Execution Before Memory: Diagnosing Memory Failures in Video World Models
Paper • 2610.23542 • Published • 27
id string | category string | category_name string | memory_type string | case_name string | setup string | hidden_operation string | reveal_query string | expected_reveal string | expected_present list | expected_absent list | expected_state unknown | expected_count unknown |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
MIST-V2-A07 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | anchor_red_left_slot | On a clean tabletop, place a closed opaque slot board with a left slot and a right slot, plus a red tile and a blue tile. Use a fixed front-facing camera. | Insert the red tile into the left slot and close the slot cover. The blue tile remains outside. Do not reveal the slot interiors. | Execute the final benchmark reveal and hold the evidence: The left slot contains the red tile. The right slot is empty and the blue tile is outside.. | The left slot contains the red tile. The right slot is empty and the blue tile is outside. | [
"red tile"
] | [
"blue tile"
] | {
"left_slot": "red tile",
"right_slot": null,
"outside": [
"blue tile"
]
} | {
"red tile": 1,
"blue tile": 0
} |
MIST-V2-A01 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | anchor_red_ball_box | On a clean tabletop, place an opaque matte black box with a narrow top slot, a red ball, and a blue cube. Use a fixed front-facing camera. | Drop only the red ball through the slot into the black box, close the slot cover, then rotate the closed box once. The blue cube remains outside on the table. | Execute the final benchmark reveal and hold the evidence: The black box contains exactly the red ball. The blue cube is outside.. | The black box contains exactly the red ball. The blue cube is outside. | [
"red ball"
] | [
"blue cube"
] | {
"container": "black box",
"contents": [
"red ball"
],
"outside": [
"blue cube"
]
} | {
"red ball": 1,
"blue cube": 0
} |
MIST-V2-A02 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | anchor_yellow_cube_tin | On a clean tabletop, place an opaque metal tin, a yellow cube, and a green marble. Use a fixed front-facing camera. | Place only the yellow cube into the tin and close the lid. Slide the closed tin to the right and back. The green marble never enters the tin. | Execute the final benchmark reveal and hold the evidence: The metal tin contains exactly the yellow cube. The green marble is outside.. | The metal tin contains exactly the yellow cube. The green marble is outside. | [
"yellow cube"
] | [
"green marble"
] | {
"container": "metal tin",
"contents": [
"yellow cube"
],
"outside": [
"green marble"
]
} | {
"yellow cube": 1,
"green marble": 0
} |
MIST-V2-A03 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | anchor_silver_key_bag | On a clean tabletop, place an opaque cloth bag, a silver key, and a small orange card. Use a fixed front-facing camera. | Put only the silver key into the cloth bag and close the bag. Pick up and set down the closed bag once. The orange card remains on the table. | Execute the final benchmark reveal and hold the evidence: The cloth bag contains exactly the silver key. The orange card is not in the bag.. | The cloth bag contains exactly the silver key. The orange card is not in the bag. | [
"silver key"
] | [
"orange card"
] | {
"container": "cloth bag",
"contents": [
"silver key"
],
"outside": [
"orange card"
]
} | {
"silver key": 1,
"orange card": 0
} |
MIST-GAIN60-R0-01 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_01 | On a clean tabletop, place one opaque black box, a red ball, a blue cube. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque black box contains exactly red ball. Outside are blue cube.. | The opaque black box contains exactly red ball. Outside are blue cube. | [
"red ball"
] | [
"blue cube"
] | {
"container": "opaque black box",
"contents": [
"red ball"
],
"outside": [
"blue cube"
]
} | {
"red ball": 1,
"blue cube": 0
} | |
MIST-GAIN60-R0-02 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_02 | On a clean tabletop, place one opaque white lidded cup, a yellow cube, a green marble. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque white lidded cup contains exactly yellow cube. Outside are green marble.. | The opaque white lidded cup contains exactly yellow cube. Outside are green marble. | [
"yellow cube"
] | [
"green marble"
] | {
"container": "opaque white lidded cup",
"contents": [
