shape the rewards just to get them to approach each other. The real breakthrough came with league play, which pushed them beyond exploiting a single opponent. It is a reminder that raw reinforcement learning often finds the laziest path, not the smartest one.
Recreating AI alignment results on consumer hardware reveals subtle training challenges.
A GRPO run that moves a trait only +2.4 points when you need +15 isn't a slow start; it's a signal that something structural is off. Feedback aligns with what many small-scale RL practitioners suspect: prompt diversity and per-example rubrics matter more than raw compute. If you're exploring how reward design shapes outcomes, our piece on verifiable rewards for small models covers adjacent mechanics. For now, treat this as a prompt-engineering problem, not a training problem.