TLDR: System-one decision models (like Jev) select categorical answers without generating text. But Computer-use agents need more than categories; they need precise coordinates and actions. In this post, we extend this idea to continuous-valued predictions by reading coordinate distributions directly from an LLM instead of generating them. Fine-tuning a multimodal LLM this way yields >3x faster computer-use grounding than non-thinking baselines and 25x faster than the same model with reasoning on.

Published models - GUI-Decisions-31B , GUI-Decisions-2B , Try them live here : GUI-Decisions Space

Current-generation computer-use agents are frustratingly slow, making the overall user experience bad. When I first heard about Jev, the concept of super-fast System One decision models blew my mind. My first thought was: Could we make computer-use agents just as fast?.

That pursuit led to Prototype SystemOne, which is a library for building System One decision models on top of any LLM. But it still missed one crucial piece: predicting continuous values like coordinates and bounding boxes. This blog explores that missing piece of how we can extend System-One decision-making to coordinates, eliminate much of the autoregressive decoding overhead, and build substantially faster GUI grounding systems

The ingredients needed are simple:

• It should be Fast: A major chunk of the latency in a GUI grounding step is spent decoding a JSON object or a tool call token by token, something like {"action": "click", "point": [412, 288]} ( with reasoning). So if we can predict these actions and coordinates in a single pass without needing to decode them as text tokens, we can get a significant speedup.
• It should be Hybrid: Because we can't truly get rid of decoding; for an action like "type", we need the text tokens. So we need a hybrid approach that decodes text tokens when needed and predicts coordinates and bounding boxes in a single pass when not.
• It should be Extensible: to leverage and extend existing multimodal LLMs already trained for computer-use/GUI grounding tasks.

This post walks through how we got there.

Figure 1: Writing vs reading a decision. Open the interactive version.

---

1. Existing LLMs are already decision models (zero-shot)

Zero-shot classification using encoder models has been around for a while. Jev's recent popularity has renewed interest in this approach. Not only for encoder models, but this has brought in a perspective shift for decoder (LLMs) models. For simple classification tasks, do we even need to spend latency on decoding / generating texts ?

Many questions an application asks a model have a closed answer: is this urgent?, which team handles this ticket?, does the user want to cancel? For these, you don't need the model to write an answer. You need its opinion, as a probability over the allowed answers.

This can be elicited from an existing LLM by building the prompt so it ends exactly where the answer goes, and reading one next-token distribution:

<system: the questions, each with single-token labels>   A: billing  B: technical  C: other
<user: the ticket>
<assistant>department: __        ← P(" A"), P(" B"), P(" C") at this position = the answer

Nothing is generated. The labels are single tokens (yes/no, A–Z, 1–9), and the probabilities over them are the answer. Inference engines like vLLM / Sglang can return log-probabilities for exactly these token IDs, so one forward pass gives a calibrated-ish distribution over the options.

There are many recent community packages, including official ones that exposes a JEV-like system-one decisions API for any vLLM or SGLang server using this approach.

Interestingly, the LLM zero-shot results are comparable to JEV on the typed-decisions benchmark (2,000 decisions, evaluation results from SystemOne) :

Figure 2: Several questions, one prefill. Open the interactive version.

1.1. Training LLMs to be decision models

Zero-shot reads are good. A small LoRA makes them much better. Because the "answer" is a distribution over a handful of label tokens, training is just cross-entropy on that distribution at the answer position.

(We tried an RL objective (RLCD) and plain cross-entropy. Cross-entropy did as well or better, so that's what we use from here on.)

This is the key design choice: the model now learns to put probability mass on the right label at a fixed position instead of writing the answer. Now we have a way to fine-tune an existing LLM to be a decent decision model. But we need to go beyond categories/classification to predict continuous values like coordinates to build computer-use agents. Which brings us to the next section.