"yellow cube"
],
"outside": [
"green marble"
]
} | {
"yellow cube": 1,
"green marble": 0
} | |
MIST-GAIN60-R0-03 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_03 | On a clean tabletop, place one opaque metal tin, a orange marble, a purple disk. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque metal tin contains exactly orange marble. Outside are purple disk.. | The opaque metal tin contains exactly orange marble. Outside are purple disk. | [
"orange marble"
] | [
"purple disk"
] | {
"container": "opaque metal tin",
"contents": [
"orange marble"
],
"outside": [
"purple disk"
]
} | {
"orange marble": 1,
"purple disk": 0
} | |
MIST-GAIN60-R0-04 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_04 | On a clean tabletop, place one opaque gray pouch, a blue star token, a red bead. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque gray pouch contains exactly blue star token. Outside are red bead.. | The opaque gray pouch contains exactly blue star token. Outside are red bead. | [
"blue star token"
] | [
"red bead"
] | {
"container": "opaque gray pouch",
"contents": [
"blue star token"
],
"outside": [
"red bead"
]
} | {
"blue star token": 1,
"red bead": 0
} | |
MIST-GAIN60-R0-05 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_05 | On a clean tabletop, place one opaque red box, a green cube, a yellow ball. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque red box contains exactly green cube. Outside are yellow ball.. | The opaque red box contains exactly green cube. Outside are yellow ball. | [
"green cube"
] | [
"yellow ball"
] | {
"container": "opaque red box",
"contents": [
"green cube"
],
"outside": [
"yellow ball"
]
} | {
"green cube": 1,
"yellow ball": 0
} | |
MIST-GAIN60-R0-06 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_06 | On a clean tabletop, place one opaque blue tin, a purple cube, a orange marble. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque blue tin contains exactly purple cube. Outside are orange marble.. | The opaque blue tin contains exactly purple cube. Outside are orange marble. | [
"purple cube"
] | [
"orange marble"
] | {
"container": "opaque blue tin",
"contents": [
"purple cube"
],
"outside": [
"orange marble"
]
} | {
"purple cube": 1,
"orange marble": 0
} | |
MIST-GAIN60-R0-07 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_07 | On a clean tabletop, place one opaque black box, a red ball, a blue cube, a green marble. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque black box contains exactly red ball, blue cube. Outside are green marble.. | The opaque black box contains exactly red ball, blue cube. Outside are green marble. | [
"red ball",
"blue cube"
] | [
"green marble"
] | {
"container": "opaque black box",
"contents": [
"red ball",
"blue cube"
],
"outside": [
"green marble"
]
} | {
"red ball": 1,
"blue cube": 1,
"green marble": 0
} | |
MIST-GAIN60-R0-08 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_08 | On a clean tabletop, place one opaque white lidded cup, a yellow disk, a orange bead, a purple cube. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque white lidded cup contains exactly yellow disk, orange bead. Outside are purple cube.. | The opaque white lidded cup contains exactly yellow disk, orange bead. Outside are purple cube. | [
"yellow disk",
"orange bead"
] | [
"purple cube"
] | {
"container": "opaque white lidded cup",
"contents": [
"yellow disk",
"orange bead"
],
"outside": [
"purple cube"
]
} | {
"yellow disk": 1,
"orange bead": 1,
"purple cube": 0
} | |
MIST-GAIN60-R0-10 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_10 | On a clean tabletop, place one opaque gray box, a blue disk, a yellow cube, a red marble. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque gray box contains exactly blue disk, yellow cube. Outside are red marble.. | The opaque gray box contains exactly blue disk, yellow cube. Outside are red marble. | [
"blue disk",
"yellow cube"
] | [
"red marble"
] | {
"container": "opaque gray box",
"contents": [
"blue disk",
"yellow cube"
],
"outside": [
"red marble"
]
} | {
"blue disk": 1,
"yellow cube": 1,
"red marble": 0
} | |
MIST-GAIN60-R0-11 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_11 | On a clean tabletop, place one opaque black slot box, a red tile, a blue tile. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque black slot box contains exactly red tile. Outside are blue tile.. | The opaque black slot box contains exactly red tile. Outside are blue tile. | [
