Datasets:
File size: 196,845 Bytes
f06250e | 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 | | | Automatically | | Correcting | | Large | Language | | Models: | | | |
| --- | ------------- | --- | ---------- | --- | ----- | -------- | --- | ------- | --- | --- | --- |
Surveying the Landscape of Diverse Automated Correction Strategies
| | | LiangmingPan, | | MichaelSaxon, | | | WendaXu, | | | | |
| --- | --- | -------------- | --- | ------------- | --- | ------- | -------- | --- | --- | --- | --- |
| | | DeepakNathani, | | XinyiWang, | | William | YangWang | | | | |
UniversityofCalifornia,SantaBarbara,USA
{liangmingpan, saxon, wendaxu, dnathani, xinyi wang}@ucsb.edu
william@cs.ucsb.edu
| | | Abstract | | | 2023b; | Wu | et al., | 2023b), | generating | inappropri- | |
| --- | --- | -------- | --- | --- | ------ | --- | ------- | ------- | ---------- | ----------- | --- |
ateorharmfulcontent(Gehmanetal.,2020;Levy
etal.,2021,2022;Shaikhetal.,2023),andfailing Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| While large | language | models | (LLMs) | have | | | | | | | |
| ----------- | -------- | ------ | ------ | ---- | --- | --- | --- | --- | --- | --- | --- |
shown remarkable effectiveness in various to trustfully follow rules and constraints (Zhuo
NLPtasks,theyarestillpronetoissuessuchas
etal.,2023;Wangetal.,2023a).Suchflawedbe-
| hallucination, | unfaithful | reasoning, | and | toxi- | | | | | | | |
| -------------- | ---------- | ---------- | --- | ----- | --- | --- | --- | --- | --- | --- | --- |
haviorshamperthetrustinLLMsandposehurdles
| city. A | promising | approach | to rectify | these | | | | | | | |
| ------- | --------- | -------- | ---------- | ----- | --- | --- | --- | --- | --- | --- | --- |
totheirreal-worldapplications(OpenAI,2023).
flawsiscorrectingLLMswithfeedback,where
Aprevailingstrategytorectifytheseundesired
theLLMitselfispromptedorguidedwithfeed-
backtofixproblemsinitsownoutput.Tech- behaviors of LLMs is learning from feedback,
mirroringatypicalhumanlearningstrategywhere
niquesleveragingautomatedfeedback—either
produced by the LLM itself (self-correction) individualsactivelyrefinetheirbehaviorsthrough
orsomeexternalsystem—areofparticularin-
| | | | | | a cycle | of | trial, error, | | and correction. | | Humans, |
| --- | --- | --- | --- | --- | ------- | --- | ------------- | --- | --------------- | --- | ------- |
terestastheymakeLLM-basedsolutionsmore
| | | | | | when | making | mistakes, | | often gather | feedback | ei- |
| --- | --- | --- | --- | --- | ---- | ------ | --------- | --- | ------------ | -------- | --- |
practicalanddeployablewithminimalhuman
| | | | | | ther | from | others or | through | self-reflection | | (Boyd |
| ------------- | ---- | -------------- | --- | ------- | ---- | ------ | --------- | --------- | --------------- | -------- | ------- |
| intervention. | This | paper provides | an | exhaus- | | | | | | | |
| | | | | | and | Fales, | 1983; | Metcalfe, | 2017; | Ferretti | et al., |
tivereviewoftherecentadvancesincorrecting
LLMswithautomatedfeedback,categorizing 2019;Londonetal.,2023;Bellha¨useretal.,2023).
Suchfeedbackoffersvaluableinsightsforhumans
| them into | training-time, | generation-time, | | and | | | | | | | |
| --------- | -------------- | ---------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
post-hoc approaches. We also identify poten- tocorrectmistakesandmodifytheirbehaviorac-
tial challenges and future directions in this cordingly.Inspiredbythisnaturallearningmech-
emergingfield.
| | | | | | anism, | extensive | | research | (Huang | et | al., 2022; |
| --- | --- | --- | --- | --- | ------ | --------- | --- | -------- | ------ | --- | ---------- |
Madaanetal.,2023;Geroetal.,2023;Jiangetal.,
| 1 Introduction | | | | | 2023) | has | been undertaken | | to | improve | LLMs |
| -------------- | --- | --- | --- | --- | ------- | --- | --------------- | --- | -------- | ------- | -------- |
| | | | | | through | the | paradigm | of | learning | from | both in- |
Recent years have seen striking empirical suc- ternalandexternalfeedback.
cesses of large language models (LLMs), as they One popular line of research involves the use
consistentlyobtainimpressiveresultsacrossadi- ofhumanfeedbacktoevaluateandrefinemodels,
verserangeofNLPbenchmarks(Guoetal.,2023; as encapsulated in the survey by Fernandes et al.
Suzgun et al., 2023; Qin et al., 2023), while also (2023). These methods typically involve direct
showcasingsurprisingabilitiesoflanguageunder- optimization of LLMs against human feedback
standing (Wei et al., 2022a; Begus et al., 2023), on their outputs (Kreutzer et al., 2018; Glaese
generation(PuandDemberg,2023;LinandChen, et al., 2022; Ouyang et al., 2022; Scheurer et al.,
2023;Lyuetal.,2023a),andreasoning(Weietal., 2023),wherehumanevaluationsofoutputquality
2022b;Kojimaetal.,2022;Dasguptaetal.,2022). serveasarewardsignaltoimprovemodelperfor-
However,thesemodelsarenotwithouttheirflaws. mance. However, this approach has two primary
LLMs are observed to intermittently display un- drawbacks: It can be costly due to the manual
desired and inconsistent behaviors such as pro- labor involved, and it lacks real-time capabilities
ducing seemingly convincing but inaccurate ashumanscannotprovideinstantfeedback.
‘‘hallucinations’’ (Lin et al., 2022; Zhang et al., To minimize the need for human intervention,
2023c; Min et al., 2023), conducting unfaithful another strategy is correcting LLMs with auto-
reasoning (Golovneva et al., 2023; Lyu et al., mated feedback. As illustrated by the conceptual
484
TransactionsoftheAssociationforComputationalLinguistics,vol.12,pp.484–506,2024.https://doi.org/10.1162/tacla00660
ActionEditor:IvanTitov.Submissionbatch:9/2023;Revisionbatch:11/2024;Published5/2024.
(cid:2)c 2024AssociationforComputationalLinguistics.DistributedunderaCC-BY4.0license.
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
Figure 1: A conceptual framework for correcting LLMs with automated feedback. We identify three parties
involvedintheprototypicalcorrectionpipelinethatareanalogoustoapatient,doctor,andtreatmentinmedicine,
respectively:ALanguageModelproducesinitialoutput,aCriticModelanalyzestheoutputandprovidesfeedback,
andaRefineModelprovidestreatmenttoeithertheoutputorthelanguagemodel.Wetaxonomizeexistingworks
using this conceptualization along five key aspects: the problem to be corrected, the source and format of the
feedback,andthestrategyandlearningmethodoftherefinemodel.
(§
frameworkinFigure1,thelanguagemodel(itera- major techniques 3), categorized as training-
tively) learns from automatically generated feed- time, generation-time, and post-hoc correction.
back signals to understand the consequences of Finally,wediscusstheconnectiontoearlierworks
(§4)andfivepotentialfuturedirections(§5).
itsactionsandadaptsitsbehaviors.Thesourceof
automatedfeedbackcanbemultifaceted,spanning
fromtheLLMitselfactingasthefeedbackmodel
| | | | | | | | 2 ConceptualFramework | | | | | | |
| --- | --- | --- | --- | --- | --- | --- | --------------------- | --- | --- | --- | --- | --- | --- |
(Madaanetal.,2023;Schicketal.,2023),asepa-
ratelytrainedfeedbackmodel(Yangetal.,2022b; For clean exposition, we first present a concep-
| Paul et | al., 2023), | readily | | available | external | tools | | | | | | | |
| ------- | ----------- | ------- | --- | ------------ | -------- | -------- | -------------- | --- | --------- | --- | ------- | ------- | --- |
| | | | | | | | tual framework | | outlining | the | overall | process | of |
| (Gou et | al., 2023; | Chen | et | al., 2023e), | to | external | | | | | | | |
correctingLLMswithfeedbackinFigure1,using
knowledge sources such as Wikipedia or the in- an analogy of medical treatment in our daily life.
ternet (Yu et al., 2023; Li et al., 2023b). Various Threepartiesareinvolvedinthisprocess:
| strategies | of | correction | have | been | proposed, | in- | | | | | | | |
| ---------- | --- | ---------- | ---- | ---- | --------- | --- | --- | --- | --- | --- | --- | --- | --- |
cludingself-training(Huangetal.,2022;Baietal.,
| | | | | | | | • Language | | Model | (Patient). | A | language | |
| ------------------------------------------- | ------ | ----------------- | ----------- | --------- | ------- | -------- | --------------------- | ----- | ------- | ----------- | ----------- | ---------- | --- |
| 2022b),generate-then-rank(Heetal.,2023;Weng | | | | | | | | M | X | → Y | | | |
| | | | | | | | model | | : | performs | | a specific | |
| et al., | 2023), | feedback-guided | | decoding | | (Yang | | | | | | | |
| | | | | | | | taskbymappinganinputx | | | | ∈ X | toanoutput | |
| et al., 2022a; | | Xie et | al., 2023), | iterative | | post-hoc | | | | | | | |
| | | | | | | | text | yˆ ∈ | Y. This | formulation | encompasses | | |
| revision | (Zhang | et al., | 2023a; | Jiang | et al., | 2023), | | | | | | | |
| | | | | | | | a wide | range | of | NLP tasks, | for | example, | in |
| etc. Recently, | | the incorporation | | of | such | strate- | | | | | | | |
summarization,xisapassage,yˆisthegener-
gieshasdemonstratedtheireffectivenessacrossa
| | | | | | | | ated | summary; | for | question-answering, | | | x is |
| ------ | --------- | ---- | -------- | --------- | --- | ----- | ---- | -------- | --- | ------------------- | --- | --- | ---- |
| myriad | of tasks, | from | question | answering | | (Peng | | | | | | | |
aquestionandyˆisthepredictedanswer.The
| et al., 2023) | | and reasoning | | (Pan et | al., | 2023) to | | | | | | | |
| ------------- | --- | ------------- | --- | ------- | ---- | -------- | --- | --- | --- | --- | --- | --- | --- |
initialgenerationyˆmayhaveproblemssuch
codegeneration(Zhangetal.,2023b)andtoxicity
ashallucinationandincorrectreasoning.
detection(Luetal.,2022).
•
In light of these advancements, our paper aims CriticModel(Doctor&Diagnosis).Acritic
| | | | | | | | | C | X × | Y → F | | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | ----- | --- | --- | --- |
to provide a comprehensive survey. We start by model : learns to generate
establishing the concept of correcting LLMs with feedback x,yˆ → c where yˆ ∼ M(x) is
automated feedback and creating a taxonomy of the output or partial output of the language
(§
the different methods 2). We then discuss the model,andcisthefeedbackofsomeformat,
485
•
e.g.,scalarvalue,ornaturallanguage.Asim- Toxic, Biased, and Harmful Content.
ple example is binary feedback of whether LLMs have been observed to occasionally
the output is good or bad given the input generate content that is toxic, biased, or
(C : X ×Y → {0,1}). harmful due to biases present in the training
data(Shaikhetal.,2023).Torectifythis,re-
•
Refine Model (Treatment). A refine model inforcement learning from human feedback
| R | : X ×Y×F | | → Y | learnstorepairanout- | | | | | | | | | |
| ----- | -------- | --- | ----------- | -------------------- | ------------ | --- | ------------------------------------ | ------- | ---- | ----------- | ----- | -------- | ------- |
| | | | | | | | (RLHF) | (Ouyang | | et al., | 2022; | Bai | et al., |
| put | x,yˆ,c | → y | based | on | the feedback | c, | | | | | | | |
| | | | new | | | | 2022a) | has | been | extensively | | employed | to |
| where | y | is | the revised | | output. Some | re- | | | | | | | |
| | | new | | | | | trainLLMstoalignmorecloselywithhuman | | | | | | |
fine models directly repair the language values, such as being helpful, honest, and
| model | M | through | fine-tuning | | or reinforce- | | | | | | | | |
| ----- | --- | ------- | ----------- | --- | ------------- | --- | --- | --- | --- | --- | --- | --- | --- |
harmless.However,RLHFisheavilydepen-
mentlearning.
| | | | | | | | dent | on high-quality | | human | feedback, | | the |
| --- | --- | --- | --- | --- | --- | --- | ---- | --------------- | --- | ----- | --------- | --- | --- |
collectionofwhichcanberesource-intensive.
| Based | on | the above | formulation, | | the | specific | | | | | | | |
| ----- | --- | --------- | ------------ | --- | --- | -------- | --- | --- | --- | --- | --- | --- | --- |
Toalleviatethis,recentwork(Luetal.,2022;
| model design | | in existing | works | | varies along | five | | | | | | | |
| ------------ | --- | ----------- | ----- | --- | ------------ | ---- | --- | --- | --- | --- | --- | --- | --- |
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
Gouetal.,2023)hasalsoexploredcollecting
crucialaxes,elaboratedinthefollowingsections.
| | | | | | | | automated | feedback | | to | identify | and | correct |
| --- | --- | --- | --- | --- | --- | --- | --------- | -------- | --- | --- | -------- | --- | ------- |
potentiallyharmfuloutputs.
2.1 WhatGetsCorrected?
| | | | | | | | 2.2 WhatIstheSourceoftheFeedback? | | | | | | |
| ------------ | --- | -------- | ----- | ---------- | ----------- | -------- | --------------------------------- | --- | --- | --- | --- | --- | --- |
| We summarize | | the | three | major | error types | of | | | | | | | |
| LLMs that | are | targeted | for | correction | in | existing | | | | | | | |
Feedbackcanbebroadlydividedintohumanfeed-
worksthroughautomatedfeedback. back and automated feedback. Fernandes et al.
| | | | | | | | (2023) provided | | a survey | on | integrating | | human |
| --- | --- | --- | --- | --- | --- | --- | --------------- | --- | -------- | --- | ----------- | --- | ----- |
•
Hallucination.AnopenchallengeforLLMs
| | | | | | | | feedback | for language | | generation. | In | our | survey, |
| --- | --- | --- | --- | --- | --- | --- | -------- | ------------ | --- | ----------- | --- | --- | ------- |
is that they often hallucinate by making up we focus on the emerging research area of auto-
| facts | or | citing | sources | that | do not exist | (Li | | | | | | | |
| ----- | ----------- | ------ | ------- | ---- | ------------ | ---- | --------------- | --- | ----- | --------- | ---------- | --- | ---- |
| | | | | | | | mated feedback, | | which | typically | originates | | from |
| et | al., 2023a; | Zhang | et | al., | 2023c). This | hal- | | | | | | | |
twosources:self-feedback(i.e.,thefeedbackorig-
| lucinated | | content | is | often | quite plausible- | | | | | | | | |
| --------- | --- | ------- | --- | ----- | ---------------- | --- | ----------- | --- | --- | ------- | ------------ | --- | ----- |
| | | | | | | | inates from | the | LLM | itself) | and external | | feed- |
sounding,makingitdifficultevenforhumans
| | | | | | | | back (i.e., | the feedback | | is derived | | from external | |
| --- | --- | --- | --- | --- | --- | --- | ----------- | ------------ | --- | ---------- | --- | ------------- | --- |
todetect(Clarketal.,2021).Toaddressthis,
models,tools,orknowledgesources).
| several | studies | | have proposed | | the collection | | | | | | | | |
| ------- | ------- | --- | ------------- | --- | -------------- | --- | --- | --- | --- | --- | --- | --- | --- |
of automated feedback on potential factual • Self-Feedback. The LLM can act as its
| inaccuracies | | by | cross-referencing | | the | gener- | | | | | | | |
| ------------ | --- | --- | ----------------- | --- | --- | ------ | --- | --- | --- | --- | --- | --- | --- |
ownfeedbackproviderbyiterativelyassess-
atedoutputwithcredibleknowledgesources. ing and refining its generated outputs until
The gathered feedback can then be utilized it meets a certain standard (Madaan et al.,
byasubsequentrefinementmodeltocorrect
| | | | | | | | 2023; | Shinn | et al., | 2023). | This | continuous | |
| --- | --- | --- | --- | --- | --- | --- | ----- | ----- | ------- | ------ | ---- | ---------- | --- |
hallucinations(Gaoetal.,2023b;Pengetal.,
| | | | | | | | self-improvement | | | strategy | has | proven | effec- |
| ------ | --- | --- | --- | --- | --- | --- | ---------------- | ----------- | -------- | -------- | ---------- | ------ | ------ |
| 2023). | | | | | | | tive | in multiple | studies, | | especially | when | ex- |
ternalfeedbackisunavailableorlimited(Ye
| • Unfaithful | | Reasoning. | | | | | | | | | | | |
| ------------ | --- | ---------- | --- | -------- | --- | ------ | --- | --- | --- | --- | --- | --- | --- |
| | | | | A number | of | recent | | | | | | | |
etal.,2023;Yanetal.,2023).
| studies | | (Ribeiro | et al., | 2023; | Lyu | et al., | | | | | | | |
| ------- | --- | -------- | ------- | ----- | --- | ------- | --- | --- | --- | --- | --- | --- | --- |
•
2023b; Golovneva et al., 2023) found that External Feedback for LLMs comes
LLMs occasionally make unfaithful reason- from other models (Yang et al., 2022b;
ing, i.e., the derived conclusion does not Lightmanetal.,2023),tools(Gouetal.,2023;
follow the previously generated reasoning Charalambous et al., 2023), knowledge
chain. To address this, existing works have sources (Gao et al., 2023b; Yu et al., 2023),
usedautomatedfeedbackfromexternaltools and evaluation metrics (Jung et al., 2022;
or models for guiding the reasoning process Wellecketal.,2023).Externalfeedbackpro-
(Xieetal.,2023;Yaoetal.,2023a),verifying videsavaluableoutsideperspectiveforiden-
the reasoning process and rectifying errors tifyingerrorsthattheLLMcannotrecognize
(He et al., 2023; Pan et al., 2023), or fine- on its own. For example, code interpreters
tuning LLMs with process-based feedback arewidelyusedinprogrammingtaskstopro-
(Huangetal.,2022;Lightmanetal.,2023). videreal-timeerrormessages;whileexternal
486
knowledge sources are used to verify the tion may be hindered by the infeasibility of
| factualaccuracy. | | | | | | fine-tuning | | giant | closed-source | | LLMs, such |
| ---------------- | --- | --- | --- | --- | --- | ----------- | ----- | -------- | ------------- | --- | ---------- |
| | | | | | | as | GPT-4 | (OpenAI, | 2023) | and | the poten- |
2.3 WhatIstheFormatoftheFeedback? tial unavailability of feedback during model
training.
| The selection | of | the feedback | format | | requires | | | | | | |
| ------------- | --- | ------------ | ------ | --- | -------- | --- | --- | --- | --- | --- | --- |
considering its expressivity, ease of collection, • Generation-time Correction.
| | | | | | | | | | | | It utilizes au- |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --------------- |
and its potential to improve systems (Fernandes tomated feedback to guide the LLM to cor-
et al., 2023). Automated feedback is commonly rect errors during generation. For example,
eitherascalarvalueorinnaturallanguage. for proof generation, several studies utilize
| | | | | | | the | automated | feedback | | of the | intermediate |
| --- | --- | --- | --- | --- | --- | --- | --------- | -------- | --- | ------ | ------------ |
•
Scalar Value Feedback. In this scenario, reasoningstepstoguidethemodeltorecover
| the critic | model | maps | the input | and | output | | | | | | |
| ----------- | ----- | ------- | --------- | --- | ------ | ------- | --------- | ---------- | ---- | ---- | -------------- |
| | | | | | | from | incorrect | generation | | and | search for the |
| | | (C | X × Y | → N | ⊆ R). | | | | | | |
| to a single | | score : | | | | optimal | | solution | in a | more | efficient way |
Scalarvaluefeedbackcanbeeasilyintegrated (Yangetal.,2022a;Lightmanetal.,2023).
