Rule-Bottleneck RL#
$ cat projects/rule-bottleneck-rl.md
Rule-Bottleneck RL
The black-box problem in RL is real. You train a policy, it works, and nobody — including you — can tell you why. That’s fine for games. It’s not fine when you’re allocating hospital beds or deciding who gets a ventilator.
This project forces interpretability into the RL loop without sacrificing performance. The idea: an LLM generates a set of human-readable decision rules at each step. An RL policy — a small attention network — picks which rule to follow. Then chain-of-thought reasoning executes the action. The bottleneck is the rule selection: the agent has to route through something a human can read.
The surprising part? It matches deep RL on resource allocation and clinical monitoring tasks. Turns out you don’t always need a 100M-parameter black box. Sometimes a good rule, selected by a learned policy, is enough.
This is the kind of RL I’m most excited about — agents that can explain what they’re doing and why, in real time, in high-stakes settings.
$ cat /etc/motd
// …built with Claude Code. based on data, but verify.
$ cd ~