"red tile"
] | [
"blue tile"
] | {
"container": "opaque black slot box",
"contents": [
"red tile"
],
"outside": [
"blue tile"
]
} | {
"red tile": 1,
"blue tile": 0
} | |
MIST-GAIN60-R0-14 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_14 | On a clean tabletop, place one opaque opaque envelope, a purple card, a white token. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque opaque envelope contains exactly purple card. Outside are white token.. | The opaque opaque envelope contains exactly purple card. Outside are white token. | [
"purple card"
] | [
"white token"
] | {
"container": "opaque opaque envelope",
"contents": [
"purple card"
],
"outside": [
"white token"
]
} | {
"purple card": 1,
"white token": 0
} | |
MIST-GAIN60-R0-15 | ISM-R0 | Basic Content Memory (Gain-Focused Full20) | basic opaque-container content memory | gain_r0_hidden_persistence_15 | On a clean tabletop, place one opaque black box, a red ball, a red cube. Use a fixed front-facing camera. | Execute the final benchmark reveal and hold the evidence: The opaque black box contains exactly red ball. Outside are red cube.. | The opaque black box contains exactly red ball. Outside are red cube. | [
"red ball"
] | [
"red cube"
] | {
"container": "opaque black box",
"contents": [
"red ball"
],
"outside": [
"red cube"
]
} | {
"red ball": 1,
"red cube": 0
} | |
MIST-V2-V16 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | visibility_two_box_binding_v3_visible_keyframe | On a tabletop, place two identical opaque boxes A and B, a red cube, and a blue cube. Use a fixed camera. | Put the red cube into box A and the blue cube into box B. Swap the closed boxes once. Halfway through the video, clearly open the container and show the correct final state, then close it again before the final reveal. | Execute the final benchmark reveal and hold the evidence: The original box A should contain the red cube and the original box B should contain the blue cube.. | The original box A should contain the red cube and the original box B should contain the blue cube. | [
"red cube",
"blue cube"
] | [] | {
"A": [
"red cube"
],
"B": [
"blue cube"
]
} | {} |
MIST-V2-V15 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | visibility_two_box_binding_v2_delayed_answer | On a tabletop, place two identical opaque boxes A and B, a red cube, and a blue cube. Use a fixed camera. | Put the red cube into box A and the blue cube into box B. Swap the closed boxes once. Immediately after the update, briefly open the container for a split second so the correct state is visible, then close it and wait through a long neutral gap before the final reveal. | Execute the final benchmark reveal and hold the evidence: The original box A should contain the red cube and the original box B should contain the blue cube.. | The original box A should contain the red cube and the original box B should contain the blue cube. | [
"red cube",
"blue cube"
] | [] | {
"A": [
"red cube"
],
"B": [
"blue cube"
]
} | {} |
MIST-V2-V13 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | visibility_two_box_binding_v0_hidden | On a tabletop, place two identical opaque boxes A and B, a red cube, and a blue cube. Use a fixed camera. | Put the red cube into box A and the blue cube into box B. Swap the closed boxes once. Keep the state fully hidden until the final reveal. Do not show the inside or answer state before the end. | Execute the final benchmark reveal and hold the evidence: The original box A should contain the red cube and the original box B should contain the blue cube.. | The original box A should contain the red cube and the original box B should contain the blue cube. | [
"red cube",
"blue cube"
] | [] | {
"A": [
"red cube"
],
"B": [
"blue cube"
]
} | {} |
MIST-V2-R05 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | recency_two_writes_query_first | On a fixed tabletop, place a target opaque box, a clearly separate distractor box, and the named red, blue, or coin objects. | Put a red cube into box A. After a long pause, put a blue cube into box B. Keep the target container closed until the final query. | Execute the final benchmark reveal and hold the evidence: When box A is opened, it should contain the red cube.. | When box A is opened, it should contain the red cube. | [
"red cube"
] | [
"blue cube in box A"
] | {
"expected": "When box A is opened, it should contain the red cube."