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| into the | training/decoding | | process | of | LLMs. | | | | | | |
| -------- | ----------------- | --- | ------- | --- | ----- | ---------- | --- | ----------- | --- | ---------- | --------- |
| | | | | | | • Post-hoc | | Correction. | | It refines | the model |
Forexample,Xieetal.(2023)usereal-value
| | | | | | | output | after | it has | been | generated, | without |
| -------- | ----- | ----------------- | ------------- | --------- | ------ | -------- | ----- | --------- | ----------- | ---------- | ---------- |
| feedback | for | each intermediate | | reasoning | | | | | | | |
| | | | | | | updating | | the model | parameters. | | This typi- |
| step to | guide | the model | in performing | | a sto- | | | | | | |
callyinvolvesaniterativeprocessofgenerat-
chasticbeamsearchfortheoptimalsolution.
| | | | | | | ing | output, | receiving | feedback, | | and refining |
| --- | --- | --- | --- | --- | --- | --- | ------- | --------- | --------- | --- | ------------ |
Despiteitsflexibility,scalarfeedbackisless
| | | | | | | output. | Post-hoc | | correction | | is more flexi- |
| --- | --- | --- | --- | --- | --- | ------- | -------- | --- | ---------- | --- | -------------- |
descriptivefordetailedcorrections.
| | | | | | | ble | as it | does not | require | training | the LLM |
| --- | --- | --- | --- | --- | --- | --- | ----- | -------- | ------- | -------- | ------- |
•
Natural Language Feedback provides or accessing its parameters. Furthermore, it
richerinformationthatcanhighlightspecific
| | | | | | | facilitates | | the incorporation | | | of more infor- |
| --- | --- | --- | --- | --- | --- | ----------- | --- | ----------------- | --- | --- | -------------- |
errors and provide nuanced suggestions for mative natural language feedback, offering
improvement. This is important for certain a more transparent and explainable self-
| applications | | such as | text editing | and | code | | | | | | |
| ------------ | --- | ------- | ------------ | --- | ---- | --- | --- | --- | --- | --- | --- |
correctionprocess.
| generation. | | For example, | Self-Debug | | (Chen | | | | | | |
| ----------- | --- | ------------ | ---------- | --- | ----- | ------------------------ | --- | --- | --- | --- | --- |
| | | | | | | 2.5 HowtoCorrecttheModel | | | | | |
etal.,2023e)usesLLMstogenerateexplana-
withFeedback?
| tions for | the | produced | code and | utilize | both | | | | | | |
| --------------- | --- | -------- | --------- | ------- | ---- | ------- | -------- | ---------- | --- | ---- | ------------- |
| the explanation | | and the | execution | results | as | | | | | | |
| | | | | | | Various | concrete | strategies | | have | been proposed |
feedbacktoenhancecodingsolutions. tocorrectLLMswithautomatedfeedback,which
| | | | | | | are tailored | to | the different | dimensions | | we men- |
| --- | --- | --- | --- | --- | --- | ------------ | --- | ------------- | ---------- | --- | ------- |
2.4 WhentoCorrecttheModel?
| | | | | | | tioned in | previous | sections. | | For | example, self- |
| --- | --- | --- | --- | --- | --- | --------- | -------- | --------- | --- | --- | -------------- |
Dependingonthetimingofusingautomatedfeed- trainingisoftenusedfortraining-timecorrection.
back to correct the model, existing work can be Generate-then-rankoftencomeswithscalarvalue
dividedintothreemajorcategories. feedback.Wewillcoverthecomprehensiveland-
scapeofself-correctionstrategiesinSection3.
| • Training-time | | Correction. | | | | | | | | | |
| --------------- | ----- | ----------- | ------ | --------- | ------ | ------------------------- | --- | --- | --- | --- | --- |
| | | | | The ideal | sce- | | | | | | |
| | | | | | | 2.6 SummaryofExistingWork | | | | | |
| nario | is to | rectify a | flawed | model | during | | | | | | |
training,priortoitsdeploymentforuse.Once Building upon the taxonomy established in the
feedback has been collected, it is directly preceding sections, we collate existing work in
used to optimize the model parameters. Hu- Table1andTable2.Wehavethreemajorselection
man feedback is typically used for training- criteriaforaworktobeincludedinthissurvey:
time correction, as exemplified by the widely 1. Automated Feedback: Explicit feedback
adopted RLHF approach (Ouyang et al., is involved to assess the quality of the model
2022). For leveraging automated feedback, output.Wefocusonautomatedfeedbackthatorig-
a common strategy is self-training (Huang inates from external models, metrics, knowledge,
et al., 2022), where the model is trained etc. However, we will cover some representative
with its own generated high-quality output worksofhumanfeedbackforcompleteness.
filteredoutbythecriticmodel.However,the 2. Model Refinement: The feedback should
practicalapplicationoftraining-timecorrec- act as a directive to enhance the LLM, either by:
487
Feedback ModelRefinement
Method Application
Source Format Strategy Learning
Training-TimeCorrection
RLHF(Ouyangetal.,2022) RewardModel Scalar RLHF RL MultipleTasks
Fine-GrainedRLHF(Wuetal.,2023a) RewardModel Scalar RLHF RL Detoxification,Long-formQA
HH-RLHF(Baietal.,2022a) RewardModel Scalar RLHF SL&RL Helpfulness,Harmlessness
MoralRLHF(Gangulietal.,2023) RewardModel Scalar RLHF RL MoralCorrection
Sparrow(Glaeseetal.,2022) RewardModel NL RLHF SL&RL Dialogue
ILF(Scheureretal.,2023) HumanFeedback NL Fine-tuning SL Summarization
ILF-Code(Chenetal.,2023a) HumanFeedback NL Fine-tuning SL CodeGeneration
SLT(Yuanetal.,2023) HumanFeedback NL Fine-tuning SL ResponseGeneration
Chain-of-Hindsight(Liuetal.,2023a) HumanFeedback NL Fine-tuning SL MultipleTasks
Crystal(Liuetal.,2023b) LanguageModel Scalar Fine-Tuning SL&RL CommonsenseReasoning
STaR(Zelikmanetal.,2022) LanguageModel NL Self-Training SL QA,Reasoning
RLAIF(Baietal.,2022b) LanguageModel NL Self-Training SL&RL Dialogue
SIRLC(Pangetal.,2023) LanguageModel NL Self-Training RL Reasoning,Translation,Summary
Self-Improve(Huangetal.,2022) LanguageModel NL Self-Training SL QA,Reasoning,NLI
AlpacaFarm(Duboisetal.,2023) LanguageModel NL Self-Training SL&RL None(IntrinsicEvaluation)
ReST(Gulcehreetal.,2023) LanguageModel NL Self-Training RL MachineTranslation
Generation-TimeCorrection
Self-Verification(Wengetal.,2023) LanguageModel Scalar Re-Ranking ICL ArithmeticReasoning
CodeT(Chenetal.,2023b) ProgramExecutor Scalar Re-Ranking ICL CodeGeneration
LEVER(Nietal.,2023) ProgramExecutor Scalar Re-Ranking SL TableQA,MathQA,Program
RR(Heetal.,2023) ExternalKnowledge Scalar Re-Ranking — Reasoning
InstructScore(Xuetal.,2023) LanguageModel NL Re-Ranking SL GenerationEvaluation
MBRDecoding(Freitagetal.,2022) ExternalMetrics Scalar Re-Ranking SL MachineTranslation
DIVERSE(Lietal.,2023d) TrainedModel Scalar Re-Ranking SL ArithmeticReasoning
PRM(Lightmanetal.,2023) RewardModel Scalar Feedback-guided SL ArithmeticReasoning
DiffusionLM(Lietal.,2022) TrainedModel Scalar Feedback-guided SL ControlledTextGeneration
Fudge(YangandKlein,2021) TrainedModel Scalar Feedback-guided SL ControlledTextGeneration
Entailer(Tafjordetal.,2022) TrainedModel Scalar Feedback-guided SL ProofGeneration
NLProofS(Yangetal.,2022a) TrainedModel Scalar Feedback-guided SL ProofGeneration
GRACE(Khalifaetal.,2023) TrainedModel Scalar Feedback-guided SL ArithmeticReasoning
CoRe(Zhuetal.,2023) TrainedModel Scalar Feedback-guided SL ArithmeticReasoning
Varshneyetal.(2023) ExternalKnowledge NL Feedback-guided ICL HallucinationDetection
MemPrompt(Madaanetal.,2022) ExternalKnowledge NL Feedback-guided ICL LexicalandEthicalReasoning
MaieuticPrompting(Jungetal.,2022) ExternalMetrics Scalar Feedback-guided ICL CommonsenseReasoning
SI(CreswellandShanahan,2022) LanguageModel Scalar Feedback-guided ICL ProofGeneration
RAP(Haoetal.,2023) LanguageModel Scalar Feedback-guided ICL Planning,Reasoning
SelfEval-Decoding(Xieetal.,2023) LanguageModel Scalar Feedback-guided ICL Arithmetic/SymbolicReasoning
SelfCheck(Miaoetal.,2023) LanguageModel NL Feedback-guided ICL ArithmeticReasoning
TreeofThoughts(Yaoetal.,2023a) LanguageModel NL/Scalar Feedback-guided ICL Games,Writing
Table1:RepresentativeworksonTraining-timeCorrectionandGeneration-TimeCorrection.
1) updating model parameters, or 2) altering the therefinement,4)whethertherefinementprocess
model’soutputduringorpostthegeneration. isiterative,and5)theapplicationofthemethod.
3.LargeLanguageModel:Weprimarilyfocus
3 Methodologies
on automated correction strategies in the era of
modern large language models. Given this focus, Inthissection,wedelveintoadetailedreviewof
wemainlyemphasizeveryrecentworkfrom2022 various correction methodologies. Depending on
and2023.However,itisimportanttoacknowledge thetimethatthecorrectionhappens,wecategorize
thattheconceptofautomatedcorrectionisnotnew them as Training-Time Correction, Generation-
and has roots in early NLP research. To provide TimeCorrection,andPost-hocCorrection.
a complete historical perspective, we provide a
succinct overview of these initial approaches to 3.1 Training-TimeCorrection
automatedcorrectioninSection4.1. Training-timecorrectionrectifiesmodelbehavior
Thesestudiesarecategorizedbasedonthethree during the training phase. We identify three typi-
strategiesintroducedinSection2.4.Wealsosum- calstrategiesshowninFigure2.Eachstrategyuti-
marize key features of each study, including: 1) lizes different forms of feedback to optimize the
thesourceoffeedback,2)theformatoffeedback, model during training: human feedback (a), a re-
3)thestrategyandlearningmethodemployedfor wardmodel(b),andautomatedfeedback(c).
488
Downloaded
from
http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf
by
guest
on
23
May
2024
Feedback ModelRefinement
Method Application
Source Format Strategy Learning Iter.
Post-hocCorrection
Self-Refine(Madaanetal.,2023) LanguageModel NL Self-Refine ICL ✓ MultipleTasks
ClinicalSV(Geroetal.,2023) LanguageModel NL Self-Refine ICL ✗ InformationExtraction
Reflexion(Shinnetal.,2023) LanguageModel NL Self-Refine RL ✓ QA,CodeGeneration
IterRefinement(Chenetal.,2023d) LanguageModel NL Self-Refine ICL ✓ MachineTranslation
Auto-Post-Editing(Raunaketal.,2023) LanguageModel NL Self-Refine ICL ✗ MachineTranslation
RCI(Kimetal.,2023) LanguageModel NL Self-Refine ICL ✓ ComputerTasks
SelFee(Yeetal.,2023) LanguageModel NL Self-Refine SL ✓ Dialogue
SelfCheckGPT(Manakuletal.,2023) LanguageModel NL Self-Refine ICL ✗ HallucinationDetection
LLMSelfDefense(Helblingetal.,2023) LanguageModel NL Self-Refine ICL ✗ HarmfulTextCorrection
Re3(Yangetal.,2022b) TrainedModel Scalar ExternalFeedback SL&ICL ✓ StoryGeneration
CodeRL(Leetal.,2022) TrainedModel Scalar ExternalFeedback RL ✗ CodeGeneration
FLIRT(Mehrabietal.,2023) TrainedModel Scalar ExternalFeedback ICL ✓ AdversarialPromptGeneration
REFINER(Pauletal.,2023) TrainedModel NL ExternalFeedback SL&ICL ✓ Reasoning,MoralStory
RL4F(Akyu¨reketal.,2023) TrainedModel NL ExternalFeedback SL&RL ✓ Planning,Summarization
Yanetal.(2023) TrainedModel NL ExternalFeedback SL ✓ SemanticParsing
Baldur(Firstetal.,2023) TrainedModel NL ExternalFeedback ICL ✓ ProofGeneration
CRITIC(Gouetal.,2023) ExternalTools NL ExternalFeedback ICL ✓ QA,Program,Toxicity
FacTool(Chernetal.,2023) ExternalTools NL ExternalFeedback ICL ✓ QA,Reasoning,Generation
MAF(Nathanietal.,2023) ExternalTools NL ExternalFeedback ICL ✓ QA,Reasoning
RARR(Gaoetal.,2023b) ExternalKnowledge NL ExternalFeedback ICL ✗ Open-DomainQA
LLM-Augmenter(Pengetal.,2023) ExternalKnowledge NL ExternalFeedback RL ✓ Open-DomainQA
Self-Checker(Lietal.,2023b) ExternalKnowledge NL ExternalFeedback ICL ✗ Fact-Checking
REFEED(Yuetal.,2023) ExternalKnowledge NL ExternalFeedback ICL ✗ QA,Dialogue
Olaussonetal.(2023) ProgramExecutor NL ExternalFeedback ICL ✓ CodeGeneration
Self-Edit(Zhangetal.,2023a) ProgramExecutor NL ExternalFeedback ICL ✓ CodeGeneration
Self-Debug(Chenetal.,2023e) ProgramExecutor NL ExternalFeedback ICL ✓ CodeGeneration
Self-Evolve(Jiangetal.,2023) ProgramExecutor NL ExternalFeedback ICL ✓ CodeGeneration
Logic-LM(Panetal.,2023) SymbolicSolver NL ExternalFeedback ICL ✓ LogicalReasoning
Self-Critique(Saundersetal.,2022) LLMs+Human NL ExternalFeedback SL ✗ Summarization
ALGO(Zhangetal.,2023b) OracleVerifier Scalar ExternalFeedback ICL ✓ CodeGeneration
Charalambousetal.(2023) BMCTool NL ExternalFeedback ICL ✗ SoftwareVerification
Self-Correction(Wellecketal.,2023) ExternalMetrics NL/Scalar ExternalFeedback SL ✓ Reasoning,Generation,Toxicity
MultiagentDebate(Duetal.,2023) LanguageModel NL ModelDebate ICL ✓ Reasoning,Factuality
LMvsLM(Cohenetal.,2023) LanguageModel NL ModelDebate ICL ✓ FactualErrorDetection
ICL-AIF(Fuetal.,2023) LanguageModel NL ModelDebate ICL ✓ BargainingGame
PRD(Lietal.,2023c) LanguageModel NL ModelDebate ICL ✓ Open-endedQA
MADRA(Wangetal.,2023b) LanguageModel NL ModelDebate ICL ✓ QA,Fact-Checking
ReConcile(Chenetal.,2023c) LanguageModel NL ModelDebate ICL ✓ Reasoning
Table2:RepresentativeworkonPost-hocCorrection.
Figure 2: Three typical strategies of training-time correction: direct optimization with human feedback (a),
trainingarewardmodelthatapproximateshumanfeedback(b),andself-trainingwithautomatedfeedback(c).
Direct Optimization with Human Feedback. 2022; Scheurer et al., 2023; Chen et al., 2023a).
In an ideal scenario, we would directly leverage However, only utilizing positive-rated data may
human feedback to optimize the model parame- constrain the model’s ability to identify and cor-
ters, following the framework in Figure 2(a): 1) rect negative attributes or errors. To address this,
CandidateoutputsaregeneratedbyLLMs,2)Hu- Chain-of-Hindsight (Liu et al., 2023a) fine-tunes
mans provide feedback or refinements on these the LLM on model outputs paired with both pos-
outputs,and3)LLMsarethendirectlyoptimized itive and negative feedback. Beyond fine-tuning,
onthecollected(outputs,feedback)tobetteralign other optimization methods are explored as well.
with human preferences. A simple strategy is to For example, Gao et al. (2023a) utilize human
fine-tunethemodelontheoutputsthatreceivepos- feedback as the reward signal and optimize the
itive feedback from human raters (Glaese et al., modelwithcontextualbanditlearning.
489
Downloaded
from
http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf
by
guest
on
23
May
2024
| Reward | Modeling | | and | RLHF. | Direct | opti- | | | | | |
| -------- | -------- | ----- | -------- | ----- | ------- | ------ | --- | --- | --- | --- | --- |
| mization | with | human | feedback | | may not | always | | | | | |
bepractical,sincecollectinghumanfeedbackcan
| be both | labor-intensive | | and | time-consuming. | | An | | | | | |
| --------- | --------------- | --- | ----- | --------------- | -------- | ----- | --- | --- | --- | --- | --- |
| efficient | alternative | | is to | train | a reward | model | | | | | |
thatemulateshumanfeedback.Oncetrained,this
| reward | model | can | provide | consistent, | | real-time | | | | | |
| ------ | ----- | --- | ------- | ----------- | --- | --------- | --- | --- | --- | --- | --- |
feedbackforeverymodeloutput,therebycircum-
ventingtheneedforconstanthumaninvolvement.
| A prominent | | example | of | this approach | | is RLHF | | | | | |
| ----------- | --- | ------- | --- | ------------- | --- | ------- | --- | --- | --- | --- | --- |
(Ouyangetal.,2022),asillustratedinFigure2(b).
| It first asks | human | | annotators | to | label the | prefer- | | | | | |
| ------------- | ----- | --- | ---------- | --- | --------- | ------- | --- | --- | --- | --- | --- |
encefordifferentLLMoutputsandthentrainthe
rewardmodeltopredictthehumanpreference.Af-
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| terward, | reinforcement | | learning | | (RL) algorithms | | | | | | |
| --------------- | ------------- | ------ | ------------ | --- | --------------- | --- | --- | --- | --- | --- | --- |
| (e.g., Proximal | | Policy | Optimization | | [Schulman | | | | | | |
etal.,2017])areemployedtooptimizethemodel.
| RLHF and | its | variants | have | proven | effective | in | | | | | |
| ---------- | ---- | -------- | ------ | ------ | ---------- | --- | --- | --- | --- | --- | --- |
| correcting | LLMs | to | become | more | beneficial | and | | | | | |
Figure3:Theillustrationsofthetwotypicalstrategies
| less harmful | (Bai | et | al., 2022a), | as | well | as instill- | | | | | |
| ------------ | ---- | --- | ------------ | --- | ---- | ----------- | --- | --- | --- | --- | --- |
ofgeneration-timecorrection:(a)Generate-then-Rank,
ingmoralcorrectness(Gangulietal.,2023).
and(b)Feedback-GuidedDecoding.
| Self-TrainingwithAutomatedFeedback. | | | | | | Re- | | | | | |
| ----------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
ward modeling still requires the collection of inaccessible, and colossal LLMs with billions
| human | feedback. | To | build | a fully | autonomous | | | | | | |
| ----- | --------- | --- | ----- | ------- | ---------- | --- | -------------- | ---- | ------------ | --------------- | --- |
| | | | | | | | of parameters. | This | necessitates | generation-time | |
self-improving agent, recent work has adopted correction methods that correct LLMs dur-
the self-training strategy that self-improves LLM ing the generation time. Two main strategies
| by bootstrapping | | its | original | outputs, | as | depicted | | | | | |
| ---------------- | ----- | --- | -------- | -------- | ------ | -------- | ---------------------- | --- | --- | --------------- | --- |
| | | | | | | | are Generate-then-Rank | | and | Feedback-Guided | |
| in Figure | 2(c). | The | language | model | itself | is used | | | | | |
Decoding.
| to provide | feedback | | for | its own | output. | STaR | | | | | |
| ---------- | -------- | --- | --- | ------- | ------- | ---- | --- | --- | --- | --- | --- |
(Zelikmanetal.,2022)leveragestheideaofchain-
| | | | | | | | Generate-then-Rank. | | This | involves | sampling |
| --- | --- | --- | --- | --- | --- | --- | ------------------- | --- | ---- | -------- | -------- |
of-thought to prompt LLM to generate answers alargenumberofcandidategenerationsandsub-
with rationales. They found that the performance sequently picking up the best generation based
of LLM can be improved by iteratively selecting on the feedback provided by the critic model, as
rationalesleadingtothecorrectanswertofurther illustrated in Figure 3(a). This approach is of-
finetune LLM. Self-training has also been used ten integrated with chain-of-thought prompting
to reduce the harmful responses of LLMs. For (Wei et al., 2022b) to tackle complex reasoning
| example, | in RLAIF | | (Bai | et al., | 2022b), | the ini- | | | | | |
| -------- | -------- | --- | ---- | ------- | ------- | -------- | --- | --- | --- | --- | --- |
tasks,suchassolvingmathwordproblems.Given
tialtoxicresponsesarecriticiquedandrevisedby an input problem x, the LLM initially generates
,···
the LLM itself following a set of human-defined multiplecandidatesolutionsy ,y .Eachso-
| | | | | | | | | | | 1 | n |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
principles. Afterward, the LLM is fine-tuned on lution y i = [z i ,a i ] comprises a reasoning path
therevisedresponses.AlpacaFarm(Duboisetal., (explanation) z leading to the predicted answer
i
2023) further shows that LLMs can self-improve a . Subsequently, the critic model C assigns a
i
withRL.ItdesignsLLMpromptstosimulatehu- plausibility score s to each candidate reason-
i
| man feedback | | in RLHF | and | shows | that | the feed- | | | | | |
| ------------ | --- | ------- | --- | ----- | ---- | --------- | -------- | -------------- | -------- | ----------- | ---- |
| | | | | | | | ing path | z i . The best | solution | is selected | from |
backiseffectiveandgreatlyreducesthecost. the scored set (z ,a ,s )n via either ranking or
| | | | | | | | | i | i i i=1 | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | ------- | --- | --- |
voting.
