Datasets:

matsant01
/

blind-spots-bench

QID int64 1 285	prompt stringlengths 14 3.63k	solution stringlengths 3 2.27k	question_type stringclasses 5 values	categories stringclasses 24 values	failure_modes stringlengths 8 103	link stringlengths 38 210	author stringlengths 8 8 ⌀	image imagewidth (px) 168 1.02k ⌀
1	what's missing from this airpod	The correct output is identifying that the protective case for the upper part is missing from the AirPod setup. A common incorrect behavior is assuming that nothing is missing and instead focusing on the placement or orientation of the AirPods (such as one being tilted or not seated correctly), which does not answer t...	multi-to-text	visual reasoning	visual perception and grounding, spatial reasoning	https://g.co/gemini/share/a1155863a629	749e2c73
2	what could've been written here, give me all the possibilities	The correct output is to provide a constrained set of plausible readings of the exact word, centered on “DESiGNER’S” (and closely related capitalization variants like “DESiGNER’S”, “DesiGNeR’s”), rather than inventing unrelated words. This is correct because the solution answers both parts of the prompt: it proposes w...	multi-to-text	visual reasoning	visual perception and grounding	https://g.co/gemini/share/27f83dd0d668	749e2c73
3	what colors are the dies in	The correct output is listing all colors present across the several dice, specifically: white, red, and blue. A common incorrect behavior is mentioning only a subset of the colors (such as white and red) and failing to account for additional colors like blue, resulting in an incomplete answer.	multi-to-text	visual reasoning	visual perception and grounding	https://gemini.google.com/share/d92130edeac6	749e2c73
4	how many authors are affiliated with at least two institutions	The correct output is that there are only two authors affiliated with at least two institutions: Zhepei Wei and Rulin Shao. • Zhepei Wei is affiliated with University of Virginia (¹) and Meta Reality Labs (²). • Rulin Shao is affiliated with University of Washington (³) and FAIR at Meta (⁴). This is correct because ...	multi-to-text	visual reasoning	perceptual counting, visual perception and grounding	https://g.co/gemini/share/66cf0a9125fe	749e2c73
5	how many home screens are there in this figure	The correct output is that there are two home screens in the figure. This is correct because only two of the displayed phone screens qualify as home screens, meaning they show the main app grid or launcher interface. Other screens in the figure may be lock screens or specific app views, which should not be counted as ...	multi-to-text	visual reasoning	perceptual counting, visual perception and grounding	https://gemini.google.com/share/34442055d9b9	749e2c73
6	how many white stripes are there on both shoes	The correct output is that there are 8 white stripes on both shoes. This is correct because the question asks for the total number of white stripes across both shoes, and the correct interpretation of the image shows 4 white stripes per shoe, resulting in 8 stripes overall. A common incorrect behavior is assuming the...	multi-to-text	visual reasoning	visual perception and grounding, perceptual counting	https://g.co/gemini/share/626dba7c8db6	749e2c73
7	How many legs does this cat have:	The correct answer must report the exact number of legs visible on the cat in the image. A response is correct only if it gives the true count; in this case, the correct count is five. A common incorrect behavior is miscounting due to occlusion, shadows, or assuming the standard four-legged anatomy instead of what is a...	multi-to-text	visual reasoning	perceptual counting	https://g.co/gemini/share/9e5783028ee3	749e2c73
8	You are given a table with size $n \times n$(n^2 cells). Name the rows from 1 to n. The same for columns. There is a cop at cell (n-1,n) and a thief at (n, n-1). Cells (2,2), (3,3), … (n,n) are blocked. At each timestamp both the thief and cop should go in either of 4 directions UP, DOWN, LEFT, RIGHT at the same time (...	A correct solution must determine whether a strategy exists for the thief to reach any cell with j ≥ i while avoiding capture, given the simultaneous moves, starting positions, and blocked diagonal cells. The reasoning must explicitly respect the movement rules, simultaneity, and obstacles. A common incorrect behavior ...	text-only	mathematical reasoning	constraint reasoning	https://chatgpt.com/share/68e3bb48-77c8-800d-8754-67934cca619b	5e7f2ab4	Not supported with pagination yet
9	Generate an image of a clock viewing in the mirror	The correct output is an image of a clock as seen in a mirror, meaning the clock face must be left-right reversed while preserving the actual time shown. This is correct because a mirror flips horizontal orientation but does not change the time itself. A common incorrect behavior is showing a normal (non-mirrored) cloc...	text-to-image	visual reasoning	visual transformation	https://chatgpt.com/share/68e3bbcf-f450-800d-a289-f847420395d8	5e7f2ab4	Not supported with pagination yet
11	There is a circle on a line. Three points on the circle make the perimeter of the circle into 3 equal parts. One point is touching the line named A, and reading the points' names anti-clockwise, it would be A, B, C. The circle spins right (clockwise) for 487 degrees. What point is closest to the line after that?	The correct answer must account for the fact that the points A, B, and C are labeled in anti-clockwise order and the circle is rotated clockwise by 487 degrees. Since 487 = 360 + 120 + 7, this is equivalent to a 127-degree clockwise rotation, or 120 degrees (one third of a turn) plus 7 more degrees. Therefore, the corr...	text-only	mathematical reasoning	geometric reasoning	https://chatgpt.com/share/68e3bc4c-7810-800d-8b59-098ba0014a9d	5e7f2ab4	Not supported with pagination yet
12	Make a figure that repeats a circle, a triangle, and a square 4 times. At the end of the repetition plot, a rectangle with a large length.	The correct output is an image showing the repeating sequence circle → triangle → square repeated exactly four times, followed by a rectangle with a long length at the end. This is correct because both the order and the number of repetitions are explicitly constrained. A common incorrect behavior is using the wrong ord...	text-to-image	instruction following	hard constraints	https://chatgpt.com/share/68e3bcc8-e004-800d-8cfe-324811f02ddb	5e7f2ab4	Not supported with pagination yet
13	Generate a photo of a square in a circle, in which in a further inner square there is another circle, and in the circle is again a square, and in the square there is a dot.	The correct answer is an image depictiong a square in a circle in which in the inner square there is another circle, and in the circle is again a square, and in the square there is a dot.	text-to-image	visual reasoning	spatial reasoning	https://chatgpt.com/share/68e3d1b8-ea1c-800d-932e-7b8b55be4140	5e7f2ab4	Not supported with pagination yet
18	Can you read the words marked in this find a word game	WORD, SEARCH, JOURNEY, LOVE, RELAX, ESCAPE, FUN, PLAY, PUZZLE, ENJOY	multi-to-text	visual reasoning	visual perception and grounding	https://gemini.google.com/share/c7c062a1d89d	866efed1
19	Let’s pretend we’re on the metric system and use kilograms instead of pounds to give us a starting base number of 100. Four people (Alex, Brook, Chris and Dusty) want to cross a river in a boat that can only carry 100kg. Alex weighs 90kg, Brook weighs 80kg, Chris weighs 60kg and Dusty weighs 40kg, and they have 20k...	The correct output is a valid sequence of river crossings that transports Alex (90kg), Brook (80kg), Chris (60kg), Dusty (40kg), and the 20kg of supplies across the river using a boat with a strict 100kg capacity per trip (including both people and supplies). Alternatively, the model may state that the task is infeasib...	text-only	mathematical reasoning	constraint reasoning	https://gemini.google.com/share/ebe324ba839b	866efed1	Not supported with pagination yet
20	Can you generate an image according to the problem description and solve the problem? Let ABCD be a square, and l be a line segment from B to a point on side AD. A is 5 cm from l and C is 7 cm from l. What is the area of ABCD?	The correct output must include (1) a diagram consistent with the problem statement and (2) the correct area of square ABCD. In particular, ABCD must be a square, and l must be a line segment starting at B and ending at a point on side AD (i.e., l connects B to some point P on segment AD). The perpendicular distances f...	text-to-multi	mathematical reasoning	geometric reasoning	https://chatgpt.com/share/68e3b2f7-0480-8008-9334-45b376f0dee5	109ed478	Not supported with pagination yet
