File size: 16,350 Bytes
3404d44 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 | ====================================================================== Model: molmo-7B-O-0924 ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 326 483 67.5% FAR counter 16 61 26.2% FAR ambiguous 24 50 48.0% CLOSE consistent 294 493 59.6% CLOSE counter 29 68 42.6% CLOSE ambiguous 22 51 43.1% ---------------------------------------------------- TOTAL consistent 620 976 63.5% TOTAL counter 45 129 34.9% Accuracy Gap (Consistent - Counter): 28.6%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 84 / 129 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_80k ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 269 483 55.7% FAR counter 19 61 31.1% FAR ambiguous 21 50 42.0% CLOSE consistent 322 493 65.3% CLOSE counter 19 68 27.9% CLOSE ambiguous 16 51 31.4% ---------------------------------------------------- TOTAL consistent 591 976 60.6% TOTAL counter 38 129 29.5% Accuracy Gap (Consistent - Counter): 31.1%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 91 / 129 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_400k ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 315 483 65.2% FAR counter 15 61 24.6% FAR ambiguous 19 50 38.0% CLOSE consistent 297 493 60.2% CLOSE counter 20 68 29.4% CLOSE ambiguous 16 51 31.4% ---------------------------------------------------- TOTAL consistent 612 976 62.7% TOTAL counter 35 129 27.1% Accuracy Gap (Consistent - Counter): 35.6%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 94 / 129 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_800k ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 319 483 66.0% FAR counter 22 61 36.1% FAR ambiguous 19 50 38.0% CLOSE consistent 317 493 64.3% CLOSE counter 22 68 32.4% CLOSE ambiguous 23 51 45.1% ---------------------------------------------------- TOTAL consistent 636 976 65.2% TOTAL counter 44 129 34.1% Accuracy Gap (Consistent - Counter): 31.1%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 85 / 129 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_2m ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 312 483 64.6% FAR counter 18 61 29.5% FAR ambiguous 20 50 40.0% CLOSE consistent 325 493 65.9% CLOSE counter 33 68 48.5% CLOSE ambiguous 24 51 47.1% ---------------------------------------------------- TOTAL consistent 637 976 65.3% TOTAL counter 51 129 39.5% Accuracy Gap (Consistent - Counter): 25.7%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 78 / 129 ====================================================================== Model: NVILA-Lite-2B ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ far consistent 242 483 50.1% far counter 15 61 24.6% far ambiguous 14 50 28.0% close consistent 236 493 47.9% close counter 20 68 29.4% close ambiguous 20 51 39.2% ------------------------------------------------------ TOTAL consistent 478 976 49.0% TOTAL counter 35 129 27.1% Accuracy Gap (Consistent - Counter): 21.8%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 94 / 129 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-80k ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 317 483 65.6% FAR counter 2 61 3.3% FAR ambiguous 11 50 22.0% CLOSE consistent 246 493 49.9% CLOSE counter 18 68 26.5% CLOSE ambiguous 15 51 29.4% ---------------------------------------------------- TOTAL consistent 563 976 57.7% TOTAL counter 20 129 15.5% Accuracy Gap (Consistent - Counter): 42.2%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 109 / 129 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-400k ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 319 483 66.0% FAR counter 12 61 19.7% FAR ambiguous 18 50 36.0% CLOSE consistent 277 493 56.2% CLOSE counter 32 68 47.1% CLOSE ambiguous 20 51 39.2% ---------------------------------------------------- TOTAL consistent 596 976 61.1% TOTAL counter 44 129 34.1% Accuracy Gap (Consistent - Counter): 27.0%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 85 / 129 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-800k ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 335 483 69.4% FAR counter 17 61 27.9% FAR ambiguous 20 50 40.0% CLOSE consistent 282 493 57.2% CLOSE counter 33 68 48.5% CLOSE ambiguous 15 51 29.4% ---------------------------------------------------- TOTAL consistent 617 976 63.2% TOTAL counter 50 129 38.8% Accuracy Gap (Consistent - Counter): 24.5%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 79 / 129 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-2m ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 324 483 67.1% FAR counter 19 61 31.1% FAR ambiguous 21 50 42.0% CLOSE