Inference Pipeline Diagram
Hover the stages. The key transition is where demonstration selection becomes curriculum design.
From Retrieval Bucket to Teaching Interface
Instead of treating demonstrations as isolated neighbors, the paper’s framing says the prompt can become a structured workspace. This view only pays off when the model can actually use the reasoning traces in sequence.
Concept Tension Matrix
Rows show design choices. Columns show the paper’s claimed consequences when prompts become long.
low effect
medium effect
high effect
Scaling Is Conditional, Not Monotonic
The paper argues that more demonstrations are not uniformly better. The visual below contrasts reasoning-friendly and non-reasoning settings while exposing growing order sensitivity as prompt length increases.
Many-Shot Scaling Curves
Toggle between outcome and instability. Mock data mirrors the episode’s reported directional pattern.
Order Sensitivity Heatmap
Cells estimate how much performance swings across random demo orderings as demonstration count rises.
Curvilinear Demonstration Selection
CDS is the paper’s answer to conceptual whiplash. Rather than choosing only nearby examples, it tries to create a smooth route through reasoning space.
Reasoning Manifold
The left path shows semantic nearest-neighbor hopping. The highlighted curve shows a staged curriculum.
Geometry Gains
Best-case improvement from smoother ordering becomes most visible at long prompt length.
Prompt Studio
Two knobs shape the mock prompt: demonstration count and procedural compatibility. The token map shows why “more” can become clutter instead of instruction.
Prompt Composition Map
Hover cells to inspect where the context window is spending tokens and which spans are actually teachable.
Predicted Outcome Dial
A compact scorecard converts the visual budget into expected accuracy, latency, and fragility.