---

2. Slot reading: Predict Coordinates without decoding

A computer-use step needs more than just classification, It needs an action (click, type, scroll…), and usually a point on the screen. A point is two continuous values - x and y coordinates. The normalized coordinates are predicted as values between 0 and 1. One way to approximate these continuous values is to discretise them into a finite number of bins.

Slots:

We give the assistant turn a fixed template of slots, each ending in a placeholder token:

action: ___ x: ___ y: ___ x2: ___ y2: ___ key: ___ sdir: ___ samt: ___

Each slot's answer is the next-token distribution at the token before its placeholder. The placeholders let every slot be potentially read from one prefill: no slot waits for another to be generated. (In practice we send one request per slot in parallel and let vLLM's / SGlang's prefix cache share the screenshot and prompt.)

Figure 3: Slot reading. Open the interactive version.

The action slot is easy: its labels are words like click. The coordinates are the interesting part. How do you make a language model output a number in one token?

Numbers as bins

We split each axis into bins and give every bin its own single-token label. The model's answer for x is a distribution over bins. We read it out as the peak, refined by its neighbours (or the expected value). Training uses soft ordinal targets: a small Gaussian over neighbouring bins, so "one bin off" is penalised less than "half a screen off".

Figure 4: Where the model wants to click: a real read of the 256-bin x and y slots. Open the interactive version.

This leads us to two important questions: how many bins and which tokens from existing vocabulary to use as bins ?

3.1 Coarse to fine: xc, then x

My first experiment was a method borrowed from robotics, where action heads often predict a coarse bin and then a refinement (e.g. binned SE(2) trajectories). Each number gets two slots:

xc: ___   ← the tens digit: which tenth of the screen (10 labels: 0–9)
x:  ___   ← the fine position: 256 bins over the full axis

At read time, we fuse them: the fine distribution is multiplied by the coarse slot's probability for the tenth each bin falls in, then refined around the peak. The coarse slot acts as a prior that keeps the fine read from jumping to the wrong side of the screen.

I know that Occam wouldn’t be happy with this approach. To make him happy we test a simpler alternative.

Figure 5: How a number becomes one token (5 steps, ending with coarse to fine). Open the interactive version.

3.2 Fine only: 256 bins in one slot

The simpler alternative is one slot per number with 256 bins. This approach is inspired by GPC-1’s 101 Bin approach. The catch is which tokens to use as the 256 labels:

• Ordinary words (apple, river, …) as bins never learned: 0.6% accuracy on ScreenSpot. The model has to unlearn 256 word meanings and has no sense that bin 117 sits next to bin 118. With a coarse slot in front (3.1), the same word bins do learn. Combining with coarse slots seem to condition the fine slots for better learning.
• Byte-like tokens (ASCII characters plus two-letter tokens like ab, zz) work well. They carry little meaning to unlearn, and we chose them to have low embedding similarity to each other (mean cosine 0.14 for the 256-label set), so the model can separate them.

We also compared our 256-bin approach to 101 bins of single-character labels of GPC-1

Figure 6: Choosing bin labels. Open the interactive version.

Our arms, all LoRA on Gemma 4 31B:

3.3 Results

Figure 7: Accuracy summary. Open the interactive version.

What we took away:

• Occam was right. Coarse-to-fine isn't needed once the labels are right. With byte labels, the single 256-bin slot (E) matches or beats coarse + fine (F). It's also simpler and ~15–20 ms faster (fewer slots). Coarse-to-fine mainly rescues bad labels (A).
• 256 bins outperform 101 on agent steps (+9.5 points), and were comparable on clicks. Clicks only need to land inside an element. Agent steps (swipe ends, small targets) reward precision.

Latency

Each measurement is a single end-to-end computer-use step with a cold screenshot, an idle GPU, and no request batching — RTX PRO 6000, vLLM, Gemma 4 31B NVFP4. Median across 80 screenshots:

"Search Wikipedia for Lewis Hamilton and tell me which Formula One team he drives for." Both correct (Ferrari): slots 4.2 s, vanilla with thinking 30.5 s.

