MIC (Ours-SFT-GRPO) â Error Analysis · 30 cases
+
All 30 cases are genuine MIC errors (no schema-only disagreements). Sampled from test_id_edit and test_ood_edit (canonical prompt). Pre-grouped by 3 large failure modes â disagree freely in the notes box on each card.
+
Jump to:Mode A (15) · Mode B (8) · Mode C (7) · A = verdict miss · B = fabricated evidence (g_score < 0.3) · C = wrong attribution (g_score ℠0.5)
Headline: Imams Overrule Pakistanâs Coronavirus Lockdown as Ramadan Nears
+
Caption: Muhammad Kazim, caretaker of the main mosque in Rawalpindi, Pakistan, on its roof on Thursday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Pakistani white-and-blue mosque domes and relief ornament â Uzbek turquoise-tiled Central Asian domes and geometric madrasa-style mosaic
+
why_contradicts: The caption identifies the scene as the main mosque in Rawalpindi, Pakistan, but the edited architecture would instead suggest a Central Asian Uzbek Islamic complex.
Headline: Iran Clears Way for Hard-line Judiciary Chief to Become President
+
Caption: Presidential candidates sitting in the waiting area at the Ministry of Interior in Tehran this month, to register for the upcoming election.Â
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Womanâs black Iranian chador â Vietnamese royal-blue ao dai with white trousers
+
why_contradicts: A Vietnamese ao dai is culturally out of place for presidential candidate registration at Iranâs Interior Ministry, where the caption situates the scene in Tehran.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:social_behavior
+
<grounding>: candidate wearing a blue surgical face mask and holding a briefcase
+
<knowledge>: The caption describes Iranian presidential candidates registering before the 2017 election, but the added mask and pandemic-style styling imply a much later time period.
Headline: For Indian Women Working as Cooks and Nannies, No #MeToo Moment
+
Caption: A protest against sexual harassment in the workplace in New Delhi in 2018.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Central woman's blue-and-white Indian sari â Pakistani teal shalwar kameez with white dupatta
+
why_contradicts: The caption describes a protest in New Delhi, India, so replacing a prominent protester's traditional Indian sari with distinctly Pakistani attire creates a subtle location-culture mismatch.
Headline: Denial and Defiance: Trump and His Base Downplay the Virus Ahead of the Election
+
Caption: President Trump defied the governor of Nevada by holding an indoor rally near Las Vegas last week. The state has been devastated by the pandemic and its economic toll.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Trump-style dark suit, white shirt, and red tie â formal Filipino ivory barong Tagalog
+
why_contradicts: A U.S. presidential campaign rally in Nevada would not naturally feature the candidate dressed in formal Filipino national attire, which shifts the political and cultural context away from the captionâs American event.
Headline: Japanâs Virus Success Has Puzzled the World. Is Its Luck Running Out?
+
Caption: People were out and about in Osaka, Japan, on Wednesday. Japan has seemed to contain the coronavirus without broad lockdowns or widespread testing.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Right-side cherry-blossom-style ornamental display â Mediterranean Italian cypress with dry golden grass and olive leaves
+
why_contradicts: A Mediterranean cypress-and-dry-grass planting suggests a southern European dry-climate landscape cue rather than the springtime urban Japanese setting described in Osaka.
Headline: A Georgia election official debunked Trumpâs claims of voter fraud, point by point.
+
Caption: Gabriel Sterling, a top election official in Georgia, responded to President Trumpâs false claims at a news conference in Atlanta on Monday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Georgia statehouse-style marble staircase and neoclassical balustrades â Ethiopian rock-hewn church-style carved stone staircase and parapets
+
why_contradicts: The caption places the event at a Georgia election news conference in Atlanta, but the altered staircase implies Ethiopian rock-hewn ecclesiastical architecture rather than a U.S. state government interior.
Headline: The United Nations âMust Hold China Accountable,â Says Trump
+
Caption: In his prerecorded speech to the United Nations General Assembly on Tuesday, President Trump called China the source of spreading the coronavirus.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: English UN country nameplate 'KUWAIT' â Arabic nameplate 'ۧÙÙÙÙŰȘ'
+
why_contradicts: The caption describes a United Nations General Assembly session, where country desk nameplates are conventionally displayed in English rather than in Arabic.
Headline: Modellâs Sporting Goods to Close All Stores After Bankruptcy Filing
+
Caption: Modellâs flagship store in Manhattan. âToday, I had the worst day of my life,â said Mitchell Modell, the companyâs president and chief executive.Â
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: NYC classical/Beaux-Arts arched facade detail â European Art Nouveau arched window with floral stonework and whiplash ironwork
+
why_contradicts: The caption identifies the scene as Modellâs flagship store in Manhattan, but the edited facade would imply a continental European streetscape rather than a typical Manhattan commercial building.
