RealityBench v1

🌐 Website | 📄 Paper: arXiv link to come

RealityBench tests whether a vision-language model can review a photo taken to document physical work. Each of the 1,378 items is one photo from one of eleven environments (seven from rooftop solar installation, four from vehicle repair intake) and one prompt. The model answers in a single JSON object:

• Acceptability: does the photo meet the stated requirement ("adequate": true/false)?
• Documentation: what does the photo show (a tape reading, a serial number, whether a lag bolt is present, ...)?

The photos are rendered in Blender from procedurally generated scenes, through a model of a phone camera (depth of field, sensor noise, missed focus and hand shake), so the ground truth for every documentation answer is read from the scene. The acceptability verdict comes from a rule per environment, tuned against human raters (below).

🏆 Leaderboard

26 models, one query per item, each at its provider's or model card's default settings, 13-23 September 2026. Sorted by acceptability F1; best model score in each column in bold.

Model	F1	precision	recall	documentation	in grammar (%)
GPT-6 Astra	0.61	0.55	0.68	0.83	100
Claude Opus 5.5	0.58	0.51	0.67	0.83	100
GPT-5.6 Sol	0.57	0.49	0.69	0.77	100
GLM-5.3 Flash	0.57	0.49	0.68	0.75	99
Gemini 3.5 Flash	0.56	0.49	0.66	0.80	97
Claude Fable 5.1	0.56	0.51	0.63	0.83	100
Claude Opus 5	0.56	0.54	0.57	0.75	100
Qwen 3.8 Max	0.55	0.43	0.78	0.82	100
GPT-6 Sol	0.55	0.50	0.61	0.78	100
Muse Spark 1.3	0.55	0.49	0.62	0.79	100
Qwen 3.5 397B-A17B	0.55	0.49	0.61	0.76	99
GPT-5.6 Terra	0.55	0.45	0.70	0.76	100
Qwen 3.8 27B	0.54	0.45	0.68	0.74	99
Grok 4.6	0.54	0.43	0.72	0.78	100
Gemini 3.8 Flash	0.54	0.48	0.62	0.84	100
Kimi K3	0.53	0.42	0.70	0.77	99
GPT-5.6 Luna	0.53	0.46	0.62	0.71	100
Gemma 4 31B	0.52	0.44	0.63	0.79	100
DeepSeek V4.1 Flash	0.50	0.39	0.68	0.68	95
Gemma 4 26B-A4B	0.48	0.39	0.64	0.77	99
Claude Sonnet 5	0.48	0.45	0.51	0.73	100
Gemini 3.1 Flash Lite	0.47	0.39	0.58	0.73	98
LFM2.5-VL-3B	0.43	0.30	0.79	0.68	99
GPT-6 Luna	0.43	0.46	0.41	0.62	100
Gemma 4 E4B	0.42	0.30	0.70	0.69	98
Qwen3.5-4B	0.39	0.31	0.50	0.66	95
Human rater*	0.59
Always "adequate"	0.33	0.19	1.00

* A single non-author rater against the scorer, on 274 rated frames rather than this set (95 % CI 0.50-0.67). Intervals, per-environment scores and every model's answers are in results/.

Example

tape_gap/img_005117. The scorer accepts this frame. Ground truth: tape reads 4.88 in where the module frame crosses it (1/8 in tolerance), true gap 4.72 in (1/4 in tolerance), tape not tilted.

Prompt

You are reviewing a photo taken by a solar installer to document their work; the photo criteria will be used to prove the work was carried out correctly. The photo we need: Tape measure hooked at the roof surface and standing up to the module frame; the number where the frame crosses the tape must be legible, and the tape should be seated flush (perpendicular to the roof). First: is this photo adequate -- does it satisfy the requirement as a reviewer would judge it? Then answer the documentation question. What number (inches) is on the tape where the module frame crosses it, what is your estimate of the true perpendicular module-to-roof gap in inches, and is the tape leaning rather than standing flush? Answer with exactly one JSON object and nothing else -- no prose, no code fence -- with exactly these keys: "adequate": true or false (a JSON boolean): whether the photo is adequate; "reading_in": a number (JSON number, no units, no quotes); "gap_in": a number (JSON number, no units, no quotes); "tilted": true or false (a JSON boolean). Use these values exactly as written; a value in any other form is scored as wrong. Answer every key even if you judge the photo inadequate: give your best reading of what is visible.