consistent 268 493 54.4% CLOSE counter 34 68 50.0% CLOSE ambiguous 20 51 39.2% ---------------------------------------------------- TOTAL consistent 592 976 60.7% TOTAL counter 53 129 41.1% Accuracy Gap (Consistent - Counter): 19.6%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 76 / 129 ====================================================================== Model: RoboRefer-2B-SFT ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 422 483 87.4% FAR counter 28 61 45.9% FAR ambiguous 34 50 68.0% CLOSE consistent 427 493 86.6% CLOSE counter 49 68 72.1% CLOSE ambiguous 37 51 72.5% ---------------------------------------------------- TOTAL consistent 849 976 87.0% TOTAL counter 77 129 59.7% Accuracy Gap (Consistent - Counter): 27.3%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 52 / 129 ====================================================================== Model: Qwen2.5-VL-3B-Instruct ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 274 483 56.7% FAR counter 15 61 24.6% FAR ambiguous 16 50 32.0% CLOSE consistent 260 493 52.7% CLOSE counter 27 68 39.7% CLOSE ambiguous 17 51 33.3% ---------------------------------------------------- TOTAL consistent 534 976 54.7% TOTAL counter 42 129 32.6% Accuracy Gap (Consistent - Counter): 22.2%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 87 / 129 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_80k ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 267 483 55.3% FAR counter 15 61 24.6% FAR ambiguous 13 50 26.0% CLOSE consistent 227 493 46.0% CLOSE counter 24 68 35.3% CLOSE ambiguous 18 51 35.3% ---------------------------------------------------- TOTAL consistent 494 976 50.6% TOTAL counter 39 129 30.2% Accuracy Gap (Consistent - Counter): 20.4%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 90 / 129 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_400k ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 275 483 56.9% FAR counter 13 61 21.3% FAR ambiguous 17 50 34.0% CLOSE consistent 238 493 48.3% CLOSE counter 22 68 32.4% CLOSE ambiguous 17 51 33.3% ---------------------------------------------------- TOTAL consistent 513 976 52.6% TOTAL counter 35 129 27.1% Accuracy Gap (Consistent - Counter): 25.4%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 94 / 129 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_800k ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 297 483 61.5% FAR counter 12 61 19.7% FAR ambiguous 16 50 32.0% CLOSE consistent 248 493 50.3% CLOSE counter 22 68 32.4% CLOSE ambiguous 15 51 29.4% ---------------------------------------------------- TOTAL consistent 545 976 55.8% TOTAL counter 34 129 26.4% Accuracy Gap (Consistent - Counter): 29.5%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 95 / 129 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_2m ====================================================================== Category Type Correct Total Accuracy ---------------------------------------------------- FAR consistent 313 483 64.8% FAR counter 6 61 9.8% FAR ambiguous 14 50 28.0% CLOSE consistent 281 493 57.0% CLOSE counter 25 68 36.8% CLOSE ambiguous 19 51 37.3% ---------------------------------------------------- TOTAL consistent 594 976 60.9% TOTAL counter 31 129 24.0% Accuracy Gap (Consistent - Counter): 36.8%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 98 / 129 ============================================================================== MODEL COMPARISON (EmbSpatialBench far/close) ============================================================================== Model Consistent Counter Gap ------------------------------------------------------------------------------ molmo-7B-O-0924 63.5% 34.9% +28.6%p molmo-7B-O-0924-data_scale_exp_80k 60.6% 29.5% +31.1%p molmo-7B-O-0924-data_scale_exp_400k 62.7% 27.1% +35.6%p molmo-7B-O-0924-data_scale_exp_800k 65.2% 34.1% +31.1%p molmo-7B-O-0924-data_scale_exp_2m 65.3% 39.5% +25.7%p ------------------------------------------------------------------------------ NVILA-Lite-2B 49.2% 27.1% +21.8%p NVILA-Lite-2B-data-scale-exp-80k 57.7% 15.5% +42.2%p NVILA-Lite-2B-data-scale-exp-400k 61.1% 34.1% +27.0%p NVILA-Lite-2B-data-scale-exp-800k 63.2% 38.8% +24.5%p NVILA-Lite-2B-data-scale-exp-2m 60.7% 41.1% +19.6%p RoboRefer-2B-SFT 87.0% 59.7% +27.3%p ------------------------------------------------------------------------------ Qwen2.5-VL-3B-Instruct 54.7% 32.6% +22.2%p Qwen2.5-VL-3B-Instruct-data_scale_exp_80k 50.6% 30.2% +20.4%p Qwen2.5-VL-3B-Instruct-data_scale_exp_400k 52.6% 27.1% +25.4%p Qwen2.5-VL-3B-Instruct-data_scale_exp_800k 55.8% 26.4% +29.5%p Qwen2.5-VL-3B-Instruct-data_scale_exp_2m 60.9% 24.0% +36.8%p |