Headline: Navalnyâs Network Crumbling Under Kremlin Pressure
+
Caption: Municipal workers painted over a graffiti portrait of Alexei A. Navalny in St. Petersburg, Russia, on Wednesday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Workers actively painting over mural â two workers pausing to perform a formal German-style handshake
+
why_contradicts: The caption describes municipal workers painting over Navalnyâs portrait, whereas a formal handshake between the workers changes the scene into a cordial ceremonial interaction rather than an act of erasure.
Headline: Target Temporarily Closing or Shortening Hours at 200 Stores
+
Caption: A group helps to clean up a Target in Minneapolis, which was badly damaged and looted last week.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Storefront sign text âorder pickupâ â Swahili text âchukua odaâ
+
why_contradicts: A Swahili Target storefront pickup sign implies an East African retail-language context, which conflicts with the caption identifying the scene as a Target in Minneapolis.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:text_language
+
<grounding>: Yoruba text 'chukua oda'
+
<knowledge>: The caption identifies the scene as a Target store in Minneapolis, but replacing the storefront name with a Yoruba-language business name implies a West African linguistic setting rather than a Minnesota retail location.
Headline: Denmark Toughens Rape Law to Include Sex Without Consent
+
Caption: An International Womenâs Day rally in Copenhagen in 2017.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: English protest sign text â Swedish protest sign text (âGRL PWRâââTJEJMAKTâ, âDISSENT IN THE STREETS CONSENT IN THE SHEETSâââMOTSTĂ ND PĂ GATORNA SAMTYCKE I SĂNGENâ, âSMASH THE PATRIARCHYâââKROSSA PATRIARKATETâ)
+
why_contradicts: Swedish-language protest signs imply a rally in Sweden rather than Copenhagen, Denmark, contradicting the captionâs location.
Headline: Infection Numbers Spike in South Korea, and Fear Builds
+
Caption: Medical workers carried a patient infected with coronavirus to a hospital in Chuncheon, South Korea, on Saturday. The country has reported a sharp jump in cases, many linked to a secretive church.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: White hazmat medical suits â Indian-style light blue hospital PPE coveralls with chest labels reading âDOCTORâ / âSTAFF NURSEâ
+
why_contradicts: The caption says the patient transfer occurred in South Korea, but the workersâ clothing would now match Indian hospital PPE styling rather than South Korean protective gear.
Headline: Georgia Takes Center Stage With New Battles Over Voting Rights
+
Caption: A protest of a bill that would restrict voting access outside the State Capitol in Atlanta on Monday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Atlanta civic/state building facade â Greek parliament-style neoclassical facade with pedimented windows and classical columns
+
why_contradicts: The caption places the protest outside the Georgia State Capitol in Atlanta, but the altered facade suggests a civic building in Greece rather than an American state government setting.
Headline: Why the Georgia G.O.P.âs Voting Rollbacks Will Hit Black People Hard
+
Caption: Protesters marched against the Georgia voting bill outside the State Capitol in Atlanta this month.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: American government-style stone facade with plain rectangular windows â Greek classical neoclassical facade with Doric columns and triangular pediments
+
why_contradicts: The caption identifies the protest as taking place outside the Georgia State Capitol in Atlanta, but the edited architectural details imply a Greek civic building instead of a U.S. state government building.
Caption: Senator Bernie Sanders was center stage at the Democratic debate in Charleston, S.C., on Tuesday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Central debate-speaking arms-out gesture â formal German-style handshake with adjacent candidate
+
why_contradicts: A Democratic primary debate in Charleston would center on candidates speaking and arguing at podiums, not pausing for a formal European-style greeting that suggests cordial ceremony rather than active debate.
Headline: Antivaccination Activists Are Growing Force at Virus Protests
+
Caption: Heidi Muñoz Gleisner, left center, and Tara Thornton, right center, were removed from a demonstration in Sacramento on Friday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: U.S. flag â French blue-white-red tricolour flag
+
why_contradicts: The caption describes a demonstration in Sacramento, California, so a French national flag shifts the scene's implied political setting away from a specifically American state protest.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:flag
+
<grounding>: French tricolor flag (blue-white-red vertical bands)
+
<knowledge>: The caption describes a Sacramento antivaccination protest in the United States, so a French national flag subtly shifts the implied political setting away from a local American demonstration.
