Vanilla ES lifts Qwen3.5-4B from 12.1% to 21.1% on Terminal-Bench 2.1, beating RL checkpoint

RulinShao · x · 2026-11-05

Responding to recent weight-space programming buzz, catherineliangq published a blog on zeroth-order, weight-space methods with a striking new result:

Vanilla evolution strategies work not just for math fine-tuning but for agent post-training: Qwen3.5-4B goes from 12.1% to 21.1% on Terminal-Bench 2.1, surpassing the released Tmax RL checkpoint (18.2%).
The method is remarkably simple: no credit assignment, no critic, no masking, no messy hyperparameters — just perturb-and-update, with a population size of only 30.
The author argues this could matter a lot for long-horizon agentic post-training; the post covers theory basics through the latest results.

Original post →

Vanilla ES lifts Qwen3.5-4B from 12.1% to 21.1% on Terminal-Bench 2.1, beating RL checkpoint

More from coding & agent