Spaces:
Sleeping
Sleeping
File size: 93,063 Bytes
a213258 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 |
--- Page 1 ---
arXiv:2505.14442v1 [cs.CL] 20 May 2025
Creative Preference Optimization
Mete Ismayilzada1,2, Antonio Laverghetta Jr.3, Simone A. Luchini3,
Reet Patel3, Antoine Bosselut1, Lonneke van der Plas 2 Roger Beaty 3
1EPFL, 2Università della Svizzera Italiana, 3Pennsylvania State University
mahammad.ismayilzada epfl.ch
Abstract
While
Large
Language
Models
(LLMs)
have demonstrated impressive performance
across natural language generation tasks,
their ability to generate truly creative con-
tent characterized by novelty, diversity, sur-
prise, and quality remains limited. Existing
methods for enhancing LLM creativity often
focus narrowly on diversity or specific tasks,
failing to address creativity s multifaceted na-
ture in a generalizable way.
In this work,
we propose Creative Preference Optimization
(CRPO), a novel alignment method that in-
jects signals from multiple creativity dimen-
sions into the preference optimization objec-
tive in a modular fashion. We train and eval-
uate creativity-augmented versions of several
models using CRPO and MUCE, a new large-
scale human preference dataset spanning over
200,000 human-generated responses and rat-
ings from more than 30 psychological creativ-
ity assessments. Our models outperform strong
baselines, including GPT-4o, on both auto-
mated and human evaluations, producing more
novel, diverse, and surprising generations while
maintaining high output quality. Additional
evaluations on NOVELTYBENCH further con-
firm the generalizability of our approach. To-
gether, our results demonstrate that directly op-
timizing for creativity within preference frame-
works is a promising direction for advancing
the creative capabilities of LLMs without com-
promising output quality.
Introduction
Large Language Models (LLMs) have made sig-
nificant progress across a broad range of natural
language generation tasks (Team et al., 2023; Zhao
et al., 2025; Bubeck et al., 2023; Wei et al., 2022;
Brown et al., 2020). However, whether LLMs ex-
hibit true human-like creativity i.e the ability to pro-
duce novel (i.e., original), high-quality (i.e. useful)
and surprising (i.e. unexpected) ideas (Simonton,
2012; Boden, 2004) remains unclear. Research on
the creativity of LLMs has found mixed results,
with some reporting that LLMs are more creative
than humans (Bellemare-Pepin et al., 2024; Zhao
et al., 2024), others reporting that they are less cre-
ative (Koivisto and Grassini, 2023; Chakrabarty
et al., 2024; Ismayilzada et al., 2024b), and some
finding their creativity to be on par with each other
(Stevenson et al., 2022; Góes et al., 2023; Gilhooly,
2024).
However, past research has also found
that the high LLM performance can be attributed
to the artificial nature of the creativity tasks (Is-
mayilzada et al., 2024a) commonly employed to
evaluate LLMs such as the Alternative Uses Task
(Guilford, 1967) or to the remarkable creativity of
human-written texts on the web (Lu et al., 2024).
Consequently, LLMs have been shown to often
lack novelty and surprise in their generations (Is-
mayilzada et al., 2024a,b; Zhang et al., 2025; Tian
et al., 2024; Chakrabarty et al., 2024) and produce
significantly less diverse content compared to hu-
mans (Padmakumar and He, 2023; Anderson et al.,
2024; Kirk et al., 2023; Xu et al., 2024; O Mahony
et al., 2024; Zhang et al., 2024; Wenger and Kenett,
2025). These tendencies limit the utility of LLMs
for creative tasks, such as story generation and cre-
ative problem solving that often require longer re-
sponses and out-of-the-box thinking (Tian et al.,
2023; Huang et al., 2024; Chen et al., 2024).
Recent research has proposed some methods for
improving creativity of LLMs, often targeting diver-
sity aspect alone (Wong et al., 2024; Hayati et al.,
2023; Chung et al., 2023; Franceschelli and Mu-
solesi, 2024; Zhang et al., 2024; Wang et al., 2024b;
Zhou et al., 2025; Lanchantin et al., 2025; Chung
et al., 2025) or focusing on a single creativity task
(Tian et al., 2023; Nair et al., 2024; Summers-Stay
et al., 2023). However, creativity is a multifaceted
ability that also encompasses novelty, surprise, and
quality and manifests itself in a wide range of tasks.
Consequently, it has been argued that methods pro-
moting creativity improvements should consider
--- Page 2 ---
prompt
preferred
response
set of
preferred
responses
Novelty
Diversity
Surprise
Quality
LM
RM
λn
λd
λs
λq
Creativity
DPO Loss
Figure 1: Our preference alignment method CRPO to improve output creativity by injecting a weighted combination
of signals from multiple creativity dimensions.
multiple dimensions of creativity together across
several creative tasks (Ismayilzada et al., 2024a).
Hence, the broader challenge of enhancing overall
creativity in LLM outputs largely remain underex-
plored.
To this end, we propose a novel approach to di-
rectly optimize for creativity in language model
generation through preference learning (Ouyang
et al., 2022; Rafailov et al., 2023). Recent works
targeting improvement in LLM creativity have
mainly focused on black-box techniques to elicit
creative outputs through input-level (e.g., prompt-
ing) (Tian et al., 2023; Mehrotra et al., 2024;
Nair et al., 2024; Summers-Stay et al., 2023) and
output-level strategies (e.g., creative decoding)
(Franceschelli and Musolesi, 2024; Meister et al.,
2023).
However, these methods are inherently
limited to the fixed creative capacity of language
models and are not designed to optimize for fine-
grained dimensions of creativity. Recently, moti-
vated by the negative impact of the preference align-
ment techniques on the diversity of LLM outputs
(Padmakumar and He, 2023; Anderson et al., 2024;
Kirk et al., 2023; O Mahony et al., 2024; West and
Potts, 2025), few works have suggested directly
modifying the preference optimization methods to
promote output diversity (Lanchantin et al., 2025;
Chung et al., 2025). Inspired by these approaches,
we design a new optimization strategy that injects
signals from multiple dimensions of creativity into
the preference modeling objective in a modular
fashion. Specifically, we integrate the novelty, di-
versity, surprise and quality dimensions of creativ-
ity into the training objective of direct preference
optimization (DPO) (Rafailov et al., 2023), with
weighted composition that allow balancing each
dimension s contribution. We call this method cre-
ative preference optimization (CRPO) and provide
its conceptual illustration in Figure 1 with full de-
tails in Section 3.
We test the efficacy of CRPO using MUCE
(Multitask Creativity Evaluation), our newly cu-
rated large-scale dataset of prompt-response pairs
annotated with human preferences across a di-
verse range of creative tasks in multiple languages.
While previous work has largely evaluated creativ-
ity improvements on a narrow range of tasks like
story generation (Chung et al., 2025; Lanchantin
et al., 2025) or creative problem solving (Tian
et al., 2023), MUCE enables us to test whether
our methods truly generalize across a diverse range
of creativity assessments. Our results show that
Llama-3.1-8B-Instruct (AI Meta, 2024) and
Mistral-7B-Instruct-v0.3 (Jiang et al., 2023)
trained using CRPO outperform the same models
trained using only supervised fine-tuning (SFT) or
DPO without any creativity injections, as well as
existing LLMs such as GPT-4o, generating more
novel, diverse, and surprising outputs than all the
baselines while maintaining high quality.
Our main contributions are as follows:
1. We introduce MUCE, a large-scale prefer-
ence dataset consisting of more than 200,000
human responses and ratings for more than
30 creativity assessments. All tasks within
MUCE are carefully chosen to provide valid
measures of creativity in humans, making
MUCE one of the largest psychologically
valid datasets of human creativity for train-
--- Page 3 ---
ing preference models.
2. We propose a novel flexible preference
alignment method CRPO that injects sig-
nals from several dimensions of creativ-
ity into the existing preference optimization
method DPO and train creativity-enhanced
versions of Llama-3.1-8B-Instruct and
Mistral-7B-Instruct-v0.3.
3. We evaluate the effectiveness of our approach
on a range of creativity tasks from MUCE, as
well as external tasks from NOVELTYBENCH
(Zhang et al., 2025), using both automated
metrics and human evaluations. Our analy-
sis shows that CRPO is a promising method
for enhancing the creative capabilities of lan-
guage models while maintaining quality.
Related Work
2.1
Large Language Model Creativity
The potential of building LLM applications for
creative industries has spurred significant research
interest on AI creativity (Bellemare-Pepin et al.,
2024), and many LLM tools marketed for assis-
tance with creative tasks have been developed in
the last few years (Wang et al., 2024b). Yet de-
bates on whether AI is capable of true creativity are
nearly as old as AI itself (Stein, 2014; Franceschelli
and Musolesi, 2024; Sæbø and Brovold, 2024),
with theoretical and philosophical arguments being
made both for and against AI creativity (Ismay-
ilzada et al., 2024a). Classic psychological theories
of creativity generally agree that, for a product to
be creative, it must be new, surprising, and valu-
able (Boden, 2004). Creative tasks are also often
characterized by high diversity (Padmakumar and
He, 2023; Shypula et al., 2025), though diversity is
only one facet of creativity (Johnson et al., 2021).
Studies on LLM creativity have yielded conflicting
findings: some suggest LLMs surpass human cre-
ativity (Bellemare-Pepin et al., 2024; Zhao et al.,
2024), others argue they fall short (Koivisto and
Grassini, 2023; Chakrabarty et al., 2024; Ismay-
ilzada et al., 2024b), while some conclude that
LLM and human creativity are roughly equivalent
(Gilhooly, 2024; Stevenson et al., 2022; Góes et al.,
2023). Some works have suggested that LLMs
lack novelty and surprise in their generations (Is-
mayilzada et al., 2024a,b; Zhang et al., 2025; Tian
et al., 2024; Chakrabarty et al., 2024) and their
seemingly remarkable creative outputs may be in
large part attributable to the remarkable creativity
of human-written texts on the web (Lu et al., 2024).
Some recent works have suggested improving the
creativity of LLMs through prompting techniques
(Tian et al., 2023; Mehrotra et al., 2024; Nair et al.,
2024; Summers-Stay et al., 2023) and decoding
strategies (Franceschelli and Musolesi, 2024; Meis-
ter et al., 2023). In this work, we instead explore
directly optimizing language models for creativity
using human preferences extracted from responses
to creativity assessments.
2.2
Preference Learning
Aligning LLMs to human preferences has proven
effective in developing models that are helpful and
useful to users, leading to the emergence of numer-
ous preference learning methods (Gao et al., 2024;
Ouyang et al., 2022; Rafailov et al., 2023). How-
ever, prior work has highlighted a lack of diversity
in LLM outputs (Anderson et al., 2024; Lanchantin
et al., 2025; Wenger and Kenett, 2025; Padmaku-
mar and He, 2023), with alignment often cited as
a contributing factor (West and Potts, 2025). In
response, recent research has explored modifica-
tions to existing preference modeling techniques
aimed at mitigating this reduction in diversity. One
notable approach, Diverse Preference Optimiza-
tion, proposes enhancing preference data creation
by selecting preference pairs based on a diversity
metric (Lanchantin et al., 2025). Another recent
method introduces a modification to the optimiza-
tion objective itself to incorporate a diversity signal
(Chung et al., 2025). Both strategies have demon-
strated effectiveness in promoting output diversity
with minimal impact on output quality. However,
as previously noted, diversity represents only one
facet of creativity; true creativity also requires the
capacity for novelty and surprise. In this work,
we present a modular preference alignment frame-
work for creativity that enables direct optimization
across multiple dimensions of creative expression.