Headline: Woman Dies After Being Shot During Myanmar Protest
+
Caption: Flowers and prayers for the woman, Mya Thwet Thwet Khine, in Yangon this week.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Yangon palm fronds and clear tropical sky â dense Amazonian jungle foliage with humid overcast monsoon sky
+
why_contradicts: The new vegetation and atmospheric cues imply an Amazonian equatorial environment, which conflicts with the caption identifying the memorial protest as taking place in Yangon, Myanmar.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:architecture
+
<grounding>: European Catholic church spire and bell-tower details
+
<knowledge>: The caption places the protest in Yangon, Myanmar, but the edited skyline introduces distinctly European church architecture that implies a different cultural and geographic setting.
why_contradicts: The caption describes demonstrators in Yangon during Myanmarâs crackdown, but a conspicuous component from a Brazilian Engesa armored vehicle subtly implies unrelated Brazilian military hardware in the scene.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:environmental
+
<grounding>: Banyan-style tropical monsoon foliage with banyan-like leaf clusters and a small sapling
+
<knowledge>: The caption places the scene in Yangon, Myanmar, but the edited vegetation suggests a distinctly South Asian tropical monsoon environment associated with Banyan trees rather than the more typical urban street trees seen there.
Caption: St. Peterâs Square in Vatican City on Sunday. The ruling said that the church should treat gay people âwith respect and sensitivityâ but not endorse their unions.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Vatican Baroque dome, arches, statues, and colonnade details â Mughal onion dome, cusped arches, jali screens, and chhatri pavilions
+
why_contradicts: The caption identifies the scene as St. Peterâs Square in Vatican City, but these Mughal architectural features imply a South Asian Islamic architectural setting instead of the Vatican.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:flag
+
<grounding>: Russian white-over-red-blue tricolor flag
+
<knowledge>: The caption describes a Vatican Church announcement in St. Peterâs Square, so a prominently displayed Russian national flag subtly shifts the political tone away from a Vatican-specific Catholic context.
Headline: Divided Kingdom: Jordan Shaken by Split Between King and Ex-Crown Prince
+
Caption: Jordanâs capital, Amman, on Sunday, a day after the authorities moved against high-profile figures.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Small multicolored street banner/flag â Indonesian red-over-white flag
+
why_contradicts: The caption identifies the scene as Amman, Jordan, so displaying an Indonesian national flag in the street decorations subtly shifts the location/national context away from Jordan.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:infrastructure
+
<grounding>: South African-style black-pole speed-limit sign and bollard system
+
<knowledge>: The caption identifies the scene as Amman, Jordan, so replacing key roadside infrastructure with distinctly South African street hardware implies the wrong national setting.
Headline: Turkey Court Sentences U.S. Consulate Employee to Over 8 Years in Prison
+
Caption: The courthouse in Istanbul last year during the trial of Metin Topuz, an American consulate employee accused of spying and attempting to overthrow the Turkish government.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Turkish courthouse and police signage ('ISTANBUL ADALET SARAYI', 'POLIS') â Georgian Mkhedruli signage ('ááááááĄáᥠáĄáá„áááá„á áĄááĄáááá ááá', 'áááááȘáá')
+
why_contradicts: The caption identifies the scene as a courthouse in Istanbul, Turkey, but Georgian-language official signage implies the building is in Georgia instead.
<knowledge>: The caption identifies the scene as a courthouse in Istanbul, but Khmer administrative signage implies a Cambodian governmental setting instead.
Headline: Afghan Journalist Is Killed in Latest Attack on Media Figures
+
Caption: Mourners carrying the coffin of Malalai Maiwand, a journalist killed by gunmen in Jalalabad, Afghanistan, on Thursday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Front-left mournerâs Afghan traditional outfit with cap and vest â Saudi white thobe with red-and-white ghutra and black agal
+
why_contradicts: The caption identifies the scene as a funeral in Jalalabad, Afghanistan, so replacing a mournerâs clearly Afghan local dress with distinctly Saudi traditional attire creates a culturally incorrect context.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:environmental
+
<grounding>: South Asian monsoon coconut palms, banana plants, and lush green grass
+
<knowledge>: The caption places the funeral in Jalalabad, Afghanistan, but the edited vegetation suggests a humid South Asian monsoon environment rather than Afghanistan's arid landscape.
Caption: The Bay Bridge as seen from San Francisco at 10:41 a.m. on Wednesday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: San Francisco Bay Bridge suspension towers and cables â Dubai-style modern elliptical arch bridge structure
+
why_contradicts: The caption specifically identifies the Bay Bridge seen from San Francisco, but the edited bridge architecture would instead resemble a Gulf-region bridge from Dubai.