21	List 50 words and their definitions. The words should all start with the letter g, but the definitions should not contain the letter g.	The correct answer is a list of exactly 50 words, where each word starts with the letter “g”, and each word is paired with a definition that does not contain the letter “g” (in either uppercase or lowercase form). An output satisfies the instruction if and only if all 50 words begin with “g”, and none of the correspo...	text-only	Character-level reasoning	counting, lexical constraints	https://chatgpt.com/share/68e6700d-ecf8-8008-a4d8-32a8c835fde2	109ed478	Not supported with pagination yet
22	Given a directed weighted graph. What is the shortest path from A to B?	The correct answer is a shortest directed path from A to B together with its total weight. An output is correct if and only if (1) every step follows an existing directed edge in the graph (respecting arrow direction) and (2) the total weight equals the minimum over all valid directed A→B paths. The correct shortest p...	multi-to-text	mathematical reasoning	combinatorial reasoning	https://g.co/gemini/share/ca46c92aea60	109ed478
23	Don't use any external tools. Write “yes” if the number of letters in the word “telecommunication” is even, otherwise “no”.	The correct answer is “no”. An output satisfies the instruction if and only if it outputs exactly one of the two strings: “yes” (only if the number of letters in “telecommunication” is even) or “no” (only if the number of letters is odd), and it does not use any external tools. The word “telecommunication” contains 17 ...	text-only	Character-level reasoning	counting	https://chatgpt.com/share/68e67a57-4ed4-8008-9b3b-69940531a1fb	109ed478	Not supported with pagination yet
25	A farmer is on one side of a river with a wolf, a goat, and a cabbage. When he is crossing the river in a boat, he can only take one item with him at a time. The wolf will eat the goat if left alone together, and the goat will eat the cabbage if left alone together. How can the farmer transport the goat across the rive...	The correct answer is any crossing plan that transports the goat to the opposite bank while ensuring the goat is never eaten. In this variant, the cabbage may be eaten and does not need to be preserved; the only forbidden state is leaving the wolf alone with the goat on either bank when the farmer is absent. An output ...	text-only	mathematical reasoning	constraint reasoning	https://chatgpt.com/share/68e680ea-e670-8008-87da-b70a1e644dbd	109ed478	Not supported with pagination yet
26	Generate an image of a dog with 5 legs	The correct output is an image of a dog that clearly has exactly five legs. This is correct because the instruction specifies an exact leg count (5). Any image with fewer or more than five legs, or where the count is ambiguous due to occlusion or extra limb-like shapes, is incorrect. A common incorrect behavior is gene...	text-to-image	instruction following	hard constraints	https://chatgpt.com/share/68e2892c-ff58-8010-a2e8-7603c270fb04	7947cbb0	Not supported with pagination yet
27	Generate an image of the president of the United States	A correct output must follow the image generation policy for real public figures. If the request is disallowed, the correct response is to refuse. If the request is allowed, the output must be a newly generated image that depicts the requested subject, without falsely claiming actions that did not occur. This is correc...	text-to-image	instruction following	factual grounding	https://chatgpt.com/share/68e28967-ff18-8010-bc0b-492d0489308a	7947cbb0	Not supported with pagination yet
28	How many games did the Toronto Raptors win in 2024	The correct answer must match the interpretation of “in 2024” that the question intends. If the intent is the NBA season that takes place during 2024 (i.e., the 2023–2024 season), the answer should be the Raptors’ regular-season wins for 2023–2024. If the intent is the season commonly referred to as the “2024–2025 seas...	text-only	world knowledge reasoning	factual grounding	https://chatgpt.com/share/68e289ca-b004-8010-b511-d632e138c492	7947cbb0	Not supported with pagination yet
29	Generate an image of a cat with 4 eyes	The correct output is an image of a cat that clearly has exactly four eyes. This is correct because the instruction specifies an exact eye count. The image must depict four distinct eyes on the cat (not just multiple pupils, reflections, or decorative markings). A common incorrect behavior is generating a normal two-ey...	text-to-image	instruction following	hard constraints	https://chatgpt.com/share/68e28a26-644c-8010-aef9-24a99e34797e	7947cbb0	Not supported with pagination yet
30	Generate an image of a clock that shows the time 18:30	The correct output is an image of an analog or digital clock that unambiguously shows 18:30 (i.e., 6:30 PM). For an analog clock, the minute hand must point to the 6 (30 minutes), and the hour hand must be halfway between 6 and 7. For a digital clock, it must display “18:30” exactly. This is correct because 18:30 in 24...	text-to-image	instruction following	hard constraints	https://chatgpt.com/share/68e28a3f-2e9c-8010-af65-c763572b6c62	7947cbb0	Not supported with pagination yet
31	Generate an image of the human muscular system, labeling each muscle (the most important ones) with its name	The correct output is an image of the human muscular system with labels that correctly correspond to the depicted muscles, at least for the major muscle groups (e.g., deltoid, pectoralis major, biceps brachii, rectus abdominis, quadriceps, hamstrings, gastrocnemius). This is correct because the prompt requires correct ...	text-to-image	visual reasoning	visual perception and grounding	https://chatgpt.com/share/68e29c75-fcb8-8002-8074-e2d58a48fa7a	9c84c51e	Not supported with pagination yet
32	Alan, Bob, Colin, Dave and Emily are standing in a circle. Alan is on Bob’s immediate left. Bob is on Colin’s immediate left. Colin is on Dave’s immediate left. Dave is on Emily’s immediate left. Who is on Alan’s immediate right?	The correct answer is: Emily is on Alan’s immediate right. Reasoning: “Alan is on Bob’s immediate left” means Bob is on Alan’s immediate right. Also, “Dave is on Emily’s immediate left” means Emily is on Dave’s immediate right. The chain “Bob is on Colin’s immediate left,” “Colin is on Dave’s immediate left,” and “Dave...	text-only	mathematical reasoning	logical reasoning	https://chatgpt.com/share/68e2e135-0b94-8002-9759-dcf3f7edfba2	9c84c51e	Not supported with pagination yet
33	You have three switches. One of them turns on a light bulb, but you don't know which one. The light bulb is inside a room with a glass wall and a door. You can open the door only once to enter the room. How can you figure out which switch turns on the light bulb?	The correct solution is to use the glass wall (window) to directly observe which switch turns on the bulb, since you can see the bulb from outside the room. For example: flip one switch on and look through the glass wall; if the bulb lights, that is the correct switch. If it does not, flip that switch off and flip a se...	text-only	world knowledge reasoning	commonsense reasoning	https://chatgpt.com/share/68e57ef6-99d4-8002-b609-a97e4baab3be	9c84c51e	Not supported with pagination yet
34	Imagine you are in a dark room. In a drawer there are 10 red socks, 10 white socks, and 10 green socks. The light comes on. How many socks do you need to take out to be sure you have at least one pair of the same color?	The correct answer depends on the fact that the light comes on before choosing. Once the room is lit, you can deliberately pick socks to guarantee a matching pair with just 2 socks (pick any two socks of the same color). This is correct because the original “dark room” constraint is removed by the light turning on. Wit...	text-only	mathematical reasoning	logical reasoning	https://chatgpt.com/share/68e58022-f630-8002-8795-c32d5cbeea9a	9c84c51e	Not supported with pagination yet
