BLINDSPOT — a no-image control for figure-understanding benchmarks

Byline: Nalandadata

A model is shown an unlabelled scientific diagram and must name every part a leader line points to. The labelled original is the answer key. The headline result is not the leaderboard — it is the no-image control: the same prompt repeated with the picture removed. Whatever the model still scores came from memory alone. Eight of nine models keep 70–100% of their score that way.

Dataset structure

This dataset contains 6,335 scored responses across 9 models, 4 conditions, and 181 benchmark items. Source images are not distributable (copyrighted textbook figures and a proprietary question bank); only scores, predicted names, and answer keys are released.

Files

File	Description
results.jsonl	One row per (item, model, condition) — 6,335 rows
items.jsonl	One row per benchmark item — 181 rows with aggregate metadata
sample/results.jsonl	10-row public preview

Schema — results.jsonl

Field	Type	Description
item_id	string	Public item code (NALANDA-N)
model	string	Model identifier, e.g. google/gemini-3.6-flash
condition	string	blind, told, desc, or no-image
f1	float	Synonym-aware name-set F1 (0–1)
precision	float	Precision component
recall	float	Recall component
n_pred	int	Number of names the model predicted
n_key	int	Number of ground-truth names
predicted_names	list[str]	The model's label list
key_names	list[str]	Ground-truth anatomical names
subject	string	Short factual description of the diagram
tier	string	Item tier (core or extended)
unsolved_core	bool	True if no baseline model solved this item
recitable	bool	True if ≥1 model scored F1 ≥ 0.5 with no image
parse_ok	bool	Whether the model's output parsed correctly

The four conditions

Condition	Model receives	Isolates
blind	Unlabelled diagram only	Visual recognition
told	Diagram + subject name	Naming, given recognition
desc	Diagram + long subject description	Does verbosity help?
no-image	Subject name, no picture	Prior knowledge alone — the control

grounded contribution = told − no-image   (percentage points)
recitation share      = no-image ÷ told   (%)

Prompts — verbatim, as sent

The condition column is only reproducible with the exact prompt text. All four prompts share the same JSON output schema; only the opening instruction differs.

blind — image attached, subject unknown:

You are an expert biology educator labeling a scientific diagram.
You are shown one biology diagram. Its parts are indicated by leader lines / arrows
(the printed labels have been erased). First work out what the diagram depicts, then
identify and name EVERY part that a leader line or arrow points to.

Rules:
- Use the single most specific, standard biology term for each part.
- Base your answer only on the image; do not invent parts.
- Output ONLY valid JSON (no markdown, no prose) in exactly this shape:
{
  "diagram_is": "<what the whole diagram depicts>",
  "labels": [ {"name": "<term>", "location_hint": "<where in image>"} ]
}

told — image attached, subject disclosed ({subject} = per-item string):

You are an expert biology educator labeling a scientific diagram.
You are shown a diagram of: {subject}. Its parts are indicated by leader lines / arrows
(the printed labels have been erased). Identify and name EVERY part that a leader line
or arrow points to.

Rules:
- Use the single most specific, standard biology term for each part.
- Base your answer only on the image; do not invent parts.
- Output ONLY valid JSON (no markdown, no prose) in exactly this shape:
{
  "diagram_is": "<what the whole diagram depicts>",
  "labels": [ {"name": "<term>", "location_hint": "<where in image>"} ]
}

desc — image attached, verbose subject description:

You are an expert biology educator labeling a scientific diagram.
You are shown a detailed, magnified educational diagram of {subject} — a biology
illustration drawn in a schematic / sectional style typical of a biology textbook.
Several of its distinct component structures, regions and parts are individually picked
out by leader lines or arrows (the printed labels have been erased). Study the whole
diagram carefully, then identify and name EVERY part that a leader line or arrow
points to.