} | {} |
MIST-V2-V14 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | visibility_two_box_binding_v1_weak_leakage | On a tabletop, place two identical opaque boxes A and B, a red cube, and a blue cube. Use a fixed camera. | Put the red cube into box A and the blue cube into box B. Swap the closed boxes once. During the hidden interval, tilt the closed container so a vague colored shadow or motion cue is visible, but no object identity is clearly revealed. | Execute the final benchmark reveal and hold the evidence: The original box A should contain the red cube and the original box B should contain the blue cube.. | The original box A should contain the red cube and the original box B should contain the blue cube. | [
"red cube",
"blue cube"
] | [] | {
"A": [
"red cube"
],
"B": [
"blue cube"
]
} | {} |
MIST-V2-R06 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | recency_two_writes_query_last | On a fixed tabletop, place a target opaque box, a clearly separate distractor box, and the named red, blue, or coin objects. | Put a red cube into box A. After a long pause, put a blue cube into box B. Keep the target container closed until the final query. | Execute the final benchmark reveal and hold the evidence: When box B is opened, it should contain the blue cube.. | When box B is opened, it should contain the blue cube. | [
"blue cube"
] | [
"red cube in box B"
] | {
"expected": "When box B is opened, it should contain the blue cube."
} | {} |
MIST-CORE60-R1-01 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | cup_row_target_starts_under_cup_1_then_swap_cup_1_and_cup_3 | On a clean tabletop, arrange the carrier objects for this scenario: cup row target starts under cup 1 then swap cup 1 and cup 3. The hidden target is a red ball. Use a fixed front-facing camera and a simple left-to-right or top-to-bottom layout. | Execute the final benchmark reveal and hold the evidence: The red ball is in or under the cup 3.. | The red ball is in or under the cup 3. | [
"red ball",
"cup 3"
] | [] | {
"target": "red ball",
"final_carrier": "cup 3",
"scenario": "cup row target starts under cup 1 then swap cup 1 and cup 3"
} | {
"red ball": 1
} | |
MIST-CORE60-R1-05 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | tray_grid_target_starts_in_front-left_cell_then_tray_shifts_right | On a clean tabletop, arrange the carrier objects for this scenario: tray grid target starts in front-left cell then tray shifts right. The hidden target is a red cube. Use a fixed front-facing camera and a simple left-to-right or top-to-bottom layout. | Execute the final benchmark reveal and hold the evidence: The red cube is in or under the front-left cell after the tray shift.. | The red cube is in or under the front-left cell after the tray shift. | [
"red cube",
"front-left cell after the tray shift"
] | [] | {
"target": "red cube",
"final_carrier": "front-left cell after the tray shift",
"scenario": "tray grid target starts in front-left cell then tray shifts right"
} | {
"red cube": 1
} | |
MIST-CORE60-R1-06 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | tray_grid_target_starts_in_back-right_cell_then_tray_rotates_half_turn | On a clean tabletop, arrange the carrier objects for this scenario: tray grid target starts in back-right cell then tray rotates half turn. The hidden target is a blue cube. Use a fixed front-facing camera and a simple left-to-right or top-to-bottom layout. | Execute the final benchmark reveal and hold the evidence: The blue cube is in or under the front-left cell after rotation.. | The blue cube is in or under the front-left cell after rotation. | [
"blue cube",
"front-left cell after rotation"
] | [] | {
"target": "blue cube",
"final_carrier": "front-left cell after rotation",
"scenario": "tray grid target starts in back-right cell then tray rotates half turn"
} | {
"blue cube": 1
} | |
MIST-CORE60-R1-09 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | vertical_tube_target_starts_in_lower_slot_and_tube_flips_once | On a clean tabletop, arrange the carrier objects for this scenario: vertical tube target starts in lower slot and tube flips once. The hidden target is a red disk. Use a fixed front-facing camera and a simple left-to-right or top-to-bottom layout. | Execute the final benchmark reveal and hold the evidence: The red disk is in or under the upper visible slot.. | The red disk is in or under the upper visible slot. | [
"red disk",
"upper visible slot"
] | [] | {
"target": "red disk",
"final_carrier": "upper visible slot",