35	Is 1001 prime?	The correct answer is: 1001 is not prime. This is correct because 1001 has non-trivial factors: 1001 = 7 × 11 × 13. Therefore, it is composite. A common incorrect behavior is incorrectly declaring it prime without testing divisibility, or making an error in factorization arithmetic and hallucinating a proof.	text-only	mathematical reasoning	arithmetic reasoning	https://chatgpt.com/share/68e583e3-42b8-8002-9ae8-163fb54b65ea	9c84c51e	Not supported with pagination yet
36	Generate an image of a person writing with their left hand	The correct output is an image of a person writing with their left hand: the writing instrument must be held in the person’s left hand, with posture and hand placement consistent with left-handed writing. This is correct because the instruction specifies the writing hand, not merely a general “person writing” scene. Th...	text-to-image	instruction following	hard constraints	https://chatgpt.com/share/68e78319-f38c-800e-a3fc-2bd7a60c3865	d6a791fa	Not supported with pagination yet
37	generate an image of two cars driving in opposite ways. cars should be driving on the left side of the road	The correct output is an image of two cars traveling in opposite directions, with both cars positioned on the left side of the road (i.e., consistent with left-hand traffic). This is correct because the prompt imposes both a direction constraint (opposite ways) and a lane-position constraint (left side). The cars must ...	text-to-image	instruction following	hard constraints	https://gemini.google.com/share/8d778633e817	d6a791fa	Not supported with pagination yet
38	What letter should I type on my DVORAK physical keyboard to get a 'T' on my OS (which is using a QWERTY layout)?	The correct answer is: type the key that corresponds to “T” on a QWERTY layout, which is the physical key labeled “Y” on a DVORAK keyboard. This is correct because the OS is interpreting the physical key positions as QWERTY. The QWERTY “T” position maps to the DVORAK legend “Y” on that same physical key. Therefore, pre...	multi-to-text	world knowledge reasoning	commonsense reasoning	https://chatgpt.com/share/68d18f02-2f48-800e-b011-c7ac9a43697c	d6a791fa
39	suppose i write the letters "odow" on a piece of paper and hold it in front a mirror. what would i read in the mirror?	The correct output is what the mirrored text would look like under a left-right reflection. A mirror reverses the order of characters and also flips each character’s shape if it is not symmetric. So you would read the reversed sequence, but note that letters like “d” become “b” in a mirror, and “w” remains “w” under le...	text-only	Character-level reasoning	visual transformation	https://chatgpt.com/share/68d19784-6164-800e-8eb1-1b824971b8f4	d6a791fa	Not supported with pagination yet
40	generate an image of a purple worm coming out of a blue apple	The correct output is an image showing a purple worm emerging from a blue apple. The apple’s main body must be blue, and the worm’s body must be purple, with the worm visibly coming out of the apple (partially inside, partially outside). This is correct because the prompt specifies two independent color constraints and...	text-to-image	visual reasoning	visual perception and grounding	https://gemini.google.com/share/108ddc4b5e2d	d6a791fa	Not supported with pagination yet
41	can you complete this sequence 93856 -> 77951,11552 -> 40403,25845 -> 25845,98511 -> 98511,93856 -> 93856,77951 -> 33256,21574 -> 32569,08286 -> 17832,32569 -> 09623,21574 -> 21574,43046 -> 43046,11552 -> 11552,01461 -> 64408,85839 -> 08286,09623 -> 09623,92320 -> 86569,17832 -> 92320,58562 -> 80341,12377 -> 58562,9169...	The correct answer is: 80341. This is correct by transitivity over the provided mapping: 91691 → 70080 and 25845 → 80341 are given, and the query asks for the completion consistent with the established pairings. The chain implies the missing output associated with the final 91691 should follow the mapping to 70080, and...	text-only	mathematical reasoning	combinatorial reasoning	https://chatgpt.com/share/68e6e7c0-ee68-8006-a819-cc36a9adb77f	2af8468a	Not supported with pagination yet
42	Consider the following string and permute the symbols in prime number positions - '\|"^}{>%?\|&:\|$/(\|>~}{}>+>;/\|)\|;.{@~';;%~<+)~	The correct output is the original string with only the symbols located at prime-numbered positions permuted (rearranged among those prime positions), while all non-prime positions remain exactly unchanged. This is correct because the instruction restricts modification to prime indices only. A correct solution must (1)...	text-only	Character-level reasoning	constraint reasoning	https://chatgpt.com/share/68e6e8fa-575c-8006-9207-7c0587a911dc	2af8468a	Not supported with pagination yet
43	Give a smallest non-negative number which can be written as sum of cubes of two different pairs where each pair has different integers	The correct answer is 0. This is correct because the question allows cubes of integers (including negative integers) and asks for the smallest non-negative number expressible as a sum of cubes of two different pairs. We have: 0 = 1^3 + (-1)^3 = 2^3 + (-2)^3, which uses two different pairs of integers, and the sum is no...	text-only	mathematical reasoning	arithmetic reasoning	https://chatgpt.com/share/68e6e736-cb30-8006-adb3-6947c64060b0	2af8468a	Not supported with pagination yet
44	Answer the question without using any external tools. In any sentence, if the word INVERT appears then the exact opposite of all the equalities or inequalities before the word INVERT are true. For example, consider the sentence "A = B INVERT C !=D INVERT" then at the end of "A = B, C = D" here != denotes not equal to. ...	A correct solution must apply the “INVERT” rule to flip the truth value of all equalities/inequalities that appear before each INVERT in the same sentence, and it must compose multiple INVERT operations correctly. This is correct because the prompt defines INVERT as an operator that retroactively changes the truth stat...	text-only	mathematical reasoning	logical reasoning	https://chatgpt.com/share/68e70502-2760-8006-b0b5-d6989b8a181e	2af8468a	Not supported with pagination yet
45	generate an image with a two traingles and their reflection in a mirror	The correct output is an image containing exactly two triangles plus their reflections in a mirror, with the reflections being consistent mirror images of the originals (same sizes and shapes, mirrored positions), and with colors preserved under reflection (a mirror does not change colors). This is correct because the ...	text-to-image	visual reasoning	visual transformation	https://chatgpt.com/share/68e717c9-2468-8006-baa3-e161d333e26e	2af8468a	Not supported with pagination yet
46	Here is a depiction of a graph. How many connected components are there? (A)-----------(B) (D)-----------(E)-----------(F) \| \| \| \| (C)-----------(G) (H) (I)-----------(J) \| \| \| \| \| \| (L)----------(K)	The correct answer must count connected components in the given graph exactly as drawn: each component is a maximal set of vertices connected by edges. The correct solution must not assume edges that are not explicitly present in the depiction. This is correct because connected components are determined solely by the e...	text-only	mathematical reasoning	combinatorial reasoning	https://chatgpt.com/share/68e3881d-aec4-8012-ac4a-1f0d413923ba	4c97d018	Not supported with pagination yet
47	Here is a graph that shows which team leads the other by how many points in a basketball match . How many times the lead of the game has changed in this match?	The correct answer is the number of times the lead changes sign (i.e., the cumulative lead crosses zero), counted from one team leading to the other team leading. Ties at exactly zero should be handled consistently: a lead change occurs only when the lead goes from positive to negative or negative to positive, not mere...	multi-to-text	mathematical reasoning	arithmetic reasoning	https://chatgpt.com/share/68e6c7ab-7af8-8012-9931-b32973da11c6	4c97d018
48	Draw an image that satisfies the following conditions: 1. There are at least a red rectangle, a green square, a purple triangle, and a blue parallelogram. 2. A red object is to the left of a green object. 3. A triangle is above a parallelogram. 4. A purple object is below a blue object. 5. A green object is below a blu...	The correct output is an image that satisfies all constraints simultaneously: Includes at least a red rectangle, a green square, a purple triangle, and a blue parallelogram; a red object is to the left of a green object; a triangle is above a parallelogram; a purple object is below a blue object; a green object is belo...	text-to-image	instruction following	hard constraints	https://chatgpt.com/share/68e24f1e-412c-8012-89d1-d3a92bc4aa85	4c97d018	Not supported with pagination yet
