README.md

---
license: other
license_name: mixed-see-licence-section
task_categories:
- image-to-video
language:
- en
tags:
- video-reasoning
- visual-domain-shift
- blender
- benchmark
pretty_name: XVReason Visual Shift
size_categories:
- n<1K
configs:
- config_name: vbvr_bench_visual_shift
  data_dir: vbvr_bench_visual_shift/first_frames
- config_name: viper
  data_dir: edited_first_frames/viper
- config_name: genvire
  data_dir: edited_first_frames/genvire
- config_name: vreasonbench
  data_dir: edited_first_frames/vreasonbench
- config_name: videothinkbench
  data_dir: edited_first_frames/videothinkbench
---

# XVReason Visual Shift

The inputs of the visual-domain-shift evaluation in *Video Reasoning Generalizes to Perception,
Embodiment, and Real-World Simulation*. It asks whether a model fine-tuned on abstract reasoning
videos (flat shapes, grids, simple colours) still reasons when the same problem is shown as a
natural scene. Every item here is a one-to-one counterpart of an item in an existing benchmark,
with the task and its answer unchanged and only the appearance changed.

| part | items | counterpart of | what changed |
|---|---:|---|---|
| `vbvr_bench_visual_shift/` | 500 | VBVR-Bench | the whole clip, re-staged in Blender |
| `edited_first_frames/viper/` | 145 | VIPER | the first frame |
| `edited_first_frames/genvire/` | 39 | Gen-ViRe | the first frame |
| `edited_first_frames/vreasonbench/` | 17 | V-ReasonBench | the first frame |
| `edited_first_frames/videothinkbench/` | 162 | VideoThinkBench (eyeballing + visual puzzles) | the first frame |

Code to generate and score: [`Zane-ZYQiu/xvreason`](https://github.com/Zane-ZYQiu/xvreason),
`bench/` (see its README, section *Visual shift*).

## `vbvr_bench_visual_shift/` — VBVR-Bench re-rendered as natural scenes

Each of VBVR-Bench's 500 items (100 task families x 5 samples) is re-staged as a Blender scene:
billiard balls instead of a bouncing dot, a hedge maze instead of a grid. The answer is rebuilt
analytically from the original item's stored state, so it is preserved by construction, and each
clip keeps the original's frame count and frame rate. Clips are 1024 x 1024 H.264.
This is the final set (revision `44dc965` of the renderers' working repo, 2026-09-25).

| file | |
|---|---|
| `bench_index.csv` | one row per VBVR-Bench item: `bench_id` (the item's VBVR-Bench path), the clip, its first frame, the natural-domain `prompt` and the original `source_prompt` |
| `first_frames/<split>/<task>/<sample>.png` | frame 0 of each clip, the image the model is conditioned on |
| `first_frames/original/<split>/<task>/<sample>.png` | the original VBVR-Bench first frame of the same item, for comparison |
| `first_frames/metadata.jsonl` | one row per item pairing the original first frame (`original`) with the re-rendered one (`edited`), plus its prompts; this is what the dataset viewer and `load_dataset` show |
| `videos/<family>/<sample>/<variant>/video.mp4` | the rendered clips (ground truth for the evaluator) |
| `inputs/<family>/<sample>/input.json` | the recovered task state each clip was rendered from |
| `metadata.csv`, `meta/<family>.jsonl` | per-clip provenance, hashes and the checks each clip passed |
| `rendering/README.md` | how to re-render a clip with the scene builders in the GitHub repo |

A model is prompted with the natural `prompt` and conditioned on the first frame. VBVR-Bench's own
evaluator finds task elements by hard-coded colour and so cannot score these clips; the natural
evaluator in the GitHub repo (`bench/score/vbvr_bench_natural/`) applies the same per-family criteria
with detectors that work on rendered scenes, and was validated against a 19-control battery.

## `edited_first_frames/` — edited first frames for four benchmarks

Each first frame is moved into a genuinely real-world scene while the task and its answer stay the
same -- a ball sketch becomes chutes and bins in a factory, tic-tac-toe becomes cinnamon sticks and
cookies on a floured board, a maze becomes a garden hedge maze.

For each item a VLM (Gemini 3.8 Flash) brainstormed real-world scenes and wrote an edit instruction;
Nano Banana applied it; a second VLM pass accepted the result only if the task logic was intact and
the image was a convincing real-world scene, with up to three attempts. Position-critical items
(geometric constructions, colour puzzles) were also checked on an overlay of the original and the
edit, and their candidate letters had to stay legible.

Each folder holds the edited images, the benchmark's original first frames under `original/`, and a `metadata.csv` pairing them with the benchmark's item `id`. `load_dataset` (and the dataset viewer) returns `original` and `edited` side by side. Everything else
about each item stays as in the original benchmark: the prompt, the reference frames, and the scoring.

## Licence

The Blender scenes, renders, first frames, edit instructions and edited images are released under
**CC BY 4.0**. Each part is derived from a third-party benchmark, and that benchmark's terms still
apply to what came from it:

- VBVR-Bench is Apache-2.0.
- VIPER is MIT.
- VideoThinkBench is MIT.
- V-ReasonBench is Apache-2.0.
- Gen-ViRe's dataset states no licence.

The render bundles use HDRIs and PBR textures from Poly Haven and ambientCG, both CC0.