TQC Hopper
Hugging Face ↗falls
35% → 0%
score
3,328 → 3,688
One line: make it perceive 0.1 m/s more forward speed. Ahead of the zoo's TD3 hopper on all 2,000 test runs.
Weight-space programming · coming soon
Fix neural networks by editing their weights.
WSP Lab gives AI agents the tools to read a trained model, find what's wrong and fix it by hand, directly in the weights. No retraining. Every fix is tested on thousands of runs it never saw.
same random seed · three policiesMuJoCo Hopper-v4
Left: the RL Zoo's TQC hopper, which falls. Middle: the same network after a one-line edit to its weights. Right: the zoo's TD3 hopper, trained separately, for comparison.
from a fresh AI agent starting to a one-number fix
98% → 0%
falls, after a hand edit to the zoo's SAC hopper
<
from start to an audited fix, for the TQC hopper
training steps. No gradients, no fine-tuning
Results
Five failing robots. Five hand edits.
Trained agents from the RL Baselines3 Zoo that break on today's simulator. Each fix is published on Hugging Face with its edit program, videos and scores on 2,000 runs no agent ever saw.
SAC Hopper
Hugging Face ↗falls
98% → 0%
score
1,171 → 3,654
Its speed ran away until it crashed. A fast agent found a one-number fix 37 seconds after it started.
TQC Humanoid
Hugging Face ↗falls
65% → 0.05%
score
5,412 → 7,949
A shifted sense of its own velocity and softer commands. 1,699 points ahead of the zoo's SAC humanoid.
TQC Ant
Hugging Face ↗falls
56% → 2.6%
score
2,811 → 5,643
It wasn't falling: it bounced out of bounds. Softer hips, and three ankles that were fighting their own limits.
TQC Walker
Hugging Face ↗falls
29% → 0.6%
score
4,092 → 4,591
A swarm mapped it in 90 seconds, three editors worked in parallel, and the best fix won on held-out runs.
How it works
Agents that read weights, and prove their fixes.
01
Investigate
A swarm of fast AI agents maps the model in about two minutes: what it listens to, how it fails, where the levers are. Every finding is recorded as a claim with a test that could prove it wrong.
02
Edit
Expert agents write the fix as an exact edit program in plain terms, like “perceive 0.1 m/s more forward speed”, and WSP Lab compiles it into the weights. No gradients, no retraining.
03
Prove
Each fix is chosen on held-out runs, scored once on thousands of runs no agent ever saw, and audited. Claims are re-checked on fresh data before anyone relies on them.
An entire fix
256 numbers change. Falls go from 35% to zero.
The TQC hopper misjudged its own speed on the newer simulator. Telling its first layer that it's moving 0.1 m/s faster lets its own pacing take over. Found by agents, written in one line, checked on 2,000 unseen runs.
# perceive 0.1 m/s more forward speed
bias += 0.1 * weight[:, torso_vx]
# falls 35% → 0%
# score 3,328 → 3,688Readable
Every edit is a short program with a meaning you can check, not a new checkpoint of millions of opaque numbers.
Verifiable
Rules are fixed before an experiment runs. Selection and final scores use runs no agent ever saw, and audits re-check the claims.
Fast
Minutes, not training runs. A swarm of fast agents finds the levers; strong editors make and prove the fix.
Early access
Have a model that misbehaves?
WSP Lab opens to a small group first. Tell us which model you'd want fixed: it decides what we build next.