Creative Preference Optimization
According to its three-criterion definition, creativity
involves the generation of novel, high-quality, and
surprising ideas (Simonton, 2012; Boden, 2004;
Runco and Jaeger, 2012). Moreover, creative out-
puts tend to be highly diverse across individuals
(Anderson et al., 2024). Therefore, to promote over-
all creativity in LLM outputs, we propose to inject
unsupervised metrics related to each dimension of
--- Page 4 ---
creativity into the loss functions of standard pref-
erence optimization methods. We use direct pref-
erence optimization (DPO) (Rafailov et al., 2023)
to illustrate our modifications to the loss function.
Recall that in the standard formulation of DPO, a
policy model (pθ) is directly optimized on a dataset
of (x, yw, yl) where x, yw and yl refer to the model
input (i.e. prompt), preferred (i.e. chosen) model
response and dispreferred (i.e. rejected) model re-
sponse, respectively. Using the ratio between the
policy model s likelihood and that of the reference
SFT model (pSFT ) as an implicit reward, the train-
ing objective of DPO is defined as follows:
lDP O
h
log σ
β log
pθ(yw x)
pSFT(yw x) β log
pθ(yl x)
pSFT(yl x)
i
LDPO E(x,yw,yl) D
lDP O
(1)
A challenge with standard preference optimiza-
tion methods is that they may significantly reduce
the diversity of the responses LLMs generate, as the
loss function encourages models to generate pre-
ferred responses even if they are not very creative
(West and Potts, 2025; Padmakumar and He, 2023;
Anderson et al., 2024; Kirk et al., 2023; Xu et al.,
2024; O Mahony et al., 2024; Zhang et al., 2024;
Wenger and Kenett, 2025). Existing approaches
to address this in the preference optimization ob-
jective have centered around curating a preference
data based on various diversity metrics (Lanchantin
et al., 2025) or incorporating extra regularization
terms that encourage diverse generations while bal-
ancing quality (Chung et al., 2025). For example,
the recently proposed Diversified DPO (DDPO)
method adds a scalar diversity term δw (i.e. diver-
sity score of the preferred response) into the DPO
loss (Chung et al., 2025):
LDDPO E(x,yw,yl) D
δwlDP O
(2)
While diversity is important for creativity, re-
search in psychology has long established that truly
creative responses also require novelty, surprise,
and quality (Boden, 2004; Barron, 1955; Simon-
ton, 2018). Therefore, we propose incorporating
metrics for each of these, alongside diversity, into
the preference loss in a modular structure, enabling
the construction of different creativity models by
combining these dimensions as needed.
LCDPO E(x,yw,yl) D
h
(λdδw λnνw λsξw λqγw)lDP O
i
(3)
In our proposed creative DPO loss, δw, νw, ξw
and γw correspond to diversity, novelty, surprise
and quality scores of the preferred response respec-
tively and λd, λn, λs and λq are hyperparameters
that control the effect of each score (we call them
injection weights). In particular, when λd 1,
λn 0, λs 0 and λq 0, we recover the DDPO
loss. While there are multiple approaches for oper-
ationalizing δw, νw, ξw and γw, we propose to use
the following metrics for each:
3.1
Diversity
We use an inverse homogenization metric from
Padmakumar and He (2023) similar to Chung et al.
(2025). Specifically, given a prompt x and a set of
(preferred) responses for x denoted as Yx, we com-
pute the diversity score of any particular preferred
response as the average pairwise semantic distance
to all the other preferred responses in Yx:
δw
Yx 1
X
yi Yx yw
semdis(yw, yi)
(4)
We use 1 cos_sim( , ) as a semantic distance
function (i.e., semdis( , )).
3.2
Novelty
We use a novelty metric similar to Karampiperis
et al. (2014) where the novelty of a text is defined
as the absolute difference between the average pair-
wise semantic distances of words in the text and
those of a reference corpus of texts. In particular,
we define the set of preferred responses to a prompt
x as a reference corpus (Yx) and define the novelty
of a preferred response as follows:
νw DSI(yw) DSI(Yx)
(5)
DSI(T)
P T
i,j 1 semdis(Ti, Tj), i j
T
(6)
Here T refers to a piece of text, Ti to the word
i in the set of unique words in T denoted as T
and DSI( ) is divergent semantic integration, the
average pairwise semantic distances of words in a
text (Johnson et al., 2022).
3.3
Surprise
We use Shannon surprise the negative log-
likelihood of the text which has been widely
used as a measure of surprise in prior work
(Bunescu and Uduehi, 2022; Modirshanechi et al.,
2022; Kuznetsova et al., 2013). More specifically,
--- Page 5 ---
given a prompt x, we define the surprise of a par-
ticular response as the exponentiated negative log-
likelihood of the response (i.e. perplexity) condi-
tioned on the prompt x and under some reference
model S as follows:
ξw 2 logPS(yw x)
(7)
3.4
Quality
Although a general quality scoring method is hard
to define, reward models that are trained to output
a high score to preferred answers can be used as
a proxy (Zhang et al., 2025; Lambert et al., 2024).
In particular, we define the quality of a preferred
response given a prompt x as the score assigned by
some reward model R: γw R(yw x).
The MUCE Dataset
To compile MUCE, we solicited data from the
global creativity research community, specifically
targeting researchers studying human creativity to
obtain data from tasks known to be valid creativity
measures. We specifically targeted datasets which
contained complete metadata, including informa-
tion about the task, language, and items that partici-
pants responded to. We gathered additional data by
performing a manual search of the Open Science
Framework database1, and only retained data from
peer-reviewed articles. In total, 43 of the data in
MUCE has never been publicly released, making
it unlikely that LLMs have seen the item-response
combinations for the majority of our tasks.
Every response in MUCE was rated for creativ-
ity by at least two raters, and in some cases up to
75 employing a missing-raters design (Forthmann
et al., 2025). While it is common practice to mea-
sure creativity using multiple independent raters,
individual raters may deliver unhelpful or noisy rat-
ings if they did not understand the task instructions,
had a different understanding of the rating criteria,
or for other reasons (Forthmann et al., 2017). To
account for this, we followed best practices for sub-
jective scoring tasks by employing Judge Response
Theory (Myszkowski and Storme, 2019) to check
for raters whose ratings were uninformative in an
information-theoretic sense. We fit JRT models to
each task within MUCE, which gave us an infor-
mation function for each rater across tasks. We
then input the results from the JRT into a genetic
algorithm (Schroeders et al., 2016) which identi-
fied a subset of raters per dataset that maximized
1https: osf.io
the per dataset rater information function.2 This
process dropped uninformative raters from each
dataset, enhancing the quality of the final creativity
ratings. The individual rater s scores were aggre-
gated via factor scores, as is best practice in creativ-
ity assessment (Silvia, 2011), and we rescaled the
factor-transformed creativity scores into the integer
range 10-50 as is done for prior work in automated
creativity assessment (Organisciak et al., 2023).
From this dataset, we create multiple data splits for
training and testing. Full details about the dataset
construction are in Appendix A.
Experiments
5.1
SFT and Preference Datasets
While our MUCE dataset contains samples for mul-
tiple languages, we focus on showing the effec-
tiveness of CRPO on the English subset in this
work and leave experiments using the full dataset
as future work. From the base English MUCE
dataset, we generate a preference dataset by creat-
ing tuples of preferred and rejected responses to the
same prompt, treating the response that received
the higher creativity score as the preferred one. Past
work has shown that data quality is one of the main
factors behind preference model performance (Liu
et al., 2024; Deng et al., 2025; Wang et al., 2024a).
Therefore, we curate a high-quality SFT dataset
of 5, 275 samples (MUCE-SFT) and preference
dataset of 42, 058 samples (MUCE-PREF) from
the base MUCE which we detail in Appendix B.
5.2
Training
Models
As
our
base
models,
we
use
Llama-3.1-8B-Instruct
(AI Meta,
2024)
and Mistral-7B-Instruct-v0.3 (Jiang et al.,
2023) and implement CRPO as described in
Section 3.
We first train our models using
supervised fine-tuning (SFT model) for a single
epoch on MUCE-SFT, and then apply preference
optimization on the SFT model using CRPO and
MUCE-PREF dataset. We train all models using
parameter-efficient tuning with LoRA using a
rank of 128 and an alpha of 256 (Hu et al., 2022).
Additional details on the training setup can be
found in Appendix C.
Creativity Injection
We compute creativity met-
ric scores for each preferred response and inject
2While ensuring that the algorithm kept at least two raters
per dataset.
--- Page 6 ---
quality
0.05
0.06
0.07
0.08
0.09
novelty
quality
0.20
0.25
0.30
0.35
0.40
0.45
diversity
quality
surprise
SFT
DPO
Llama-3.1-8B
Gemini-2.0
GPT-4o
Claude-3.7
CrPO-nov
CrPO-div
CrPO-sur
CrPO-nov-qua
CrPO-div-qua
CrPO-sur-qua
CrPO-qua
CrPO-nov-div-sur
CrPO-cre
Figure 2:
Results on held-out evaluation suite from MUCE across all baselines and our models using
Llama-3.1-8B-Instruct as a base model. nov, div, sur, qua, cre denote novelty, diversity, surprise, qual-
ity, and creativity, respectively. Results are averaged across tasks. Mistral-7B-Instruct-v0.3 results can be
found in Appendix Figure 6.
them into the DPO objective function as described
in Section 3. Since each metric is on a differ-
ent scale and we would like to combine the ef-
fects of different injections, we normalize each
score to a range of [0, 1] before injection.
We
vary the injection weights λd, λn, λs, λq accord-
ingly3 to train different suites of creative mod-
els. As novelty and diversity measures require
a reference set to compute against, we adopt a
prompt-level granularity and consider the set of
responses for a given prompt as the reference cor-
pus similar to prior work (Chung et al., 2025).
We use the jina-embeddings-v3 model (Sturua
et al., 2024) to compute text embeddings for
all metrics that rely on semantic distance. For
surprise, we use instruction-tuned Gemma-2-27B
(Google, 2024a) as our reference surprise model
S. While our creativity preference dataset is al-
ready high-quality, we also experiment with in-
jecting external quality signals to study its inter-
action with other creativity dimensions. Hence,
for the quality measure, we employ an existing re-
ward model Skywork-Reward-Gemma-27B-v0.2
(Liu et al., 2024) that is one of the top-performing
models on RewardBench (Lambert et al., 2024) as
our reference reward model R.
5.3
Evaluation
Tasks and Metrics
We evaluate all models
across several dimensions of creativity on held-out
prompts of various tasks and two held-out tasks.
3For example, to train a novelty model, we set λn 1
and others to 0 whereas for novelty and quality model we set
λn 1 and λq 1.
More specifically, we use 6 held-out prompts from
Real-Life Creative Problem Solving, Alternate Uses
of Objects, Design Solutions, Hypothesis Genera-
tion, and Metaphors tasks, and 9 prompts from two
held-out tasks of Poems and Sentence Completion.
For each prompt, we generate 16 responses from
each model by varying the temperature, topp, and
topk decoding parameters. Our final held-out eval-
uation suite contains 224 samples. We evaluate the
responses on the dimensions of novelty, diversity,
and surprise using the metrics described in Sec-
tion 3. Additionally, to study the tradeoff between
creativity and quality, we train a reward model
on our preference dataset using instruction tuned
Gemma-2-9b (Google, 2024a) and use it to score
the overall quality of model generations. More
details about the evaluation setup can be found in
Appendix D.