"scenario": "vertical tube target starts in lower slot and tube flips once"
} | {
"red disk": 1
} | |
MIST-CORE60-R1-10 | ISM-R1 | Carrier / Position Binding Memory (Gain-Focused Full20) | object carrier or position binding memory | vertical_tube_target_starts_in_upper_slot_and_tube_flips_once | On a clean tabletop, arrange the carrier objects for this scenario: vertical tube target starts in upper slot and tube flips once. The hidden target is a blue disk. Use a fixed front-facing camera and a simple left-to-right or top-to-bottom layout. | Execute the final benchmark reveal and hold the evidence: The blue disk is in or under the lower visible slot.. | The blue disk is in or under the lower visible slot. | [
"blue disk",
"lower visible slot"
] | [] | {
"target": "blue disk",
"final_carrier": "lower visible slot",
"scenario": "vertical tube target starts in upper slot and tube flips once"
} | {
"blue disk": 1
} | |
MIST-V2-S03 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | structural_remove_two_of_three | On a fixed tabletop, place one opaque black box or covered state device and the named colored objects beside it. The inside must not be visible except when the prompt explicitly says to show an initial state. | Drop a red token, a blue token, and a green token into the opaque box. Close the box. The red token and blue token exit through the side door one by one. Keep the container or device closed after the update and add a short pause with a neutral hand movement nearby. | Execute the final benchmark reveal and hold the evidence: The box should contain only the green token.. | The box should contain only the green token. | [
"green token"
] | [
"red token",
"blue token"
] | {
"box_contents": [
"green token"
],
"removed": [
"red token",
"blue token"
]
} | {
"red token": 0,
"blue token": 0,
"green token": 1
} |
MIST-V2-S04 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | structural_remove_one_of_three | On a fixed tabletop, place one opaque black box or covered state device and the named colored objects beside it. The inside must not be visible except when the prompt explicitly says to show an initial state. | Drop a red token, a blue token, and a green token into the opaque box. Close the box. Only the red token exits through the side door. Keep the container or device closed after the update and add a short pause with a neutral hand movement nearby. | Execute the final benchmark reveal and hold the evidence: The box should contain the blue token and the green token.. | The box should contain the blue token and the green token. | [
"blue token",
"green token"
] | [
"red token"
] | {
"box_contents": [
"blue token",
"green token"
],
"removed": [
"red token"
]
} | {
"red token": 0,
"blue token": 1,
"green token": 1
} |
MIST-V2-S07 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | structural_toggle_once | On a fixed tabletop, place one opaque black box or covered state device and the named colored objects beside it. The inside must not be visible except when the prompt explicitly says to show an initial state. | Show a small indicator device starting OFF. Insert one white toggle token into the hidden slot, then close the slot. Keep the container or device closed after the update and add a short pause with a neutral hand movement nearby. | Execute the final benchmark reveal and hold the evidence: The final indicator should be ON.. | The final indicator should be ON. | [
"ON indicator"
] | [
"OFF indicator"
] | {
"indicator": "ON",
"toggles": 1
} | {
"ON indicator": 1,
"OFF indicator": 0
} |
MIST-V2-S02 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | structural_remove_blue | On a fixed tabletop, place one opaque black box or covered state device and the named colored objects beside it. The inside must not be visible except when the prompt explicitly says to show an initial state. | Drop a red ball and a blue ball into the opaque box. Close the box. Then let exactly the blue ball roll out through a small side door while the inside remains hidden. Keep the container or device closed after the update and add a short pause with a neutral hand movement nearby. | Execute the final benchmark reveal and hold the evidence: The box should contain only the red ball. The blue ball should be outside.. | The box should contain only the red ball. The blue ball should be outside. | [
"red ball"
] | [
"blue ball"
] | {
"box_contents": [
"red ball"
],
"removed": [
"blue ball"
]
} | {
"red ball": 1,
"blue ball": 0