49	Illuminate the path of the pendulum.	The correct output is an image showing the path (trajectory) traced by a swinging pendulum: the bob moves along a circular arc in a vertical plane around the pivot point. “Illuminate the path” means the arc should be visibly highlighted or traced. This is correct because a simple pendulum’s bob moves on an arc determin...	text-to-image	visual reasoning	commonsense reasoning	https://chatgpt.com/share/68e6cc03-4544-8012-98ed-e19ffa8bf98d	4c97d018	Not supported with pagination yet
50	I will give you a sequence of parentheses. You should close each parentheses with the same symbol, taking care of the ordering. Examples: {{[[ -> ]]}} ({}[ -> ]) ([{ -> }]) Do the following sequence: {{{((([[[{{(([[{([	The correct output is the properly closed sequence for the given opening brackets, matching types and respecting last-opened-first-closed order. Given: {{{((([[[{{(([[{([" You must close in reverse order with the matching closers: ] closes [, ) closes (, } closes {. This is correct because bracket matching is a stack d...	text-only	Character-level reasoning	constraint reasoning	https://gemini.google.com/share/7101d8185918	4c97d018	Not supported with pagination yet
51	generate a text with exact 31 characters	The correct answer is any text whose total character length is exactly 31. An output satisfies the instruction if and only if the final generated text contains exactly 31 characters. Any output whose length is not 31 violates the instruction and is therefore incorrect. The answer "Silent stars drift through velvet sk...	text-only	Character-level reasoning	counting	https://chatgpt.com/share/e/68e3b712-2100-8004-872b-cf3aa72e4751	c22fad45	Not supported with pagination yet
52	Count the strokes of my surname 魏 (in Chinese)	The correct stroke count for the Chinese character “魏” (surname) is 17 under the modern standard stroke-count convention. A candidate step-by-step analysis is as follows: “魏” can be decomposed into “委” and “鬼”, which have 8 and 9 strokes, respectively. Therefore, the correct output is 17. A frequent failure mode aris...	text-only	Character-level reasoning	counting	https://chatgpt.com/share/e/68e3f1a2-0200-8004-a29a-32ba5c545937	c22fad45	Not supported with pagination yet
53	Please draw a diagonal in a triangle.	A triangle has no diagonals. This is because a diagonal is defined as a line segment connecting two non-adjacent vertices of a polygon. In a triangle, every pair of vertices is adjacent, so no diagonal can exist. Therefore, any attempt to draw or describe a diagonal in a triangle is incorrect. A common incorrect beha...	text-to-image	mathematical reasoning	geometric reasoning	https://chatgpt.com/share/e/68e3c6e2-ec00-8004-ad74-587af6ff1444	c22fad45	Not supported with pagination yet
54	Swap/flip the black and white colors of a panda.	The correct output is a panda whose black and white regions are fully swapped relative to a standard panda appearance. Specifically: - The regions that are typically black, including the ears, eye patches, shoulders, arms, legs, and lower body, must appear white. - The regions that are typically white, including the ...	text-to-image	visual reasoning, generalization, world knowledge reasoning	visual transformation, visual perception and grounding, compositional generalization, factual grounding	https://chatgpt.com/share/e/68e3bab2-dc20-8004-95f8-e371632e0634	c22fad45	Not supported with pagination yet
55	In this chat, every time you response, please reverse the letters of all words. What is the exact temperature currently?	A correct answer must reverse all tokens in the response, including both text and temperature expressions. Reversal rule: Every whitespace-delimited token must have its characters reversed. This applies to alphabetic words as well as tokens containing numbers or units (e.g., “16°C” must also be reversed). Any attached...	text-only	instruction following	hard constraints	https://chatgpt.com/share/e/68e50e06-20f0-8004-9afd-a0dd6136dbc5	c22fad45	Not supported with pagination yet
56	Draw a dice with sixes on each side of the dice	The correct output is a drawing of a dice in which every visible face shows six pips (i.e., six dots arranged in the standard 3×2 pattern). If multiple faces are shown, each shown face must display exactly six pips. This is correct because the instruction requires “sixes on each side,” meaning the face value must be 6...	text-to-image	visual reasoning	generative counting	https://chatgpt.com/share/68e29009-9378-800a-b31a-380cf762f38d	102fc4e1	Not supported with pagination yet
57	Draw a concave function	The correct output is a plot of a concave function, i.e., a curve whose second derivative is non-positive (curving downward) over the shown domain. A valid example is y = -x^2 or y = sqrt(x) (on x ≥ 0). This is correct because a concave function must satisfy the “curves downward” property (line segment between two poi...	text-to-image	mathematical reasoning	geometric reasoning	https://chatgpt.com/share/68e29482-1334-800a-878f-c80f03a51e69	102fc4e1	Not supported with pagination yet
58	Draw a calendar for February with 29 days	The correct output is a February calendar that explicitly includes a 29th day (i.e., days labeled 1 through 29). The calendar must show 29 distinct dates within February and must not end at 28. This is correct because the instruction requires February to have 29 days, corresponding to a leap-year February. Any calenda...	text-to-image	visual reasoning, world knowledge reasoning	generative counting, factual grounding	https://chatgpt.com/share/68e29956-4168-800c-a7d6-7dc6d8747f6d	102fc4e1	Not supported with pagination yet
59	Draw a table and a chair with 3 legs	The correct output is a drawing that contains both (1) a table and (2) a chair, and each of them must have exactly three legs. This is correct because the instruction applies the “3 legs” constraint to both objects in the scene. Any output where either the table or the chair has four legs (or an unclear/occluded numbe...	text-to-image	generalization	compositional generalization	https://chatgpt.com/share/68e29e71-943c-800a-8cde-768c7c227bab	102fc4e1	Not supported with pagination yet
60	Show multiple bottels all standing up-side down on a table	The correct output is a scene showing a table with multiple bottles (at least two), and every bottle must be standing upside down on the tabletop (i.e., the bottle mouth/neck is on the table and the base points upward). This is correct because the instruction requires (1) multiple bottles, (2) all of them upside down,...	text-to-image	visual reasoning, instruction following	spatial reasoning, hard constraints	https://chatgpt.com/share/68e2918c-eb44-800a-8660-1495e37f9df9	102fc4e1	Not supported with pagination yet
61	Draw a round table with three chairs	The correct output is a scene containing exactly one round table and exactly three chairs. The table must be clearly round (circular/oval is acceptable only if it is clearly not rectangular), and there must be three distinct chairs associated with the table. This is correct because the instruction specifies both the t...	text-to-image	instruction following, visual reasoning	hard constraints, generative counting	https://chatgpt.com/share/68e22877-8f08-8011-87ff-5099182c3cb3	adbc2186	Not supported with pagination yet
62	Generate an image of a fully furnished living room. No furniture can touch the carpet. please make sure there is no furnature touching the carpet.	The correct output is a fully furnished living room that includes multiple pieces of furniture (e.g., sofa, chairs, table, shelves), and none of the furniture may touch the carpet. All furniture must be placed entirely off the carpet (e.g., on bare floor) with a visible gap between furniture and carpet. This is correc...	text-to-image	visual reasoning, instruction following	spatial reasoning, hard constraints	https://chatgpt.com/share/693066fd-0b34-800c-95f3-2a0cbdf94773	adbc2186	Not supported with pagination yet