Baselines
As baselines, we use the base mod-
els Llama-3.1-8B-Instruct (AI Meta, 2024)
and Mistral-7B-Instruct-v0.3 (Jiang et al.,
2023), SFT models which are the base mod-
els supervised fine-tuned on MUCE-SFT, vanilla
DPO model trained on top of the SFT model
using the MUCE-PREF dataset without any
creativity injections and three closed-source
instruction-tuned LLMs, namely GPT-4o (OpenAI,
2024), Claude-3.7-Sonnet (Anthropic, 2025),
and Gemini-2.0-Flash (Google, 2024b).
CRPO Models
We train several CRPO mod-
els corresponding to the different dimensions of
creativity. More specifically, for each dimension,
we train a model that is injected with a signal for
--- Page 7 ---
0.0
0.5
1.0
1.5
2.0
n
0.088
0.090
0.092
0.094
0.096
0.098
0.100
novelty
0.0
0.5
1.0
1.5
2.0
d
0.36
0.38
0.40
0.42
0.44
0.46
0.48
diversity
0.0
0.5
1.0
1.5
2.0
s
surprise
0.0
0.5
1.0
1.5
2.0
q
5.6
5.4
5.2
5.0
4.8
4.6
4.4
quality
Figure 3: Effect of injection weights for each dimension. Results are averaged across three seed runs.
the given dimension and another model that is in-
jected with a signal for both the given dimension
(e.g. CRPO-nov) and the quality dimension (e.g.
CRPO-nov-qua). We train the latter models to un-
derstand the tradeoff between other dimensions of
creativity and the quality that has been reported in
previous research (Zhang et al., 2025; Lanchantin
et al., 2025; Chung et al., 2025). Additionally, we
train two creative models that inject all dimensions
of creativity (denoted as CRPO-cre) and all ex-
cept quality (denoted as CRPO-nov-div-sur). In
all these experiments, λ injection weights are set
to 1 for simplicity. We perform a more detailed
analysis of these hyperparameters in Section 6.1.
Results
Figure 2 summarizes performance on our held-
out evaluation suite across creativity dimensions
for all baselines and CRPO models using the
Llama-3.1-8B-Instruct as a base. Results for
Mistral-7B-Instruct-v0.3 can be found in Ap-
pendix Figure 6 and follows the same trends. First,
we observe a clear separation between existing
instruction-tuned LLMs and our models: while
the former cluster around high quality but low nov-
elty, diversity, and surprise, our models achieve
high scores across all four dimensions. Second,
for each creativity dimension, the model trained
with that specific injection outperforms others on
the same metric, confirming the effectiveness of
targeted optimization, without a considerable drop
in quality.
Models
that
combine
a
creativity
signal
with
an
external
quality
signal
(CRPO-{nov,div,sur}-qua) improve in quality
but show reduced performance on the targeted di-
mension, illustrating a trade-off. The same pattern
holds when comparing the CRPO-nov-div-sur
model to the full CRPO-cre model, further
highlighting the balance between quality and other
facets of creativity. Interestingly, the vanilla DPO
model, without any creativity injections, already
outperforms existing LLM baselines, demon-
strating the strength of our preference dataset.
Still, most of our creativity-optimized models
significantly surpass DPO across all dimensions.
Finally, the SFT model performs worst in quality
and shows only comparable performance on other
dimensions, reinforcing prior findings (Chung
et al., 2025) about the limited generalizability of
supervised fine-tuning in creative tasks, where no
single correct answer exists.
Overall, our results show that CrPO enhances
multiple aspects of creativity with minimal im-
pact on quality, offering a flexible and effective
framework for creativity alignment in LLMs.
6.1
Effect of Injection Weights
While we set all injection weights to 1 for sim-
plicity in our main evaluations, we also study the
effect of the different injection values on the perfor-
mance of models across dimensions. In particular,
we vary the injection weights from 0 to 2.0 with
an increment of 0.5 for all dimensions and report
the averaged results across three seed runs in Fig-
ure 3. We observe that across most dimensions,
an injection weight of 0.5 yields the greatest per-
formance gains, with further increases resulting in
diminishing returns or slight performance degrada-
tion. In terms of quality, the injection weight of
1.0 results in the highest performance. Neverthe-
less, any weight above 0 consistently outperforms
the model without any injection with minimal drop
in quality (Appendix Figure 8). We suggest tun-
ing these values depending on the training dataset,
underlying task, and the base model for the best
performance.
6.2
Human Evaluation
In addition to automated metrics, we conduct a
human evaluation to assess the real-world effec-
tiveness of our approach. Due to the high cost
--- Page 8 ---
DPO
GPT-4o
Llama-3.1-8B
SFT
Baseline Models
CrPO-cre
CrPO-nov-div-sur
CrPO-nov
CrPO-div
CrPO-sur
Our Models
50.0
43.8
56.2
93.8
56.2
56.2
68.8
100.0
37.5
37.5
37.5
75.0
68.8
37.5
18.8
100.0
43.8
56.2
43.8
93.8
Win Rates ( ) for Human Evaluation - Creativity
Figure 4: Human evaluation results measured by win
rates. Participants were asked to make a pairwise com-
parison between our models and baselines with respect
to the overall creativity.
of human studies, we focus on the overall cre-
ativity dimension using a single task (Sentence
Completion), 4 prompts, 4 baselines (SFT, DPO,
Llama-3.1-8B-Instruct, and GPT-4o), and 5
CRPO variants (nov, div, sur, nov-div-sur,
cre). In a blind pairwise setup, participants com-
pared responses from a baseline and a CRPO
model for creativity, unaware that the texts were
AI-generated. A total of 320 comparisons were
collected with balanced sampling across models.
Additional details are in Appendix D.1.
Figure 4 presents the win rates. The CRPO-
nov-div-sur model consistently outperforms all
baselines, particularly Llama-3.1-8B-Instruct,
by a wide margin. In contrast, the full CRPO-cre
model lags slightly, reflecting the creativity quality
tradeoff seen in automated evaluations. Notably,
CRPO models achieve especially strong gains over
SFT, reinforcing previous findings.
3.00
3.25
3.50
3.75
4.00
4.25
4.50
4.75
5.00
quality
novelty
SFT
DPO
Llama-3.1-8B
Gemini-2.0
GPT-4o
Claude-3.7
CrPO-nov
CrPO-div
CrPO-sur
CrPO-nov-qua
CrPO-div-qua
CrPO-sur-qua
CrPO-qua
CrPO-nov-div-sur
CrPO-cre
Figure 5: Evaluation results on NOVELTYBENCH, using
the novelty and quality metrics defined in Zhang et al.
(2025).
6.3
NOVELTYBENCH Evaluation
While we demonstrate the effectiveness of our ap-
proach on the MUCE held-out set using automated
metrics, we also evaluate generalization on external
benchmarks using the recently introduced NOVEL-
TYBENCH (Zhang et al., 2025). This benchmark
includes tasks spanning randomness, factual knowl-
edge, creative writing, and subjectivity. Following
the recommended evaluation setup, we benchmark
all baselines and CRPO variants on a curated 100-
prompt subset, using the benchmark s novelty and
quality metrics. Full details are in Appendix D.2.
Figure 5 shows novelty vs. quality scores across
all models and tasks.
As in our internal eval-
uation, we observe a clear separation: existing
LLM baselines cluster around lower novelty and
variable quality, while our models consistently
achieve high scores on both dimensions.
No-
tably, although our models outperform SFT on nov-
elty, the SFT model surprisingly achieves higher
quality beating both baselines by a large mar-
gin and our models by a smaller one. This aligns
with findings from NOVELTYBENCH (Zhang et al.,
2025), where smaller models like Gemma-2-2B-it
and Llama-3.1-8B-Instruct often surpass larger
ones in quality.
Overall, our models set a new state-of-the-art
on the NOVELTYBENCH leaderboard in terms of
novelty. 4
Conclusion
We introduce CRPO, a flexible methodology for
enhancing the creativity of LLMs. Leveraging a
novel large-scale human preference dataset focused
on creativity, we show that models aligned with
CRPO produce generations that are not only novel,
diverse, and surprising, but also high in quality
on both our held-out evaluation suite and the
external NOVELTYBENCH dataset. Human evalua-
tions further confirm that raters consistently judge
our model s outputs to be more creative than those
of several strong baselines, highlighting the po-
tential of our approach to boost LLM creativity.
While our experiments focus on smaller models
such as Llama-3.1-8B and an English-only dataset,
future work could explore the scalability of CRPO
to larger models, multilingual settings and other
preference optimization methods.
4https: novelty-bench.github.io
--- Page 9 ---
Limitations
Due to constraints on both computational resources
and budget for human studies, we were unable to
evaluate CRPO on any languages other than En-
glish. Multilingual creativity assessment using gen-
erative AI remains a challenging problem and an
active area of research (Haase et al., 2025). While
we believe our data represents a valuable resource
for the community, future work will need to test
our methods in multilingual settings to ensure mul-
tilingual generalization. These compute constraints
also prevented us from evaluating CRPO on larger
open-weight models, making scaling trends diffi-
cult to predict. We retained only samples with full
agreement for the creativity score when training our
models. While this aligns with best practices for
creativity measurement in psychology (Cseh and
Jeffries, 2019), it may also mask genuine sources of
rater disagreement that should be modeled. Finally,
we acknowledge that, much like other datasets used
to align LLMs, the preferences represented by our
annotator population likely do not reflect the full
range of human preferences, which could bias our
models generations (Yeh et al., 2024). We believe
that the large-scale and multilingual nature of our
preference data likely makes it one of the most rep-
resentative creativity datasets currently available,
but stress that future work should consider issues of
bias and fairness more carefully for LLM creativity
assessment.
Ethical Considerations
We emphasize that our models should not be used
for safety-critical applications, as the relationship
between creativity and alignment with other val-
ues remains underexplored. Notably, our dataset
contains responses to tests of malevolent creativity
that are by definition unsafe for models to generate.
We also observed qualitatively that CRPO mod-
els were more likely to generate unsafe or toxic
responses even to prompts that did not explicitly
request such behaviors. We believe that our data
is valuable for red-teaming evaluations on tasks re-
quiring creativity, and that aligning models on these
malevolent responses could be beneficial for under-
standing how malicious actors might use creativity-
enhanced models to execute unsafe goals. How-
ever, we also acknowledge the ethical concerns that
the release of our models and datasets would raise,
and believe that restricting access to only those
which have signed a license agreement is the best
approach for balancing safety with continued sci-
entific advancement. While we believe our results
demonstrate how aligning LLMs with carefully de-
signed human creativity datasets can significantly
improve the novelty and diversity of their genera-
tions, it remains unclear how to both optimize for
creativity while preserving guardrails that prevent
unsafe behavior.
We also acknowledge the broader debates around
the valid use of AI in social-behavioral research
(Sun et al., 2025) and concerns surrounding AI au-
tomation of industries requiring creativity (Wilkin-
son, 2023) in which our work is situated. While
the over-reliance on AI for creative tasks to the
detriment of human welfare is a legitimate con-
cern, AI has also been acknowledged for its poten-
tial to enhance human creativity above and beyond
what might be possible otherwise (de Chantal et al.,
2025). Creativity is a vital skill for future knowl-
edge workers to master (Forum, 2025), and we
believe that enhancing the creativity of AI is an
important prerequisite for developing AI systems
capable of training humans to be more creative.