} |
MIST-GAIN60-R2-01 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_add_red_ball | On a clean tabletop, place one opaque black update box and the named colored objects. Initial contents before the update: blue cube. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly blue cube, red ball.. | The final state contains exactly blue cube, red ball. | [
"blue cube",
"red ball"
] | [] | {
"container": "opaque black update box",
"initial": [
"blue cube"
],
"added": [
"red ball"
],
"removed": [],
"final": [
"blue cube",
"red ball"
]
} | {
"blue cube": 1,
"red ball": 1
} | |
MIST-GAIN60-R2-02 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_add_green_marble | On a clean tabletop, place one opaque black update box and the named colored objects. Initial contents before the update: yellow cube. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly yellow cube, green marble.. | The final state contains exactly yellow cube, green marble. | [
"yellow cube",
"green marble"
] | [] | {
"container": "opaque black update box",
"initial": [
"yellow cube"
],
"added": [
"green marble"
],
"removed": [],
"final": [
"yellow cube",
"green marble"
]
} | {
"yellow cube": 1,
"green marble": 1
} | |
MIST-GAIN60-R2-03 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_add_blue_disk | On a clean tabletop, place one opaque black update box and the named colored objects. Initial contents before the update: orange bead. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly orange bead, blue disk.. | The final state contains exactly orange bead, blue disk. | [
"orange bead",
"blue disk"
] | [] | {
"container": "opaque black update box",
"initial": [
"orange bead"
],
"added": [
"blue disk"
],
"removed": [],
"final": [
"orange bead",
"blue disk"
]
} | {
"orange bead": 1,
"blue disk": 1
} | |
MIST-GAIN60-R2-04 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_add_purple_cube | On a clean tabletop, place one opaque black update box and the named colored objects. Initial contents before the update: red ball. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly red ball, purple cube.. | The final state contains exactly red ball, purple cube. | [
"red ball",
"purple cube"
] | [] | {
"container": "opaque black update box",
"initial": [
"red ball"
],
"added": [
"purple cube"
],
"removed": [],
"final": [
"red ball",
"purple cube"
]
} | {
"red ball": 1,
"purple cube": 1
} | |
MIST-GAIN60-R2-05 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_remove_red_ball | On a clean tabletop, place one opaque black update box and the named colored objects. Initial contents before the update: red ball, blue cube. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly blue cube.. | The final state contains exactly blue cube. | [
"blue cube"
] | [
"red ball"
] | {
"container": "opaque black update box",
"initial": [
"red ball",
"blue cube"
],
"added": [],
"removed": [
"red ball"
],
"final": [
"blue cube"
]
} | {
"blue cube": 1,
"red ball": 0
} | |
MIST-GAIN60-R2-07 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_remove_blue_disk | On a clean tabletop, place one opaque black update box and the named colored objects. Initial contents before the update: blue disk, orange bead. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly orange bead.. | The final state contains exactly orange bead. | [
"orange bead"
] | [
"blue disk"
] | {
"container": "opaque black update box",
"initial": [
"blue disk",
"orange bead"
],
"added": [],
"removed": [
"blue disk"
],
"final": [
"orange bead"
]
} | {
"orange bead": 1,
"blue disk": 0
} | |
MIST-GAIN60-R2-09 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_replace_red_with_green | On a clean tabletop, place one opaque black update box and the named colored objects. Initial contents before the update: red ball. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly green ball.. | The final state contains exactly green ball. | [
"green ball"
] | [
"red ball"
] | {
"container": "opaque black update box",
"initial": [
"red ball"
],
"added": [
"green ball"
],
"removed": [
"red ball"
],
"final": [
"green ball"
]
} | {
"green ball": 1,
"red ball": 0
} | |
MIST-GAIN60-R2-10 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_replace_blue_with_yellow | On a clean tabletop, place one opaque black update box and the named colored objects. Initial contents before the update: blue cube. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly yellow cube.. | The final state contains exactly yellow cube. | [
"yellow cube"
] | [
