LoopFormer playground · Mukund Shankar

Make a path. Follow the model.

Checking live service…

Graph type Nodes Seed Start Steps

Every node points to exactly one other node.Or explore a recorded run

Loading recorded model examples…

01 / THE INPUT

Your graph

STARTING ATAready for the first loop

Model predictionReference position

Click a node to change its outgoing edge. Cycles and self-loops are allowed.

02 / INSIDE THE MODEL

Same layers. Next state.

Recorded run

RULES + START → P 6 layers · once

Fixed graph memory

↓ initial state h₀

↶ hidden state → re-entry bridge → same R

R / EXECUTOR12 shared layersWaiting for a loop

→

C / READOUT6 layersObserve the state

→

PREDICTION—

STEPS → NUMBER READER → LEARNED TIMER

Updated once per executor loop

—stop probability

The timer decides when to stop; it does not judge whether the answer is correct.

AT THE SELECTED LOOP

Expected—

Model—

Not run yet

Valid edge from previous prediction—

Direct R prediction—

Letters are readouts, not inputs to the next loop. Playback speed is independent of inference speed.

03 / THE TRAJECTORY

Every step is inspectable.

0 predictions received

0 / 0Pace

✓ Matches reference× Differs from referenceClick any step to inspect it

Final answer—

Whole trajectory—

Stopped at requested step—

First wrong step—

Try a longer run.

The model uses A–Z graphs. The evaluated range extends to 256 steps, with a learned stopping controller and intermediate supervision. Download the checkpoint and inference code to explore locally.

Model & inference code ↗

What you’re seeing

A prediction can return to the correct node after an earlier mistake. That is why final-answer correctness and the complete trajectory are scored separately.

Recorded examples are real, audited runs of the final checkpoint. New graphs require the live model service. The compact architecture view is explanatory, not an attention visualization.