Acknowledgements
Mete and Lonneke gratefully acknowledge the sup-
port of the Swiss National Science Foundation
(grant 205121_207437: C - LING). R.E.B. is sup-
ported by grants from the US National Science
Foundation [DRL-1920653; DRL-240078; DUE-
2155070].
References
Sergio Agnoli, Giovanni E Corazza, and Mark A
Runco. 2016. Estimating creativity with a multiple-
measurement approach within scientific and artistic
domains. Creativity Research Journal, 28(2):171
176.
AI Meta. 2024. Llama 3 model card.
Barrett R Anderson, Jash Hemant Shah, and Max
Kreminski. 2024. Homogenization effects of large
language models on human creative ideation.
In
Proceedings of the 16th conference on creativity
cognition, pages 413 425.
Anthropic. 2025. Claude 3.7 sonnet and claude code.
Frank Barron. 1955. The disposition toward original-
ity. The Journal of Abnormal and Social Psychology,
51(3):478.
Roger Beaty, Robert A Cortes, Simone Luchini, John D
Patterson, Boris Forthmann, Brendan S Baker, Bap-
tiste Barbot, Mariale Hardiman, and Adam Green.
--- Page 10 ---
2024. The scientific creative thinking test (sctt): Re-
liability, validity, and automated scoring. PsyArxiv
Preprints.
Roger E Beaty and Dan R Johnson. 2021. Automating
creativity assessment with semdis: An open platform
for computing semantic distance. Behavior research
methods, 53(2):757 780.
Antoine Bellemare-Pepin, François Lespinasse, Philipp
Thölke, Yann Harel, Kory Mathewson, Jay A Ol-
son, Yoshua Bengio, and Karim Jerbi. 2024. Diver-
gent creativity in humans and large language models.
arXiv preprint arXiv:2405.13012.
Margaret A Boden. 2004. The creative mind: Myths
and mechanisms. Routledge.
Tom Brown, Benjamin Mann, Nick Ryder, Melanie
Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind
Neelakantan, Pranav Shyam, Girish Sastry, Amanda
Askell, and 1 others. 2020. Language models are
few-shot learners. Advances in neural information
processing systems, 33:1877 1901.
Sébastien Bubeck, Varun Chandrasekaran, Ronen El-
dan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Pe-
ter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg,
Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro,
and Yi Zhang. 2023. Sparks of artificial general in-
telligence: Early experiments with gpt-4. Preprint,
arXiv:2303.12712.
Razvan C. Bunescu and Oseremen O. Uduehi. 2022.
Distribution-based measures of surprise for creative
language: Experiments with humor and metaphor.
Proceedings of the 3rd Workshop on Figurative Lan-
guage Processing (FLP).
Tuhin Chakrabarty, Philippe Laban, Divyansh Agar-
wal, Smaranda Muresan, and Chien-Sheng Wu. 2024.
Art or artifice? large language models and the false
promise of creativity. In Proceedings of the 2024
CHI Conference on Human Factors in Computing
Systems, pages 1 34.
Soma Chaudhuri, Alan Pickering, and Joydeep Bhat-
tacharya. 2025. Evaluating poetry: Navigating the
divide between aesthetical and creativity judgments.
The Journal of Creative Behavior, 59(1):e683.
Qi Chen, Bowen Zhang, Gang Wang, and Qi Wu.
2024. Weak-eval-strong: Evaluating and eliciting
lateral thinking of llms with situation puzzles. arXiv
preprint arXiv:2410.06733.
John Joon Young Chung, Ece Kamar, and Saleema
Amershi. 2023.
Increasing diversity while main-
taining accuracy: Text data generation with large
language models and human interventions. arXiv
preprint arXiv:2306.04140.
John Joon Young Chung, Vishakh Padmakumar, Melissa
Roemmele, Yuqian Sun, and Max Kreminski.
2025.
Modifying large language model post-
training for diverse creative writing. arXiv preprint
arXiv:2503.17126.
Katherine N Cotter, Jean E Pretz, and James C Kaufman.
2016. Applicant extracurricular involvement predicts
creativity better than traditional admissions factors.
Psychology of Aesthetics, Creativity, and the Arts,
10(1):2.
Genevieve M Cseh and Karl K Jeffries. 2019. A scat-
tered cat: A critical evaluation of the consensual as-
sessment technique for creativity research. Psychol-
ogy of Aesthetics, Creativity, and the Arts, 13(2):159.
Pier Luc de Chantal, Roger Beaty, Antonio Laverghetta,
Jimmy Pronchick, John Patterson, Peter Organisciak,
Katarzyna Potega vel Zabik, Baptiste Barbot, and
Maciej Karwowski. 2025. Artificial intelligence en-
hances human creativity through real-time evaluative
feedback.
Xun Deng, Han Zhong, Rui Ai, Fuli Feng, Zheng Wang,
and Xiangnan He. 2025. Less is more: Improving
llm alignment via preference data selection. arXiv
preprint arXiv:2502.14560.
Paul V DiStefano, John D Patterson, and Roger E Beaty.
2024. Automatic scoring of metaphor creativity with
large language models. Creativity Research Journal,
pages 1 15.
Paul V DiStefano, Daniel Zeitlen, Janet Rafner, Pier-
Luc de Chantal, Aoran Peng, Scarlett Miller, and
Roger Beaty. 2025. Evaluating ai s ideas: The role
of individual creativity and expertise in human-ai
co-creativity.
Angela Fan, Mike Lewis, and Yann Dauphin. 2018.
Hierarchical neural story generation. In Proceedings
of the 56th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers),
pages 889 898, Melbourne, Australia. Association
for Computational Linguistics.
Li Fan, Kaixiang Zhuang, Xueyang Wang, Jingyi Zhang,
Cheng Liu, Jing Gu, and Jiang Qiu. 2023. Explor-
ing the behavioral and neural correlates of seman-
tic distance in creative writing. Psychophysiology,
60(5):e14239.
Boris Forthmann, Benjamin Goecke, and Roger E Beaty.
2025. Planning missing data designs for human rat-
ings in creativity research: A practical guide. Cre-
ativity Research Journal, 37(1):167 178.
Boris Forthmann, Heinz Holling, Nima Zandi, Anne
Gerwig, Pınar Çelik, Martin Storme, and Todd
Lubart. 2017. Missing creativity: The effect of cogni-
tive workload on rater (dis-) agreement in subjective
divergent-thinking scores. Thinking Skills and Cre-
ativity, 23:129 139.
World Economic Forum. 2025. Future of jobs report.
Giorgio Franceschelli and Mirco Musolesi. 2024. Cre-
ative beam search: Llm-as-a-judge for improving re-
sponse generation. arXiv preprint arXiv:2405.00099.
--- Page 11 ---
Bofei Gao, Feifan Song, Yibo Miao, Zefan Cai,
Zhe Yang, Liang Chen, Helan Hu, Runxin Xu,
Qingxiu Dong, Ce Zheng, Wen Xiao, Ge Zhang,
Daoguang Zan, Keming Lu, Bowen Yu, Dayiheng
Liu, Zeyu Cui, Jian Yang, Lei Sha, and 5 others.
2024. Towards a unified view of preference learn-
ing for large language models: A survey. Preprint,
arXiv:2409.02795.
Ken Gilhooly. 2024. Ai vs humans in the aut: Simula-
tions to llms. Journal of Creativity, 34(1):100071.
Benjamin Goecke, Paul V DiStefano, Wolfgang As-
chauer, Kurt Haim, Roger Beaty, and Boris Forth-
mann. 2024a. Automated scoring of scientific cre-
ativity in german. The Journal of Creative Behavior,
58(3):321 327.
Benjamin Goecke, Selina Weiss, and Oliver Wilhelm.
2024b. Driving factors of individual differences in
broad retrieval ability: Gr is more than the sum of its
parts. Journal of Experimental Psychology: Learn-
ing, Memory, and Cognition.
Fabrício Góes, Piotr Sawicki, Marek Grzes, Marco
Volpe, and Jacob Watson. 2023. Pushing gpt s cre-
ativity to its limits: Alternative uses and torrance
tests. In ICCC.
Google. 2024a. Gemma 2: Improving open language
models at a practical size.
Google. 2024b. Introducing gemini 2.0: our new ai
model for the agentic era.
Sylvain Gugger, Lysandre Debut, Thomas Wolf, Philipp
Schmid, Zachary Mueller, Sourab Mangrulkar, Marc
Sun, and Benjamin Bossan. 2022. Accelerate: Train-
ing and inference at scale made simple, efficient and
adaptable.
https: github.com huggingface
accelerate.
J.P. Guilford. 1967. The Nature of Human Intelligence.
McGraw-Hill series in psychology. McGraw-Hill.
Jennifer Haase, Paul H. P. Hanel, and Sebastian Pokutta.
2025. S-dat: A multilingual, genai-driven frame-
work for automated divergent thinking assessment.
Preprint, arXiv:2505.09068.
Shirley Anugrah Hayati, Minhwa Lee, Dheeraj Ra-
jagopal, and Dongyeop Kang. 2023. How far can
we extract diverse perspectives from large language
models? arXiv preprint arXiv:2311.09799.
Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2023.
Debertav3: Improving deberta using electra-style pre-
training with gradient-disentangled embedding shar-
ing. Preprint, arXiv:2111.09543.
Ruizhi He, Kaixiang Zhuang, Lijun Liu, Ke Ding,
Xi Wang, Lei Fu, Jiang Qiu, and Qunlin Chen. 2022.
The impact of knowledge on poetry composition: An
fmri investigation. Brain and language, 235:105202.
Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and
Yejin Choi. 2019. The curious case of neural text
degeneration. arXiv preprint arXiv:1904.09751.
Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan
Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang,
Weizhu Chen, and 1 others. 2022. Lora: Low-rank
adaptation of large language models. ICLR, 1(2):3.
Shulin Huang, Shirong Ma, Yinghui Li, Mengzuo
Huang, Wuhe Zou, Weidong Zhang, and Haitao
Zheng. 2024. LatEval: An interactive LLMs evalu-
ation benchmark with incomplete information from
lateral thinking puzzles. In Proceedings of the 2024
Joint International Conference on Computational
Linguistics, Language Resources and Evaluation
(LREC-COLING 2024), pages 10186 10197, Torino,
Italia. ELRA and ICCL.
Mete Ismayilzada, Debjit Paul, Antoine Bosselut,
and Lonneke van der Plas. 2024a.
Creativity in
ai:
Progresses and challenges.
arXiv preprint
arXiv:2410.17218.
Mete Ismayilzada, Claire Stevenson, and Lonneke
van der Plas. 2024b. Evaluating creative short story
generation in humans and large language models.
arXiv preprint arXiv:2411.02316.
Albert Q. Jiang, Alexandre Sablayrolles, Arthur Men-
sch, Chris Bamford, Devendra Singh Chaplot, Diego
de las Casas, Florian Bressand, Gianna Lengyel, Guil-
laume Lample, Lucile Saulnier, Lélio Renard Lavaud,
Marie-Anne Lachaux, Pierre Stock, Teven Le Scao,
Thibaut Lavril, Thomas Wang, Timothée Lacroix,
and William El Sayed. 2023. Mistral 7b. Preprint,
arXiv:2310.06825.
Dan R Johnson, Andrew S Cuthbert, and Mara E Tynan.
2021. The neglect of idea diversity in creative idea
generation and evaluation. Psychology of Aesthetics,
Creativity, and the Arts, 15(1):125.
