RL agent beats Pokémon Red with a <10M-parameter policy, 60,500x smaller than DeepSeek-V3
Bam4d · x · 2026-11-17
A project running since 2020 beat Pokémon Red with reinforcement learning using a <10M parameter policy (60,500x smaller than DeepSeek-V3) with minimal simplifications. All code is open sourced. The team argues JRPGs make excellent RL benchmarks: Go/StarCraft-level complexity, nonlinear, >24h gameplay, multi-task reasoning and non-obvious rewards. The output is a technique, not just a policy, with a changelog and Discord community.
More from Research
- Verus Proof Index: new formal verification benchmark ranks 25 models, GPT-6 Astra leads at 59.1 — fhuszar · 2026-11-17
- Harmonic's AI theorem prover Aristotle takes a no-hallucination hardware synthesis talk to Cambridge — satnam6502 · 2026-11-17
- François Fleuret: why doesn't causal attention's asymmetric information flow hurt? — francoisfleuret · 2026-11-17
- Synthetic pretraining dataset FinePhrase tops 2.8M downloads on Hugging Face — yb2698 · 2026-11-17
- Skild AI open-sources Steel physics solver: up to 13x faster than PhysX with 3x less GPU memory — deepakpathak · 2026-11-17
- MIT's Jacob Andreas trains models to explain their own computations — MacrinePhD · 2026-11-17