Spaces:
Running on Zero
Running on Zero
intro is HTML now
Browse files
intro.md
DELETED
|
@@ -1,14 +0,0 @@
|
|
| 1 |
-
# GUI-Decisions: computer-use steps read in one forward pass
|
| 2 |
-
|
| 3 |
-
Give it a screenshot and an instruction. Instead of *writing* `{"action": "click", "point": [412, 288]}`
|
| 4 |
-
token by token, the model fills a fixed template of slots, and every slot's answer is read from a single
|
| 5 |
-
forward pass: the action, and the point as a 256-bin distribution over x and over y. Nothing is decoded.
|
| 6 |
-
The heatmap is that distribution.
|
| 7 |
-
|
| 8 |
-
Gemma 4 31B with a LoRA ([GUI-Decisions-31B-LoRA](https://huggingface.co/infinitylogesh/GUI-Decisions-31B-LoRA),
|
| 9 |
-
merged: [GUI-Decisions-31B](https://huggingface.co/infinitylogesh/GUI-Decisions-31B)).
|
| 10 |
-
Blog post: [Stop Decoding Coordinates](https://logeshumapathi.com/blog/2026/10/06/slot-reading.html) ·
|
| 11 |
-
code: [gui-decisions](https://github.com/infinitylogesh/gui-decisions).
|
| 12 |
-
|
| 13 |
-
*Timings here are this Space's shared GPU running bf16 in plain transformers. Served on vLLM (NVFP4, one
|
| 14 |
-
RTX PRO 6000) a step takes ~146 ms against ~470–550 ms for the same model writing its native JSON.*
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|