Rules:
- Use the single most specific, standard biology term for each part.
- Base your answer only on the image; do not invent parts.
- Output ONLY valid JSON (no markdown, no prose) in exactly this shape:
{
  "diagram_is": "<what the whole diagram depicts>",
  "labels": [ {"name": "<term>", "location_hint": "<where in image>"} ]
}

no-image — no image attached, subject disclosed:

⚠️ The no-image prompt is byte-identical to told — it still says "You are shown a diagram of" and "Base your answer only on the image." No image is attached. This is intentional, not a transcription error: the goal was to hold the prompt constant and vary only the presence of the image, to isolate what the model could score from prior knowledge alone. A model that notices the contradiction and declines gets the same F1 as one that bluffs — the metric cannot distinguish them (see Finding 6).

You are an expert biology educator labeling a scientific diagram.
You are shown a diagram of: {subject}. Its parts are indicated by leader lines / arrows
(the printed labels have been erased). Identify and name EVERY part that a leader line
or arrow points to.

Rules:
- Use the single most specific, standard biology term for each part.
- Base your answer only on the image; do not invent parts.
- Output ONLY valid JSON (no markdown, no prose) in exactly this shape:
{
  "diagram_is": "<what the whole diagram depicts>",
  "labels": [ {"name": "<term>", "location_hint": "<where in image>"} ]
}

Decode settings

Provider	Settings
Google (Gemini)	temperature=0, thinkingBudget=1024 tokens, maxOutputTokens=8000
OpenAI (gpt-4o, gpt-4o-mini)	temperature=0, max_tokens=8000
OpenAI (gpt-5.6-luna)	reasoning_effort=low, max_completion_tokens=8000
Anthropic (claude-opus-5)	output_config.effort=high, max_tokens=24000†
Anthropic (claude-sonnet-4)	temperature=0, max_tokens=8000

† max_tokens was raised from 8,000 to 24,000 after 14 items truncated mid-JSON at the lower limit; those items were re-run.

Terminology

Every term this dataset uses. The column values in results.jsonl are only meaningful with these definitions.

The task

Term	Meaning
leader line	The thin line or arrow drawn from a printed label to the part of the figure it names. Erasing the label leaves the line, which makes the question answerable: the line says "name this" without saying what it is.
answer key	The list of part names printed on the original figure.
marker	A letter (A, B, C…) printed on the figure itself, used instead of leader lines by some items.

The four conditions

Condition	Image?	Subject named?	Question it answers
blind	✅	✗	Can it work out what the diagram is, unaided?
told	✅	✅	Given the subject, can it read the leader lines?
desc	✅	✅ described at length	Does verbose framing add anything?
no-image	✗	✅	How much can it answer from memory alone?

The metrics

Term	Meaning
name-set F1	The primary score. Compares the set of names the model produced against the set in the answer key — synonym-aware, so loop of Henle matches Henle's loop and alveoli matches alveolus. Position is ignored; only which names were produced. Reported as a percentage.
marker accuracy	The harder secondary score, for items with printed markers: was the right name attached to the right letter? Much lower than name-set F1 for every model — producing the vocabulary is easy, placing it is not.
grounded contribution	told − no-image, in percentage points. How much of the score actually required the picture. This is the headline metric.
recitation share	no-image ÷ told, as a percentage of the score. How much of the score came from prior knowledge instead.
pts vs %	A gap is written +17.7 pts; a share is written 71%. Reading one as the other inverts the finding.

The item sets

Term	Meaning
BLINDSPOT-181	The benchmark: 181 items kept because no baseline model scored 100% on them under told. A failure-only set — see Caveat 7.1.
BLINDSPOT-core	The 161 items no model ever solved, in any condition.
unsolved_core	Per-item flag — True if no model ever solved this item.
recitable	Per-item flag — True if at least one model scored F1 ≥ 0.5 with no image. 120 of 181 qualify.
tier	Which harvest an item came from: core (question/solution image pairs), extended (textbook figures), patternc (marker-keyed). Provenance, not difficulty.
Control A / Control B	The two selection-bias controls (Caveat 7.1). A = items excluded because a model solved them; B = pool figures never run at all.

Identifiers

Term	Meaning
NALANDA-1 … NALANDA-329	Public item IDs. Deliberately opaque: the internal IDs encode source book and page, which cannot be published. Ordered by a salted hash, not alphabetically, so the numbering does not cluster by source.