Results with smaller models:

We validated the same recipe on a 2B GUI model (GUI-Owl-1.5-2B, computer-use data only):

Same accuracy as vanilla GUI-Owl at native resolution, at 30% of the latency. (Vanilla GUI-Owl's click numbers use its agent prompt, not its dedicated grounding prompt.)

---

Caveats

1. Of course, this alone doesn't completely fix computer-use agents’ latency issue: it just speeds up grounding, and latency would increase as turns increase in an agentic session; multi-turn/complex tasks would also require reasoning for the agents to do well. These still add significant latency.
2. Other changes that made it read fast. The first version took ~240 ms per step. Three fixes brought it to ~146 ms:

What's next

• Faster text steps. Typing and answering still wait for the base model to write. We want to overlap that call with the slot read.
• Rethinking reasoning: Decision models have successfully raised a contrarian question: Is reasoning necessary for all the tasks? We should probably explore that further for multi-turn computer use agent tasks and do away with reasoning for the steps that don’t truly need it. **
• Multi-step computer use : As mentioned in the caveats, this blog post focuses only on speeding up the grounding step of the computer use task. Combining this with approaches to train a multi-step computer-use agent would be an interesting next step.

---

Appendix:

A.1 setup

• Model. Gemma 4 31B (bf16 for training, NVFP4 for serving with vLLM 0.30). LoRA r16 on the language model.
• Training. Full-vocabulary cross-entropy on soft ordinal targets, loss normalized per number, one adapter for all tasks.
• Data. AGUVIS (stage 2), Wave-UI clicks and boxes, RefCOCO (boxes, masks). Evaluation on ScreenSpot, held-out AGUVIS (a clean set from rows 6,000+ of every source), RefCOCO val, COCO val2017 and Fashionpedia.

A.2. Beyond clicks: boxes and segmentation

I wanted to experiment with extending the slot reading approach to bounding boxes and segmentation masks, with the following templates:

box:   cx: ___ cy: ___ w: ___ h: ___                 (sizes on a log scale)
mask:  cx: ___ cy: ___ r0: ___ r1: ___ … r23: ___    (24 rays from a centre)

• Boxes are a centre plus a size. The sizes are log-warped, because UI elements and objects span three orders of magnitude.
• Masks are a centre plus 24 ray lengths ("how far the object reaches in each direction"), drawn as a polygon. The centre is the point deepest inside the object, so it always falls inside it.

Figure 9: Boxes and masks as slots. Open the interactive version.

The 24-ray polygon gets the object's extent roughly right but misses thin parts. SAM 2 is very good at edges when told where to look. So we read a box and a centre in one pass, and hand them to SAM 2 as its prompt.

Figure 10: Box → SAM 2 → mask. Open the interactive version.

Results

Boxes on general objects. The first box task was trained on UI elements only. On photos it found the right object (centre inside the object 97% of the time) but made boxes far too small. Adding 5k RefCOCO photo boxes (E3) fixed that, without hurting computer use (clean test: 0.659 vs 0.654):

Masks. On their own, the 24 rays reach ~0.56 mIoU. The polygon gets the extent roughly right but cuts off thin parts (legs, arms). Handing our box and centre to SAM 2 as a prompt fixes the edges:

SAM 2 adds ~20 ms per image plus a few ms per prompt.

---

Citation

If you found this useful, please cite it as:

Umapathi, Logesh Kumar. (Oct 2026). "GUI-Decisions - Stop Decoding Coordinates for 3x Faster Computer-Use Grounding". hf.co/blog/infinitylogesh/gui-decisions-stop-decoding-coordinates

@article{umapathi2026slotreading,
  title   = "GUI-Decisions - Stop Decoding Coordinates for 3x Faster Computer-Use Grounding",
  author  = "Umapathi, Logesh Kumar",
  journal = "hf.co/blog",
  year    = "2026",
  month   = "Oct",
  url     = "hf.co/blog/infinitylogesh/gui-decisions-stop-decoding-coordinates"
}