Dan Richard Johnson, J. Kaufman, Brendan S. Baker,
John D. Patterson, Baptiste Barbot, Adam E. Green,
Janet G. van Hell, Evan S. Kennedy, Grace F Sulli-
van, Christa L. Taylor, Thomas Ward, and Roger E.
Beaty. 2022. Divergent semantic integration (dsi):
Extracting creativity from narratives with distribu-
tional semantic modeling. Behavior Research Meth-
ods, 55:3726 3759.
Hansika Kapoor, Hreem Mahadeshwar, Sarah Rezaei,
Roni Reiter-Palmon, and James C Kaufman. 2024.
The ties that bind: Low morals, high deception, and
dark creativity. Creativity Research Journal, pages
1 20.
Pythagoras Karampiperis, Antonis Koukourikos, and
Evangelia Koliopoulou. 2014. Towards machines for
measuring creativity: The use of computational tools
in storytelling activities. In 2014 IEEE 14th Interna-
tional Conference on Advanced Learning Technolo-
gies, pages 508 512.
--- Page 12 ---
Robert Kirk, Ishita Mediratta, Christoforos Nalmpantis,
Jelena Luketina, Eric Hambro, Edward Grefenstette,
and Roberta Raileanu. 2023. Understanding the ef-
fects of rlhf on llm generalisation and diversity. arXiv
preprint arXiv:2310.06452.
Mika Koivisto and Simone Grassini. 2023. Best hu-
mans still outperform artificial intelligence in a cre-
ative divergent thinking task.
Scientific reports,
13(1):13601.
Polina Kuznetsova, Jianfu Chen, and Yejin Choi. 2013.
Understanding and quantifying creativity in lexical
composition. In Conference on Empirical Methods
in Natural Language Processing.
Nathan Lambert, Valentina Pyatkin, Jacob Morrison,
LJ Miranda, Bill Yuchen Lin, Khyathi Chandu,
Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi,
and 1 others. 2024. Rewardbench: Evaluating re-
ward models for language modeling. arXiv preprint
arXiv:2403.13787.
Jack Lanchantin, Angelica Chen, Shehzaad Dhuliawala,
Ping Yu, Jason Weston, Sainbayar Sukhbaatar, and
Ilia Kulikov. 2025. Diverse preference optimization.
arXiv preprint arXiv:2501.18101.
Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Ju-
jie He, Chaojie Wang, Shuicheng Yan, Yang Liu,
and Yahui Zhou. 2024. Skywork-reward: Bag of
tricks for reward modeling in llms. arXiv preprint
arXiv:2410.18451.
Ximing Lu, Melanie Sclar, Skyler Hallinan, Niloofar
Mireshghallah, Jiacheng Liu, Seungju Han, Allyson
Ettinger, Liwei Jiang, Khyathi Chandu, Nouha Dziri,
and 1 others. 2024. Ai as humanity s salieri: Quan-
tifying linguistic creativity of language models via
systematic attribution of machine text against web
text. arXiv preprint arXiv:2410.04265.
Simone A Luchini, Nadine T Maliakkal, Paul V DiS-
tefano, Antonio Laverghetta Jr, John D Patterson,
Roger E Beaty, and Roni Reiter-Palmon. 2025. Auto-
mated scoring of creative problem solving with large
language models: A comparison of originality and
quality ratings. Psychology of Aesthetics, Creativity,
and the Arts.
Pronita Mehrotra, Aishni Parab, and Sumit Gulwani.
2024. Enhancing creativity in large language mod-
els through associative thinking strategies. arXiv
preprint arXiv:2405.06715.
Clara Meister, Tiago Pimentel, Gian Wiher, and Ryan
Cotterell. 2023. Locally typical sampling. Transac-
tions of the Association for Computational Linguis-
tics, 11:102 121.
Alireza Modirshanechi, Johanni Brea, and Wulfram
Gerstner. 2022. A taxonomy of surprise definitions.
Journal of Mathematical Psychology, 110:102712.
Nils Myszkowski and Martin Storme. 2019. Judge re-
sponse theory? a call to upgrade our psychometrical
account of creativity judgments. Psychology of Aes-
thetics, Creativity, and the Arts, 13(2):167.
Lakshmi Nair, Evana Gizzi, and Jivko Sinapov. 2024.
Creative problem solving in large language and vi-
sion models - what would it take? In Findings of the
Association for Computational Linguistics: EMNLP
2024, pages 11978 11994, Miami, Florida, USA.
Association for Computational Linguistics.
OpenAI. 2024.
Gpt-4o system card.
Preprint,
arXiv:2410.21276.
Peter Organisciak, Selcuk Acar, Denis Dumas, and
Kelly Berthiaume. 2023. Beyond semantic distance:
Automated scoring of divergent thinking greatly im-
proves with large language models. Thinking Skills
and Creativity, 49:101356.
Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida,
Carroll Wainwright, Pamela Mishkin, Chong Zhang,
Sandhini Agarwal, Katarina Slama, Alex Ray, and 1
others. 2022. Training language models to follow in-
structions with human feedback. Advances in neural
information processing systems, 35:27730 27744.
Laura O Mahony, Leo Grinsztajn, Hailey Schoelkopf,
and Stella Biderman. 2024. Attributing mode col-
lapse in the fine-tuning of large language models. In
ICLR 2024 Workshop on Mathematical and Empiri-
cal Understanding of Foundation Models.
Vishakh Padmakumar and He He. 2023. Does writ-
ing with language models reduce content diversity?
arXiv preprint arXiv:2309.05196.
John D Patterson, Hannah M Merseal, Dan R Johnson,
Sergio Agnoli, Matthijs Baas, Brendan S Baker, Bap-
tiste Barbot, Mathias Benedek, Khatereh Borhani,
Qunlin Chen, and 1 others. 2023. Multilingual se-
mantic distance: Automatic verbal creativity assess-
ment in many languages. Psychology of Aesthetics,
Creativity, and the Arts, 17(4):495.
Corinna Perchtold-Stefan, Hansika Kapoor, James C
Kaufman, Hreem Mahadeshwar, and Alison Fernan-
des. 2024. Development and neuronal validation of
the dark creativity deception battery (dcdb).
Corinna M Perchtold-Stefan, Christian Rominger, Ilona
Papousek, and Andreas Fink. 2023. Functional eeg
alpha activation patterns during malevolent creativity.
Neuroscience, 522:98 108.
Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo-
pher D Manning, Stefano Ermon, and Chelsea Finn.
2023.
Direct preference optimization: Your lan-
guage model is secretly a reward model. Advances in
Neural Information Processing Systems, 36:53728
53741.
Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase,
and Yuxiong He. 2020. Zero: Memory optimizations
toward training trillion parameter models. In SC20:
--- Page 13 ---
International Conference for High Performance Com-
puting, Networking, Storage and Analysis, pages 1
16. IEEE.
Tuval Raz, Simone Luchini, Roger Beaty, and Yoed
Kenett. 2024. Bridging the measurement gap: A
large language model method of assessing open-
ended question complexity. In Proceedings of the
Annual Meeting of the Cognitive Science Society, vol-
ume 46.
Mark A Runco and Garrett J Jaeger. 2012. The standard
definition of creativity. Creativity research journal,
24(1):92 96.
Solve Sæbø and Helge Brovold. 2024. On the stochas-
tics of human and artificial creativity. arXiv preprint
arXiv:2403.06996.
Keita Saito, Akifumi Wachi, Koki Wataoka, and
Youhei Akimoto. 2023.
Verbosity bias in prefer-
ence labeling by large language models. Preprint,
arXiv:2310.10076.
Janika Saretzki, Rosalie Andrae, Boris Forthmann, and
Mathias Benedek. 2024. Investigation of response ag-
gregation methods in divergent thinking assessments.
The Journal of Creative Behavior.
Ulrich Schroeders, Oliver Wilhelm, and Gabriel Olaru.
2016. Meta-heuristics in short scale construction:
Ant colony optimization and genetic algorithm. PloS
one, 11(11):e0167110.
Alexander Shypula, Shuo Li, Botong Zhang, Vishakh
Padmakumar, Kayo Yin, and Osbert Bastani. 2025.
Evaluating the diversity and quality of llm generated
content. arXiv preprint arXiv:2504.12522.
Paul J Silvia. 2011. Subjective scoring of divergent
thinking: Examining the reliability of unusual uses,
instances, and consequences tasks. Thinking Skills
and Creativity, 6(1):24 30.
Dean Keith Simonton. 2012. Taking the us patent of-
fice criteria seriously: A quantitative three-criterion
creativity definition and its implications. Creativity
research journal, 24(2-3):97 106.
Dean Keith Simonton. 2018. Defining creativity: Don t
we also need to define what is not creative?
The
Journal of Creative Behavior, 52(1):80 90.
Morris I Stein. 2014. Stimulating creativity: Individual
procedures. Academic Press.
Claire E. Stevenson, Iris Smal, Matthijs Baas, Raoul
Grasman, and Han L. J. van der Maas. 2022. Putting
gpt-3 s creativity to the (alternative uses) test. In
ICCC.
Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram,
Michael Günther, Bo Wang, Markus Krimmel, Feng
Wang, Georgios Mastrapas, Andreas Koukounas, An-
dreas Koukounas, Nan Wang, and Han Xiao. 2024.
jina-embeddings-v3: Multilingual embeddings with
task lora. Preprint, arXiv:2409.10173.
Douglas Summers-Stay, Stephanie M. Lukin, and
Clare R. Voss. 2023. Brainstorm, then select: a gen-
erative language model improves its creativity score.
Huaman Sun, Jiaxin Pei, Minje Choi, and David Jur-
gens. 2025. Sociodemographic prompting is not yet
an effective approach for simulating subjective judg-
ments with llms. In Proceedings of the 2025 Confer-
ence of the Nations of the Americas Chapter of the
Association for Computational Linguistics: Human
Language Technologies (Volume 2: Short Papers),
pages 845 854.
Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-
Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan
Schalkwyk, Andrew M Dai, Anja Hauth, Katie Mil-
lican, and 1 others. 2023.
Gemini: a family of
highly capable multimodal models. arXiv preprint
arXiv:2312.11805.
Yufei Tian, Tenghao Huang, Miri Liu, Derek Jiang,
Alexander Spangher, Muhao Chen, Jonathan May,
and Nanyun Peng. 2024. Are large language models
capable of generating human-level narratives? arXiv
preprint arXiv:2407.13248.
Yufei Tian, Abhilasha Ravichander, Lianhui Qin, Ro-
nan Le Bras, Raja Marjieh, Nanyun Peng, Yejin Choi,
Thomas L Griffiths, and Faeze Brahman. 2023. Mac-
gyver: Are large language models creative problem
solvers? arXiv preprint arXiv:2311.09682.
Binghai Wang, Rui Zheng, Lu Chen, Zhiheng Xi, Wei
Shen, Yuhao Zhou, Dong Yan, Tao Gui, Qi Zhang,
and Xuan-Jing Huang. 2024a.
Reward modeling
requires automatic adjustment based on data quality.
In Findings of the Association for Computational
Linguistics: EMNLP 2024, pages 4041 4064.
Tiannan Wang, Jiamin Chen, Qingrui Jia, Shuai Wang,
Ruoyu Fang, Huilin Wang, Zhaowei Gao, Chunzhao
Xie, Chuou Xu, Jihong Dai, and 1 others. 2024b.
Weaver: Foundation models for creative writing.
arXiv preprint arXiv:2401.17268.
Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel,
Barret Zoph, Sebastian Borgeaud, Dani Yogatama,
Maarten Bosma, Denny Zhou, Donald Metzler, and
1 others. 2022. Emergent abilities of large language
models. arXiv preprint arXiv:2206.07682.
Selina Weiss, Benjamin Goecke, and Oliver Wilhelm.
2024. How much retrieval ability is in originality?
The Journal of Creative Behavior, 58(3):370 387.
Selina Weiss, Sally Olderbak, and Oliver Wilhelm. 2023.
Conceptualizing and measuring ability emotional cre-
ativity. Psychology of Aesthetics, Creativity, and the
Arts.
Emily Wenger and Yoed Kenett. 2025. We re different,
we re the same: Creative homogeneity across llms.
arXiv preprint arXiv:2501.19361.
Peter West and Christopher Potts. 2025. Base models
beat aligned models at randomness and creativity.
Preprint, arXiv:2505.00047.
--- Page 14 ---
Alissa Wilkinson. 2023. Hollywood s writers are on
strike. here s why that matters.
Justin Wong, Yury Orlovskiy, Michael Luo, Sanjit A
Seshia, and Joseph E Gonzalez. 2024. Simplestrat:
Diversifying language model generation with stratifi-
cation. arXiv preprint arXiv:2410.09038.
Weijia Xu, Nebojsa Jojic, Sudha Rao, Chris Brockett,
and Bill Dolan. 2024. Echoes in ai: Quantifying
lack of plot diversity in llm outputs. arXiv preprint
arXiv:2501.00273.
Min-Hsuan Yeh, Leitian Tao, Jeffrey Wang, Xuefeng
Du, and Yixuan Li. 2024. How reliable is human
feedback for aligning large language models? arXiv
preprint arXiv:2410.01957.
Yuhua Yu, Lindsay Krebs, Mark Beeman, and Vicky T
Lai. 2024. Exploring how generating metaphor via
insight versus analysis affects metaphor quality and
learning outcomes. Cognitive science, 48(8):e13488.
Yiming Zhang, Harshita Diddee, Susan Holm, Hanchen
Liu, Xinyue Liu, Vinay Samuel, Barry Wang, and
Daphne Ippolito. 2025. Noveltybench: Evaluating
creativity and diversity in language models. arXiv
preprint arXiv:2504.05228.
Yiming Zhang, Avi Schwarzschild, Nicholas Carlini,
Zico Kolter, and Daphne Ippolito. 2024. Forcing
diffuse distributions out of language models. arXiv
preprint arXiv:2404.10859.
Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang,
Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen
Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen
Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang,
Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, and
3 others. 2025. A survey of large language models.
Preprint, arXiv:2303.18223.
Yunpu Zhao, Rui Zhang, Wenyi Li, Di Huang, Jiaming
Guo, Shaohui Peng, Yifan Hao, Yuanbo Wen, Xing
Hu, Zidong Du, and 1 others. 2024. Assessing and
understanding creativity in large language models.
arXiv preprint arXiv:2401.12491.
Kuan Lok Zhou, Jiayi Chen, Siddharth Suresh, Reuben
Narad, Timothy T Rogers, Lalit K Jain, Robert D
Nowak, Bob Mankoff, and Jifan Zhang. 2025. Bridg-
ing the creativity understanding gap: Small-scale
human alignment enables expert-level humor ranking
in llms. arXiv preprint arXiv:2502.20356.
Aleksandra Zieli nska, Peter Organisciak, Denis Du-
mas, and Maciej Karwowski. 2023. Lost in trans-
lation? not for large language models: Automated
divergent thinking scoring performance translates to
non-english contexts. Thinking Skills and Creativity,
50:101414.
A
MUCE Dataset
We compiled data by means of crowdsourcing and
data mining of the open-source data sharing plat-
form OSF. We crowdsourced from the global cre-
ativity research community by means of direct re-
quests and posts on academic listservs. In our call
for data-sharing, we requested data relating to any
creativity responses that were provided by human
participants and scored for creativity by human
raters. We specifically requested that the datasets
include scores from each rater, rather than com-
posite creativity scores, to determine rating data
quality for each submission. As part of our inclu-
sion criteria, we further requested that researchers
provide information relating to: (a) the creativity
task, (b) the item associated with each response, (c)
the construct that was rated, and (d) the language
of the task. We further asked researchers to provide
a statement on whether they agreed to making their
data open-source. In terms of data mining through
the OSF platform, we first searched through a se-
ries of relevant keywords (e.g., creativity task ,
originality score ). We only retained sub-datasets
from credible sources, which were associated with
a citable peer-reviewed article, and which included
all the required data relating to our inclusion crite-
ria.
After removing responses that didn t meet our
inclusion criteria, our dataset amounted to 321,572
human-rated and language-based creativity re-
sponses. The dataset was thus cleaned by standard-
izing the naming for each variable except for the
responses. We then removed responses for having
been rated by fewer than 2 human judges. Dupli-
cate responses were also removed, by retaining a
single exemplar for responses that appeared twice
within a specific item and task.
To enhance the reliability of human creativity
ratings across the numerous datasets, we optimized
the selection of raters by applying a meta-heuristic
algorithm. Specifically, we applied a Genetic Algo-
rithm (Schroeders et al., 2016). The GA operates
through iterative selection, crossover, and muta-
tion processes, mirroring the principles of natural
selection, and in our case to identify the optimal
subsets of raters for each dataset. In each itera-
tion, candidate solutions that is, combinations of
raters were evaluated based on a predefined fit-
ness function that prioritized the maximization of
empirical reliability (rxx) within a graded response
model (GRM) and hence in line to judge response
--- Page 15 ---
theory. For sub-datasets involving decimal-based
scales, individual ratings were rounded to the near-
est integer value (rounding up if containing a deci-
mal .5) to meet the requirements of the GRM.
Rater subsets demonstrating superior reliability
were selected, recombined, and modified through
random perturbations to prevent premature con-
vergence to suboptimal solutions. This approach
ensured that the selected raters provided consistent
and informative judgments while reducing noise
introduced by inconsistent or uninformative ratings.
By automating the selection process through GA,
we opted for maximal comparability in the selec-
tion process across datasets. Previous research has
demonstrated the utility of GA in psychometric op-
timization tasks, particularly in balancing brevity
and measurement precision while maintaining con-
struct validity. In the present study, GA facilitated
a systematic and data-driven refinement of rater
selection, arguably enhancing the overall quality of
creativity ratings.
After dropping uninformative raters in each sub-
dataset, we again removed any rows containing less
than 2 ratings due to rater removal. Afterwards, we
used the new rater subsets per dataset and computed
factor scores for each given response that were used
as creativity scores. We calculated factor scores
via a GRM model, ran separately over each sub-
dataset, to derive a single creativity score for each
response. Finally, we applied min-max scaling on
each sub-dataset to transform ratings into a range of
10 to 50, with intervals of 1. This step was applied
to ensure that ratings would only constitute a single
token in length, to lessen the burden of predicting
multi-token labels by the LLMs.
We then withheld all responses in the Spanish
language from our final dataset and assigned them
to an out-of-distribution-language (OOD-l) set. Re-
sponses from the OOD-l set were not included in
the training data of MUCE, allowing us to test
whether the model could generalize to creative re-
sponses in an unseen language. We selected Span-
ish as it would allow for a fair test of generalizabil-
ity given: (1) Spanish tends to be a high-resource
language within the pre-training of modern LLMs,
(2) it is similar to other Latin-root languages in our
training data (e.g., Italian), (3) responses in Span-
ish spanned multiple creativity tasks, and (4) the
language spanned a limited number of responses in
our total dataset. We further withheld all responses
from two highly-naturalistic tasks, the Poem and
Alternative Title Generation, and assigned these
to an out-of-distribution task (OOD-t) set. We se-
lected these tasks as they made up a limited portion
of the total dataset and would provide a test of
MUCE s performance on unseen naturalistic cre-
ativity tasks.
We then randomly selected items within each
task and assigned them to an out-of-distribution
item (OOD-i) set. We identified candidate items
that corresponded to 5 or less of the responses
within a task. Then, for tasks that contained 20 or
more total items, we randomly assigned 2 of these
items to our OOD-i set. For tasks that contained
fewer than 20 total items, we instead randomly as-
signed 1 of these items to the OOD-i set. Finally,
we split the remaining responses in our dataset
into training, validation, and out-of-distribution re-
sponses (OOD-r) sets according to an 80 10 10
split. We grouped responses into unique combina-
tions of sub-dataset, task, language, item, and rat-
ing label, then randomly assigned responses within
each combination to each of the sets, ensuring an
equal representation of responses associated with
each of these variables within the training, valida-
tion, and OOD-r sets. Table 1 contains the final
dataset statistics for MUCE. Tables 6 and 7 contain
the descriptions and data statistics for each task in
MUCE. Tables 8, 9, 10, 11, and 12 list some exam-
ple prompts and low-rated and high-rated responses
for each task from MUCE.
B
SFT and Preference Datasets
Past work has shown that data quality is one of
the main factors behind preference model perfor-
mance (Liu et al., 2024; Deng et al., 2025; Wang
et al., 2024a). In particular, the margin in the score
(i.e. reward margin) between the preferred and re-
jected response may influence the performance of
the model, since training pairs with smaller mar-
gins are likely to contain annotation noise and be
more difficult to learn. We experiment with dif-
ferent reward margins and choose a margin of 5
for the final experiments as it showed a balance
between mitigating annotator noise and creating
a dataset with nuanced preferences. Additionally,
to ensure a high-quality preference dataset, first
we filter the base MUCE dataset and select only
the samples that have a full agreement from all
annotators. Then we filter out all samples that
have a rating below 20 and limit the number of
pairings between samples to 10. This results in
a final preference training dataset of 42, 058 sam-
--- Page 16 ---
quality
0.05
0.06
0.07
0.08
novelty
quality
0.20
0.25
0.30
0.35
0.40
0.45
diversity
quality
surprise
SFT
DPO
Mistral-7B
Gemini-2.0
GPT-4o
Claude-3.7
CrPO-nov
CrPO-div
CrPO-sur
CrPO-nov-qua
CrPO-div-qua
CrPO-sur-qua
CrPO-qua
CrPO-nov-div-sur
CrPO-cre
Figure 6:
Results on held-out evaluation suite from MUCE across all baselines and our models using
Mistral-7B-Instruct-v0.3 as a base model. nov, div, sur, qua, cre denote novelty, diversity, surprise, quality,
and creativity, respectively. Results are averaged across tasks.
Total
Train
Dev
Test
OOD-i
OOD-l
OOD-t
samples
245,030
183,973
23,254
22,419
6,253
4,719
4,412
tasks
languages
prompts
Table 1: Detailed statistics for each split of MUCE.
Human Evaluation Instructions
In this study, you will be presented with two
responses to a creative task. Your job is to
select the response that you believe is the
most creative. Please base your judgment
only on the creativity of the ideas not on
how long or detailed the response is. A
shorter response can be more creative than
a longer one, and vice versa. Focus on how
original, unique, and innovative the idea
feels to you. There are no right or wrong
answers we re interested in your opinion.
Figure 7: Rater instructions for the human evaluation.
ples (MUCE-PREF). We also create a high-quality
instruction-tuning dataset from MUCE-PREF by
pairing the prompts with all preferred responses
that have a rating above 30 resulting in a dataset
of 5, 275 samples (MUCE-SFT). Tables 2 and 3
contain the statistics for these datasets.