"blue cube"
] | {
"container": "opaque black update box",
"initial": [
"blue cube"
],
"added": [
"yellow cube"
],
"removed": [
"blue cube"
],
"final": [
"yellow cube"
]
} | {
"yellow cube": 1,
"blue cube": 0
} | |
MIST-GAIN60-R2-12 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_replace_yellow_with_blue | On a clean tabletop, place one opaque black update box and the named colored objects. Initial contents before the update: yellow disk. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly blue disk.. | The final state contains exactly blue disk. | [
"blue disk"
] | [
"yellow disk"
] | {
"container": "opaque black update box",
"initial": [
"yellow disk"
],
"added": [
"blue disk"
],
"removed": [
"yellow disk"
],
"final": [
"blue disk"
]
} | {
"blue disk": 1,
"yellow disk": 0
} | |
MIST-GAIN60-R2-13 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_stack_red_below_blue | On a clean tabletop, place one transparent vertical tube with an opaque sleeve cover and the named colored objects. Initial contents before the update: empty tube or empty box. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly red disk below blue disk.. | The final state contains exactly red disk below blue disk. | [
"red disk below blue disk"
] | [] | {
"container": "transparent vertical tube with an opaque sleeve cover",
"initial": [],
"added": [
"red disk",
"blue disk"
],
"removed": [],
"final": [
"red disk below blue disk"
]
} | {
"red disk below blue disk": 1
} | |
MIST-GAIN60-R2-14 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_stack_yellow_below_green | On a clean tabletop, place one transparent vertical tube with an opaque sleeve cover and the named colored objects. Initial contents before the update: empty tube or empty box. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly yellow disk below green disk.. | The final state contains exactly yellow disk below green disk. | [
"yellow disk below green disk"
] | [] | {
"container": "transparent vertical tube with an opaque sleeve cover",
"initial": [],
"added": [
"yellow disk",
"green disk"
],
"removed": [],
"final": [
"yellow disk below green disk"
]
} | {
"yellow disk below green disk": 1
} | |
MIST-GAIN60-R2-15 | ISM-R2 | Operation-History State Update Memory (Gain-Focused Full20) | operation-history add/remove state memory | gain_r2_stack_orange_below_purple | On a clean tabletop, place one transparent vertical tube with an opaque sleeve cover and the named colored objects. Initial contents before the update: empty tube or empty box. | Execute the final benchmark reveal and hold the evidence: The final state contains exactly orange disk below purple disk.. | The final state contains exactly orange disk below purple disk. | [
"orange disk below purple disk"
] | [] | {
"container": "transparent vertical tube with an opaque sleeve cover",
"initial": [],
"added": [
"orange disk",
"purple disk"
],
"removed": [],
"final": [
"orange disk below purple disk"
]
} | {
"orange disk below purple disk": 1
} |
MIST
43 Internal MIST case cards for delayed reveal. Each row names the scene setup, the hidden operation, the reveal query, and the state the last frame must show.
A low memory score is often a missed action. The scene can already be set, and the hidden interval can stay leak-free, while the asked change is never carried out. These cards are the task record for that split: setup, the hidden change, and reveal.
M3-Agent is a test-time rewrite of the missed move. It does not add a trained memory module. Keep-best is an oracle over the same score that reports the result.
Load
from datasets import load_dataset
cards = load_dataset("Wayne-King/MIST", split="test")
data/mist_cards.json is the same table in the cases shape read by the public control loop.
Links
- Paper: https://arxiv.org/abs/2610.23542
- Hugging Face paper page: https://huggingface.co/papers/2610.23542
- Project page: https://hku-mmlab.github.io/MIST/
- Code: https://github.com/HKU-MMLab/MIST
Fields
| Field | Meaning |
|---|---|
id |
Case id |
category |
ISM-R0 content, ISM-R1 binding, or ISM-R2 operation history |
setup |
Scene before the change |
hidden_operation |
The change that must happen while it stays hidden |
reveal_query |
The later request to show the result |
expected_reveal |
What the last frame must show |
expected_present / expected_absent / expected_state / expected_count |
Scoring checks |
The reveal text does not name the hidden answer as a gold sentence. expected_* is the scoring key.
- Downloads last month
- 35