3.2 Generation-TimeCorrection
VariouscriticmodelshavebeenusedforLLM
Correcting LLMs at training time is ideal but output verification. DIVERSE (Li et al., 2023d)
not always feasible because it can be resource- trains a binary verifier based on DeBERTa (He
intensive or even impractical for many LLMs, et al., 2021) to rate each reasoning path. Weng
e.g., closed-source LLMs where weights are etal.(2023)introducedatraining-freecriticmodel
490
based on the consistency between forward and et al., 2023) or by making text perturbations
backward reasoning. In a different vein, RR (He onpositivesamples(Yangetal.,2022a).
| et al., | 2023) | used a | critic | model | to assess | rea- | | | | | | | |
| ------- | ----- | ------ | ------ | ----- | --------- | ---- | --- | --- | --- | --- | --- | --- | --- |
•
| | | | | | | | External | Metric: | | Several | studies | also | lever- |
| ------ | ----------------- | --- | --- | ---------- | ---------- | --- | -------- | -------- | ------- | ------- | ------- | -------- | ------ |
| soning | path faithfulness | | by | retrieving | supporting | | | | | | | | |
| | | | | | | | age | external | metrics | to | re-rank | or guide | text |
informationfromaknowledgebase.Incodegen-
generationwithoutadditionalmodeltraining,
| eration, | LEVER | (Ni | et al., | 2023) | uses | a veri- | | | | | | | |
| ------------ | ----- | ------- | --------- | ----- | -------- | ------- | ---- | -------- | ------- | --- | ----- | ---- | ------ |
| | | | | | | | such | as using | minimum | | Bayes | risk | decod- |
| fier trained | on | program | execution | | results. | CodeT | | | | | | | |
ing(Freitagetal.,2022),attributeclassifiers
| (Chen et | al., 2023b) | | similarly | employs | | dual exe- | | | | | | | |
| -------- | ----------- | --- | --------- | ------- | --- | --------- | ---------- | --- | ------- | ----- | ---- | --- | ------ |
| | | | | | | | (Dathathri | | et al., | 2020; | Yang | and | Klein, |
cutionagreementtoselectthebestcodesolution.
| | | | | | | | 2021), | and | Gaussian | | denoising | (Li | et al., |
| ---------------------------------------------- | --- | --- | --------- | ------- | -------- | --------- | -------- | --- | ---------- | --- | --------- | --------- | ------- |
| Feedback-Guided | | | Decoding. | Despite | | its effi- | 2022). | | | | | | |
| ciency,thegenerate-then-rankstrategyhasseveral | | | | | | | • | | | | | | |
| | | | | | | | External | | Knowledge: | | External | knowledge | |
| limitations: | 1) | The | critic | model | provides | only | | | | | | | |
sourceshavealsobeenusedtoprovidefeed-
coarse-grained,output-levelfeedback,2)Thelong
| | | | | | | | back. | Varshney | | et al. | (2023) | use Wikipedia | |
| --- | --- | --- | --- | --- | --- | --- | ----- | -------- | --- | ------ | ------ | ------------- | --- |
lengthoftheoutputcancomplicateitsqualityas- Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| | | | | | | | to validate | | and | correct | each | generated | sen- |
| --- | --- | --- | --- | --- | --- | --- | ----------- | --- | --- | ------- | ---- | --------- | ---- |
sessment,and3)ItrequirestheLLMtowaituntil
tence,whichisthenreinsertedforfurthergen-
theentireoutputisgeneratedforanycorrections.
eration.Alternatively,MemPrompt(Madaan
| The | feedback-guided | | decoding | | strategy | shown | | | | | | | |
| --- | --------------- | --- | -------- | --- | -------- | ----- | --- | --- | --- | --- | --- | --- | --- |
etal.,2022)utilizesapoolofprioruserfeed-
| in Figure | 3(b) | overcomes | | the above | limitations | | | | | | | | |
| --------- | ---- | --------- | --- | --------- | ----------- | --- | ---- | -------- | --- | ---- | ---------- | ----- | --- |
| | | | | | | | back | to guide | the | text | generation | based | on |
byusingstep-levelfeedbackforfine-grainedcon-
thecurrentquery’sintent.
| trol during | generation. | | Each | output | y is | split into | | | | | | | |
| ----------- | ----------- | --- | ---- | ------ | ---- | ---------- | --- | --- | --- | --- | --- | --- | --- |
| | | | | | ,··· | | • | | | | | | |
multiple reasoning steps y = [o ,o ,o ]. A Self-Evaluation: For better flexibility, meth-
| | | | | 1 | 2 | n | | | | | | | |
| ------------ | --------- | --- | ---- | ---- | ------------- | --- | --- | ---- | ------------------ | --- | --- | ---- | ------- |
| | | | | | | | ods | such | as Tree-of-Thought | | | (Yao | et al., |
| critic model | evaluates | | each | step | o t , guiding | al- | | | | | | | |
gorithms like beam search to explore the output 2023a) and Guided-decoding (Xie et al.,
space systematically and correct early mistakes. 2023) use the LLM itself as the critic model
This strategy also helps alleviate the reason- by prompting it to evaluate each individ-
inginconsistencyproblem(Zelikmanetal.,2022; ual reasoning step, avoiding the need for
CreswellandShanahan,2022),i.e.,incorrectrea- fine-tuningtask-specificverifier.
soningleadstocorrectfinalanswer.Thisstrategy
| | | | | | | | Different | strategies | | are adopted | | to control | the |
| -------- | ------- | --- | ------ | ----- | ---- | -------- | --------- | ---------- | --- | ----------- | --- | ---------- | --- |
| has been | adopted | in | recent | works | like | Tree-of- | | | | | | | |
Thought (Yao et al., 2023a), GRACE (Khalifa decoding process with the help of the step-level
| | | | | | | | critic model. | Tree-of-Thought | | | uses | breadth-first | |
| -------------- | --- | ------- | ------ | ------- | ------ | ------- | --------------- | --------------- | --------- | ----- | ----- | ------------- | -------- |
| et al., 2023), | | and RAP | (Hao | et al., | 2023), | which | | | | | | | |
| | | | | | | | and depth-first | | searches, | while | GRACE | | (Khalifa |
| vary mainly | | in the | critic | model | they | employ, | | | | | | | |
categorized into methods involving human feed- et al., 2023) and Xie et al. (2023) employ beam
back, trained verifiers, external metrics, external search. CoRe (Zhu et al., 2023) and RAP (Hao
| | | | | | | | et al., 2023) | use | Monte | Carlo | Tree | Search | for a |
| --- | --- | --- | --- | --- | --- | --- | ------------- | --- | ----- | ----- | ---- | ------ | ----- |
knowledge,andself-evaluation.
balancebetweenexplorationandexploitation.
•
RewardModelfromHumanFeedback:Stud-
| | | | | | | | 3.3 Post-hocCorrection | | | | | | |
| ----- | ----------- | ------- | -------- | --------------- | ------------ | ----- | ---------------------- | --- | --- | --------------- | --- | ---------- | --- |
| ies | like Uesato | | et al. | (2022) | and Lightman | | | | | | | | |
| et | al. (2023) | collect | | human-annotated | | step- | | | | | | | |
| | | | | | | | The effectiveness | | of | generation-time | | correction | |
| level | feedback | | to train | a more | robust | re- | | | | | | | |
hingesonthecriticmodel’sabilitytogiveprecise
| ward | model, | | which | improves | reasoning | | | | | | | | |
| ---- | ------ | --- | ----- | -------- | --------- | --- | -------- | ---------------- | --- | --- | -------- | ------------- | --- |
| | | | | | | | feedback | for intermediate | | | outputs, | a challenging | |
faithfulness.
| | | | | | | | task in holistic | | NLP | evaluations | like | summariza- | |
| --- | --- | --- | --- | --- | --- | --- | ---------------- | --- | --- | ----------- | ---- | ---------- | --- |
• Trained Verifier: To reduce the cost of hu- tion.Thismotivatesthepost-hoccorrectionmeth-
man annotations, some work (Yang et al., ods, where both critic and refinement models act
2022a; Tafjord et al., 2022; Li et al., 2023d; only after generating the complete output. Post-
Khalifa et al., 2023) uses automated meth- hoccorrectionallowsformorediversenaturallan-
ods to generate training data for obtaining guage feedback, ranging from specific diagnostic
a step-wise verifier. Positive examples are reports to broader writing suggestions. As shown
derived from ground-truth reasoning paths, in Figure 4, we categorize the key post-hoc cor-
while negative examples are synthesized by rection strategies into Self-Correction, Correction
proposing an alignment algorithm (Khalifa withExternalFeedback,andMulti-AgentDebate.
491
Figure4:Threepost-hoccorrectionstrategies:self-correction(a),externalfeedback(b),multi-agentdebate(c).
Self-Correction. In ‘‘Self-Correction’’, a sin- tools,includingtrainedmodels,codeinterpreters,
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
gle LLM both generates and refines its output. and search engines, offer specialized feedback to
As shown in Figure 4(a), the LLM first pro- addressLLMconstraints.
| duces an | output | and then | acts | as | its critic | for | | | | | | |
| -------- | ------ | -------- | ---- | --- | ---------- | --- | --- | --- | --- | --- | --- | --- |
•
iterative refinements. This process continues un- Code Interpreter. In code generation, mod-
| til the | output | obtains | an acceptable | | quality | or | | | | | | |
| ------- | ------ | ------- | ------------- | --- | ------- | --- | -------- | --------- | ------ | --- | ----------- | --- |
| | | | | | | | els like | Self-Edit | (Zhang | et | al., 2023a) | and |
a pre-specified number of iterations is reached. Self-Evolve employ program executors to
Self-Refine (Madaan et al., 2023) introduced an provide feedback from executed test cases.
effective framework using one LLM guided Others, like Self-Debug (Chen et al., 2023e)
by varied prompts for the roles of generation, andALGO(Zhangetal.,2023b),explorede-
critic, and refinement, respectively. Clinical Self- tailedfeedbackmechanismsusingunittests,
Verification (Gero et al., 2023) applies this to program explanations, or comparison with
| extract | clinical | data, refining | | by spotting | | missing | | | | | | |
| ------- | -------- | -------------- | --- | ----------- | --- | ------- | --------- | ------ | --------- | --- | ------------ | --- |
| | | | | | | | reference | oracle | programs. | | Charalambous | |
elements and verifying data accuracy. Reflexion et al. (2023) use Bounded Model Checking
(Shinn et al., 2023) extends the method, adding forsoftwareverification.
| a ‘‘long-term | | memory’’ | to recall | past | errors | and | | | | | | |
| ------------- | --- | -------- | --------- | ---- | ------ | --- | --- | --- | --- | --- | --- | --- |
• LogicReasoner.Logic-LM(Panetal.,2023)
integratingdiversefeedbackforms.
| | | | | | | | and Baldur | (First | et | al., 2023) | harness | ex- |
| ---------------------- | ---------- | ------- | ------- | --------------- | --------- | --- | ----------- | --------- | --- | ---------- | ---------------- | --- |
| Though | beneficial | | in many | text-generation | | | | | | | | |
| | | | | | | | ternallogic | reasoners | | and | proof assistants | to |
| tasks, self-correction | | usually | | demands | powerful, | | | | | | | |
refineLLMoutputs,usingerrormessagesas
| large-scale | LLMs | for effectiveness, | | | which | sacri- | | | | | | |
| ----------------- | ---- | ------------------ | --- | --- | ------ | ------ | -------- | ----------- | --- | --------- | --- | -------- |
| | | | | | | | feedback | for logical | | reasoning | and | theorem- |
| fices efficiency. | | As observed | | by | Madaan | et al. | | | | | | |
proofgeneration.
| (2023), | smaller | models | often | falter | in | refining, | | | | | | |
| ------- | ------- | ------ | ----- | ------ | --- | --------- | --- | --- | --- | --- | --- | --- |
even with correct feedback. A possible solution • ExternalKnowledgeisusedtoensurefactual
| involves | explicitly | training | models | | for this | self- | | | | | | |
| ---------- | ---------- | -------- | ------ | --- | ---------- | ----- | -------- | ------ | ------- | ------ | ---- | ------- |
| | | | | | | | accuracy | of the | output. | Models | like | RARR |
| correction | process. | SelFee | (Ye | et | al., 2023) | pro- | | | | | | |
| | | | | | | | (Gao et | al., | 2023b), | REFEED | (Yu | et al., |
posestrainingamodeltoemulatetheself-correction
| | | | | | | | 2023), | and LLM-Augmenter | | | (Peng | et al., |
| --- | --- | --- | --- | --- | --- | --- | ------ | ----------------- | --- | --- | ----- | ------- |
process by generating output, feedback, and a re- 2023)promptLLMstoquestiontheiroutputs.
finedsolutioninanauto-regressivemanner.They
| | | | | | | | An external | retriever | | then | searches | for rele- |
| -------- | -------- | ---- | --- | ------- | -------- | --- | -------------- | --------- | ----- | ------- | --------- | --------- |
| use more | powerful | LLMs | to | provide | feedback | | | | | | | |
| | | | | | | | vant evidence, | | which | is used | to refine | out- |
and refinement data, with data collection facili- puts.FACTOOL(Chernetal.,2023)extends
tatedthroughChatGPT.
| | | | | | | | this approach | | to a wider | range | of | tasks, in- |
| ------------ | --- | ------------ | --- | ------------------- | --- | --- | ------------- | ---- | ----------- | ------------ | --- | ---------- |
| | | | | | | | cluding | code | generation, | mathematical | | rea- |
| Models/Tools | | as Feedback. | | In self-correction, | | | | | | | | |
soning,andscientificliteraturereview.
| the quality | of | the feedback | is | constrained | | by the | | | | | | |
| ----------- | --- | ------------ | --- | ----------- | --- | ------ | --- | --- | --- | --- | --- | --- |
•
inherentlimitationsofLLMs,suchastheinability TrainedModel.Researchhasfine-tunedspe-
to access up-to-date information, take actions, or cializedcriticmodelstoprovidefeedbackfor
perform precise mathematical reasoning. To en- iterative refinement alongside more power-
hancefeedbackquality,recentresearchleverages ful language models. For example, CodeRL
external tools, as shown in Figure 4(b). These (Le et al., 2022) treats program synthesis
492
as a reinforcement learning task and trains ity, Fu et al. (2023) mimicked real-world human
a critic model whose output optimizes the interactions, like a buyer-seller scenario, show-
main model. REFINER (Paul et al., 2023) casingtheversatilityofmulti-agentdebates.
usesacritiquemodeltoprovidefeedbackon
| an intermediate | | representation, | | suitable | for | | | | | | |
| --------------- | --- | --------------- | --- | -------- | --- | --- | --- | --- | --- | --- | --- |
4 Discussion
| refining | larger | models | like | ChatGPT. | Simi- | | | | | | |
| -------- | ------ | --------- | ------- | -------- | -------- | --- | --- | --- | --- | --- | --- |
| larly, | RL4F | (Akyu¨rek | et al., | 2023) | trains a | | | | | | |
4.1 PriorResearchonAutomatedCorrection
criticviareinforcementlearning,fine-tuning
| | | | | | | In our survey, | | we primarily | examine | | the auto- |
| --- | --- | --- | --- | --- | --- | -------------- | --- | ------------ | ------- | --- | --------- |
itwithpolicyoptimization.Theeffectiveness
is gauged by comparing the refined output’s mated correction strategies in the era of modern
accuracytogroundtruth.Inadversarialcon- large language models. However, the idea of
‘‘correctingthemodelwithautomatedfeedback’’
texts,feedbackfromcontentfilterscanguide
the generation of better adversarial exam- has been a longstanding practice in diverse NLP
ples, like how FLIRT (Mehrabi et al., 2023) tasks. Recognizing these early works provides a
deeperhistoricalinsightintotheevolutionofself- Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| leverages | image | classifier | | signals | to guide | | | | | | |
| -------------- | ----------- | ----------- | --- | ------- | -------- | ---------- | -------- | ------------ | ---- | ------------ | ---------- |
| | | | | | | correction | methods | within | NLP. | Next, | we briefly |
| LLMs | in creating | adversarial | | prompts | for | | | | | | |
| auditpurposes. | | | | | | discuss | the NLP | applications | | where | automated |
| | | | | | | correction | has been | effectively | | implemented, | and |
•
Integrating Multiple Tools. Broadening the we discuss how these early works link to the
ideaoftool-assistedfeedback,CRITIC (Gou automated correction strategies defined in this
| et al., | 2023) | unifies | various | tools, | such as | | | | | | |
| ------- | ----- | ------- | ------- | ------ | ------- | --- | --- | --- | --- | --- | --- |
survey.
| code | interpreters, | search | engines, | | and LLM | | | | | | |
| ---- | ------------- | ------ | -------- | --- | ------- | --- | --- | --- | --- | --- | --- |
feedback, offering a multifaceted feedback MachineTranslation. Theconceptofpost-hoc
approach.
| | | | | | | self-correction | has | deep roots | in | the field | of ma- |
| --- | --- | --- | --- | --- | --- | ----------------- | ------------ | ---------- | ----- | ----------- | ------- |
| | | | | | | chine translation | | (MT), | where | it is often | called |
| | | | | | | Automatic | Post-Editing | (APE) | | (do Carmo | et al., |
3.4 Multi-AgentDebate
| | | | | | | 2021). A | long | line of prior | work | trains | models |
| --- | --- | --- | --- | --- | --- | -------- | ---- | ------------- | ---- | ------ | ------ |
Besidesintegratingtools,recentresearchhasalso to fix translation errors by either learning from
explored the debate approach among multiple human correction data (Alabau et al., 2014) or
LLMs, inspired by the idea that multiple per- from synthetic training data (Lee et al., 2021).
spectives can converge to an improved solution. To minimize the cost of data collection, recent
Multiple LLM instances debate their individ- work(Chenetal.,2023d;Raunaketal.,2023)has
ual answers over several rounds, aiming for a leveragedthein-contextlearningabilityofLLMs
consensus. for post-editing translations. As well as post-hoc
Duetal.(2023)trialedthisinarithmeticreason- methods, training-time correction (Unanue et al.,
ing.Agents,orLLMduplicates,presentindividual 2021)anddecoding-timecorrection(Freitagetal.,
solutions and justifications. In the debate phase, 2022)arealsoadoptedbypriorworks.
theseresponsesareaggregatedandusedascontext
| for each agent | to | revise its | original | answer. | After | Summarization. | | | | | |
| -------------- | --- | ---------- | -------- | ------- | ----- | -------------- | --- | -------- | --- | --------- | ----- |
| | | | | | | | | The idea | of | automated | model |
several iterations, they typically reach a consen- correction has been commonly used in summa-
sus, showing superior performance compared to rization to ensure the factuality of the generated
PRD
self-correction. (Li et al., 2023c) furthered summary. There are two mainstream methods:
this by introducing the peer rank algorithm to 1)training-timecorrectionthatimposesfactuality
optimizetheconsensusprocess.Itconsiderspair- constraints during training (Liu and Liu, 2021;
wisepreferencesbetweenallpossibleanswerpairs WanandBansal,2022;Scheureretal.,2023),and
fromindividualLLMsandusesthesepreferences 2) post-hoc correction that post-edits generated
togenerateafinalrankingofmodels. summaries to correct factual errors (Falke et al.,
In addition to reasoning tasks, LM vs LM 2019;Caoetal.,2020;Saundersetal.,2022).Re-
(Cohen et al., 2023) employed this debate ap- centworkhasinvestigatedusing RLtorefinethe
proach for factual error detection, where a gener- modelguidedbyautomatedfeedbackfromeither
atingLLMmakesaclaimandanexaminingLLM rewardmodels(Akyu¨reketal.,2023)orlanguage
checks for inaccuracies. Extending its applicabil- models(Pangetal.,2023).