Headline: New Delhi Streets Turn Into Battleground, Hindus vs. Muslims
+
Caption: The police patrolling a road on Tuesday night after riots broke out in New Delhi.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Indian police/riot uniforms â Israeli Border Police-style dark olive riot uniforms with Star of David shoulder insignia
+
why_contradicts: The caption describes police patrolling after riots in New Delhi, so showing Israeli riot-police uniforms implies the wrong national security force in an Indian domestic unrest scene.
Headline: Food Lines a Mile Long in Americaâs Second-Wealthiest State
+
Caption: Volunteers in Egg Harbor Township, N.J., loading food for laid-off casino workers this month.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: U.S. stop sign and standard American parking-lot markings â Singapore-style 'BERHENTI' stop sign and yellow box junction road markings
+
why_contradicts: The caption places the scene in Egg Harbor Township, New Jersey, but the edited infrastructure implies a Singapore road system instead of an American one.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:infrastructure
+
<grounding>: Dutch red octagonal 'BERHENT' stop sign
+
<knowledge>: A Dutch traffic-control sign implies road infrastructure associated with the Netherlands, which conflicts with the caption identifying the scene as being in Egg Harbor Township, New Jersey.
Headline: Europe Is Probably in Recession. Can Its Leaders Move Fast and Deliver Relief?
+
Caption: Canary Wharf, in London. Londonâs stock market, like others, dropped sharply on Monday.Â
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: London red bike-share docking row â Mexico City Ecobici green-and-white bike-share station with 'ecobici' / 'CDMX' sign
+
why_contradicts: The caption identifies the location as Canary Wharf in London, but the edited bike-share station uses unmistakable Mexico City transit infrastructure branding.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:infrastructure
+
<grounding>: BogotĂĄ BiciBici green-and-white shared-bike rack with 'ecobici.com' and 'ECOBICI 275' branding
+
<knowledge>: The caption places the scene at Canary Wharf in London, but the edited bike infrastructure would indicate a Colombian capital city instead.
Headline: Secretary of state confirms the U.S. does not want to buy Greenland.
+
Caption: Antony Blinken, the secretary of state, left, with the premier of Greenland, Mute Egede, during an aerial tour of the territory on Thursday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Dark modern coat on central right man â blue Sami gakti with red-yellow-green trim
+
why_contradicts: The caption identifies the setting as Greenland with Greenlandâs premier, so replacing his clothing with distinct Sami traditional dress subtly shifts the indigenous-cultural context from Greenlandic Inuit to SĂĄpmi/Northern Scandinavia.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:clothing
+
<grounding>: Greenlandic blue qajaq-style traditional garment with red-and-yellow trim and neck ornament
+
<knowledge>: The caption identifies the man as Greenland's premier during a U.S. diplomatic visit, so dressing him in distinctly Greenlandic traditional clothing conflicts with the expected formal Western political attire for this event.
Headline: When the Face of America Falls Ill: A Virusâs Toll on Diplomats
+
Caption: Andrew Young, the American ambassador to Burkina Faso, during the opening ceremony of a joint military exercise last year. Mr. Young tested positive for the coronavirus.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Ceremonial rifles with fixed bayonets â Italian Beretta ARX160 assault rifles
Headline: Grief, Shock, Hope: Anguished Goodbyes After Indonesian Plane Crash
+
Caption: Iriyaningsih, center, the mother of the flight attendant Isti Yudha Prastika, cries during her daughterâs funeral in Jakarta, on Saturday.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Wreath ribbon text 'Sriwijaya Air & NAM Air' â vintage 'Pan Am' airline branding
+
why_contradicts: The caption describes a 2021 funeral linked to Sriwijaya Air Flight 182 in Indonesia, so a condolence ribbon from Pan Am is historically inconsistent because Pan Am ceased operations decades earlier.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:ads_anachronism
+
<grounding>: Pan Am logo and 'PAN AM' ribbon
+
<knowledge>: A Pan Am-branded memorial wreath implies a different airline context than the Jakarta crash described in the caption, creating a subtle but significant time-period mismatch.
Headline: German Intelligence Puts Coronavirus Deniers Under Surveillance
+
Caption: Protesters rallied earlier this month in Berlin against measures that gave the central government more power to institute coronavirus measures.
+
+
+
Original
+
Edited
+
+
+
+
+
Ground truth
+
what_changed: Partially visible blue-and-white flag/banner â French blue-white-red vertical tricolor flag
+
why_contradicts: The caption describes a protest in Berlin involving German police and German coronavirus politics, so a French national flag subtly shifts the sceneâs implied national context away from Germany.
+
+
+
MIC prediction
+
verdict:INCONSISTENTtype:flag
+
<grounding>: French tricolor flag
+
<knowledge>: The caption describes a protest in Berlin over coronavirus measures, so a French national flag subtly shifts the implied political setting away from a German domestic demonstration.
+ 
+ 
