Turn ML ideas into
interactive explanations.

Give readers a figure they can explore. Edit a preset, add your explanation and sources, then share it in an article, presentation, or documentation page.

Explore presets See a finished example

Free with LiminalML attribution · No account · No code

Real Q-learning. Change how much the agent explores.Edit this visual

Where PPO stops rewarding a bigger step

Pause and take one update at a time. Follow the probabilities and the clipped-term count as PPO stops rewarding larger changes.

Clipping removes a term’s direct gradient. Other outcomes can still move its probability.

Edit this example

Study PPO step by step

Model and source

Six fixed rollout observations with illustrative advantages. Clip range ε = 0.20; logit step size η = 0.60. This shows the actor surrogate, not a full PPO training run.

Softmax couples the outcomes, so a clipped term’s probability can still change due to other terms. This isolates the actor surrogate without a value function, entropy bonus, or fresh rollout collection.

Proximal Policy Optimization Algorithms

PPO clipping · actor surrogateε = 0.20

A mechanism, a figure,
or a system.

Explore an ML mechanism, plot your measurements, or map a pipeline. Presets include explanations and editable inputs; replace sample data before making empirical claims.

Clipping changes the incentive.

Change the clip range and inspect where the policy surrogate stops rewarding movement in the same direction.

There is information beyond the top answer.

Reveal a teacher’s softened distribution and watch a student learn its relative preferences.

Let the evidence speak.

Multi-series plots with uncertainty bands. Paste observations, name the axes, and compose a research figure.

Write the connections.

Describe a pipeline in a few lines. Nodes arrange themselves into a crisp, publication-style diagram.

Explore all presets

Follow the mechanism into modern ML.

Start with a concrete question. Read the primary paper, explore an available visual, or work through the topic in Study.

How reasoning models learn

Separate demonstrations, group-relative rewards, and teacher signals. Inspect a small GRPO mechanism before studying the full training pipeline.

Explore the GRPO mechanism DeepSeek-R1 · 2025, revised 2026

Where inference time goes

A draft model proposes tokens; the target verifies and corrects them. Learn why the correction rule matters for preserving the target distribution.

Edit the decoding diagram Speculative decoding · ICML 2023

When more thinking helps

Compare search, verification, and revision at inference time. More compute is a budget to allocate; its usefulness depends on the problem and the method.

Study this topic Scaling test-time compute · 2024

Studio mechanisms use small illustrative models. The decoding diagram shows data flow. Neither reproduces a paper’s training runs or benchmark results.

Make it yours.
Then put it to work.

Your drafts stay in this browser. Download a project file to keep an editable backup.

Shape the explanation.

Edit the data, labels, and narrative. Add the source, methods, and limits that help readers interpret your figure.

Check it in context.

Preview the figure in a slide, document, or article layout. These are placement previews; the download contains the figure itself.

Share the finished artifact.

Download a PNG, supported SVG figure, or offline HTML for free. Share and embed links contain a snapshot. All free outputs keep the LiminalML credit.

Create an explanation Want to learn a concept? Try Study

Turn ML ideas into
interactive explanations.

Give readers a figure they can explore. Edit a preset, add your explanation and sources, then share it in an article, presentation, or documentation page.

Explore presets See a finished example

Free with LiminalML attribution · No account · No code

Real Q-learning. Change how much the agent explores.Edit this visual

Where PPO stops rewarding a bigger step

Pause and take one update at a time. Follow the probabilities and the clipped-term count as PPO stops rewarding larger changes.

Clipping removes a term’s direct gradient. Other outcomes can still move its probability.

Edit this example

Study PPO step by step

Model and source

Six fixed rollout observations with illustrative advantages. Clip range ε = 0.20; logit step size η = 0.60. This shows the actor surrogate, not a full PPO training run.

Proximal Policy Optimization Algorithms

PPO clipping · actor surrogateε = 0.20

A mechanism, a figure,
or a system.

Explore an ML mechanism, plot your measurements, or map a pipeline. Presets include explanations and editable inputs; replace sample data before making empirical claims.

Clipping changes the incentive.

Change the clip range and inspect where the policy surrogate stops rewarding movement in the same direction.

There is information beyond the top answer.

Reveal a teacher’s softened distribution and watch a student learn its relative preferences.

Let the evidence speak.

Multi-series plots with uncertainty bands. Paste observations, name the axes, and compose a research figure.

Write the connections.

Describe a pipeline in a few lines. Nodes arrange themselves into a crisp, publication-style diagram.

Explore all presets

Follow the mechanism into modern ML.

Start with a concrete question. Read the primary paper, explore an available visual, or work through the topic in Study.

How reasoning models learn

Separate demonstrations, group-relative rewards, and teacher signals. Inspect a small GRPO mechanism before studying the full training pipeline.

Explore the GRPO mechanism DeepSeek-R1 · 2025, revised 2026

Where inference time goes

A draft model proposes tokens; the target verifies and corrects them. Learn why the correction rule matters for preserving the target distribution.

Edit the decoding diagram Speculative decoding · ICML 2023

When more thinking helps

Compare search, verification, and revision at inference time. More compute is a budget to allocate; its usefulness depends on the problem and the method.

Study this topic Scaling test-time compute · 2024

Studio mechanisms use small illustrative models. The decoding diagram shows data flow. Neither reproduces a paper’s training runs or benchmark results.

Make it yours.
Then put it to work.

Your drafts stay in this browser. Download a project file to keep an editable backup.

Shape the explanation.

Edit the data, labels, and narrative. Add the source, methods, and limits that help readers interpret your figure.

Check it in context.

Preview the figure in a slide, document, or article layout. These are placement previews; the download contains the figure itself.

Share the finished artifact.

Download a PNG, supported SVG figure, or offline HTML for free. Share and embed links contain a snapshot. All free outputs keep the LiminalML credit.

Create an explanation Want to learn a concept? Try Study

Turn ML ideas intointeractive explanations.

Where PPO stops rewarding a bigger step

A mechanism, a figure,or a system.

Clipping changes the incentive.

There is information beyond the top answer.

Let the evidence speak.

Write the connections.

Follow the mechanism into modern ML.

How reasoning models learn

Where inference time goes

When more thinking helps

Make it yours.Then put it to work.

Shape the explanation.

Check it in context.

Share the finished artifact.

Turn ML ideas intointeractive explanations.

Where PPO stops rewarding a bigger step

A mechanism, a figure,or a system.

Clipping changes the incentive.

There is information beyond the top answer.

Let the evidence speak.

Write the connections.

Follow the mechanism into modern ML.

How reasoning models learn

Where inference time goes

When more thinking helps

Make it yours.Then put it to work.

Shape the explanation.

Check it in context.

Share the finished artifact.

Turn ML ideas into
interactive explanations.

A mechanism, a figure,
or a system.

Make it yours.
Then put it to work.

Turn ML ideas into
interactive explanations.

A mechanism, a figure,
or a system.

Make it yours.
Then put it to work.