493
Semantic Parsing. The use of external feed- Hong et al., 2023; Tyen et al., 2023; Valmeekam
backinsemanticparsing,particularlyforText-to- et al., 2023; Ke et al., 2023). Most findings in-
SQL tasks, has shown significant effectiveness. dicate that LLMs struggle to rectify their initial
Execution-guided semantic parsing is a notable mistakes,andtheirperformanceevenworsensaf-
approachwherethefeedbackfromexecutingpar- ter self-correction. This issue arises because the
tial SQL queries guides the search for plausible quality of the model’s self-generated feedback is
complete SQL programs. Additionally, earlier bounded by its existing knowledge and abilities.
works also explored training separate discrimi- Therefore, internal feedback may not offer any
nativemodelseithertorerankthegeneratedSQL extraadvantageforimprovingtheresults;itmight
queries (Bogin et al., 2019; Kelkar et al., 2020) even steer the model away from the correct an-
ortopredictspecificSQLcomponents(Xuetal., swer.Preventingsuchmis-guidanceiscrucialfor
2017; Yu et al., 2018; Lee, 2019). The effective- successfulself-correction(Huangetal.,2023).
ness of these generation-time correction tech- In contrast, the use of external feedback for
niquesislargelyattributabletotheeaseofdefining automated correction has shown more promise.
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
intermediatefeedbackinsemanticparsing. Numerous studies (Pan et al., 2023; Chen et al.,
2023a;Gouetal.,2023;Huangetal.,2023)report
Proof Generation. Automated correction has positive outcomes when LLMs leverage high-
beenwellstudiedandimplementedforproofgen-
qualityfeedbackfromexternalsources.However,
| eration | (Saha | et al., | 2020; | Tafjord | et | al., 2021). | | | | | | | |
| ------- | ----- | ------- | ----- | ------- | --- | ----------- | ------------ | -------- | --- | -------- | --- | ----------- | --- |
| | | | | | | | high-quality | external | | feedback | is | unavailable | in |
External feedback from natural language infer- manyreal-worldapplications.Thisconstraintnar-
ence(NLI)arecommonlyusedtospoterrorsasa
| | | | | | | | rows down | the | scope | of automated | | correction | to |
| --------- | --- | ----------- | --- | ---- | ----- | -------- | --------- | --- | ----- | ------------ | --- | ---------- | --- |
| heuristic | for | correction, | and | as a | means | to score | | | | | | | |
onlythosetaskswherepreciseandreadilyobtain-
| the quality | (Yang | et | al., 2022a; | Golovneva | | et al., | | | | | | | |
| ----------- | ----- | --- | ----------- | --------- | --- | ------- | ------------- | --- | -------- | ------- | ---- | ------------- | --- |
| | | | | | | | able external | | feedback | exists, | such | as arithmetic | |
2023). However, there are some open questions reasoning,semanticparsing,andcodegeneration.
| regarding | the | quality | of | NLI-based | | feedback | | | | | | | |
| --------- | --- | ------- | --- | --------- | --- | -------- | --- | --------- | ----- | --- | ----- | ------ | ------ |
| | | | | | | | The | empirical | study | by | Huang | et al. | (2023) |
(SrikanthandRudinger,2022;Saxonetal.,2023).
| | | | | | | | highlighted | multi-agent | | debate | as | an effective | |
| ---------- | ------- | ----------- | --- | -------- | ------- | ---------- | ----------- | ------------- | ----------- | ---------- | --------- | ------------ | ----- |
| | | | | | | | method | for automated | | correction | in | LLMs. | How- |
| Open-Ended | | Generation. | | Post-hoc | | correction | | | | | | | |
| | | | | | | | ever, the | observed | improvement | | primarily | | stems |
| is often | adopted | to improve | | the | quality | of open- | | | | | | | |
fromthemodel-drivenvotingprocessamongdif-
| ended | text | generation | (Wang | | et | al., 2017; | | | | | | | |
| --------- | ---------- | ---------- | ------------- | --- | --------- | ----------- | ------------- | -------- | ---------- | -------- | -------- | ---------------- | --- |
| | | | | | | | ferent LLMs, | | rather | than | from | self-correction. | |
| Holtzman | et | al., 2018; | Sagarkar | | et | al., 2018), | | | | | | | |
| | | | | | | | This approach | | represents | another | | successful | in- |
| such as | correcting | toxic | outputs, | | enhancing | the | | | | | | | |
| | | | | | | | stance of | learning | through | | external | feedback, | as |
| narrative | quality | in story | generation, | | and | refining | | | | | | | |
| | | | | | | | each LLM | benefits | | from the | input | provided | by |
| response | generation | | in dialogues. | | For | example, | | | | | | | |
otherLLMsinthedebate.
| Holtzman | et | al. (2018) | proposed | | a | framework | | | | | | | |
| --------- | --- | ---------- | ----------- | --- | ---------------- | --------- | --- | --- | --- | --- | --- | --- | --- |
| to refine | the | generic, | repetitive, | | and inconsistent | | | | | | | | |
5 ResearchGapsandFutureDirections
textsbycomposingacommitteeofdiscriminators
to provide multi-aspect feedback. Given the sub- 5.1 TheoreticalJustifications
| jectivity | involved | in | assessing | the | outputs, | recent | | | | | | | |
| --------- | -------- | ------- | --------- | ------------- | -------- | -------- | ---------- | ------------ | -------- | ---- | ---------------- | -------- | ----- |
| | | | | | | | First of | all, whether | | LLMs | can self-correct | | with- |
| works | started | to use | detailed, | natural | | language | | | | | | | |
| | | | | | | | out any | external | feedback | is | still an | ongoing | de- |
| feedback | and | utilize | LLMs | for iterative | | post-hoc | | | | | | | |
| | | | | | | | bate, with | both | positive | and | negative | outcomes | |
refinement.
| | | | | | | | reported. | Numerous | | studies | have | discovered | |
| --- | --- | --- | --- | --- | --- | --- | -------------------- | -------- | ----- | ------- | -------- | ---------- | ------- |
| | | | | | | | that self-correction | | often | brings | negative | | effects |
4.2 WhenDoesAutomated
(Huangetal.,2023;Tyenetal.,2023),whilesome
CorrectionWork?
| | | | | | | | research | indicates | that | the | effectiveness | | of self- |
| --- | --- | --- | --- | --- | --- | --- | -------- | --------- | ---- | --- | ------------- | --- | -------- |
Despite the relative infancy of this emerging repair is only seen in GPT-4 (Olausson et al.,
field, recent studies have explored the efficacy 2023). Although these empirical studies provide
of automated correction in LLMs. Notably, in- valuable insights, more fundamental theoretical
trinsic self-correction—where the model cor- research is needed to gain a mechanistic under-
rects its initial output based solely on its inherent standing of self-correction. Key research ques-
capabilities—has generally shown disappointing tions include: Can LLMs truly recognize their
results (Huang et al., 2023; Stechly et al., 2023; own errors without external feedback? What is
494
the upper bound of intrinsic self-correction? An- can enhance themselves through self-training on
swerstothosequestionsmightcloselyassociated positivelyevaluatedoutputs,theyoftenfocusona
with LLMs’ capacity to exhibit metacognitive single,one-timecorrectionprocess.Theresilience
awareness, i.e., their understanding of their own of this self-training in continuous settings is not
knowledge and uncertainties (Kadavath et al., well-understood. Continual learning poses chal-
2022). The concept of calibration—how well a lenges like catastrophic forgetting (Kirkpatrick
model’s predictions match observed outcomes— et al., 2016), where new skills impair old ones.
isalsocrucial(Linetal.,2023). It’s uncertain if such issues could plague con-
While language models demonstrate some tinually self-improving LLMs, e.g., correcting
capacityforself-feedback,achievingsuperiorper- one behavior may unintentionally alter a previ-
formanceoftennecessitatesincorporatingexternal ouslycorrectedbehavior.Combiningvariousself-
feedback. This ties into the alignment of lan- correction techniques for continual improvement
guage models, an area still not fully understood. also warrants exploration. Integrating immediate
For example, in RLHF, the choice of the metric post-hoccorrectionswithlong-cycletraining-time
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
to minimize between the reward model output corrections—using the former to gather data
and the final model output significantly impacts and the latter to periodically address recurrent
downstream task performance (Go et al., 2023), problems—couldbeapromisingapproach.
| yet this | aspect remains | underexplored | | in | many | | | | | |
| ------------- | -------------- | ------------- | ---- | -------- | ---- | ----------------------------------- | --- | --- | --- | --- |
| | | | | | | 5.4 Self-CorrectionwithModelEditing | | | | |
| applications. | Determining | the | best | approach | to | | | | | |
auto-generate instructive prompts for tasks like Recent advancements in model editing (Sinitsin
outputevaluationisalsoanopenchallenge. et al., 2020; Cao et al., 2021; Yao et al., 2023b)
| | | | | | | aim to adjust | the model’s | behavior | for examples | |
| --- | --- | --- | --- | --- | --- | ------------- | ----------- | -------- | ------------ | --- |
5.2 BenchmarkingAutomatedCorrection within the editing scope while leaving its perfor-
WhileLLMautomatedcorrectionhasseenempir- manceforout-of-scopeexamplesunaltered.Ithas
ical advancements across applications, there is a been applied to update LLMs’ outdated knowl-
lack of solid quantitative metrics to evaluate this edge (Lee et al., 2022; Onoe et al., 2023) and
capability. Comprehensive evaluations comparing address false associations (Murty et al., 2022;
various strategies on criteria like effectiveness, Tannoetal.,2022).Thougheffectiveinadjusting
complexity, and potential limits are still needed. LLMs’factualknowledge,challengeslikelimited
Future studies could develop evaluation frame- generalization(Yaoetal.,2023b)andunintended
works considering variables such as task com- side effects persist (Hoelscher-Obermaier et al.,
| plexity, | degree of initial | error, | improvement | | in | 2023). | | | | |
| -------- | ----------------- | ------ | ----------- | --- | --- | ------ | --- | --- | --- | --- |
qualityafterautomatedcorrection,etc. We believe model editing offers great poten-
Setting benchmarks to diagnose automated tial for LLM self-correction. It enables accurate,
correction is another potential research avenue. fine-grained corrections without full-scale re-
Diagnosticdatasetswouldofferstandardizedeval- training. Analyzing the impact of these model
uations of LLMs and their correction strategies, edits could yield insights into self-correction.
fosteringthedevelopmentofmoreprecisemodels. Techniques mitigating model editing’s side ef-
| | | | | | | fects (Hoelscher-Obermaier | | et al., | 2023) | may |
| --- | --- | --- | --- | --- | --- | -------------------------- | --- | ------- | ----- | --- |
5.3 ContinualSelf-Improvement
| | | | | | | also enhance | self-correction. | We | anticipate | fu- |
| --- | --- | --- | --- | --- | --- | ------------ | ---------------- | --- | ---------- | --- |
Another promising yet under-explored area of tureresearchtoincreasinglymergemodelediting
LLM self-correction is the idea of continual, with LLM self-correction, a relatively untouched
| life-long | self-improvement. | | As LLMs | are | inte- | | | | | |
| --------- | ----------------- | --- | ------- | --- | ----- | --- | --- | --- | --- | --- |
domain.
| grated into | varied and | evolving | scenarios, | | their | | | | | |
| ----------- | ------------- | ------------ | ---------- | --- | ----- | ------------------------------ | --- | --- | --- | --- |
| | | | | | | 5.5 Multi-modalSelf-Correction | | | | |
| capacity | for sustained | adaptability | becomes | | cru- | | | | | |
cial. This mirrors the notion of continual (life- Self-correction strategies have been well-tested
long) learning (Wang et al., 2023c), suggesting on the textual modality, where both the model
LLMs should consistently assess outputs, rectify outputsandthefeedbackareintextualform.The
mistakes,updateknowledge,andadjustdecision- recentsurgeinmulti-modaldatausage,including
| making. | | | | | | image,audio,andvideomodalities,presentsentic- | | | | |
| ------- | --- | --- | --- | --- | --- | --------------------------------------------- | --- | --- | --- | --- |
While recent studies like Huang et al. (2022) ingopportunitiesforexpansion.Theseincludethe
and Zelikman et al. (2022) indicate that LLMs exploration of self-correction capabilities within
495
multi-modal LLMs, the incorporation of visual CASMACAT: A computer-assisted translation
feedback, and improving vision-language tasks workbench. In Proceedings of the 14th Con-
throughself-correction. ference of the European Chapter of the Asso-
| | | | | | | | ciation | for Computational | | | Linguistics | | (EACL), |
| ------------ | ------ | ---------- | --- | ------------- | --- | ---- | ------------ | ----------------- | ----------------------- | ----------- | ----------- | ------------ | ------- |
| 6 Conclusion | | | | | | | pages | 25–28. | The | Association | | for Computer | |
| | | | | | | | Linguistics. | | https://doi.org/10.3115 | | | | |
| In this | paper, | we present | a | comprehensive | | sur- | /v1/E14-2007 | | | | | | |
veyofself-correctinglargelanguagemodelswith
| | | | | | | | Yuntao | Bai, | Andy | Jones, | Kamal | | Ndousse, |
| ------------ | --------- | --------------- | --- | ----------- | --------- | ---- | ------ | ------- | ---- | ------ | ----- | --------- | -------- |
| automated | feedback. | | We | categorize | and | ana- | | | | | | | |
| | | | | | | | Amanda | Askell, | Anna | Chen, | Nova | DasSarma, | |
| lyze various | | self-correction | | strategies, | including | | | | | | | | |
training-time, generation-time, and post-hoc cor- Dawn Drain, Stanislav Fort, Deep Ganguli,
| | | | | | | | Tom | Henighan, | | Nicholas | Joseph, | | Saurav |
| --------- | --- | ------------ | -------------- | ---- | --------- | ----- | --------- | --------- | --- | -------- | ------- | --- | -------- |
| rections. | We | also connect | recent | work | with | prior | | | | | | | |
| | | | | | | | Kadavath, | Jackson | | Kernion, | | Tom | Conerly, |
| research | and | discuss | the applicable | | scenarios | for | | | | | | | |
automatedcorrection.Finally,weoutlinefivepo- Sheer El Showk, Nelson Elhage, Zac Hatfield-
Dodds,DannyHernandez,TristanHume,Scott
tentialfuturedirectionsandassociatedchallenges Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| | | | | | | | Johnston, | Shauna | | Kravec, | Liane | Lovitt, | Neel |
| --- | --- | --- | --- | --- | --- | --- | --------- | ------ | --- | ------- | ----- | ------- | ---- |
inthisfield.Ourgoalwiththispaperistoprovide
| | | | | | | | Nanda, | Catherine | Olsson, | | Dario | Amodei, | Tom |
| --------------- | --- | --- | ------ | -------- | --- | ------- | ------ | --------- | ------- | --- | ----- | ------- | --- |
| a comprehensive | | and | useful | resource | for | readers | | | | | | | |
interested in the development of this rapidly B.Brown, JackClark,SamMcCandlish,Chris
| | | | | | | | Olah, | Benjamin | Mann, | | and | Jared | Kaplan. |
| ------------- | ------- | ------- | ------- | ------------ | ---- | ------ | ------ | -------- | ----- | ------- | --- | -------- | ------- |
| evolving | domain. | To | aid in | this effort, | we | create | | | | | | | |
| | | | | | | | 2022a. | Training | a | helpful | and | harmless | assis- |
| a continually | | updated | reading | list | in a | GitHub | | | | | | | |
repository as follows: https://github.com tant with reinforcement learning from human
feedback.CoRR,abs/2204.05862.