Results

Model	blind	told	desc	no-image	grounded	recited
google/gemini-3.1-pro-preview	55.8%	60.2%	60.2%	42.5%	+17.7	71%
google/gemini-3.6-flash	53.8%	59.5%	59.3%	41.4%	+18.1	70%
anthropic/claude-opus-5	42.2%	55.1%	53.5%	39.9%	+15.2	72%
google/gemini-2.5-pro	43.7%	52.6%	52.5%	41.9%	+10.7	80%
openai/gpt-5.6-luna	33.7%	50.7%	51.5%	4.5%	+46.2	9%
google/gemini-2.5-flash	39.4%	50.0%	49.5%	39.7%	+10.3	79%
anthropic/claude-sonnet-4	24.3%	45.2%	44.0%	—	—	—
openai/gpt-4o	31.4%	43.1%	44.3%	34.8%	+8.3	81%
openai/gpt-4o-mini	21.4%	41.3%	39.7%	41.1%	+0.2	100%

anthropic/claude-sonnet-4 has no no-image row: the model is retired from Anthropic's API and cannot be re-run. desc is complete for all nine.

Key findings

#	Finding
1	Recognition is the bottleneck, not naming — every model gains from being told the subject
2	Verbose context adds nothing — 9 models, every difference within ±1.6 pts
3	The recognition gap is closing, unevenly — Anthropic −8.0 pts, Google −5.0, OpenAI −2.9
4	8 of 9 models keep 70–100% of their score with no image
4b	Recitation is older, not newer — 2.5-era models recite 80–100%, current ones ~70%
4c	For gpt-4o-mini the image contributes nothing (+0.2 pts, CI contains zero)
5	120 of 181 items (66%) are answerable from memory by at least one model
6	Bluffing and abstention land at the same F1 — gpt-5.6-luna declines 81%, claude-opus-5 never declines
7	Inter-model agreement is not correctness — 31% of consensus names are absent from the key
8	Selection bias tested with two controls — it does not explain the finding (69–73%)
9	The reversal is statistically solid: +28.6 pts, 95% CI +23.2 to +33.6

Caveats

#	Caveat
7.1	Failure-only set — items were kept because baseline models failed them. Tested with two controls; the finding holds.
7.2	Mixed API routes — baseline models ran via OpenRouter, controls via native APIs. For the four baseline models the grounded contribution subtracts a native no-image score from an OpenRouter with-image score.
7.3	8 items reproduce the figure at small scale. Excluding them changes nothing.
7.5	Item-sampling variance measured by bootstrap: ±4 pts. The top two models are a statistical tie.
7.6	Biology only. Untested on other domains.
7.7	Contamination not separated from general knowledge — these are published textbook figures.

Why no images?

All 181 benchmark items come from copyrighted textbooks and a proprietary question bank. There is no subset whose figures are releasable. Item IDs are anonymised public codes (NALANDA-N); the mapping to source books is private and will not be released.

Scoring

Responses are scored with a synonym-aware name-set F1. The scorer handles:

• Latin/Greek plural forms (alveolus/alveoli, atrium/atria)
• Synonyms (mitral valve = bicuspid valve = left AV valve)
• Fuzzy matching for morphological variants (tight guards: ≥7 chars, first 3 identical, ≤3 length difference, ≥0.86 similarity)

Scoring scripts (score_ext.py, score_grounding.py) are included in the full dataset repository.

Item provenance — three patterns

Pattern	Key source	How the question was made	n
A	Solution image already labelled	Nothing — a true pair	20
B	Born-digital textbook figure	Label text erased from PDF; leader lines kept	144
C	Markers printed on question image	Key parsed from solution prose	17

Citation

@dataset{nalandadata2026blindspot,
  title   = {{BLINDSPOT}: a no-image control for figure-understanding benchmarks},
  author  = {Nalandadata},
  year    = {2026},
  url     = {https://huggingface.co/datasets/Nalandadata/BLINDSPOT}
}