====================================================================== Model: molmo-7B-O-0924 ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 364 390 93.3% Depth counter 52 71 73.2% Depth ambiguous 91 139 65.5% Distance consistent 89 165 53.9% Distance counter 276 374 73.8% Distance ambiguous 46 61 75.4% ------------------------------------------------------ TOTAL consistent 453 555 81.6% TOTAL counter 328 445 73.7% Accuracy Gap (Consistent - Counter): 7.9%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 117 / 445 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_80k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 314 390 80.5% Depth counter 39 71 54.9% Depth ambiguous 73 139 52.5% Distance consistent 96 165 58.2% Distance counter 267 374 71.4% Distance ambiguous 42 61 68.9% ------------------------------------------------------ TOTAL consistent 410 555 73.9% TOTAL counter 306 445 68.8% Accuracy Gap (Consistent - Counter): 5.1%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 139 / 445 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_400k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 354 390 90.8% Depth counter 37 71 52.1% Depth ambiguous 89 139 64.0% Distance consistent 106 165 64.2% Distance counter 274 374 73.3% Distance ambiguous 45 61 73.8% ------------------------------------------------------ TOTAL consistent 460 555 82.9% TOTAL counter 311 445 69.9% Accuracy Gap (Consistent - Counter): 13.0%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 134 / 445 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_800k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 351 390 90.0% Depth counter 48 71 67.6% Depth ambiguous 93 139 66.9% Distance consistent 112 165 67.9% Distance counter 272 374 72.7% Distance ambiguous 41 61 67.2% ------------------------------------------------------ TOTAL consistent 463 555 83.4% TOTAL counter 320 445 71.9% Accuracy Gap (Consistent - Counter): 11.5%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 125 / 445 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_2m ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 355 390 91.0% Depth counter 53 71 74.6% Depth ambiguous 116 139 83.5% Distance consistent 119 165 72.1% Distance counter 322 374 86.1% Distance ambiguous 47 61 77.0% ------------------------------------------------------ TOTAL consistent 474 555 85.4% TOTAL counter 375 445 84.3% Accuracy Gap (Consistent - Counter): 1.1%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 70 / 445 ====================================================================== Model: NVILA-Lite-2B ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 294 390 75.4% Depth counter 29 71 40.8% Depth ambiguous 92 139 66.2% Distance consistent 92 165 55.8% Distance counter 181 374 48.4% Distance ambiguous 41 61 67.2% ------------------------------------------------------ TOTAL consistent 386 555 69.5% TOTAL counter 210 445 47.2% Accuracy Gap (Consistent - Counter): 22.4%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 235 / 445 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-80k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 284 390 72.8% Depth counter 35 71 49.3% Depth ambiguous 78 139 56.1% Distance consistent 95 165 57.6% Distance counter 228 374 61.0% Distance ambiguous 42 61 68.9% ------------------------------------------------------ TOTAL consistent 379 555 68.3% TOTAL counter 263 445 59.1% Accuracy Gap (Consistent - Counter): 9.2%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 182 / 445 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-400k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 319 390 81.8% Depth counter 42 71 59.2% Depth ambiguous 85 139 61.2% Distance consistent 89 165 53.9% Distance counter 269 374 71.9% Distance ambiguous 44 61 72.1% ------------------------------------------------------ TOTAL consistent 408 555 73.5% TOTAL counter 311 445 69.9% Accuracy Gap (Consistent - Counter): 3.6%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 134 / 445 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-800k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 330 390 84.6% Depth counter 48 71 67.6% Depth ambiguous 91 139 65.5% Distance consistent 87 165 52.7% Distance counter 295 374 78.9% Distance ambiguous 46 61 75.4% ------------------------------------------------------ TOTAL consistent 417 555 75.1% TOTAL counter 343 445 77.1% Accuracy Gap (Consistent - Counter): -1.9%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 102 / 445 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-2m ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 380 390 97.4% Depth counter 67 71 94.4% Depth ambiguous 116 139 83.5% Distance consistent 119 165 72.1% Distance counter 347 374 92.8% Distance ambiguous 57 61 93.4% ------------------------------------------------------ TOTAL consistent 499 555 89.9% TOTAL counter 414 445 93.0% Accuracy Gap (Consistent - Counter): -3.1%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 31 / 445 ====================================================================== Model: RoboRefer-2B-SFT ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 384 390 98.5% Depth counter 68 71 95.8% Depth ambiguous 122 139 87.8% Distance consistent 132 165 80.0% Distance counter 354 374 94.7% Distance ambiguous 57 61 93.4% ------------------------------------------------------ TOTAL consistent 516 555 93.0% TOTAL counter 422 445 94.8% Accuracy Gap (Consistent - Counter): -1.9%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 23 / 445 ====================================================================== Model: Qwen2.5-VL-3B-Instruct ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 301 390 77.2% Depth counter 38 71 53.5% Depth ambiguous 83 139 59.7% Distance consistent 87 165 52.7% Distance counter 247 374 66.0% Distance ambiguous 27 61 44.3% ------------------------------------------------------ TOTAL consistent 388 555 69.9% TOTAL counter 285 445 64.0% Accuracy Gap (Consistent - Counter): 5.9%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 160 / 445 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_80k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 273 390 70.0% Depth counter 39 71 54.9% Depth ambiguous 76 139 54.7% Distance consistent 77 165 46.7% Distance counter 255 374 68.2% Distance ambiguous 37 61 60.7% ------------------------------------------------------ TOTAL consistent 350 555 63.1% TOTAL counter 294 445 66.1% Accuracy Gap (Consistent - Counter): -3.0%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 151 / 445 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_400k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 259 390 66.4% Depth counter 38 71 53.5% Depth ambiguous 75 139 54.0% Distance consistent 62 165 37.6% Distance counter 235 374 62.8% Distance ambiguous 30 61 49.2% ------------------------------------------------------ TOTAL consistent 321 555 57.8% TOTAL counter 273 445 61.3% Accuracy Gap (Consistent - Counter): -3.5%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 172 / 445 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_800k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 242 390 62.1% Depth counter 38 71 53.5% Depth ambiguous 72 139 51.8% Distance consistent 55 165 33.3% Distance counter 233 374 62.3% Distance ambiguous 19 61 31.1% ------------------------------------------------------ TOTAL consistent 297 555 53.5% TOTAL counter 271 445 60.9% Accuracy Gap (Consistent - Counter): -7.4%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 174 / 445 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_2m ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 243 390 62.3% Depth counter 36 71 50.7% Depth ambiguous 72 139 51.8% Distance consistent 61 165 37.0% Distance counter 233 374 62.3% Distance ambiguous 19 61 31.1% ------------------------------------------------------ TOTAL consistent 304 555 54.8% TOTAL counter 269 445 60.4% Accuracy Gap (Consistent - Counter): -5.7%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 176 / 445 ============================================================================== MODEL COMPARISON ============================================================================== Model Consistent Counter Gap ------------------------------------------------------------------------------ molmo-7B-O-0924 81.6% 73.7% +7.9%p molmo-7B-O-0924-data_scale_exp_80k 73.9% 68.8% +5.1%p molmo-7B-O-0924-data_scale_exp_400k 82.9% 69.9% +13.0%p molmo-7B-O-0924-data_scale_exp_800k 83.4% 71.9% +11.5%p molmo-7B-O-0924-data_scale_exp_2m 85.4% 84.3% +1.1%p NVILA-Lite-2B 69.5% 47.2% +22.4%p NVILA-Lite-2B-data-scale-exp-80k 68.3% 59.1% +9.2%p NVILA-Lite-2B-data-scale-exp-400k 73.5% 69.9% +3.6%p NVILA-Lite-2B-data-scale-exp-800k 75.1% 77.1% -1.9%p NVILA-Lite-2B-data-scale-exp-2m 89.9% 93.0% -3.1%p RoboRefer-2B-SFT 93.0% 94.8% -1.9%p Qwen2.5-VL-3B-Instruct 69.9% 64.0% +5.9%p Qwen2.5-VL-3B-Instruct-data_scale_exp_80k 63.1% 66.1% -3.0%p Qwen2.5-VL-3B-Instruct-data_scale_exp_400k 57.8% 61.3% -3.5%p Qwen2.5-VL-3B-Instruct-data_scale_exp_800k 53.5% 60.9% -7.4%p Qwen2.5-VL-3B-Instruct-data_scale_exp_2m 54.8% 60.4% -5.7%p