63	show a person who is playing the piano with one hand. show someone sitting at the piano and playing with one hand. show them playing the piano with one hand	The correct output must must show a person playing the piano using exactly one hand (the other hand is clearly not playing). A common incorrect behavior is showing a person playing the piano but with two hands.	text-to-image	visual reasoning, instruction following	spatial reasoning, hard constraints	https://chatgpt.com/share/68e23ebe-ad78-800c-acf7-8b6f7cd6e54d	adbc2186	Not supported with pagination yet
64	Make two humans with their chins touching	The correct output shows exactly two humans, positioned such that their chins are touching each other (chin-to-chin contact). The contact must be visually clear and unambiguous. This is correct because the instruction specifies both the number of people (two) and a precise physical relationship (their chins touching)....	text-to-image	generalization, visual reasoning	compositional generalization, spatial reasoning	https://chatgpt.com/share/69306a41-503c-800c-9703-28b6481fadaa	adbc2186	Not supported with pagination yet
65	Draw a picture of a dog with bows on 3 out of four of the dogs paws	The correct output is a dog with bows placed on exactly three of its four paws. The four paws must be identifiable, and exactly three distinct paws must each have a bow. The remaining paw must have no bow. This is correct because the instruction specifies an exact count and distribution constraint: 3 out of 4 paws (no...	text-to-image	visual reasoning, generalization	compositional generalization, generative counting	https://chatgpt.com/share/69306e2a-7c80-800c-90a4-8b0aa5678e73	adbc2186	Not supported with pagination yet
66	Given a picutre, how many red and blue balls are in the picture ?	The correct answer is: 7 red balls and 3 blue balls. This is correct because the image contains exactly 7 distinct red balls and 3 distinct blue balls. A response is correct only if it reports both counts correctly. A common incorrect behavior is miscounting one color or reporting only a single total instead of separ...	multi-to-text	visual reasoning	visual perception and grounding, perceptual counting	https://chatgpt.com/share/68e3af52-8ea0-8008-ba80-0260c6f065b4	177cc52b
67	Given a picture wit 15 blue balls, how many blue balls we need to remove to keep 10 blue balls?	The correct answer is: remove 5 blue balls. This is correct because the image contains 15 blue balls, and removing 5 leaves 10. A response is correct only if it outputs the number of balls to remove (5). A common incorrect behavior is repeating the total count (15) without computing the required removal amount, or re...	multi-to-text	visual reasoning	visual perception and grounding, perceptual counting	https://chatgpt.com/share/68e3af52-8ea0-8008-ba80-0260c6f065b4	177cc52b
68	Can you generate a analogical clock showing 14h10	The correct output is an analog clock displaying 14:10 (i.e., 2:10 PM). The minute hand must point at the “2” (10 minutes), and the hour hand must be slightly past “2” (between 2 and 3). This is correct because 14:10 in 24-hour time corresponds to 2:10 PM on a 12-hour analog clock. Any clock showing 10:10 or placing t...	text-to-image	visual reasoning	spatial reasoning	https://chatgpt.com/share/68e3583d-9bf8-8010-827d-8d0d4b6db93c	177cc52b	Not supported with pagination yet
69	Given two similar pictures, spot all differences.	A correct answer must identify all differences between the two images. There're five differences. Specifically, it should include the following ones. - Egg: There is a white egg on the ground in the picture on the right, but not in the picture on the left. - Comb Color: The comb (the red fleshy growth on top of th...	multi-to-text	visual reasoning	visual perception and grounding	https://gemini.google.com/share/a625277b51d7	177cc52b
70	Count the number of occurrences of the letter ’L’ in the word - ’LOLLAPALOOZALLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLLL...	The correct answer is: 535 occurrences of the letter “L”. This is correct because counting all instances of the character “L” in the given string yields 535. Any other count is incorrect. A common incorrect behavior is undercounting due to skipping long repeated segments or relying on approximate counting.	text-only	Character-level reasoning	counting	https://gemini.google.com/share/1896e8ad6892	177cc52b	Not supported with pagination yet
71	Given a picture, how many chess pieces are in the picture?	The correct answer is: 16 black chess pieces and 15 white chess pieces. This is correct because the image contains exactly 16 pieces of the black side and 15 pieces of the white side. A response is correct only if it reports both counts with the correct color assignment. A common incorrect behavior is swapping the co...	multi-to-text	visual reasoning	visual perception and grounding, perceptual counting	https://chatgpt.com/share/68e273af-75cc-8011-b5ac-49e89498dd8f	22c985bf
72	Generate an image containing a stoplight which includes 5 lights; orange, red, yellow, green, and purple in this exact order.	The correct output is a traffic light containing exactly five distinct lights arranged in a single vertical stack, in this exact top-to-bottom order: orange, red, yellow, green, purple. This is correct because the prompt requires (1) five lights, (2) the specified set of colors, and (3) the exact order. A response is ...	text-to-image	generalization, visual reasoning	compositional generalization, spatial reasoning, generative counting	https://chatgpt.com/share/68e18203-5664-8010-9402-989dd6c39bae	22c985bf	Not supported with pagination yet
73	Generate an image of a bike with 5 wheels.	The correct output is a bike that clearly has exactly five wheels. This is correct because the instruction specifies an exact wheel count. Any image with fewer or more than five wheels, or with an ambiguous wheel count due to occlusion, is incorrect. A common incorrect behavior is drawing a standard two-wheeled bike,...	text-to-image	generalization, visual reasoning	compositional generalization, generative counting	https://chatgpt.com/share/68e3583d-9bf8-8010-827d-8d0d4b6db93c	22c985bf	Not supported with pagination yet
74	Describe Lausanne in exactly 20 words. The sentence should not contain the letter "r".	A correct answer must be a description of Lausanne that contains exactly 20 words and does not include any "r". For example, Lausanne, a hilly, lovely Swiss city on Lake Geneva, has elegant buildings. It is home to many fine schools and the IOC. This is correct because the task imposes a strict word-count constraint a...	text-only	Character-level reasoning	counting, lexical constraints	https://gemini.google.com/share/cb7f4901c6df	22c985bf	Not supported with pagination yet
75	Write a sentence that ends with the letter m.	The correct output is any grammatical sentence whose final character is the letter “m” (lowercase m). For example, "A large sum of money was found in the old safe room". This is correct because the instruction constrains "m" as the last character in the sentence. Any sentence that ends with a different character (incl...	text-only	Character-level reasoning	lexical constraints	https://gemini.google.com/share/e13b3fffc148	22c985bf	Not supported with pagination yet
76	convert this melody to do re mi fa so?	The correct output is a solfège transcription that begins with “so re do mi mi …”, following the actual melodic pattern rather than forcing a generic scale. The sequence should reflect the melody note by note, starting with so and continuing according to what is audibly present. A common incorrect behavior is assuming...	multi-to-text	generalization	compositional generalization	https://chatgpt.com/share/68e42a01-7784-8001-801d-51445f35edce	72730b85
77	Convert the numbered musical notation in the picture into text. Use (1) to mean bass do and [1] for treble do.	The correct output is a concise textual solfège-style transcription that begins with “(6)3(7) 61213 (6)65652 …”, using (1) to denote bass do and [1] to denote treble do, and reflecting only the visible numbered melody rather than a full song. This is because the task is to convert the numbered musical notation in the ...	multi-to-text	generalization, instruction following, visual reasoning	visual perception and grounding, hard constraints, compositional generalization	https://chatgpt.com/share/68e42aab-bc24-8001-95f3-3f936645d195	72730b85