/teacherpeterpan/self-correction-llm
| -papers. | | | | | | | Yuntao Bai, | Saurav | Kadavath, | | Sandipan | | Kundu, |
| -------- | --- | --- | --- | --- | --- | --- | ----------- | ------ | --------- | --- | -------- | --- | ------ |
AmandaAskell,JacksonKernion,AndyJones,
| Acknowledgments | | | | | | | Anna | Chen, | Anna | Goldie, | Azalia | Mirhoseini, | |
| --------------- | --- | --- | --- | --- | --- | --- | ------- | --------- | ---- | ------- | ------ | ----------- | --- |
| | | | | | | | Cameron | McKinnon, | | Carol | Chen, | Catherine | |
This work was supported by the National Sci- Olsson, Christopher Olah, Danny Hernandez,
ence Foundation (award #2048122). The views Dawn Drain, Deep Ganguli, Dustin Li, Eli
| expressed | are | those | of the | authors | and | do not | | | | | | | |
| --------- | --- | ----- | ------ | ------- | --- | ------ | ------------- | --- | ----- | ------ | ----- | ----- | ----- |
| | | | | | | | Tran-Johnson, | | Ethan | Perez, | Jamie | Kerr, | Jared |
reflect the official policy or position of the US Mueller,JeffreyLadish,JoshuaLandau,Kamal
government. Thanks to Xinyuan Lu for assisting Ndousse, Kamile Lukosiute, Liane Lovitt,
withtheGithubreadinglistrepo. Michael Sellitto, Nelson Elhage, Nicholas
| | | | | | | | Schiefer, | Noem´ı | Mercado, | | Nova | DasSarma, | |
| --- | --- | --- | --- | --- | --- | --- | --------- | -------- | -------- | --- | ------- | --------- | ------- |
| | | | | | | | Robert | Lasenby, | Robin | | Larson, | Sam | Ringer, |
References
| | | | | | | | Scott | Johnston, | Shauna | | Kravec, | Sheer | El |
| ---------- | ----- | --------- | ----- | --------- | ------- | ------ | --------- | --------------- | ------ | ----- | ------- | ---------- | ------- |
| | | | | | | | Showk, | Stanislav | | Fort, | Tamera | | Lanham, |
| Afra Feyza | | Akyu¨rek, | Ekin | Akyu¨rek, | | Ashwin | | | | | | | |
| | | | | | | | Timothy | Telleen-Lawton, | | | Tom | Conerly, | Tom |
| Kalyan, | Peter | Clark, | Derry | Tanti | Wijaya, | and | | | | | | | |
| | | | | | | | Henighan, | Tristan | | Hume, | Samuel | R. Bowman, | |
NiketTandon.2023.RL4F:Generatingnatural
ZacHatfield-Dodds,BenMann,DarioAmodei,
languagefeedbackwithreinforcementlearning
| | | | | | | | Nicholas | Joseph, | | Sam | McCandlish, | | Tom |
| ------------- | ---- | ------------- | -------- | ----------- | ----------- | ------- | -------- | ------- | ------------ | ------- | ----------- | ------------ | --- |
| for repairing | | model | outputs. | In | Proceedings | | | | | | | | |
| | | | | | | | Brown, | and | Jared | Kaplan. | 2022b. | Constitu- | |
| of the | 61st | Annual | Meeting | of | the | Associ- | | | | | | | |
| | | | | | | | tional | AI: | harmlessness | | from | AI feedback. | |
| ation | for | Computational | | Linguistics | | (ACL), | | | | | | | |
CoRR,abs/2212.08073.
| pages | 7716–7733. | | https://doi.org/10 | | | | | | | | | | |
| ----- | ---------- | --- | ------------------ | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
.18653/v1/2023.acl-long.427
| | | | | | | | Gasper | Begus, | Maksymilian | | Dabkowski, | | and |
| ------ | ------- | --------- | ----- | --- | ------- | ----- | ------ | ------- | ----------- | ----- | ---------- | --- | ---- |
| | | | | | | | Ryan | Rhodes. | 2023. | Large | linguistic | | mod- |
| Vicent | Alabau, | Christian | Buck, | | Michael | Carl, | | | | | | | |
els:Analyzingtheoreticallinguisticabilitiesof
| Francisco | | Casacuberta, | | Mercedes | | Garc´ıa- | | | | | | | |
| --------- | --- | ------------ | --- | -------- | --- | -------- | --- | --- | --- | --- | --- | --- | --- |
LLMs.CoRR,abs/2305.00948.
| Mart´ınez, | | Ulrich | Germann, | Jesu´s | Gonza´lez- | | | | | | | | |
| ---------- | --- | ------ | -------- | ------ | ---------- | --- | --- | --- | --- | --- | --- | --- | --- |
Rubio, Robin L. Hill, Philipp Koehn, Luis A. Henrik Bellha¨user, Charlotte Dignath, and Maria
Leiva, Bartolome´ Mesa-Lao, Daniel Ortiz- Theobald. 2023. Daily automated feedback
Mart´ınez, Herve Saint-Amand, Germa´n enhances self-regulated learning: A longitudi-
Sanchis-Trilles, and Chara Tsoukala. 2014. nal randomized field experiment. Frontiers in
496
Psychology, 14:1125873. https://doi.org Justin Chih-Yao Chen, Swarnadeep Saha, and
/10.3389/fpsyg.2023.1125873, PubMed: Mohit Bansal. 2023c. Reconcile: Round-table
| 37275690 | | | | | | | conference | improves | | reasoning | via | consensus | |
| -------- | --- | --- | --- | --- | --- | --- | ---------- | -------- | --- | --------- | --- | --------- | --- |
amongdiverseLLMs.CoRR,abs/2309.13007.
| Ben Bogin, | Matt | Gardner, | and | Jonathan | | Berant. | | | | | | | |
| ---------- | ---- | -------- | --- | -------- | --- | ------- | --- | --- | --- | --- | --- | --- | --- |
2019. Global reasoning over database struc- PinzhenChen,ZhichengGuo,BarryHaddow,and
tures for text-to-SQL parsing. In Proceedings Kenneth Heafield. 2023d. Iterative translation
of the 2019 Conference on Empirical Meth- refinementwithlargelanguagemodels.CoRR,
| ods in | Natural | Language | Processing | | and | the | abs/2306.03856. | | | | | | |
| ----------------- | ---------- | ---------- | ------------------ | --- | --- | ----- | --------------- | ------- | -------------- | -------- | --------- | --------- | ---- |
| 9th International | | | Joint Conference | | on | Natu- | | | | | | | |
| | | | | | | | Xinyun Chen, | Maxwell | | Lin, | Nathanael | Scha¨rli, | |
| ral Language | | Processing | (EMNLP-IJCNLP), | | | | | | | | | | |
| | | | | | | | and Denny | Zhou. | 2023e. | Teaching | | large | lan- |
| pages | 3659–3664. | | https://doi.org/10 | | | | | | | | | | |
| | | | | | | | guage models | | to self-debug. | | CoRR, | abs/2304 | |
.18653/v1/D19-1378
.05128.
| Evelyn | M. Boyd | and | Ann | W. | Fales. | 1983. | | | | | | | |
| ------ | ------- | --- | --- | --- | ------ | ----- | ------------- | ------ | ------ | ----- | ----- | ------ | --- |
| | | | | | | | I-Chun Chern, | Steffi | Chern, | Shiqi | Chen, | Weizhe | |
Reflective learning: Key to learning from Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| | | | | | | | Yuan, Kehua | | Feng, Chunting | | Zhou, | Junxian | |
| ------------------ | --- | ------- | ------------------ | --- | -------- | --- | ----------- | --- | -------------- | ----------- | ----- | ------- | ----- |
| experience. | | Journal | of Humanistic | | Psychol- | | | | | | | | |
| | | | | | | | He, Graham | | Neubig, | and Pengfei | | Liu. | 2023. |
| ogy, 23(2):99–117. | | | https://doi.org/10 | | | | | | | | | | |
Factool:FactualitydetectioningenerativeAI–
.1177/0022167883232011
| | | | | | | | a tool | augmented | framework | | for | multi-task | |
| --- | --- | --- | --- | --- | --- | --- | ------ | --------- | --------- | --- | --- | ---------- | --- |
MengCao,YueDong,JiapengWu,andJackieChi
| | | | | | | | and multi-domain | | scenarios. | | CoRR, | abs/2307 | |
| ----------- | --- | ----- | ------------- | ---------- | --- | --- | ---------------- | --- | ---------- | --- | ----- | -------- | --- |
| Kit Cheung. | | 2020. | Factual error | correction | | for | | | | | | | |
.13528.
abstractivesummarizationmodels.InProceed-
ingsofthe2020ConferenceonEmpiricalMeth- ElizabethClark,TalAugust,SofiaSerrano,Nikita
| | | | | | | | Haduong, | Suchin | Gururangan, | | and | Noah | A. |
| --- | --- | --- | --- | --- | --- | --- | -------- | ------ | ----------- | --- | --- | ---- | --- |
odsinNaturalLanguageProcessing(EMNLP),
| | | | | | | | Smith. | 2021. | All that’s | ‘human’ | | is not | gold: |
| ----- | ---------- | --- | ------------------ | --- | --- | --- | ------ | ----- | ---------- | ------- | --- | ------ | ----- |
| pages | 6251–6258. | | https://doi.org/10 | | | | | | | | | | |
.18653/v1/2020.emnlp-main.506 Evaluatinghumanevaluationofgeneratedtext.
| | | | | | | | In Processings | | of the | 59th | Annual | Meeting | of |
| ------- | ------- | ----------- | --------- | -------- | ----------- | ------ | --------------- | ----- | ----------------- | --------------- | ------ | ----------- | --- |
| Nicola | De Cao, | Wilker | Aziz, | and | Ivan | Titov. | | | | | | | |
| | | | | | | | the Association | | for Computational | | | Linguistics | |
| 2021. | Editing | factual | knowledge | | in language | | | | | | | | |
| | | | | | | | (ACL), | pages | 7282–7296. | https://doi.org | | | |
| models. | In | Proceedings | of | the 2021 | Confer- | | | | | | | | |
/10.18653/v1/2021.acl-long.565
| ence | on Empirical | | Methods | in Natural | | Lan- | | | | | | | |
| ---- | ------------ | --- | ------- | ---------- | --- | ---- | --- | --- | --- | --- | --- | --- | --- |
guageProcessing(EMNLP),pages6491–6506. Roi Cohen, May Hamri, Mor Geva, and Amir
Globerson.2023.LMvsLM:Detectingfactual
https://doi.org/10.18653/v1/2021
| | | | | | | | errors via | cross | examination. | | CoRR, | abs/2305 | |
| --- | --- | --- | --- | --- | --- | --- | ---------- | ----- | ------------ | --- | ----- | -------- | --- |
.emnlp-main.522
| | | | | | | | .13281. | https://doi.org/10.18653/v1 | | | | | |
| ------- | ------------- | --- | ------- | -------- | --- | ----- | ------- | --------------------------- | --- | --- | --- | --- | --- |
| Yiannis | Charalambous, | | Norbert | Tihanyi, | | Ridhi | | | | | | | |
/2023.emnlp-main.778
| Jain, | Youcheng | Sun, | Mohamed | Amine | | Ferrag, | | | | | | | |
| ----- | -------- | ---- | ------- | ----- | --- | ------- | ---------------- | --- | ---------- | --- | --------- | --- | ----- |
| | | | | | | | Antonia Creswell | | and Murray | | Shanahan. | | 2022. |
andLucasC.Cordeiro.2023.Anewerainsoft-
| | | | | | | | Faithful | reasoning | using | large | language | | mod- |
| ---- | --------- | ------- | ------------ | --- | -------- | --- | -------- | --------- | ----- | ----- | -------- | --- | ---- |
| ware | security: | Towards | self-healing | | software | | | | | | | | |
vialarge languagemodelsandformalverifica- els.CoRR,abs/2208.14271.
tion.CoRR,abs/2305.14752.
IshitaDasgupta,AndrewK.Lampinen,Stephanie
AngelicaChen,Je´re´myScheurer,TomaszKorbak, C. Y. Chan, Antonia Creswell, Dharshan
Kumaran,JamesL.McClelland,andFelixHill.
| Jon Ander | | Campos, | Jun Shern | Chan, | Samuel | | | | | | | | |
| --------- | --- | ------- | --------- | ----- | ------ | --- | --- | --- | --- | --- | --- | --- | --- |
R.Bowman,KyunghyunCho,andEthanPerez. 2022.Languagemodelsshowhuman-likecon-
2023a. Improving code generation by training tent effects on reasoning. CoRR, abs/2207
.07051.
| with | natural | language | feedback. | | CoRR, | abs/ | | | | | | | |
| ---- | ------- | -------- | --------- | --- | ----- | ---- | --- | --- | --- | --- | --- | --- | --- |
2303.16749.
| | | | | | | | Sumanth Dathathri, | | Andrea | Madotto, | | Janice | Lan, |
| --- | --- | --- | --- | --- | --- | --- | ------------------ | --- | ------ | -------- | --- | ------ | ---- |
BeiChen,FengjiZhang,AnhNguyen,Daoguang Jane Hung, Eric Frank, Piero Molino, Jason
Zan, Zeqi Lin, Jian-Guang Lou, and Weizhu Yosinski, and Rosanne Liu. 2020. Plug and
Chen.2023b.Codet:Codegenerationwithgen- play language models: A simple approach to
eratedtests.InProceedingsofthe11thInterna- controlled text generation. In Proceedings of
tionalConferenceonLearningRepresentations the 8th International Conference on Learning
| (ICLR). | | | | | | | Representations(ICLR). | | | | | | |
| ------- | --- | --- | --- | --- | --- | --- | ---------------------- | --- | --- | --- | --- | --- | --- |
497
Fe´lix do Carmo, Dimitar Shterionov, Joss Markus Freitag, David Grangier, Qijun Tan, and
Moorkens, Joachim Wagner, Murhaf Hossari, Bowen Liang. 2022. High quality rather than
Eric Paquin, Dag Schmidtke, Declan Groves, high model probability: Minimum bayes risk
and Andy Way. 2021. A review of the state- decoding with neural metrics. Transactions of
of-the-art in automatic post-editing. Machine the Association for Computational Linguistics
Translation, 35(2):101–143. https://doi (TACL),pages811–825.https://doi.org
| .org/10.1007/s10590-020-09252-y, | | | | | | | /10.1162/tacl_a_00491 | | | | | |
| -------------------------------- | --- | --- | --- | --- | --- | --- | --------------------- | --- | --- | --- | --- | --- |
PubMed:34720417
| | | | | | | | Yao Fu, | Hao Peng, | Tushar | | Khot, | and Mirella |
| --- | --- | --- | --- | --- | --- | --- | ------- | --------- | ------ | --- | ----- | ----------- |
Yilun Du, Shuang Li, Antonio Torralba, Joshua Lapata. 2023. Improving language model ne-
B. Tenenbaum, and Igor Mordatch. 2023. Im- gotiationwithself-playandin-contextlearning
| proving | factuality | and | reasoning | | in language | | | | | | | |
| ------- | ---------- | --- | --------- | --- | ----------- | --- | --- | --- | --- | --- | --- | --- |
fromAIfeedback.CoRR,abs/2305.10142.
| models | through | multiagent | | debate. | | CoRR, | | | | | | |
| ------ | ------- | ---------- | --- | ------- | --- | ----- | --- | --- | --- | --- | --- | --- |
DeepGanguli,AmandaAskell,NicholasSchiefer,
abs/2305.14325.
ThomasI.Liao,KamileLukosiute,AnnaChen,
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
Yann Dubois, Xuechen Li, Rohan Taori, Tianyi Anna Goldie, Azalia Mirhoseini, Catherine
| Zhang, | Ishaan | Gulrajani, | | Jimmy | Ba, | Carlos | | | | | | |
| ------ | ------ | ---------- | --- | ----- | --- | ------ | --- | --- | --- | --- | --- | --- |
Olsson,DannyHernandez,DawnDrain,Dustin
| Guestrin, | Percy | Liang, | | and Tatsunori | | B. | | | | | | |
| --------- | ----- | ------ | --- | ------------- | --- | --- | ------- | ------------- | --- | ----- | ------ | ------- |
| | | | | | | | Li, Eli | Tran-Johnson, | | Ethan | Perez, | Jackson |
Hashimoto. 2023. Alpacafarm: A simulation Kernion, Jamie Kerr, Jared Mueller, Joshua
framework for methods that learn from human Landau,KamalNdousse,KarinaNguyen,Liane
feedback.CoRR,abs/2305.14387.
Lovitt,MichaelSellitto,NelsonElhage,Noem´ı
| | | | | | | | Mercado, | Nova | DasSarma, | | Oliver | Rausch, |
| ------------- | -------- | ---------- | ----- | -------- | --------- | --- | -------- | -------- | --------- | ------- | --------- | ----------- |
| Tobias Falke, | Leonardo | | F. R. | Ribeiro, | Prasetya | | | | | | | |
| | | | | | | | Robert | Lasenby, | Robin | Larson, | | Sam Ringer, |
| Ajie Utama, | | Ido Dagan, | and | Iryna | Gurevych. | | | | | | | |
| | | | | | | | Sandipan | Kundu, | | Saurav | Kadavath, | Scott |
2019.Rankinggeneratedsummariesbycorrect-
ness:Aninterestingbutchallengingapplication Johnston, Shauna Kravec, Sheer El Showk,
| | | | | | | | Tamera | Lanham, | | Timothy | Telleen-Lawton, | |
| ----------- | -------- | ------ | ---------- | --- | ----------- | --- | ------------- | ------- | ------- | ------- | --------------- | ----------- |
| for natural | language | | inference. | In | Proceedings | | | | | | | |
| | | | | | | | Tom Henighan, | | Tristan | Hume, | | Yuntao Bai, |
| of the | 57st | Annual | Meeting | of | the Associ- | | | | | | | |
ZacHatfield-Dodds,BenMann,DarioAmodei,
| ation | for Computational | | | Linguistics | (ACL), | | | | | | | |
| ----- | ----------------- | ------------------ | --- | ----------- | ------ | --- | -------- | ----------- | --- | ----- | ----------- | ------------- |
| | | | | | | | Nicholas | Joseph, | | Sam | McCandlish, | Tom |
| pages | 2214–2220. | https://doi.org/10 | | | | | | | | | | |
| | | | | | | | Brown, | Christopher | | Olah, | Jack | Clark, Samuel |
.18653/v1/P19-1213
| | | | | | | | R. Bowman, | | and Jared | Kaplan. | | 2023. The ca- |
| ------------------ | --- | ---- | ------- | --- | ---- | ---- | ---------- | --------- | --------------- | ------- | --- | ------------- |
| Patrick Fernandes, | | Aman | Madaan, | | Emmy | Liu, | | | | | | |
| | | | | | | | pacity | for moral | self-correction | | | in large lan- |
Anto´nio Farinhas, Pedro Henrique Martins, guagemodels.CoRR,abs/2302.07459.
| Amanda | Bertsch, | Jose´ | G. C. | de Souza,Shuyan | | | | | | | | |
| ------ | -------- | ----- | ----- | --------------- | --- | --- | --- | --- | --- | --- | --- | --- |
Zhou, Tongshuang Wu, Graham Neubig, and GeGao,Hung-TingChen,YoavArtzi,andEunsol
| | | | | | | | Choi. | 2023a. | Continually | | improving | extrac- |
| -------- | -------------- | ----------- | ------- | -------- | -------- | ---- | ------- | ------ | ----------- | --------- | --------- | ---------- |
| Andre´ | F. T. Martins. | | 2023. | Bridging | the | gap: | | | | | | |
| | | | | | | | tive QA | via | human | feedback. | | CoRR, abs/ |
| A survey | on | integrating | (human) | | feedback | for | | | | | | |
natural language generation. CoRR, abs/2305 2305.12473. https://doi.org/10.18653
/v1/2023.emnlp-main.27
| .00955. | https://doi.org/10.1162/tacl | | | | | | | | | | | |
| ------- | ---------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
a00626
| | | | | | | | Luyu Gao, | Zhuyun | | Dai, | Panupong | Pasupat, |
| -------- | --------- | -------- | --- | ------ | ------- | --- | --------- | ------ | ---- | ---- | -------- | --------- |
| | | | | | | | Anthony | Chen, | Arun | | Tejasvi | Chaganty, |
| Emanuela | Ferretti, | Kristina | | Rohde, | Gregory | P. | | | | | | |
Moore, and Thierry Daboval. 2019. Catch the Yicheng Fan, Vincent Y. Zhao, Ni Lao,
moment: The power of turning mistakes into HongraeLee,Da-ChengJuan,andKelvinGuu.
‘precious’ learning opportunities. Paediatrics 2023b. Rarr: Researching and revising what
& Child Health, 24(3):156–159. https:// language models say, using language models.
doi.org/10.1093/pch/pxy102, PubMed: In Proceedings of the 61th Annual Meeting of
| 31111832 | | | | | | | the Association | | for | Computational | | Linguistics |
| ------------ | ------ | --- | ----- | ----- | ------- | --- | --------------- | --------------------------- | --- | ------------- | --- | ----------- |
| | | | | | | | (ACL). | https://doi.org/10.18653/v1 | | | | |
| Emily First, | Markus | N. | Rabe, | Talia | Ringer, | and | | | | | | |
/2023.acl-long.910
| Yuriy | Brun. | 2023. Baldur: | | Whole-proof | | gen- | | | | | | |
| ----- | ----- | ------------- | --- | ----------- | --- | ---- | --- | --- | --- | --- | --- | --- |
eration and repair with large language models. Samuel Gehman, Suchin Gururangan, Maarten
CoRR, abs/2303.04910. https://doi.org Sap, Yejin Choi, and Noah A. Smith. 2020.
/10.1145/3611643.3616243 RealToxicityPrompts: Evaluating neural toxic
498
degeneration in language models. In Find- training (rest) for language modeling. CoRR,
| ings of | the | Association | | for Computational | | abs/2308.08998. | | | | | | |
| ------------ | ----- | ----------- | ----- | ----------------- | ---------- | --------------- | -------- | ------ | ------ | --- | ----- | ----- |
| Linguistics: | EMNLP | | 2020, | pages | 3356–3369. | | | | | | | |
| | | | | | | Biyang | Guo, Xin | Zhang, | Ziyuan | | Wang, | Minqi |
https://doi.org/10.18653/v1/2020
| | | | | | | Jiang, | Jinran Nie, | Yuxuan | | Ding, | Jianwei | Yue, |
| --- | --- | --- | --- | --- | --- | ------ | ----------- | ------ | --- | ----- | ------- | ---- |
.findings-emnlp.301
| | | | | | | and Yupeng | Wu. | 2023. | How | close | is | chatgpt |
| --- | --- | --- | --- | --- | --- | ---------- | --- | ----- | --- | ----- | --- | ------- |
Zelalem Gero, Chandan Singh, Hao Cheng, tohumanexperts?Comparisoncorpus,evalua-
Tristan Naumann, Michel Galley, Jianfeng tion,anddetection.CoRR,abs/2301.07597.