C
Training
We follow a training setup similar to Chung
et al. (2025) and use Llama-3.1-8B-Instruct
and Mistral-7B-Instruct-v0.3 (Jiang et al.,
2023) as our base models. Using these models,
we train an SFT, DPO and several CRPO models.
We train all models using parameter-efficient tun-
ing with LoRA using a rank of 128 and an alpha
of 256 (Hu et al., 2022). All training was done
using HuggingFace TRL library5 with Accelerate
(Gugger et al., 2022) and DeepSpeed ZeRO-2 (Ra-
jbhandari et al., 2020) on NVIDIA A100 GPUs with
gradient checkpointing.
SFT model is trained on the MUCE-SFT dataset
for a single epoch with a batch size of 2 per GPU
using a gradient accumulation size of 4 and context
size of 1024. We use a cosine scheduler with a
half-cycle warmup and maximum learning rate of
3e 5. Final model achieves 85 mean token
accuracy on the validation set.
DPO and CRPO models are trained using the
SFT model as a base on our MUCE-PREF dataset
for a single epoch with a batch size of 8 per GPU
using a gradient accumulation size of 8 and context
size of 1024. We use a linear scheduler with a
learning rate of 5e 6. All final models achieve
over 82 reward accuracy on the validation set.
5https: huggingface.co docs trl en index
--- Page 17 ---
30.5
31.0
31.5
32.0
32.5
33.0
33.5
34.0
34.5
quality
0.090
0.092
0.094
0.096
0.098
novelty
CrPO-nov with different injection weights
lambda
0.5
1.0
1.5
2.0
0.0
quality
0.36
0.38
0.40
0.42
0.44
0.46
diversity
CrPO-div with different injection weights
lambda
0.5
1.0
1.5
2.0
0.0
quality
surprise
CrPO-sur with different injection weights
lambda
0.5
1.0
1.5
2.0
0.0
Figure 8: Effect of injection weights for each dimension on the quality score. Results are averaged across three seed
runs.
D
Evaluation
For each prompt in our held-out evaluation suite,
we generate a total of 16 responses for every model
by sampling 4 responses for each of the following
four decoding setups that induce high randomness
using various sampling techniques (Fan et al., 2018;
Holtzman et al., 2019):
1. temperature 0.7, topp 0.95
2. temperature 0.9, topp 0.99
3. temperature 0.7, topk 50
4. temperature 0.8, topp 0.97
Moreover, as the existing instruction-tuned LLMs
tend to produce verbose outputs (Saito et al., 2023),
in order to minimize the length bias, we add further
instructions in the prompt, constraining the output
length in terms of the number of sentences and
words. We compute the constraint values based
on the median number of words and sentences of
responses per task from our training dataset. Table
4 lists an example evaluation prompt for each task.
Table 5 lists an example response from all models
to a single prompt.
D.1
Human Evaluation
Since we have multiple model responses per
prompt, instead of randomly choosing a response,
for each prompt, we choose top 4 model responses
measured by the overall automated creativity score
which we define as the sum of normalized novelty,
diversity, surprise and quality scores. This setup en-
sures that models are compared to each other with
their best outputs. We recruited 15 participants on
Prolific6 to complete the study, requiring that they
reside in the U.S. and have an approval rating of
at least 90 . Ethics board approval was received
from the Pennsylvania State University IRB for this
study. We provided participants with a definition of
creativity, and instructed them not to focus on the
length or detail of the response when rating. Figure
7 lists the instructions given to raters for evaluating
creativity. We additionally included a comprehen-
sion check where participants were quizzed about
the task instructions, to help catch careless partici-
pants. Raters who failed this check were excluded
from further analysis. All raters were compensated
adequately with at least a minimum payment of 9
per hour. Final win rates are calculated for each
response pair based on the majority vote across
participants. The inter-rater agreement computed
using Krippendorff s alpha was 0.463, indicating a
moderate agreement.
D.2
NOVELTYBENCH Evaluation
NOVELTYBENCH is a recently introduced bench-
mark to measure how well language models can
generate novel and high-quality answers to user re-
quests involving subjectivity, randomness, and cre-
ativity (Zhang et al., 2025). We use a 100-sample
subset of their benchmark that is manually curated
by the authors and contains four distinct categories
where diversity and novelty are expected:
Randomness: prompts that involve random-
izing over a set of options. Example: Roll a
make-believe 20-sided die.
Factual Knowledge: prompts that request un-
derspecified factual information, which allow
6https: www.prolific.com
--- Page 18 ---
Task
prompts
samples
Real-Life
Creative
Problem
Solving
5,601
Question
Asking
Malevolent
Problems
Metaphors
Alternate
Uses
of
Objects Task
4,388
Design Solu-
tions
1,366
Essays
Stories
1,498
Consequences 5
10,865
Experiment
Design
5,640
Hypothesis
Generation
5,260
Research
Questions
5,832
Associations
Total
42,058
Table 2: MUCE-PREF training dataset details.
many valid answers. Example: List a capital
city in Africa.
Creative Writing: prompts that involve gen-
erating a creative form of text, including po-
etry, and story-writing. Example: Tell me a
riddle.
Subjectivity: prompts that request subjective
answers or opinions. Example: What s the
best car to get in 2023?
Additionally, the paper proposes new metrics to
measure novelty and quality (i.e. utility) that are
different than ours. To compute novelty, they pro-
pose a method that learns to partition the output
space into equivalence classes from human annota-
tions. Each class represents one unique generation
that is roughly equivalent to the others in the same
class and different from the generations in other
classes. They consider a functional equivalence
that defines two generations to be different if and
only if a user who has seen one generation would
Task
prompts
samples
Real-Life
Creative
Problem
Solving
Question
Asking
Malevolent
Problems
Metaphors
Alternate
Uses
of
Objects Task
Design Solu-
tions
Essays
Stories
Consequences 5
1,315
Experiment
Design
Hypothesis
Generation
Research
Questions
Associations
Total
5,275
Table 3: MUCE-SFT training dataset details.
likely benefit from seeing the other. To this end, the
authors annotated 1,100 pairs of generations condi-
tioned on prompts from NOVELTYBENCH sampled
from a diverse set of models. From these annotated
pairs, they used 1,000 for training and fine-tuned a
deberta-v3-large model (He et al., 2023) to predict
binary functional equivalence between two genera-
tions. With the equivalence classifier, they partition
the output space into equivalence classes. Then
they define the novelty as the distinctk metric that
is the number of equivalence classes in a partition
of k sample generations from a language model:
distinctk : {ci i [k]}
(8)
To compute quality, they consider a model of
user behavior that describes how users interact with
and consume language model generations. They
assume that the user has a patience level p [0, 1]:
after observing each additional generation, they
have a probability p of requesting an additional
generation from the language model and observing
the next generation, and a probability 1 p of
--- Page 19 ---
Task
Prompt
Real-Life Creative Problem Solving
Come up with an original and creative solution for
the following real-world problem: Clara, a junior
pre-med student, is working part-time and taking a
15 hour credit load at school. ... skipped ... Please
limit your response to 4 sentences and at most 75
words.
Alternate Uses of Objects
Come up with an original and creative use for the
following object: rope. Please limit your response to
1 sentence and at most 17 words.
Design Solutions
Come up with an original and creative solution to
reduce the amount of litter in public spaces and pro-
mote waste reduction and recycling. Please limit
your response to 2 sentences and at most 36 words.
Hypothesis Generation
Come up with an original and creative scientific hy-
pothesis for the following scenario: You notice that
dogs seem to like one of your friends, but cats seem
to like another friend. What hypotheses do you have
about why that is? Please limit your response to 1
sentence and at most 22 words.
Metaphors
Come up with an original and creative metaphoric
equivalent for the concept described below: Stomata
are tiny openings or pores found on the underside of a
plant leaf. They are used for gas exchange, enabling
the intake of carbon dioxide and release of oxygen..
Please limit your response to 1 sentence and at most
10 words.
Poems
Come up with an original and creative poem about
the following concept: choice. Please limit your
response to 5 sentences and at most 150 words.
Sentence Completion
Finish the sentence with an original and creative end-
ing: When I got on the school bus....Please respond
in one sentence.
Table 4: MUCE dataset held-out item and task evaluation prompts.
stopping interacting with the model. Then they
compute the quality of a sequence of generations
as the cumulative utility:
utilityk : 1 p
1 pk
k
X
i 1
pi 1 1[ci cj, j i] ui
(9)
To
compute
the
utility
of
individ-
ual
generations,
they
also
use
the
Skywork-Reward-Gemma-2-27B-v0.2
(Liu
et al., 2024) model.
To benchmark our models, we follow their rec-
ommended setup for evaluation. In particular, we
set the number of generations to 10 per model and
the patience level to 0.8 and use their trained clas-
sifier for output space partition.
--- Page 20 ---
Model
Response
Human
played hide-and-seek by forming into different ob-
jects and trying not to be found.
Llama-3.1-8B-Instruct
At a sleepover, we transformed our host s base-
ment into a mystical Dreamscape where we wove a
tapestry of secrets, stories, and whispered promises
that only the darkness could keep.
Gemini-2.0-Flash
At a sleepover we...built a pillow fort so magnifi-
cent, it accidentally opened a portal to a realm where
dreams manifested as sentient, fluffy creatures de-
manding to be tickled.
Claude-3.7-Sonnet
At a sleepover we constructed an elaborate blanket
fortress complete with secret passages, flashlight con-
stellations on the ceiling, and an unspoken pact to
guard each other s midnight confessions until the end
of time.
GPT-4o
At a sleepover we transformed the living room into
a magical fort of pillows and fairy lights, where we
whispered secrets and told stories until the first light
of dawn.
SFT
We played a game of Twister where we put out our
phones to see who was the best twister.
DPO
conducted an experiment to see if a person s person-
ality is changed with an altered state of conscious-
ness.
CRPO-nov
tried to break the record for most consecutive hours
without speaking and we discovered we could all
hear each other s thoughts.
CRPO-nov-qua
Participated in an experiment where we tested the
effects of sleep deprivation on the human mind.
CRPO-div
construct a space shuttle that takes us to the moon
and from there we can launch our dream rockets.
CRPO-div-qua
recreated the conditions of a 19th century under-
ground railroad and had to map out our escape to
Canada.
CRPO-sur
Operate on each other to implant a permanent
adrenaline gland.
CRPO-sur-qua
created an underwater laboratory within our inflat-
able pool to collect the evidence we found of alien
life.
CRPO-qua
began to master the art of telekinesis by competi-
tively tossing each other s pillows across the room.
CRPO-nov-div-sur
Built a rollercoaster out of air mattresses and then did
a hot-wheel car-launch into the trenches and caught
the crash on GoPro cameras.
CRPO-cre
Created an experiment to see if our dreams could be
manipulated and transfer to one another.
Table 5: Example model (and gold human) responses to the prompt At a sleepover we ... from the Sentence
Completion task.