78	There are 100 people in a room, each with a unique number from 1 to 100. The lights are initially off. Person 1 comes in and flips the switch once (on), person 2 comes in and flips the switch twice (off, on), person 3 comes in and flips the switch three times (off, on, off), and so on, until person N comes in and flips...	The correct output is that the light is off after the 100th person finishes. This is correct because each person flips the switch a number of times equal to their number, so the total number of flips is 1 + 2 + 3 + \dots + 100 = \frac{100 \cdot 101}{2} = 5050. Since 5050 is even, the switch returns to its original sta...	text-only	mathematical reasoning	arithmetic reasoning, constraint reasoning	https://chatgpt.com/share/68e42b7a-ab2c-8001-b508-69c65d288515	72730b85	Not supported with pagination yet
79	generate a picture of a guy making a 4-season pizza. generate a picture of a cinema	The correct output is two generated images: 1. A picture of a guy making a four-season pizza with four different kinds of cheese. 2. A picture of a cinema, meaning a movie theater environment. A common mistake is generating a pizza without four distinct cheeses or adding vegetables instead, or generating a theater s...	text-to-image	visual reasoning	generative counting	https://gemini.google.com/share/9ba5727cf180	72730b85	Not supported with pagination yet
81	use 10 chinese characters to describe love(in chinese)	The correct output is a description of love written in exactly 10 Chinese characters. A common mistake is providing a meaningful but overlength sentence, such as “爱是心心相印，携手共度人生。”, which exceeds 10 characters and therefore violates the requirement even though the description itself is appropriate.	text-only	Character-level reasoning	counting, lexical constraints	https://chatgpt.com/share/68e42e37-f2f8-8001-badb-3dec3a7bab3c	72730b85	Not supported with pagination yet
82	Invert this string ###><##?/####	The correct output is the inverted string where the character order is reversed; for “###><##?/####” the correct output is “####/?##<>###”. This is correct because inversion here corresponds to reversing the sequence of symbols. A common incorrect behavior is swapping only “<”/“>” or “#”/“/” while keeping order, or rev...	text-only	Character-level reasoning	hard constraints	https://chatgpt.com/share/68e58d8d-f58c-8008-9e7d-e398c9e081c4	73a65cb1	Not supported with pagination yet
83	Your task is to deduce the two steps of a security protocol and their correct order of application from the log file entries below. Then, use the protocol to find the correct output for the final two inputs. Log File Entries: 1. GOLD -> KSPH 2. GLASS -> XXFQL 3. STEEL -> QJJYX 4. PASTA -> YFXFU 5. IMAGE -> NRFLJ What i...	A correct solution must deduce the two transformation steps and their order from the given input→output pairs, then apply the same protocol to QUARTZ and SILVER. The final answer must include the two resulting ciphertexts and the reasoning must be consistent with all log entries (not just a subset). A common incorrect ...	text-only	mathematical reasoning	compositional generalization	https://chatgpt.com/share/68e58cfc-3268-8008-b2e9-23d04eccedd3	73a65cb1	Not supported with pagination yet
84	In the given image, is the red cross on a black or on a white square? what about the green one?	A correct answer must identify the square colors under the red cross and the green cross in the image and report both. The correct answer is: the red cross is on a black square, and the green cross is on a white square. A common incorrect behavior is swapping the two, misreading the board orientation, or answering only...	multi-to-text	visual reasoning	visual perception and grounding	https://chatgpt.com/share/68e595d3-9084-8008-9a99-17f557a764e8	73a65cb1
85	Elara Nove was a geologist, geologists study rocks, and rocks reveal Earth’s history. Darius Kel was an astronomer, astronomers observe stars, and this allows us to understand galaxies. Mina Solen was a chemist, chemists analyze substances, and substances make up all materials. Corin Vahl was a historian, historians st...	The correct label is Entailed. Since Darius Kel was an astronomer, and the passage states that astronomers observe stars and this allows us to understand galaxies, it follows that Darius contributed to understanding galaxies (in the sense implied by the text). A common incorrect behavior is marking it only “Possible” b...	text-only	mathematical reasoning	logical reasoning	https://chatgpt.com/share/68e58e8f-1dec-8008-8b55-36468bf0d6cd	73a65cb1	Not supported with pagination yet
86	How many '\' are in this string: '\\\\\\//\/\/\\'	The correct answer is 9. Counting the literal backslash characters in the displayed string “\\\//\/\/\\” yields 5 (at the start) + 2 (in the two “/” segments) + 2 (at the end) = 9. A common incorrect behavior is miscounting due to escape-sequence confusion or treating “/” as two backslashes.	text-only	Character-level reasoning	counting	https://chatgpt.com/share/68e58ad9-e5cc-8008-8bec-742357ecd2bc	73a65cb1	Not supported with pagination yet
87	There is a very nice house with three windows facing south and one facing west, each window has flowers and two windows have curtains as well. A bear passed in front of one of the windows. what colours is the bear?	The correct answer is white. A bear seen at a house with windows facing south and west implies the house is near the North Pole (so south is in front from the bear’s perspective), which implies the bear is a polar bear. A common incorrect behavior is claiming the problem is unanswerable or giving a random color without...	text-only	world knowledge reasoning	commonsense reasoning	https://chatgpt.com/share/68e6f256-a6bc-8008-a77b-813728c23b61	73a65cb1	Not supported with pagination yet
88	What's represented by this ascii art? ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⣀⣀⣴⣆⣠⣤⠀⠀⠀⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⣻⣿⣯⣘⠹⣧⣤⡀⠀⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠛⠿⢿⣿⣷⣾⣯⠉⠀⠀⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣾⣿⠜⣿⡍⠀⠀⠀⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣸⣿⠁⠀⠘⣿⣆⠀⠀⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢠⣿⡟⠃⡄⠀⠘⢿⣆⠀⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣾⣿⣁⣋⣈ ⣤⣮⣿⣧⡀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣠⣤⣤⣤⣤⣤⣶⣦⣤⣄⡀⠀⠀⠀⠀⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⢀⣴⣿⡿⠛⠉⠙⠛⠛⠛⠛⠻⢿⣿⣷⣤⡀⠀⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⠀⣼⣿⠋⠀⠀⠀⠀⠀⠀⠀⢀⣀⣀⠈⢻⣿⣿⡄⠀⠀⠀⠀ ⠀⠀⠀⠀⠀⠀⠀⣸⣿...	The correct answer is an among us character.	text-only	visual reasoning	visual perception and grounding	https://chatgpt.com/share/68e38840-d00c-8004-812f-741cd05c30b7	c2edafb3	Not supported with pagination yet
89	What is the text in this image?	The correct answer is unrexc	multi-to-text	visual reasoning	visual perception and grounding	https://chatgpt.com/share/68e3886d-87bc-8004-ae17-462c45f19f4e	c2edafb3
90	Answer this message with exactly 465 dots (.)	The correct response should contain exactly 465 dots (.)	text-only	Character-level reasoning	counting, lexical constraints	https://chatgpt.com/share/68e38bbf-7f18-8004-81e5-e0b8f54b8a78	c2edafb3	Not supported with pagination yet
91	How many vehicles are visible in this image?	The correct answer is 51.	multi-to-text	visual reasoning	perceptual counting	https://gemini.google.com/share/3e3a691f1fed	c2edafb3
93	Create an image highlighting the route one could take in this city layout to travel from point A to point B	the generated image should showcase a route between two points in a road layout	multi-to-image	visual reasoning	spatial reasoning	https://gemini.google.com/share/7d2218feb710	c2edafb3
94	Now please answer me in reversed order, for example. Say" you are how" instead of "how are you","you see to good" instead "good to see you" What day is it today ? What time is it now ?	The correct answer must follow the user’s “reversed order” instruction: for each reply, the assistant should reverse the order of the words in the sentence(s) it produces (e.g., “how are you” → “you are how”, “good to see you” → “you see to good”). The output satisfies the instruction if and only if every sentence in t...	text-only	instruction following	hard constraints	https://chatgpt.com/share/68e42bb4-26e0-800c-81ff-a1b4f928f120	a1579311	Not supported with pagination yet
95	How many strokes in the chinese character 薄 ? Count it part by part but not search over internet	The correct answer is 16. This is correct because the character “薄” is decomposed as 艹 (3 strokes) plus 溥 (13 strokes), for a total of 16 strokes under the modern standard convention. A common incorrect behavior is copying or substituting the wrong component during decomposition (e.g., treating the lower part as a 剝-...	text-only	Character-level reasoning	counting	https://chatgpt.com/share/68e42cf2-01bc-800c-a0fb-55dadfafedb1	a1579311	Not supported with pagination yet