Gao,andHoifungPoon.2023.Self-verification
| | | | | | | Shibo Hao, | Yi | Gu, | Haodi | Ma, | Joshua | Jiahua |
| -------- | -------- | -------- | --- | ----------- | ------- | ---------- | ---- | ----- | ----- | --- | ------ | ------ |
| improves | few-shot | clinical | | information | extrac- | | | | | | | |
| | | | | | | Hong, | Zhen | Wang, | Daisy | Zhe | Wang, | and |
tion.CoRR,abs/2306.00024.
| | | | | | | Zhiting | Hu. 2023. | | Reasoning | | with | language |
| --- | --- | --- | --- | --- | --- | ------- | --------- | --- | --------- | --- | ---- | -------- |
Amelia Glaese, Nat McAleese, Maja Trebacz, model is planning with world model. CoRR,
John Aslanides, Vlad Firoiu, Timo Ewalds, abs/2305.14992. https://doi.org/10
Maribeth Rauh, Laura Weidinger, Martin J. .18653/v1/2023.emnlp-main.507
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| Chadwick, | Phoebe | Thacker, | | Lucy | Campbell- | | | | | | | |
| ----------- | ---------- | -------- | ------- | ------- | ------------ | -------- | ------------ | ---------- | ---------- | ----- | -------- | -------- |
| | | | | | | Hangfeng | He, Hongming | | Zhang, | | and Dan | Roth. |
| Gillingham, | Jonathan | | Uesato, | Po-Sen | Huang, | | | | | | | |
| | | | | | | 2023. | Rethinking | with | retrieval: | | Faithful | large |
| Ramona | Comanescu, | | Fan | Yang, | Abigail See, | | | | | | | |
| | | | | | | language | model | inference. | | CoRR, | | abs/2301 |
| Sumanth | Dathathri, | Rory | Greig, | Charlie | Chen, | | | | | | | |
.00303.
| Doug | Fritz, | Jaume | Sanchez | Elias, | Richard | | | | | | | |
| ------ | ------------ | ----- | -------- | --------- | ------- | --------- | --- | -------- | --- | ---- | -------- | ---- |
| Green, | Sona Mokra´, | | Nicholas | Fernando, | Boxi | | | | | | | |
| | | | | | | Pengcheng | He, | Xiaodong | | Liu, | Jianfeng | Gao, |
Wu, Rachel Foley, Susannah Young, Iason and Weizhu Chen. 2021. Deberta: Decoding-
Gabriel, William Isaac, John Mellor, Demis enhanced bert with disentangled attention. In
| Hassabis, | Koray | Kavukcuoglu, | | | Lisa Anne | | | | | | | |
| --------- | ----- | ------------ | --- | --- | --------- | ----------- | --- | --- | ----------------- | --- | --- | ------- |
| | | | | | | Proceedings | of | The | 9th International | | | Confer- |
Hendricks,andGeoffreyIrving.2022.Improv- enceonLearningRepresentations(ICLR).
| ing alignment | | of dialogue | | agents | via targeted | | | | | | | |
| ------------- | --- | ----------- | --- | ------ | ------------ | --- | --- | --- | --- | --- | --- | --- |
humanjudgements.CoRR,abs/2209.14375. Alec Helbling, Mansi Phute, Matthew Hull, and
DuenHorngChau.2023.LLMselfdefense:By
Dongyoung Go, Tomasz Korbak, Germa´n self examination, LLMs know they are being
Kruszewski, Jos Rozen, Nahyeon Ryu, and tricked.CoRR,abs/2308.07308.
| Marc | Dymetman. | 2023. | | Aligning | language | | | | | | | |
| ---- | --------- | ----- | --- | -------- | -------- | --- | --- | --- | --- | --- | --- | --- |
JasonHoelscher-Obermaier,JuliaPersson,Esben
| models | with preferences | | through | | f-divergence | | | | | | | |
| ------ | ---------------- | --- | ------- | --- | ------------ | ----- | ------- | -------- | --- | ---- | ------ | ----- |
| | | | | | | Kran, | Ioannis | Konstas, | and | Fazl | Barez. | 2023. |
minimization.CoRR,abs/2302.08215.
| | | | | | | Detecting | edit | failures | in | large | language | mod- |
| --- | --- | --- | --- | --- | --- | --------- | ---- | -------- | --- | ----- | -------- | ---- |
Olga Golovneva, Moya Chen, Spencer Poff, els: An improved specificity benchmark. In
Martin Corredor, Luke Zettlemoyer, Maryam Findings of the Association for Computational
Fazel-Zarandi, and Asli Celikyilmaz. 2023. Linguistics: ACL 2023, pages 11548–11559.
ROSCOE: A suite of metrics for scoring step- https://doi.org/10.18653/v1/2023
by-step reasoning. In Proceedings of the 11th .findings-acl.733
| International | | Conference | on | Learning | Repre- | | | | | | | |
| ----------------- | --- | ---------- | --- | -------- | ------ | ------------- | --------- | --- | ----- | ------- | --- | ------- |
| sentations(ICLR). | | | | | | Ari Holtzman, | | Jan | Buys, | Maxwell | | Forbes, |
| | | | | | | Antoine | Bosselut, | | David | Golub, | and | Yejin |
ZhibinGou,ZhihongShao,YeyunGong,Yelong Choi. 2018. Learning to write with coopera-
Shen, Yujiu Yang, Nan Duan, and Weizhu tive discriminators. In Proceedings of the 56th
Chen. 2023. CRITIC: Large language models AnnualMeetingoftheAssociationforCompu-
canself-correctwithtool-interactivecritiquing. tational Linguistics (ACL), pages 1638–1649.
| CoRR,abs/2305.11738. | | | | | | https://doi.org/10.18653/v1/P18 | | | | | | |
| -------------------- | --- | --- | --- | --- | --- | ------------------------------- | --- | --- | --- | --- | --- | --- |
-1152
| Caglar Gulcehre, | | Tom | Le | Paine, | Srivatsan | | | | | | | |
| ---------------- | --- | --- | --- | ------ | --------- | --- | --- | --- | --- | --- | --- | --- |
Srinivasan,KseniaKonyushkova,LotteWeerts, Ruixin Hong, Hongming Zhang, Xinyu Pang,
Abhishek Sharma, Aditya Siddhant, Alex Dong Yu, and Changshui Zhang. 2023. A
Ahern, Miaosen Wang, Chenjie Gu, Wolfgang closer look at the self-verification abilities of
Macherey, Arnaud Doucet, Orhan Firat, and large language models in logical reasoning.
Nando de Freitas. 2023. Reinforced self- CoRR,abs/2311.07954.
499
Jie Huang, Xinyun Chen, Swaroop Mishra, Muhammad Khalifa, Lajanugen Logeswaran,
Huaixiu Steven Zheng, Adams Wei Yu, Moontae Lee, Honglak Lee, and Lu Wang.
Xinying Song, and Denny Zhou. 2023. Large 2023. Discriminator-guided multi-step reason-
language models cannot self-correct reasoning ing with language models. CoRR, abs/2305
| yet.CoRR,abs/2310.01798. | | | | | | | .14934. | | | | | |
| ------------------------ | --- | --- | --- | --- | --- | --- | ------- | --- | --- | --- | --- | --- |
Jiaxin Huang, Shixiang Shane Gu, Le Hou, Geunwoo Kim, Pierre Baldi, and Stephen
| | | | | | | | McAleer. | 2023. | Language | | models | can solve |
| ------ | --- | ------ | ----- | --- | ------- | --- | -------- | ----- | -------- | --- | ------ | --------- |
| Yuexin | Wu, | Xuezhi | Wang, | | Hongkun | Yu, | | | | | | |
computertasks.CoRR,abs/2303.17491.
| and Jiawei | Han. | 2022. | Large | language | | mod- | | | | | | |
| ---------- | ------------- | ----- | ----- | --------------- | --- | ---- | ------------------ | --- | --- | ------ | -------- | ---- |
| els can | self-improve. | | CoRR, | abs/2210.11610. | | | | | | | | |
| | | | | | | | James Kirkpatrick, | | | Razvan | Pascanu, | Neil |
https://doi.org/10.18653/v1/2023
| | | | | | | | C. Rabinowitz, | | Joel | Veness, | | Guillaume |
| -------------- | --- | --- | --- | --- | --- | --- | -------------- | ------ | ----- | -------- | ------ | --------- |
| .emnlp-main.67 | | | | | | | Desjardins, | Andrei | | A. Rusu, | Kieran | Milan, |
| | | | | | | | John | Quan, | Tiago | Ramalho, | | Agnieszka |
ShuyangJiang,YuhaoWang,andYuWang.2023.
| | | | | | | | Grabska-Barwinska, | | | Demis | Hassabis, | Claudia |
| ----------- | --- | ---- | --------- | --------- | --- | --- | ------------------ | --- | --- | ----- | --------- | ------- |
| Selfevolve: | A | code | evolution | framework | | via | | | | | | |
Clopath,DharshanKumaran,andRaiaHadsell. Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
largelanguagemodels.CoRR,abs/2306.02907.
| | | | | | | | 2016. Overcoming | | | catastrophic | | forgetting in |
| --- | --- | --- | --- | --- | --- | --- | ---------------- | --- | --- | ------------ | --- | ------------- |
Jaehun Jung, Lianhui Qin, Sean Welleck, neuralnetworks.CoRR,abs/1612.00796.
| Faeze | Brahman, | | Chandra | | Bhagavatula, | | | | | | | |
| -------------- | -------- | ------------- | ------- | -------------- | ------------ | ------ | --------------- | ----- | ----------- | ------ | ------ | ------------- |
| | | | | | | | Takeshi Kojima, | | Shixiang | | Shane | Gu, Machel |
| Ronan Le | Bras, | and | Yejin | Choi. | 2022. | Maieu- | | | | | | |
| | | | | | | | Reid, Yutaka | | Matsuo, | and | Yusuke | Iwasawa. |
| tic prompting: | | Logically | | consistent | reasoning | | | | | | | |
| | | | | | | | 2022. | Large | language | models | | are zero-shot |
| with recursive | | explanations. | | In Proceedings | | of | | | | | | |
| | | | | | | | reasoners. | In | Proceedings | | of the | 2022 Annual |
the 2022 Conference on Empirical Methods Conference on Neural Information Processing
| in Natural | Language | | Processing | | (EMNLP), | | | | | | | |
| ---------- | -------- | --- | ---------- | --- | -------- | --- | --- | --- | --- | --- | --- | --- |
Systems(NeurIPS).
| pages 1266–1279. | | https://doi.org/10 | | | | | | | | | | |
| ---------------- | --- | ------------------ | --- | --- | --- | --- | --------------- | --- | ------- | --- | -------- | ------ |
| | | | | | | | Julia Kreutzer, | | Shahram | | Khadivi, | Evgeny |
.18653/v1/2022.emnlp-main.82
Matusov,andStefanRiezler.2018.Canneural
Saurav Kadavath, Tom Conerly, Amanda Askell, machinetranslationbeimprovedwithuserfeed-
Tom Henighan, Dawn Drain, Ethan Perez, back?InProceedingsofthe2018Conferenceof
Nicholas Schiefer, Zac Hatfield-Dodds, Nova theNorthAmericanChapteroftheAssociation
DasSarma, Eli Tran-Johnson, Scott Johnston, for Computational Linguistics: Human Lan-
Sheer El Showk, Andy Jones, Nelson Elhage, guage Technologies (NAACL-HIT). https://
Tristan Hume, Anna Chen, Yuntao Bai, Sam doi.org/10.18653/v1/N18-3012
Bowman,StanislavFort,DeepGanguli,Danny
HungLe,YueWang,AkhileshDeepakGotmare,
| Hernandez, | Josh | Jacobson, | | Jackson | Kernion, | | | | | | | |
| ---------- | ------- | ----------- | ------- | -------- | -------- | ------- | -------------- | ---------- | -------------- | ----------- | -------- | ------------- |
| | | | | | | | Silvio | Savarese, | and | Steven | Chu-Hong | Hoi. |
| Shauna | Kravec, | Liane | Lovitt, | Kamal | Ndousse, | | | | | | | |
| | | | | | | | 2022. | Coderl: | Mastering | | code | generation |
| Catherine | Olsson, | Sam | Ringer, | Dario | Amodei, | | | | | | | |
| | | | | | | | through | pretrained | models | | and deep | reinforce- |
| Tom Brown, | | Jack | Clark, | Nicholas | | Joseph, | | | | | | |
| | | | | | | | ment learning. | | In Proceedings | | | of the Annual |
| Ben Mann, | Sam | McCandlish, | | Chris | Olah, | and | | | | | | |
| | | | | | | | Conference | on | Neural | Information | | Processing |
JaredKaplan.2022.Languagemodels(mostly)
Systems(NeurIPS).
knowwhattheyknow.CoRR,abs/2207.05221.
DongjunLee.2019.Clause-wiseandrecursivede-
PeiKe,BosiWen,ZhuoerFeng,XiaoLiu,Xuanyu
| | | | | | | | coding | for complex | | and cross-domain | | text-to- |
| --- | --- | --- | --- | --- | --- | --- | ------ | ----------- | --- | ---------------- | --- | -------- |
Lei, Jiale Cheng, Shengyuan Wang, Aohan SQL generation. In Proceedings of the 2019
Zeng,YuxiaoDong,HongningWang,JieTang,
| | | | | | | | Conference | on | Empirical | | Methods | in Natural |
| ---------- | ------ | --- | ----- | ------------ | --- | ----- | ---------- | --- | --------- | --- | ------- | ---------- |
| and Minlie | Huang. | | 2023. | CritiqueLLM: | | Scal- | | | | | | |
LanguageProcessingandthe9thInternational
ingLLM-as-criticforeffectiveandexplainable Joint Conference on Natural Language Pro-
evaluationoflargelanguagemodelgeneration.
| | | | | | | | cessing | (EMNLP-IJCNLP), | | | pages | 6045–6051. |
| --- | --- | --- | --- | --- | --- | --- | ------- | --------------- | --- | --- | ----- | ---------- |
CoRR,abs/2311.18702.
https://doi.org/10.18653/v1/D19
-1624
| Amol Kelkar, | Rohan | Relan, | | Vaishali | Bhardwaj, | | | | | | | |
| ------------ | ----- | ------ | --- | -------- | --------- | --- | --- | --- | --- | --- | --- | --- |
Saurabh Vaichal, and Peter Relan. 2020. Kyungjae Lee, Wookje Han, Seung-won
Bertrand-dr: Improving text-to-sql using a dis- Hwang, Hwaran Lee, Joonsuk Park, and
criminativere-ranker.CoRR,abs/2002.00557. Sang-WooLee.2022.Plug-and-playadaptation
500
for continuously-updated QA. In Findings of Yifei Li, Zeqi Lin, Shizhuo Zhang, Qiang Fu,
theAssociationforComputationalLinguistics: Bei Chen, Jian-Guang Lou, and Weizhu Chen.
ACL2022,pages438–447.https://doi.org 2023d.Makinglanguagemodelsbetterreason-
/10.18653/v1/2022.findings-acl.37 ers with step-aware verifier. In Proceedings
| | | | | | | | of the | 61st | Annual | Meeting | | of the | Associ- |
| ----------- | ---- | --------- | ----- | ------------ | --- | --------- | ------ | ---------- | ------------- | ------------------ | ----------- | ------ | ------- |
| WonKee | Lee, | Baikjin | Jung, | Jaehun | | Shin, and | | | | | | | |
| | | | | | | | ation | for | Computational | | Linguistics | | (ACL), |
| Jong-Hyeok | | Lee. | 2021. | Adaptation | | of back- | | | | | | | |
| | | | | | | | pages | 5315–5333. | | https://doi.org/10 | | | |
| translation | to | automatic | | post-editing | | for syn- | | | | | | | |
.18653/v1/2023.acl-long.291
| thetic data | generation. | | In | Proceedings | | of the | | | | | | | |
| ----------- | ----------- | --- | --- | ----------- | --- | ------ | --- | --- | --- | --- | --- | --- | --- |
16th Conference of the European Chapter of Hunter Lightman, Vineet Kosaraju, Yura Burda,
| the Association | | for | Computational | | Linguistics | | | | | | | | |
| --------------- | --- | --- | ------------- | --- | ----------- | --- | -------- | -------- | --- | ----- | ------ | ----- | ---- |
| | | | | | | | Harrison | Edwards, | | Bowen | Baker, | Teddy | Lee, |
(EACL), pages 3685–3691. https://doi Jan Leike, John Schulman, Ilya Sutskever, and
.org/10.18653/v1/2021.eacl-main Karl Cobbe. 2023. Let’s verify step by step.
.322 CoRR,abs/2305.20050.
Sharon Levy, Emily Allaway, Melanie Subbiah, Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| | | | | | | | Stephanie | Lin, | Jacob | Hilton, | and | Owain | Evans. |
| --------- | ----------- | ------- | ------- | ------------- | ----------- | -------- | --------- | ----------------- | ----------- | ------------------ | ----------- | ----------- | -------- |
| Lydia | Chilton, | Desmond | | Patton, | | Kathleen | | | | | | | |
| | | | | | | | 2022. | TruthfulQA: | | Measuring | | how | models |
| McKeown, | and | William | | Yang | Wang. | 2022. | | | | | | | |
| | | | | | | | mimic | human | falsehoods. | | In | Proceedings | of |
| SafeText: | A benchmark | | | for exploring | | physical | | | | | | | |
| | | | | | | | the 60th | Annual | | Meeting | | of the | Associa- |
| safety in | language | | models. | In | Proceedings | of | | | | | | | |
| | | | | | | | tion | for Computational | | | Linguistics | | (ACL), |
| the 2022 | Conference | | on | Empirical | | Methods | | | | | | | |
| | | | | | | | pages | 3214–3252. | | https://doi.org/10 | | | |
in Natural Language Processing (EMNLP), .18653/v1/2022.acl-long.229
| pages 2407–2421. | | | https://doi.org/10 | | | | | | | | | | |
| ---------------- | --- | --- | ------------------ | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
.18653/v1/2022.emnlp-main.154 Yen-Ting Lin and Yun-Nung Chen. 2023.
LLM-eval:Unifiedmulti-dimensionalautomatic
| Sharon Levy, | Michael | | Saxon, | and | William | Yang | | | | | | | |
| ------------ | ------- | ------------- | ------ | ------------ | ------- | ---- | ---------- | --- | --------------- | --- | ------------- | --- | ---- |
| | | | | | | | evaluation | | for open-domain | | conversations | | with |
| Wang. | 2021. | Investigating | | memorization | | of | | | | | | | |
largelanguagemodels.CoRR,abs/2305.13711.
conspiracytheoriesintextgeneration.InFind-
ings of the Association for Computational Lin- Zhen Lin, Shubhendu Trivedi, and Jimeng Sun.
guistics:ACL-IJCNLP2021,pages4718–4729, 2023.Generatingwithconfidence:Uncertainty
Online.AssociationforComputationalLinguis- quantification for black-box large language
tics. https://doi.org/10.18653/v1 models.CoRR,abs/2305.19187.
