VLM Test-Time Sampling Likelihood-Landscape Diagram

Overview

A side-by-side conceptual diagram of two ways to get diverse answers from a vision-language model at test time. It is an editable PowerPoint diagram adapted from the original paper figure under CC BY 4.0.

What the diagram shows

Each half shows a 3D likelihood landscape (a shaded quad-mesh surface over semantic dimensions 1 and 2, with a higher-likelihood axis) whose peaks are labelled dog, cat, mountain, hiking and forest regions; an image-question card ("What animal is in the image?") feeds the starting state h⁽⁰⁾ through a dashed arrow. Temperature sampling (baseline): three "Base VLM" trajectories climb the same dog peak, a legend marks trajectories and sampled outputs, a soft coral callout says the samples stay in the same region, and the sampled-answers strip reads A dog / A dog / A dog / A cute dog. Computation-path sampling (proposed): four "Effective VLM 1–4" trajectories in different colours end on different peaks with coloured output dots, a green callout says the samples explore multiple regions, and the answers strip reads A dog / A golden retriever / A dog in the mountain / A hiking dog. Header arrows, a dashed centre divider and a one-line caption under each half complete the comparison. Every box, arrow, label and math term is a native shape you can recolor, resize and relabel for your own slides, posters or papers.

File format

The download contains .pptx (1 slide, 15.26 × 5.90 in; opens in PowerPoint, Keynote, WPS and LibreOffice Impress; Google Slides works too, but complex freeforms may shift slightly), .svg and .drawio. Includes preview.png (full resolution) and a 1200 px web preview.

Shapes and editing

The file has 1748 native editable shapes: 64 auto-shapes, 49 text boxes, 1622 freeforms, 13 lines/connectors (4 small groups). No embedded images. Everything is vector. You can edit all text and math (h⁽⁰⁾), every quad of both likelihood surfaces and their base shadows, the axis arrows and labels, the region labels, the image-question cards and their placeholder photos, the dashed input arrows, every trajectory curve with its arrowhead and label, the start and output dots, both legends, the two callouts, the sampled-answer chips and dashed strips, the header arrows, the centre divider and the caption lines.

Embedded images and fonts

This file has no embedded images; everything is vector. Each likelihood surface is a native mesh of small four-sided freeforms (784 per surface), so it recolours and scales as vector but is heavy to edit quad by quad; each surface is grouped so you can move or resize it in one go. The photo in each question card is a simplified flat vector placeholder (sky, ridge, trees, meadow and a cartoon dog), not the paper's photo; paste your own image over it if you need a real sample. All text, body and formulas alike, uses Times New Roman, which is preinstalled on Windows and macOS (LibreOffice substitutes the metric-compatible Liberation Serif). Sub- and superscripts are real formatted text runs, not equation objects. The SVG keeps text as text; the draw.io file is a shape-by-shape conversion and may need small nudges. This is a faithful redraw for reuse, not the authors' original source file.