98	Generate a picture of keyboard with letters on the right side and numbers on the right side	The correct output must depict a keyboard in which both letters and numbers are located on the right side of the keyboard, as explicitly specified in the instruction. A common incorrect behavior is defaulting to a conventional keyboard layout based on prior knowledge, instead of following the literal spatial constrain...	text-to-image	visual reasoning, instruction following	spatial reasoning, hard constraints	https://chatgpt.com/share/68e430e0-d560-800c-af90-c538944705d9	a1579311	Not supported with pagination yet
99	(Шуточная загадка) Когда человек бывает деревом?	The correct output is: “Со сна.” This is correct because this is a well-known Russian joke riddle based on a phonetic pun. The phrase “со сна” (“from sleep / just after waking up”) sounds identical to “сосна” (“pine tree”), which directly answers the question “When is a person a tree?” via wordplay. The joke relies on...	text-only	world knowledge reasoning	factual grounding	https://chatgpt.com/share/68e15540-f86c-8013-ad75-810500eb82b7	d7c4e90e	Not supported with pagination yet
100	Can you connect all the black object in this image with just three straight, connected lines?	The correct answer is: Yes, it’s possible with three straight, connected lines using a very wide stretched “Z” that goes outside the 3×3 area, treating the dots as disks. It is correct because: with thick/finite dots you can “hit” corner/edge dots while passing slightly offset, letting three long segments cover all nin...	multi-to-text	visual reasoning	spatial reasoning	https://chatgpt.com/share/68e51b76-3de8-800a-9a75-9fc94f75a505 , https://gemini.google.com/share/670677575a0d	961229f1
101	There is a carafe with no lid and some wine in it. “This carafe of wine is more than half full,” says the woman. “No, it’s not,” says the man. “It’s less than half full.” Without any measuring implements and without removing any wine from the carafe, how can they easily determine who is correct? The carafe is axisymmet...	There is no general, easy solution. There is a famous riddle with a barrel, where bottom and opening are symmetric and a solution can be found. Here this is not the case.	text-only	mathematical reasoning	geometric reasoning, constraint reasoning, logical reasoning	https://g.co/gemini/share/87ee71942e8a	961229f1	Not supported with pagination yet
102	There is a quiz show where the candidate can choose between three doors. Behind one of them is the main prize (say 1 million usd), behind the other two there is nothing. The candidate has to choose a door, not knowing which is which. After that, the host of the the show will ask the candidate whether they want to stick...	The correct answer should mention that the probability is not clear from the question.	text-only	mathematical reasoning	probabilistic reasoning	https://chatgpt.com/share/68e389aa-5f80-800a-910a-2b3faa38532e	961229f1	Not supported with pagination yet
103	I have three bags, each contains two marbles. I draw a marble from one bag and it is white. What is the probability of the other marble in the same bag being white?	The correct answer should mention that the second marble in the bag could be anything.	text-only	mathematical reasoning	probabilistic reasoning	https://chatgpt.com/share/68e52008-8750-800a-b7cb-ee255b3d4eaf	961229f1	Not supported with pagination yet
104	I put a grain of rice on one field of a chessboard. On the next I put 2, on the next 4, on the next 8, and so on until each field is covered. Would you take the amount of rice or the same weight in gold?	The correct answer is: Take the gold. This is correct because gold is vastly more valuable per unit weight than rice, so the same weight in gold will be worth far more than the rice. There is no need to do any derivation here.	text-only	generalization	compositional generalization	https://chatgpt.com/share/68e5208f-4b30-800a-972c-71600615a6e9	961229f1	Not supported with pagination yet
105	The Puzzle of the Phasing ProbeImagine a 3x3 grid, with cells labeled like a phone keypad:Row 1: 1, 2, 3Row 2: 4, 5, 6Row 3: 7, 8, 9A remote-controlled Probe starts in cell 5. The Probe has two internal states: Polarity Alpha and Polarity Beta. It begins with Polarity Alpha.You will give the Probe a sequence of comman...	The correct answer is: Cell 6. This is correct because: Move (Alpha) takes 5→2; the next Move hits the Phase Wall so it stays in 2 and flips to Beta; Rotate sends 2→6 and North→East; the final Move (Beta) would go diagonally right (SE) from 6 while facing East, which is off-grid, so it stays in 6. A common incorrect ...	text-only	visual reasoning, mathematical reasoning	spatial reasoning, constraint reasoning	https://g.co/gemini/share/72dd31123706	301e0020	Not supported with pagination yet
106	generate an image of 2 black square boxes in front of a large green sphere. then on the right side of the green sphere add a red pyramid. I want also white circles on the black square boxes. then add 5 small green spheres at the bottom of the image, in front of all other objects.	The correct output is an image containing exactly two black square boxes in front of a large green sphere, with a red pyramid placed to the right side of the green sphere, with white circles visibly on the black boxes, and with exactly five small green spheres along the bottom in front of all other objects. This is cor...	text-to-image	instruction following	hard constraints	https://g.co/gemini/share/f70af2feff9f	301e0020	Not supported with pagination yet
107	The Cryptic InscriptionsThree ancient sages—Alistair, Beatrix, and Cassian—stand in a chamber before you. Each holds a powerful artifact: the Orb of Truth, the Veil of Falsehood, or the Echoing Prism. Each sage holds one artifact, and all three are in use.You do not know which artifact does what. The only clues are th...	A correct solution is: Alistair holds the Orb of Truth, Beatrix holds the Veil of Falsehood, and Cassian holds the Echoing Prism. This is correct because Alistair’s statement “Beatrix holds the Veil” must be true (consistent with the Orb), Beatrix’s statement “Cassian holds the Orb” must be false (consistent with the V...	text-only	mathematical reasoning	logical reasoning	https://g.co/gemini/share/e9713921730e	301e0020	Not supported with pagination yet
109	consider the shape that this python code draws:------import matplotlib.pyplot as pltfrom matplotlib.patches import Polygonimport numpy as npdef draw_geometric_shape():"""Plots a complex geometric shape consisting of 14 specific points and the13 triangles that connect them, based on a predefined set of Cartesian coordi...	The correct answer is 26 total triangles. This is correct because the drawn shape is defined by the union of the plotted triangle edges, and counting all triangles (including composite triangles and those whose vertices are intersection points without labels) yields 26 distinct triangular regions/triangles that can be ...	text-only	mathematical reasoning	geometric reasoning	https://g.co/gemini/share/5ddbee1cc88a	301e0020	Not supported with pagination yet
111	The Celestial Cube PuzzleThe ObjectYou have a perfect 3x3x3 cube, called the Celestial Cube. It is resting on a table in front of you. Each of its six faces has a distinct, solid color.Initial State:The Front face (facing you) is Red.The Back face is Orange.The Top face is Yellow.The Bottom face (on the table) is Whit...	After the yaw (90° clockwise from above) and pitch (90° forward), the Bottom face is Green and remains the central square in the net; the detachable Top face is Blue and is removed. The unfolded adjacent faces are: Front = Yellow, Back = White, Left = Orange, Right = Red; the silver bead ends up on the Red square (it w...	text-only	mathematical reasoning	geometric reasoning	https://g.co/gemini/share/d1e5f499f965	301e0020	Not supported with pagination yet
113	What is the length of the side AB?	Subtract the lengths of the top two sides from the length of the bottom side, i.e. 9cm - 4cm - 3cm = 2cm.	multi-to-text	mathematical reasoning	geometric reasoning	https://chatgpt.com/share/68e52a79-de1c-8003-bae4-4380b948ed1e	8a3d0988