/2021.findings-acl.416
| | | | | | | | Hao Liu, | Carmelo | | Sferrazza, | and | Pieter | Abbeel. |
| --- | --- | --- | --- | --- | --- | --- | -------- | ------- | --- | ---------- | --- | ------ | ------- |
Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, 2023a. Chain of hindsight aligns language
| Jian-Yun | Nie, | and | Ji-Rong | Wen. | 2023a. | Hal- | | | | | | | |
| -------- | ---- | --- | ------- | ---- | ------ | ---- | --- | --- | --- | --- | --- | --- | --- |
modelswithfeedback.CoRR,abs/2302.02676.
| ueval: | A large-scale | | hallucination | | evaluation | | | | | | | | |
| --------- | ------------- | ----- | ------------- | --- | ---------- | ----- | -------- | ---- | --------- | --- | --------- | --- | -------- |
| | | | | | | | Jiacheng | Liu, | Ramakanth | | Pasunuru, | | Hannaneh |
| benchmark | for | large | language | | models. | CoRR, | | | | | | | |
abs/2305.11747. Hajishirzi, Yejin Choi, and Asli Celikyilmaz.
| | | | | | | | 2023b. | Crystal: | | Introspective | | reasoners | rein- |
| ------- | ------------- | --- | ------------- | --- | --- | ------- | -------- | ---------- | -------------- | ------------- | ------------ | ----------- | ------- |
| Miaoran | Li, Baolin | | Peng, | and | Zhu | Zhang. | | | | | | | |
| | | | | | | | forced | with | self-feedback. | | In | Proceedings | of |
| 2023b. | Self-checker: | | Plug-and-play | | | modules | | | | | | | |
| | | | | | | | the 2023 | Conference | | | on Empirical | | Methods |
for fact-checking with large language models. in Natural Language Processing (EMNLP),
CoRR,abs/2305.14623.
| | | | | | | | pages | 11557–11572. | | https://doi.org/10 | | | |
| --- | --- | --- | --- | --- | --- | --- | ----- | ------------ | --- | ------------------ | --- | --- | --- |
Ruosen Li, Teerth Patel, and Xinya Du. 2023c. .18653/v1/2023.emnlp-main.708
| PRD: Peer | rank | and | discussion | | improve | large | | | | | | | |
| --------- | ----- | ----- | ------------ | --- | ------- | ----- | --------- | --------- | ------- | --- | ----------- | ----- | --------- |
| | | | | | | | Yixin Liu | and | Pengfei | | Liu. | 2021. | Simcls: A |
| language | model | based | evaluations. | | CoRR, | abs/ | | | | | | | |
| | | | | | | | simple | framework | | for | contrastive | | learning |
2307.02762.
| | | | | | | | of abstractive | | summarization. | | | In Proceedings | |
| --- | --- | --- | --- | --- | --- | --- | -------------- | --- | -------------- | --- | --- | -------------- | --- |
XiangLi,JohnThickstun,IshaanGulrajani,Percy of the 59th Annual Meeting of the Asso-
Liang, and Tatsunori B. Hashimoto. 2022. ciation for Computational Linguistics and
Diffusion-lmimprovescontrollabletextgener- the 11th International Joint Conference on
ation.InProceedingsoftheAnnualConference Natural Language Processing (ACL/IJCNLP),
on Neural Information Processing Systems pages 1065–1072. https://doi.org/10
| (NeurIPS). | | | | | | | .18653/v1/2021.acl-short.135 | | | | | | |
| ---------- | --- | --- | --- | --- | --- | --- | ---------------------------- | --- | --- | --- | --- | --- | --- |
501
Manuel London, Valerie I. Sessa, and Loren Rahul Gupta. 2023. Flirt: Feedback loop
A. Shelley. 2023. Developing self-awareness: in-contextredteaming.CoRR,abs/2308.04265.
| Learning | processes | for | self-and | interpersonal | | | | | | | | |
| ---------- | --------- | ------------------ | -------- | -------------- | --------- | --------------- | ------ | ----- | -------------- | -------- | ----------- | ------- |
| | | | | | | Janet Metcalfe. | | 2017. | | Learning | from | errors. |
| growth. | Annual | Review | of | Organizational | | | | | | | | |
| | | | | | | Annual | Review | | of Psychology, | | 68:465–489. | |
| Psychology | | and Organizational | | | Behavior, | | | | | | | |
https://doi.org/10.1146/annurev
| 10:261–288. | | https://doi.org/10.1146 | | | | | | | | | | |
| ----------- | --- | ----------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
-psych-010416-044022,PubMed:27648988
/annurev-orgpsych-120920-044531
| | | | | | | Ning Miao, | Yee | Whye | Teh, | and | Tom | Rainforth. |
| ---------- | ---- | -------- | ---- | ------- | ----- | ---------- | ---------- | ---- | ----- | ---- | --- | ---------- |
| Ximing Lu, | Sean | Welleck, | Jack | Hessel, | Liwei | | | | | | | |
| | | | | | | 2023. | Selfcheck: | | Using | LLMs | to | zero-shot |
Jiang, Lianhui Qin, Peter West, Prithviraj checktheirownstep-by-stepreasoning.CoRR,
Ammanabrolu,andYejinChoi.2022.QUARK:
abs/2308.00436.
| Controllable | | text generation | | with | reinforced | | | | | | | |
| ------------ | --- | --------------- | ----------- | ------ | ---------- | ----- | ------ | ------- | --- | -------- | ----- | -------- |
| | | | | | | Sewon | Min, | Kalpesh | | Krishna, | Xinxi | Lyu, |
| unlearning. | | In Proceedings | | of the | Annual | | | | | | | |
| | | | | | | Mike | Lewis, | Wen-tau | | Yih, | Pang | Wei Koh, |
| Conference | on | Neural | Information | | Processing | | | | | | | |
MohitIyyer,LukeZettlemoyer,andHannaneh
Systems(NeurIPS).
Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| | | | | | | Hajishirzi. | | 2023. | Factscore: | | Fine-grained | |
| -------- | ---- | --------- | --- | ------- | ----- | ----------- | ---------- | ----- | ---------- | --- | ------------ | ------- |
| Chenyang | Lyu, | Jitao Xu, | and | Longyue | Wang. | | | | | | | |
| | | | | | | atomic | evaluation | | of factual | | precision | in long |
2023a. New trends in machine translation us- form text generation. CoRR, abs/2305.14251.
inglargelanguagemodels:Caseexampleswith
https://doi.org/10.18653/v1/2023
chatgpt.CoRR,abs/2305.01181.https://doi
.emnlp-main.741
.org/10.18653/v1/2023.emnlp-main.1036
| | | | | | | Shikhar | Murty, | Christopher | | D. | Manning, | Scott |
| --- | --- | --- | --- | --- | --- | ------- | ------ | ----------- | --- | --- | -------- | ----- |
Qing Lyu, Shreya Havaldar, Adam Stein, Li M. Lundberg, and Marco Tu´lio Ribeiro. 2022.
| Zhang, | Delip | Rao, | Eric Wong, | | Marianna | | | | | | | |
| ----------- | ----- | ----- | --------------- | --- | -------- | ------ | ----- | ---- | --- | ------------ | --- | -------- |
| | | | | | | Fixing | model | bugs | | with natural | | language |
| Apidianaki, | and | Chris | Callison-Burch. | | 2023b. | | | | | | | |
patches.InProceedingsofthe2022Conference
Faithful chain-of-thought reasoning. CoRR, on Empirical Methods in Natural Language
https://doi.org/10.18653
| abs/2301.13379. | | | | | | Processing | | (EMNLP), | | pages | 11600–11613. | |
| --------------- | --- | --- | --- | --- | --- | ---------- | --- | -------- | --- | ----- | ------------ | --- |
/v1/2023.ijcnlp-main.20
https://doi.org/10.18653/v1/2022
.emnlp-main.797
| Aman Madaan, | | Niket Tandon, | | Peter | Clark, and | | | | | | | |
| ------------ | --- | ------------- | --- | ----- | ---------- | --- | --- | --- | --- | --- | --- | --- |
Yiming Yang. 2022. Memory-assisted prompt Deepak Nathani, David Wang, Liangming
editingtoimproveGPT-3afterdeployment.In Pan, and William Wang. 2023. MAF: Multi-
| Proceedings | | of the 2022 | Conference | | on Em- | | | | | | | |
| ----------- | --- | ----------- | ---------- | --- | ------ | ------ | -------- | --- | ------------- | --- | --------- | --- |
| | | | | | | aspect | feedback | | for improving | | reasoning | in |
piricalMethodsinNaturalLanguageProcess- large language models. In Proceedings of
ing(EMNLP),pages2833–2861.https://doi the 2023 Conference on Empirical Methods
.org/10.18653/v1/2022.emnlp-main.183
| | | | | | | in Natural | | Language | | Processing | | (EMNLP), |
| ------------ | --- | ------------- | --- | ------- | ------ | ---------- | ---------- | -------- | ------------------ | ---------- | --- | -------- |
| | | | | | | pages | 6591–6616. | | https://doi.org/10 | | | |
| Aman Madaan, | | Niket Tandon, | | Prakhar | Gupta, | | | | | | | |
.18653/v1/2023.emnlp-main.407
| Skyler | Hallinan, | Luyu | Gao, | Sarah | Wiegreffe, | | | | | | | |
| ------ | --------- | ---- | ---- | ----- | ---------- | --- | --- | --- | --- | --- | --- | --- |
Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Ansong Ni, Srini Iyer, Dragomir Radev, Ves
| Yiming | Yang, | Sean | Welleck, | Bodhisattwa | | | | | | | | |
| ------ | ----- | ---- | -------- | ----------- | --- | --------- | ------- | --- | ---- | ---- | -------- | ------ |
| | | | | | | Stoyanov, | Wen-tau | | Yih, | Sida | I. Wang, | and Xi |
Prasad Majumder, Shashank Gupta, Amir VictoriaLin.2023.LEVER:Learningtoverify
Yazdanbakhsh, and Peter Clark. 2023. Self- language-to-codegenerationwithexecution.In
| refine: | Iterative | refinement | with | self-feedback. | | | | | | | | |
| -------------------- | --------- | ---------- | ---- | -------------- | --- | ---------------------------- | --- | ------ | ---- | ------------- | --- | ------- |
| | | | | | | Proceedings | | of the | 40th | International | | Confer- |
| CoRR,abs/2303.17651. | | | | | | enceonMachineLearning(ICML). | | | | | | |
Potsawee Manakul, Adian Liusie, and Mark Theo X. Olausson, Jeevana Priya Inala,
J. F. Gales. 2023. Selfcheckgpt: Zero-resource ChenglongWang,JianfengGao,andArmando
black-boxhallucinationdetectionforgenerative
| | | | | | | Solar-Lezama. | | 2023. | Demystifying | | | GPT self- |
| --- | --- | --- | --- | --- | --- | ------------- | --- | ----- | ------------ | --- | --- | --------- |
largelanguagemodels.CoRR,abs/2303.08896. repair for code generation. CoRR, abs/2306
| https://doi.org/10.18653/v1/2023 | | | | | | .09896. | | | | | | |
| -------------------------------- | --- | --- | --- | --- | --- | ------- | --- | --- | --- | --- | --- | --- |
.emnlp-main.557
| | | | | | | Yasumasa | Onoe, | Michael | | J. Q. | Zhang, | Shankar |
| --- | --- | --- | --- | --- | --- | -------- | ----- | ------- | --- | ----- | ------ | ------- |
Ninareh Mehrabi, Palash Goyal, Christophe Padmanabhan, Greg Durrett, and Eunsol Choi.
Dupuy, Qian Hu, Shalini Ghosh, Richard 2023. Can lms learn new entities from de-
Zemel, Kai-Wei Chang, Aram Galstyan, and scriptions? Challenges in propagating injected
502
knowledge. In Proceedings of the 61st An- Chengwei Qin, Aston Zhang, Zhuosheng Zhang,
nual Meeting of the Association for Compu- Jiaao Chen, Michihiro Yasunaga, and Diyi
tational Linguistics (ACL), pages 5469–5485. Yang. 2023. Is chatgpt a general-purpose nat-
https://doi.org/10.18653/v1/2023 ural language processing task solver? CoRR,
| .acl-long.300 | | | | | | abs/2302.06476. | | | | | | |
| ------------- | ----- | ----- | --------- | ------- | ----- | --------------- | --- | ---- | -------- | ---- | ----- | ------ |
| | | | | | | Vikas Raunak, | | Amr | Sharaf, | Hany | | Hassan |
| OpenAI. | 2023. | GPT-4 | technical | report. | CoRR, | | | | | | | |
| | | | | | | Awadallah, | and | Arul | Menezes. | | 2023. | Lever- |
abs/2303.08774.
| | | | | | | aging GPT-4 | | for automatic | | translation | | post- |
| --- | --- | --- | --- | --- | --- | ----------- | --- | ------------- | --- | ----------- | --- | ----- |
editing.CoRR,abs/2305.14878.https://doi
| Long Ouyang, | | Jeffrey | Wu, | Xu Jiang, | Diogo | | | | | | | |
| ------------ | --- | ------- | --- | --------- | ----- | --- | --- | --- | --- | --- | --- | --- |
Almeida, Carroll L. Wainwright, Pamela .org/10.18653/v1/2023.findings-emnlp
| Mishkin, | Chong | Zhang, | Sandhini | | Agarwal, | .804 | | | | | | |
| -------- | ----- | ------ | -------- | --- | -------- | ---- | --- | --- | --- | --- | --- | --- |
Katarina Slama, Alex Ray, John Schulman, Danilo Neves Ribeiro, Shen Wang, Xiaofei Ma,
| Jacob | Hilton, | Fraser | Kelton, | Luke | Miller, | | | | | | | |
| ----- | ------- | ------ | ------- | ---- | ------- | --- | --- | --- | --- | --- | --- | --- |
HenryZhu,RuiDong,DeguangKong,Juliette
| Maddie | Simens, | Amanda | | Askell, | Peter | | | | | | | |
| ------ | ------- | ------ | --- | ------- | ----- | --- | --- | --- | --- | --- | --- | --- |
Burger,AnjelicaRamos,WilliamYangWang, Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| Welinder, | Paul | F. Christiano, | | Jan | Leike, and | | | | | | | |
| --------- | ---- | -------------- | --- | --- | ---------- | ------- | ------ | ------ | -------- | --- | ---- | ------ |
| | | | | | | Zhiheng | Huang, | George | Karypis, | | Bing | Xiang, |
Ryan Lowe. 2022. Training language models and Dan Roth. 2023. STREET: A multi-task
tofollowinstructionswithhumanfeedback.In
| | | | | | | structured | reasoning | | and | explanation | | bench- |
| --- | --- | --- | --- | --- | --- | ---------- | --------- | --- | --- | ----------- | --- | ------ |
ProceedingsoftheAnnualConferenceonNeu-
| | | | | | | mark. In | Proceedings | | of | the | 11th Interna- | |
| --- | --- | --- | --- | --- | --- | -------- | ----------- | --- | --- | --- | ------------- | --- |
ralInformationProcessingSystems(NeurIPS). tionalConferenceonLearningRepresentations
(ICLR).
| Liangming | Pan, | Alon Albalak, | | Xinyi | Wang, and | | | | | | | |
| --------- | ---- | ------------- | --- | ----- | --------- | --- | --- | --- | --- | --- | --- | --- |
William Yang Wang. 2023. Logic-LM: Em- Manasvi Sagarkar, John Wieting, Lifu Tu, and
| | | | | | | Kevin Gimpel. | | 2018. | Quality | signals | | in gen- |
| --- | --- | --- | --- | --- | --- | ------------- | --- | ----- | ------- | ------- | --- | ------- |
poweringlargelanguagemodelswithsymbolic
solvers for faithful logical reasoning. CoRR, erated stories. In Proceedings of the Seventh
abs/2305.12295.https://doi.org/10.18653 Joint Conference on Lexical and Computa-
| | | | | | | tional Semantics | | (SEM@NAACL-HLT | | | | 2018), |
| --- | --- | --- | --- | --- | --- | ---------------- | --- | -------------- | --- | --- | --- | ------ |
/v1/2023.findings-emnlp.248
| | | | | | | pages 192–202. | | https://doi.org/10.18653 | | | | |
| ---------- | ----- | -------- | ----- | --- | ----------- | -------------- | --- | ------------------------ | --- | --- | --- | --- |
| Jing-Cheng | Pang, | Pengyuan | Wang, | | Kaiyuan Li, | | | | | | | |
/v1/S18-2024
| Xiong-Hui | | Chen, Jiacheng | | Xu, | Zongzhang | | | | | | | |
| ---------------- | --- | -------------- | ------------- | -------- | --------- | ---------------- | ----- | ----- | ------------- | ------ | --------- | --- |
| | | | | | | Swarnadeep | Saha, | | Sayan | Ghosh, | Shashank | |
| Zhang, | and | Yang Yu. | 2023. | Language | model | | | | | | | |
| | | | | | | Srivastava, | and | Mohit | Bansal. | 2020. | PRover: | |
| self-improvement | | by | reinforcement | | learning | | | | | | | |
| | | | | | | Proof generation | | for | interpretable | | reasoning | |
contemplation.CoRR,abs/2305.14483.
| | | | | | | over rules. | In | Proceedings | | of the | 2020 | Confer- |
| ------------ | ---- | ------------ | --- | ------ | -------- | ---------------- | --------- | ----------- | -------- | ------ | -------- | ------- |
| | | | | | | ence on | Empirical | | Methods | in | Natural | Lan- |
| Debjit Paul, | Mete | Ismayilzada, | | Maxime | Peyrard, | | | | | | | |
| | | | | | | guage Processing | | | (EMNLP), | pages | 122–136. | |
BeatrizBorges,AntoineBosselut,RobertWest,
and Boi Faltings. 2023. REFINER: Reason- https://doi.org/10.18653/v1/2020
.emnlp-main.9
| ing feedback | | on intermediate | | representations. | | | | | | | | |
| ------------ | --- | --------------- | --- | ---------------- | --- | --- | --- | --- | --- | --- | --- | --- |
CoRR,abs/2304.01904. William Saunders, Catherine Yeh, Jeff Wu,
| | | | | | | Steven | Bills, | Long | Ouyang, | Jonathan | | Ward, |
| ------ | ------------ | ------- | --------- | --- | ------- | ------ | ------ | ---- | ------- | -------- | --- | ----- |
| Baolin | Peng, Michel | Galley, | Pengcheng | | He, Hao | | | | | | | |
andJanLeike.2022.Self-critiquingmodelsfor
Cheng,YujiaXie,YuHu,QiuyuanHuang,Lars
| | | | | | | assisting | human | evaluators. | | CoRR, | abs/2206 | |
| ------ | ---- | ---------- | ----- | --- | -------- | --------- | ----- | ----------- | --- | ----- | -------- | --- |
| Liden, | Zhou | Yu, Weizhu | Chen, | and | Jianfeng | | | | | | | |
.05802.
| Gao. | 2023. | Check your | facts | and | try again: | | | | | | | |
| --------- | ----- | ---------- | ------ | --- | ----------- | -------------- | --- | ----- | ----- | ----- | --- | --- |
| | | | | | | Michael Saxon, | | Xinyi | Wang, | Wenda | Xu, | and |
| Improving | large | language | models | | with exter- | | | | | | | |
WilliamYangWang.2023.PECO:Examining
nalknowledgeandautomatedfeedback.CoRR,
| | | | | | | single sentence | | label | leakage | in | natural | lan- |
| --- | --- | --- | --- | --- | --- | --------------- | --- | ----- | ------- | --- | ------- | ---- |
abs/2302.12813.
| | | | | | | guage inference | | datasets | | through | progressive | |
| --- | --- | --- | --- | --- | --- | --------------- | --- | -------- | --- | ------- | ----------- | --- |
Dongqi Pu and Vera Demberg. 2023. Chatgpt vs evaluation of cluster outliers. In Proceedings
human-authoredtext:Insightsintocontrollable of the 17th Conference of the European Chap-
textsummarizationandsentencestyletransfer. ter of the Association for Computational Lin-
In Proceedings of the 61st Annual Meeting of guistics (EACL), pages 3053–3066. https://
theAssociationforComputationalLinguistics: doi.org/10.18653/v1/2023.eacl-main
| StudentResearchWorkshop(ACL),pages1–18. | | | | | | .223 | | | | | | |
| --------------------------------------- | --- | --- | --- | --- | --- | ---- | --- | --- | --- | --- | --- | --- |
503
Je´re´my Scheurer, Jon Ander Campos, Tomasz Chung, Aakanksha Chowdhery, Quoc Le, Ed
Korbak, Jun Shern Chan, Angelica Chen, Chi, Denny Zhou, and Jason Wei. 2023. Chal-
KyunghyunCho,andEthanPerez.2023.Train- lenging big-bench tasks and whether chain-
inglanguagemodelswithlanguagefeedbackat of-thought can solve them. In Findings of the
scale.CoRR,abs/2303.16755. Association for Computational Linguistics:
| | | | | | | | ACL 2023, | pages | 13003–13051. | | | https://doi | |
| ------------ | --- | ---- | ------ | -------- | --- | ------ | --------- | ----- | ------------ | --- | --- | ----------- | --- |
| Timo Schick, | | Jane | A. Yu, | Zhengbao | | Jiang, | | | | | | | |
.org/10.18653/v1/2023.findings-acl.824
| Fabio Petroni, | | Patrick | S. | H. Lewis, | | Gautier | | | | | | | |
| -------------- | --- | ------- | --- | --------- | --- | ------- | --- | --- | --- | --- | --- | --- | --- |
Izacard,QingfeiYou,ChristoforosNalmpantis,
| | | | | | | | Oyvind Tafjord, | | Bhavana | Dalvi, | and | Peter | Clark. |
| ------- | ------ | --- | --------- | --- | ------- | ----- | --------------- | ------------ | ------- | ---------- | --- | ------------- | ------ |
| Edouard | Grave, | and | Sebastian | | Riedel. | 2023. | | | | | | | |
| | | | | | | | 2021. | ProofWriter: | | Generating | | implications, | |
PEER:Acollaborativelanguagemodel.InPro-
| | | | | | | | proofs, | and | abductive | statements | | over | natu- |
| -------- | ------ | ---- | ------------- | --- | ---------- | --- | ------- | --- | --------- | ---------- | --- | ---- | ----- |
| ceedings | of the | 11th | International | | Conference | | | | | | | | |
rallanguage.InFindingsoftheAssociationfor
onLearningRepresentations(ICLR).