Model	adequate	reading_in	gap_in	tilted
GPT-6 Astra	false ✗	5 ✓	4 ✗	true ✗
Claude Opus 5.5	false ✗	5.1 ✗	4.8 ✓	true ✗
Gemini 3.8 Flash	true ✓	5.5 ✗	5.5 ✗	false ✓

Usage

from datasets import load_dataset
ds = load_dataset("relantic/realitybench", split="test")   # send prompt + image as one user message

python run_model.py --base-url http://localhost:8000/v1 --model <your model> --out predictions.jsonl   # any OpenAI-compatible endpoint
python score.py predictions.jsonl            # standard library only; lines are {"id": ..., "response": "<reply text>"}
python score.py results/lfm2.5-vl-3b.jsonl   # check your setup: prints that model's leaderboard row

Acceptability

Each prompt states a short natural-language requirement for the photo. The scorer turns it into checks (in frame, unoccluded, distance, view angle, resolution, sharpness, contrast), and a frame is acceptable when it passes all of them. Acceptability judges only the photo, so a photo of faulty work can still be acceptable. Raters disagree on these requirements, so the scorer encodes one interpretation of each: on held-out frames it agrees with the majority of six raters on 82 % of frames, while a single rater agrees with the others' majority on 76-92 %.

Scoring rules, items and fields

Metrics. F1: positives are the frames the scorer accepts; precision on the 934 random frames with an acceptability score, recall on those plus the positive strata. A reply with no JSON boolean under adequate gives no verdict and is left out. Documentation: mean per-field score on scorer-acceptable frames (exact match after normalisation for text; tape 1/8 in, gap 1/4 in, slope 0.2 degrees), except tape leaning, which is scored on every tape_gap frame. In grammar: answers with exactly the requested keys and types; reported, not scored.

Items. Random frames come from each environment's camera-pose distribution; the positive stratum adds acceptable frames where they are rare and counts for recall only. Counts below 100 are frames left out because their verdict depended only on an ambiguous term (43) or withdrawn for a texture defect (ro_board).

environment	documentation question	random frames	of which acceptable	positive stratum
breaker_label	main breaker amps; every circuit's name and amps	100	5	50
car_plate	plate text and state	100	21	--
car_view	which face (front, rear, driver, passenger); body style	87	23	--
damage_view	damage kind (dent, scratch, rust); severity	86	24	--
equipment_label	brand, model, serial; on a module, microinverter or inverter	100	6	50
horizon_peak	horizon visible? level?	100	19	50
rail_attachment	lag bolt present? flashing present?	100	12	50
ro_board	RO number, VIN	79	14	42
slope_gauge	gauge reading (degrees); on the module?	91	25	50
tape_gap	tape reading at the module frame; true gap (in); tape leaning?	93	17	50
wire_adequacy	wire touching the roof?	100	39	--

Fields. id, image, environment, stratum (random / positive), prompt (send as is), acceptable (the scorer's verdict), acceptability_scored, documentation_asked, ground_truth (JSON string; also in items.jsonl).

Caveats. car_plate and 13 rail_attachment frames have no acceptability score, and 22 damage_view frames that do not show the damage have no documentation score. body_style in ro_board and own_roof_blocks in horizon_peak are asked but not scored. Our runs allowed 8,192 output tokens, thinking included.

Real tape-measure photos (tape_real)

To check that tape reading, one of the hardest rendered tasks, transfers to real photos: 133 readings on 84 photos of a tape measure held against everyday objects, with the truth from a caliper or a careful human tape read. A reading counts when it is within 1/8 in.

tape = load_dataset("relantic/realitybench", "tape_real", split="test")   # then: python score.py --tape predictions.jsonl

Scores on the real photos (22 models)

A reply cut by the output-token cap is left out of that model's rate, so the number of scored readings varies.

Model	within 1/8 in	readings scored
GPT-6 Astra	0.44	133
Claude Opus 5.5	0.35	133
GPT-6 Sol	0.33	133
GPT-5.6 Sol	0.33	133
Claude Fable 5.1	0.29	133
GLM-5.3 Flash	0.27	79
Claude Opus 5	0.23	133
Gemini 3.8 Flash	0.21	133
GPT-5.6 Luna	0.19	133
Gemini 3.5 Flash	0.17	133
Qwen 3.5 397B-A17B	0.13	125
Grok 4.6	0.13	133
Qwen 3.8 27B	0.13	118
Gemma 4 31B	0.11	133
Gemma 4 26B-A4B	0.11	133
Gemini 3.1 Flash Lite	0.11	133
GPT-6 Luna	0.11	133
GPT-5.6 Terra	0.10	133
Claude Sonnet 5	0.08	133
Gemma 4 E4B	0.05	133
LFM2.5-VL-3B	0.03	133
Qwen3.5-4B	0.02	133

Limitations

Single frames rather than video; the scorer is one reading of each requirement; two kinds of work; general-purpose models zero-shot at provider defaults. See the paper.

Contact

support@relantic.com, including for access to the environments (this repository holds the rendered benchmark only).

Citation

@misc{perrault2026realitybench,
  title  = {RealityBench: Simulated Environments to Evaluate a Vision Language Model's Ability to Document Human Physical Work},
  author = {Perrault, Andrew and Takiar, Anmol and Wienecki, Dominik and Nandi, Arnab},
  year   = {2026},
}

License

Frames, tape-measure photos, ground truth and model results: CC BY 4.0 (LICENSE-DATA). Code: MIT (LICENSE). The frames were rendered from third-party CC0 and CC BY 4.0 assets, credited in ATTRIBUTION.md. © 2026 Relantic, Inc.