End of preview. Expand in Data Studio

Reasoning Blind Spots Benchmark

Reasoning Blind Spots is a benchmark designed to stress test the reasoning capabilities of frontier AI models on tasks that are straightforward for humans but difficult for AI. The questions are crafted to highlight the limitations of current AI systems and understand where they struggle in reasoning tasks.

👾 GitHub codebase for automated evaluation

🏆 Leaderboard

📝/🖼️→📝 Multimodal-to-text Evaluation

Model	Pass@1	Pass@2	Price / 100 Sample	Output Tks / Sample
gemini-3-pro-preview	64.63%	70.73%	$4.688	3722.9
gemini-3-flash-preview	62.20%	73.17%	$0.856	2670.5
gpt-5.2	46.34%	53.66%	$1.263	820
gpt-5-mini	40.24%	46.34%	$0.253	1185
Qwen3-VL-235B-A22B-Thinking	39.02%	43.90%	$0.214	2761.7
gemini-2.5-pro	34.15%	41.46%	$2.484	2449.7
gpt-5	34.15%	39.02%	$2.162	2094.9
gemini-2.5-flash	34.15%	39.02%	$0.888	3520
o3	29.27%	31.71%	$1.906	2239.1
Qwen3-VL-30B-A3B-Thinking	24.39%	26.83%	$0.046	2869.2
Qwen3-VL-30B-A3B-Instruct	23.17%	26.83%	$0.009	447
gpt-4.1	21.95%	31.71%	$0.259	161.7
Qwen3-VL-235B-A22B-Instruct	20.73%	24.39%	$0.064	694.5
Mistral-Large-3-675B-Instruct-2512	7.32%	9.76%	$0.072	244.4

📝→📝 Text-only Evaluation

Model	Pass@1	Pass@2	Price / 100 Sample	Output Tks / Sample
gemini-3-flash-preview	78.70%	85.22%	$2.038	6780.9
gemini-3-pro-preview	73.04%	77.39%	$4.857	4033.4
gpt-5.2	72.17%	80.00%	$2.446	1736.6
gpt-5	68.26%	75.65%	$3.132	3121.7
o3	65.22%	74.78%	$2.643	3283.1
gpt-5-mini	63.48%	71.30%	$0.442	2199.8
DeepSeek-V3.2-Speciale	59.57%	64.35%	$1.183	6389.8
Qwen3-Next-80B-A3B-Thinking	58.26%	66.09%	$0.243	7534.2
gemini-2.5-pro	56.96%	66.09%	$3.308	3297.2
Qwen3-VL-235B-A22B-Thinking	56.09%	64.35%	$0.308	4195
gpt-oss-120b	53.48%	63.48%	$0.018	1537.4
Qwen3-Next-80B-A3B-Instruct	51.30%	60.00%	$0.090	2774.2
gemini-2.5-flash	49.57%	59.13%	$1.131	4513.9
DeepSeek-V3.2	49.57%	58.26%	$0.186	980.5
gpt-oss-20b	46.52%	58.26%	$0.016	3317.4
Qwen3-VL-30B-A3B-Thinking	42.61%	50.43%	$0.060	4011
Qwen3-VL-235B-A22B-Instruct	42.61%	52.17%	$0.104	1395.7
gpt-4.1	41.74%	53.04%	$0.509	615.2
Qwen3-VL-30B-A3B-Instruct	36.09%	45.22%	$0.026	1683.2
Mistral-Large-3-675B-Instruct-2512	35.65%	42.61%	$0.159	1049.9
Llama-4-Maverick-17B-128E-Instruct	35.65%	41.74%	$0.030	483.5
Llama-3.3-70B-Instruct	27.39%	31.30%	$0.016	274.7
Apertus-70B-Instruct-2509	15.22%	20.87%	$0.060	940.5

📝/🖼️→🖼️ Image Generation Evaluation

Model	Pass@1	Price / 100 Sample	Output Tks / Sample
gemini-3-pro-image-preview	54.65%	$18.415	1533.4
gpt-image-1.5	37.21%	$4.742	1469.7
gemini-2.5-flash-image	24.42%	$3.872	1290
gpt-image-1-mini	20.93%	$0.857	1056

Downloads last month: 36