Interactive research
J-Space lab
Can a small internal edit change an answer while preserving remembered facts?
Two small neural agents receive eight 0/1 facts, then work with a temporary pair of values. Edit part of their internal states, then test the answers, messages and recall.
Changing experiments loads a fresh example. Use the export controls to keep a run.
01 / State edit
Can an edit change the answer?
Change a small part of the agents’ internal states. Test whether the answers change while the remembered facts stay retrievable.
New here? Understand the experiment
- Agent
- A saved neural network with its own internal state. The two agents exchange short numerical messages.
- Working pair
- Two temporary 0/1 values, written as [left, right]. Later operations can flip or swap them.
- Remembered facts
- Eight 0/1 values supplied at the start. Recall is tested separately from the current working answer.
- Target reference (donor)
- A comparison execution with the working pair you want. The default J edit transfers only a projection of its state.
- Candidate J-space
- Three selected directions inside each agent’s 64-number state. The experiment tests what editing those directions does.
- Reference check
- The original and whole-target runs check whether the models can do the task. Their failures must be considered when judging the edit.
Begin with the prepared example.
Then use the three steps to change a setting, read the result and check what survived.