--- Page 21 ---
Task
Description
Dataset Sources
prompts
samples
Real-Life
Creative
Problem
Solving
Produce solutions for ev-
eryday problems
(Luchini
et
al.,
2025;
Kapoor et al., 2024; Saret-
zki et al., 2024; Perchtold-
Stefan et al., 2024)
33,340
Alternate Ti-
tles Genera-
tion
Produce alternative titles
for widely known books or
movies
(Agnoli et al., 2016)
2,986
Question
Asking
Produce questions about
everyday objects
(Raz et al., 2024)
3,566
Poems
Produce poems about ev-
eryday concepts
(Fan et al., 2023; Chaud-
huri et al., 2025; He et al.,
2022)
2,580
Design Solu-
tions
Produce solutions to real-
world design problems
(DiStefano et al., 2025)
10,818
Combining
Objects
Produce
combinations
of everyday objects to
achieve a goal
(Weiss et al., 2023)
4,494
Plot
Titles
Generation
Produce titles for story
plots
(Weiss
et
al.,
2023;
Goecke
et
al.,
2024b;
Weiss et al., 2024)
1,832
Instances
of Common
Concepts
Produce instances related
to everyday adjectives
(Organisciak et al., 2023)
2,474
Experiment
Design
Produce experiment de-
signs to test scientific hy-
potheses
(Beaty
et
al.,
2024;
Goecke et al., 2024a)
4,893
Associations
Produce word associations
(Beaty and Johnson, 2021)
1,004
Emotional
Trials
Produce
feelings
one
might have in a given
situation
(Weiss et al., 2023)
Invent Nick-
names
Produce nicknames for ev-
eryday concepts and ob-
jects
(Weiss et al., 2023)
Situation Re-
description
Produce redescriptions of
negative situations into
positive situations
(Weiss et al., 2023)
Alternate
Uses
of
Objects Task
Produce alternate uses for
everyday objects
(Patterson et al., 2023;
Zieli nska et al., 2023; Or-
ganisciak et al., 2023)
88,155
Stories
Produce short stories from
three word prompts
(Luchini et al., 2025; Ag-
noli et al., 2016; Fan et al.,
2023; He et al., 2022)
2,757
Table 6: MUCE dataset details broken down by task (Part 1).
--- Page 22 ---
Task
Description
Dataset Sources
prompts
samples
Malevolent
Problems
Produce ideas on how to
take revenge on or sabo-
tage a wrongdoer
(Perchtold-Stefan et al.,
2023; Kapoor et al., 2024;
Perchtold-Stefan
et
al.,
2024)
16,536
Metaphors
Produce metaphors to de-
scribe scenarios
(DiStefano et al., 2024; Yu
et al., 2024)
13,210
Essays
Produce essays on a topic
(Cotter et al., 2016)
Consequences Produce possible conse-
quences to scenarios
(Weiss et al., 2024, 2023;
Goecke et al., 2024b)
24,874
Sentence
Completion
Produce endings to incom-
plete sentences
(Organisciak et al., 2023)
2,629
Hypothesis
Generation
Produce
scientific
hy-
potheses
for
specific
observations
(Beaty
et
al.,
2024;
Goecke et al., 2024a)
18,455
Research
Questions
Produce research ques-
tions relating to scenarios
(Beaty
et
al.,
2024;
Goecke et al., 2024a)
5,161
Composites
Produce composite words
from a prompt word
(Weiss et al., 2023)
Evoking
Emotional
Responses
from People
Produce ways to evoke
emotional responses in
people as a TV producer
(Weiss et al., 2023)
Emotions in
Everyday Sit-
uations
Produce
emotions
you
might feel in response to
everyday situations
(Weiss et al., 2023)
Table 7: MUCE dataset details broken down by task (Part 2).
--- Page 23 ---
Task
Example prompt
Example low rating re-
sponse
Example high rating re-
sponse
Real-Life
Creative
Problem
Solving
Becky is a college stu-
dent who works part-time
at Mark s Pizzeria. Mark,
the owner of the restau-
rant, has treated Becky
very well. He gave her a
job that she needs to help
pay her rent when no other
business would employ
her because she was ar-
rested for shoplifting three
years ago.
Mark also
lets Becky work around
her school schedule, and
has asked if she wants to
be a shift manager in the
summers. Becky s room-
mate Jim also works at the
pizzeria, but Jim has been
causing a lot of problems
at work. He always avoids
doing his job, treats cus-
tomers rudely, and makes
a lot of mistakes with or-
ders.
Jim recently be-
gan stealing food from the
pizzeria. Two days ago the
pizzeria was short- staffed,
so Jim and Becky were
the only employees left at
closing time.
Jim made
10 extra pizzas and took
them home to a party he
was hosting without pay-
ing for them. Becky feels
like she needs to do some-
thing about Jim s behav-
ior.
However, Becky is
hesitant to tell Mark about
Jim because Jim is a good
friend to Becky.
Becky
also needs Jim to have a
job so he can pay his por-
tion of their rent. Becky
does not know what to
do..
Morally the right thing for
Becky to do would be to
tell her boss.
However,
to be a good friend would
to be not to tell on Jim.
The only creative solution
to this problem would to
be to try and talk to Jim
one on one. Give Jim the
decision of whether or nt
he wants Becky to inform
their boss of what he has
been doing. As a friend he
should understand where
Becky is coming from and
want to take the strain off
her.
Becky should first dis-
cuss this with Jim, and
tell him that he needs to
either pay for the pizzas
or he needs to go to the
boss, and admit what he
has done.
He will get
caught in the end because
eventually the ingredients
will be missed. The boss
may unerstand, and per-
haps will allow him to
work off the pizzas some-
how. Maybe he could help
out cleaning up around the
restaurant. If Jim will not
tell his boss Becky should
tell him.
She wouldn t
necessarily have to come
right out and tell on her
coworker she could come
up with a way for the boss
to catch him at it. If he
does it once Jim will more
than likely do it again. She
could tell the boss to check
on the inventory.
She
could have other people
who might have been at
the party come tell her
boss bout it. If all of that
fails, she should just tell
Mark about Jim stealing
the pizzas.
Table 8: MUCE dataset examples (Part 1).
--- Page 24 ---
Task
Example prompt
Example low rating re-
sponse
Example high rating re-
sponse
Question
Asking
pencil
How big is it?
How many great ideas
have started with a pen-
cil?
Poems
childhood
Twinkle, Twinkle little
star....ect
Red Rover, Red Rover
Is my childhood over? I
don t feel quite grown up
I still laugh at "I CUP"
I play slide with my sis-
ter and still call my fourth
grade teacher "mister" I
suppose, even still, my
childhood is over even if
I can still play red rover
red rover
Design Solu-
tions
Develop as many design
ideas as you can to reduce
air pollution in cities.
Walk
use 3d printing as an in-
nivating way of building
houses as it reduces labour
and
Combining
Objects
Paint sign
paper, ballpoint pen
beetroot
juice,
quark
cheese
Plot
Titles
Generation
Now spoke
A completely normal ev-
eryday life
VR glasses charger defec-
tive
Instances
of Common
Concepts
soft
something that is not
hard
a futuristic ball that turns
really fuzzy and comfy at
places it gets contact to
Experiment
Design
You think some animals
have a sense of humor that
humans don t usually un-
derstand. How could you
test that hypothesis?
observe
tickle your dog to see
how he acts when he s
laughing. then, observe
your dog throughout the
day and note when he is
laughing.
you may be-
gin to pick up on moments
where he does things that
are funny to him.
Associations
expert
winner
ace
Emotional
Trials
You have a date tonight,
and once again your dress
didn t get ready in time at
the laundry.
worried, afraid, sad
Anger, panic, anticipa-
tion
Invent Nick-
names
plate
porcelain
Shrunken UFO
Table 9: MUCE dataset examples (Part 2).
--- Page 25 ---
Task
Example prompt
Example low rating re-
sponse
Example high rating re-
sponse
Alternate
Uses
of
Objects Task
knife
weapon
make up "knife charac-
ters" and create a movie
Stories
petrol-diesel-pump
I needed to fuel my car
before we could start the
long drive. I drove to the
petrol station. i went to the
pump and fuel my car with
diesel. new i was ready for
the task ahead
Manly Merde was a truck
driver looking for trouble.
He pulled into the Casino
in the back where the
drivers go. He took a swig
of whisky and walked to
the petrol station, grabbed
the pump and spurt diesel
into the air like hydro-
carbon fountain.
He let
out a big belly laugh and
screamed, "Let the revo-
lution begin!" And that
is how the trucker wars
started.
Malevolent
Problems
Your professor in class
announces an award for
the person who comes up
with the best solution for
a project. By chance, an-
other student leaves their
notebook behind in class.
You read their ideas and
believe that they are the
best. You decide to turn
them in as your own; how-
ever you know that if the
other student submits the
same solution, there will
be a problem.
I will not do the above
render their notebook un-
readable by dropping wa-
ter at the last moment
Metaphors
The hot tea is...
boiling
liquid fire
Consequences What would be the result
if society no longer used
money, and instead traded
goods and services?
Banks would be unneces-
sary.
People (especially cou-
ples) would stop fighting
so much about financial is-
sues
Sentence
Completion
It started raining and...
I got wet
because I was covered in
oil, I began to levitate, and
all the witnesses called me
the next coming of some
sort of goddess.
Table 10: MUCE dataset examples (Part 3).
--- Page 26 ---
Task
Example prompt
Example low rating re-
sponse
Example high rating re-
sponse
Hypothesis
Generation
On a field trip, you drive
past a massive field with
hundreds of large holes
visible as far as the eye
can see.
What hypothe-
ses do you have about
what purpose the holes
may serve?
the holes resulted over
time and nature
the holes are for animals
giving birth.
Essays
dream project
I don t really know what
carreer path I want to fol-
low.
I just want a job
where I can help people
and get a good pay check
so I can support myt fu-
ture endevors. I want to
do something that no one
has ever done before in a
way no one has ever seen.
I want to inspire a genera-
tion to work on a better fu-
ture for everybody. I guess
what I really want is to be
remembered as an icon. i
want to be someone that
people look up to.
I want to go into foren-
sic science when I gradu-
tate. Therefore, my dream
project is to discover the
perfect device that can
help solve every crime
scene. This device would
be able to analyize the
crime scene and tell us
exactly how many people
died and how they died.
It would then collect ev-
idence samples such as
blood.
Next, it would
use what the information
it found at the crime scene
to help make up questions
the dectectives would ask
the suspects. It would use
it s technology to come up
with questions that only
the murderer could an-
swer. Later on, back at the
lab, it would help discover
whose blood the samples
belonged to. In the end,
the only human power
that would be needed was
someone to arrest the con-
victed person and the peo-
ple to help clean up the
crime scene.
Table 11: MUCE dataset examples (Part 4).
--- Page 27 ---
Task
Example prompt
Example low rating re-
sponse
Example high rating re-
sponse
Situation Re-
description
You notice how your col-
league first treats another
employee very kindly and
then shortly afterwards
starts talking negatively
behind his back
It would be nice if you
were older
I ll talk to them. Then I ll
have to work less
Alternate Ti-
tles Genera-
tion
The Betrothed
renzo and lucia
Plague, Honor and Love
in Baroque Brianza
Research
Questions
You travel to a jungle
that contains no human
life and is completely un-
known to the scientific
community. What scien-
tific questions could you
ask about this jungle?
How many people will
come with me?
Do these species share
a common characterisitic
that humans don t have?
Composites
jitters
Exam jitters
Easter bunny missing jit-
ters
Evoking
Emotional
Responses
from People
Describe how you would
make people look down on
others
I
will
always
scream
loudly
I would divide the au-
dience into two groups
and give one group a rub-
ber glove as headgear and
the other group a tiara or
crown made of real gold.
Emotions in
Everyday Sit-
uations
You re at work. A glance
at the clock tells you
that you re about to finish
work and start your long-
awaited weekend.
I feel happy
I feel sorry for my desk
chair, which is unused
over the weekend and
stands alone in the office.
Table 12: MUCE dataset examples (Part 5). |