ComputationalLinguistics:ACL-IJCNLP2021,
| John Schulman, | | Filip | Wolski, | Prafulla | Dhariwal, | | | | | | | | |
| -------------- | --- | ----- | ------- | -------- | --------- | --- | ----- | ---------- | --- | ------------------ | --- | --- | --- |
| | | | | | | | pages | 3621–3634. | | https://doi.org/10 | | | |
Alec Radford, and Oleg Klimov. 2017. Prox- .18653/v1/2021.findings-acl.317 Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| imal policy | | optimization | | algorithms. | | CoRR, | | | | | | | |
| --------------- | ---------- | ------------ | -------- | ----------- | ------- | ----- | ----------------------------------------- | ----- | ------------ | --------- | ------ | ------------- | --- |
| abs/1707.06347. | | | | | | | OyvindTafjord,BhavanaDalviMishra,andPeter | | | | | | |
| | | | | | | | Clark. | 2022. | Entailer: | Answering | | questions | |
| Omar Shaikh, | | Hongxin | Zhang, | | William | Held, | | | | | | | |
| | | | | | | | with faithful | | and truthful | | chains | of reasoning. | |
| Michael | Bernstein, | | and Diyi | Yang. | 2023. | On | | | | | | | |
InProceedingsofthe2022ConferenceonEm-
secondthought,let’snotthinkstepbystep!Bias
piricalMethodsinNaturalLanguageProcess-
andtoxicityinzero-shotreasoning.InProceed-
| | | | | | | | ing (EMNLP), | | pages | 2078–2093. | | https:// | |
| ------- | -------- | ------ | ------- | --- | ------ | ----- | ------------ | --- | ----- | ---------- | --- | -------- | --- |
| ings of | the 61st | Annual | Meeting | | of the | Asso- | | | | | | | |
doi.org/10.18653/v1/2022.emnlp-main
| ciation | for Computational | | | Linguistics | | (ACL), | | | | | | | |
| ------- | ----------------- | --- | --- | ----------- | --- | ------ | --- | --- | --- | --- | --- | --- | --- |
.134 | pages 4454–4470. | | | https://doi.org/10 | | | | | | | | | | |
| ---------------- | --- | --- | ------------------ | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
.18653/v1/2023.acl-long.244
| | | | | | | | Ryutaro Tanno, | | Melanie | F. | Pradier, | Aditya | V. |
| --- | --- | --- | --- | --- | --- | --- | -------------- | --- | ------- | --- | -------- | ------ | --- |
Noah Shinn, Federico Cassano, Beck Labash, Nori, and Yingzhen Li. 2022. Repairing neu-
Ashwin Gopinath, Karthik Narasimhan, and ral networks by leaving the right past behind.
ShunyuYao.2023.Reflexion:Languageagents In Proceedings of the 2022 Annual Confer-
with verbal reinforcement learning. CoRR, ence on Neural Information Processing Sys-
| abs/2303.11366. | | | | | | | tems(NeurIPS). | | | | | | |
| --------------- | --- | -------- | ------------ | --- | --- | ------ | -------------- | --- | --- | --- | --- | --- | --- |
| Anton Sinitsin, | | Vsevolod | Plokhotnyuk, | | | Dmitry | | | | | | | |
GladysTyen,HassanMansoor,PeterChen,Tony
| V. Pyrkin, | Sergei | Popov, | | and Artem | Babenko. | | | | | | | | |
| ---------- | ------ | ------ | --- | --------- | -------- | --- | -------- | ------ | -------- | --- | ----- | ---- | ------ |
| | | | | | | | Mak, and | Victor | Carbune. | | 2023. | LLMs | cannot |
2020.Editableneuralnetworks.InProceedings
| | | | | | | | find reasoning | | errors, | but | can | correct | them! |
| ---------- | ------------- | --- | ---------- | --- | --- | ------ | -------------- | --- | ------- | --- | --- | ------- | ----- |
| of the 8th | International | | Conference | | on | Learn- | | | | | | | |
CoRR,abs/2311.08516.
ingRepresentations(ICLR).
NehaSrikanthandRachelRudinger.2022.Partial- JonathanUesato,NateKushman,RamanaKumar,
| | | | | | | | H. Francis | Song, | Noah | Y. | Siegel, | Lisa | Wang, |
| --------------- | --- | ---- | ---- | --- | ------ | ------- | ---------- | ----- | ---- | --- | ------- | ---- | ----- |
| input baselines | | show | that | NLI | models | can ig- | | | | | | | |
nore context,but theydon’t. InProceedingsof Antonia Creswell, Geoffrey Irving, and Irina
the 2022 Conference of the North American Higgins. 2022. Solving math word problems
| | | | | | | | with process- | | and | outcome-based | | feedback. | |
| ------- | ------ | ----------- | --- | --- | ------------- | --- | ------------- | --- | --- | ------------- | --- | --------- | --- |
| Chapter | of the | Association | | for | Computational | | | | | | | | |
Linguistics: Human Language Technologies CoRR,abs/2211.14275.
| (NAACL-HLT), | | pages | 4753–4763. | | https:// | | | | | | | | |
| ------------ | --- | ----- | ---------- | --- | -------- | --- | --- | --- | --- | --- | --- | --- | --- |
InigoJauregiUnanue,JacobParnell,andMassimo
doi.org/10.18653/v1/2022.naacl-main
.350 Piccardi. 2021. Berttune: Fine-tuning neural
| | | | | | | | machine | translation | | with | bertscore. | | In Pro- |
| ------------- | ----------- | ----- | ------------ | ------- | ------------ | ---- | ----------- | ------------- | ------------- | ------ | ----------- | --- | ------- |
| Kaya Stechly, | Matthew | | Marquez, | | and Subbarao | | | | | | | | |
| | | | | | | | ceedings | of | the 59th | Annual | Meeting | | of the |
| Kambhampati. | | 2023. | GPT-4 | doesn’t | know | it’s | | | | | | | |
| | | | | | | | Association | for | Computational | | Linguistics | | and |
| wrong: | An analysis | | of iterative | | prompting | for | | | | | | | |
| | | | | | | | the 11th | International | | Joint | Conference | | on |
reasoningproblems.CoRR,abs/2310.12397.
| | | | | | | | Natural | Language | Processing | | (ACL/IJCNLP), | | |
| --- | --- | --- | --- | --- | --- | --- | ------- | -------- | ---------- | --- | ------------- | --- | --- |
Mirac Suzgun, Nathan Scales, Nathanael Scha¨rli, pages915–924.https://doi.org/10.18653
| Sebastian | Gehrmann, | | Yi | Tay, | Hyung | Won | /v1/2021.acl-short.115 | | | | | | |
| --------- | --------- | --- | --- | ---- | ----- | --- | ---------------------- | --- | --- | --- | --- | --- | --- |
504
Karthik Valmeekam, Matthew Marquez, and abilities of large language models. CoRR,
| Subbarao | Kambhampati. | | | 2023. | Can large | lan- | abs/2206.07682. | | | | | | |
| -------- | ------------ | --- | --- | ----- | --------- | ---- | --------------- | --- | --- | --- | --- | --- | --- |
guagemodelsreallyimprovebyself-critiquing
| | | | | | | | Jason Wei, | Xuezhi | | Wang, | Dale | Schuurmans, | |
| --- | --- | --- | --- | --- | --- | --- | ---------- | ------ | --- | ----- | ---- | ----------- | --- |
theirownplans?CoRR,abs/2310.08118.
| | | | | | | | Maarten | Bosma, | Brian | | Ichter, | Fei | Xia, Ed H. |
| --- | --- | --- | --- | --- | --- | --- | ------- | ------ | ----- | --- | ------- | --- | ---------- |
NeerajVarshney,WenlinYao,HongmingZhang, Chi, Quoc V. Le, and Denny Zhou. 2022b.
Jianshu Chen, and Dong Yu. 2023. A stitch Chain-of-thought prompting elicits reasoning
| in time | saves | nine: | Detecting | | and | mitigat- | | | | | | | |
| ------- | ----- | ----- | --------- | --- | --- | -------- | -------- | -------- | --- | ------- | --- | ----------- | --- |
| | | | | | | | in large | language | | models. | In | Proceedings | of |
ing hallucinations of LLMs by validating low- the Annual Conference on Neural Information
confidencegeneration.CoRR,abs/2307.03987. ProcessingSystems(NeurIPS).
DavidWanandMohitBansal.2022.Factpegasus:
| | | | | | | | Sean Welleck, | | Ximing | Lu, | Peter | West, | Faeze |
| ---------------- | ---- | -------------- | --- | --------- | -------------- | --- | ------------- | -------- | ---------------- | ----- | ---------- | ----------- | --------- |
| Factuality-aware | | pre-training | | and | fine-tuning | | | | | | | | |
| | | | | | | | Brahman, | Tianxiao | | Shen, | Daniel | | Khashabi, |
| for abstractive | | summarization. | | | In Proceedings | | | | | | | | |
| | | | | | | | and Yejin | Choi. | 2023. | | Generating | | sequences |
| of the | 2022 | Conference | of | the North | American | | | | | | | | |
| | | | | | | | by learning | | to self-correct. | | In | Proceedings | of |
Chapter of the Association for Computational Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
The11thInternationalConferenceonLearning
| Linguistics: | | Human | Language | | Technologies | | | | | | | | |
| ------------ | --- | ----- | -------- | --- | ------------ | --- | --- | --- | --- | --- | --- | --- | --- |
Representations(ICLR).
| (NAACL-HLT), | | pages | 1010–1028. | | https:// | | | | | | | | |
| ------------ | --- | ----- | ---------- | --- | -------- | --- | --- | --- | --- | --- | --- | --- | --- |
doi.org/10.18653/v1/2022.naacl-main Yixuan Weng, Minjun Zhu, Fei Xia, Bin Li,
.74 Shizhu He, Kang Liu, and Jun Zhao. 2023.
| | | | | | | | Large | language | models | | are | better | reasoners |
| ------------ | ------ | ----- | ------- | ------- | ---- | ------- | ----- | ------------------ | ------ | ----- | --- | --------------- | --------- |
| Boxin Wang, | Weixin | | Chen, | Hengzhi | Pei, | Chulin | | | | | | | |
| | | | | | | | with | self-verification. | | CoRR, | | abs/2212.09561. | |
| Xie, Mintong | | Kang, | Chenhui | Zhang, | | Chejian | | | | | | | |
https://doi.org/10.18653/v1/2023
| Xu, Zidi | Xiong, | Ritik | Dutta, | Rylan | Schaeffer, | | | | | | | | |
| -------- | ------ | ----- | ------ | ----- | ---------- | --- | --- | --- | --- | --- | --- | --- | --- |
.findings-emnlp.167
| Sang | T. Truong, | | Simran | Arora, | | Mantas | | | | | | | |
| ---- | ---------- | --- | ------ | ------ | --- | ------ | --- | --- | --- | --- | --- | --- | --- |
Mazeika, Dan Hendrycks, Zinan Lin, Yu Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri,
Cheng, Sanmi Koyejo, Dawn Song, and Bo Alane Suhr, Prithviraj Ammanabrolu, Noah
Li.2023a.Decodingtrust:Acomprehensiveas-
| | | | | | | | A. Smith, | | Mari | Ostendorf, | | and | Hannaneh |
| -------- | --- | --------------- | --- | --- | --- | ------- | ----------- | ------ | ---- | ------------ | --- | ----- | -------- |
| sessment | of | trustworthiness | | in | GPT | models. | | | | | | | |
| | | | | | | | Hajishirzi. | 2023a. | | Fine-grained | | human | feed- |
CoRR,abs/2306.11698. back gives better rewards for language model
training.CoRR,abs/2306.01693.
| Haotian | Wang, | Xiyuan | | Du, | Weijiang | Yu, | | | | | | | |
| --------- | ----- | ------ | ------ | -------- | -------- | ------- | -------- | --- | ----- | ---- | ------ | ----- | ---- |
| Qianglong | Chen, | Kun | Zhu, | Zheng | Chu, | Lian | | | | | | | |
| | | | | | | | Zhaofeng | Wu, | Linlu | Qiu, | Alexis | Ross, | Ekin |
| Yan, and | Yi | Guan. | 2023b. | Apollo’s | | oracle: | | | | | | | |
Akyu¨rek,BoyuanChen,BailinWang,Najoung
| Retrieval-augmented | | | reasoning | | in multi-agent | | | | | | | | |
| ------------------- | --- | --- | --------- | --- | -------------- | --- | ---- | ----- | -------- | --- | ---- | ---- | ------ |
| | | | | | | | Kim, | Jacob | Andreas, | and | Yoon | Kim. | 2023b. |
debates.CoRR,abs/2312.04854.
| | | | | | | | Reasoning | or | reciting? | Exploring | | the | capabili- |
| --- | --- | --- | --- | --- | --- | --- | --------- | --- | --------- | --------- | --- | --- | --------- |
Liyuan Wang, Xingxing Zhang, Hang Su, and tiesandlimitationsoflanguagemodelsthrough
Jun Zhu. 2023c. A comprehensive survey of counterfactualtasks.CoRR,abs/2307.02477.
| continual | learning: | | Theory, | method | and | appli- | | | | | | | |
| --------- | --------- | --- | ------- | ------ | --- | ------ | --------- | ----- | ---------- | --- | ----- | --- | -------- |
| | | | | | | | Yuxi Xie, | Kenji | Kawaguchi, | | Yiran | | Zhao, Xu |
cation.CoRR,abs/2302.00487.
| | | | | | | | Zhao, | Min-Yen | Kan, | Junxian | | He, | and Qizhe |
| ---------- | ---- | ----- | --- | ------ | --- | ----- | ----- | ------- | ------------- | ------- | -------- | --- | --------- |
| Tong Wang, | Ping | Chen, | and | Boyang | Li. | 2017. | | | | | | | |
| | | | | | | | Xie. | 2023. | Decomposition | | enhances | | reason- |
Predicting the quality of short narratives from ingviaself-evaluationguideddecoding.CoRR,
| social | media. | In Proceedings | | | of the | Twenty- | abs/2305.00633. | | | | | | |
| ------------------- | ------ | -------------- | ----- | ---------- | ---------- | ------- | --------------- | ----------- | ------ | ----- | -------- | --------- | ---- |
| Sixth International | | | Joint | Conference | on | Arti- | | | | | | | |
| | | | | | | | Wenda | Xu, Danqing | | Wang, | | Liangming | Pan, |
| ficial Intelligence | | (IJCAI), | | pages | 3859–3865. | | | | | | | | |
| | | | | | | | Zhenqiao | Song, | Markus | | Freitag, | William | Yang |
https://doi.org/10.24963/ijcai.2017
| /539 | | | | | | | Wang, | and Lei | Li. | 2023. | INSTRUCTSCORE: | | |
| ---------- | --- | ---- | ----- | ---------- | --- | ----- | --------- | ----------- | --- | --------- | -------------- | ----- | -------- |
| | | | | | | | Towards | explainable | | text | generation | | evalua- |
| Jason Wei, | Yi | Tay, | Rishi | Bommasani, | | Colin | | | | | | | |
| | | | | | | | tion with | automatic | | feedback. | | CoRR, | abs/2305 |
Raffel,BarretZoph,SebastianBorgeaud,Dani
| | | | | | | | .14282. | https://doi.org/10.18653/v1 | | | | | |
| --------- | --- | ------- | ------ | --- | ----- | ----- | ------- | --------------------------- | --- | --- | --- | --- | --- |
| Yogatama, | | Maarten | Bosma, | | Denny | Zhou, | | | | | | | |
/2023.emnlp-main.365
| Donald | Metzler, | | Ed | H. Chi, | Tatsunori | | | | | | | | |
| ------ | -------- | --- | --- | ------- | --------- | --- | --- | --- | --- | --- | --- | --- | --- |
Hashimoto, Oriol Vinyals, Percy Liang, Jeff Xiaojun Xu, Chang Liu, and Dawn Song. 2017.
Dean, and William Fedus. 2022a. Emergent Sqlnet: Generating structured queries from
505
natural language without reinforcement learn- Tao Yu, Michihiro Yasunaga, Kai Yang,
ing.CoRR,abs/1711.04436. Rui Zhang, Dongxu Wang, Zifan Li, and
| | | | | | | Dragomir | Radev. | 2018. | SyntaxSQLNet: | | | Syn- |
| -------- | ------- | ----------- | --- | ------ | ---------- | -------- | ----------- | ----- | ------------- | -------------- | --- | ------ |
| Hao Yan, | Saurabh | Srivastava, | | Yintao | Tai, Sida | | | | | | | |
| | | | | | | tax tree | networks | | for complex | | and | cross- |
| I. Wang, | Wen-tau | Yih, | and | Ziyu | Yao. 2023. | | | | | | | |
| | | | | | | domain | text-to-SQL | | task. | In Proceedings | | of |
Learningtosimulatenaturallanguagefeedback
| | | | | | | the 2018 | Conference | | on | Empirical | Methods | |
| --- | --- | --- | --- | --- | --- | -------- | ---------- | --- | --- | --------- | ------- | --- |
forinteractivesemanticparsing.InProceedings
| | | | | | | in Natural | Language | | Processing | | (EMNLP), | |
| --------- | ------------- | ------ | ------- | ----------- | ----------- | ---------------- | -------- | ------------------ | ---------- | --- | -------- | --- |
| of the | 61th | Annual | Meeting | of | the Associ- | | | | | | | |
| | | | | | | pages 1653–1663. | | https://doi.org/10 | | | | |
| ation for | Computational | | | Linguistics | (ACL), | | | | | | | |
.18653/v1/D18-1193
| pages | 3149–3170. | https://doi.org/10 | | | | | | | | | | |
| ----- | ---------- | ------------------ | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
.18653/v1/2023.acl-long.177 Wenhao Yu, Zhihan Zhang, Zhenwen Liang,
| | | | | | | Meng Jiang, | and | Ashish | Sabharwal. | | 2023. | Im- |
| ----------- | ---------- | --------- | ------- | --------- | ------ | ----------- | -------- | ------ | ---------- | --- | ------------- | --- |
| Kaiyu Yang, | | Jia Deng, | | and Danqi | Chen. | | | | | | | |
| | | | | | | proving | language | models | | via | plug-and-play | |
| 2022a. | Generating | | natural | language | proofs | | | | | | | |
with verifier-guided search. In Proceedings of retrievalfeedback.CoRR,abs/2305.14002.
the 2022 Conference on Empirical Methods Downloaded from http://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00660/2369509/tacl_a_00660.pdf by guest on 23 May 2024
| | | | | | | Weizhe Yuan, | | Kyunghyun | | Cho, | and | Jason |
| --- | --- | --- | --- | --- | --- | ------------ | --- | --------- | --- | ---- | --- | ----- |
in Natural Language Processing (EMNLP), Weston. 2023. System-level natural language
https://doi.org/10.18653
| pages 89–105. | | | | | | feedback.CoRR,abs/2306.13588. | | | | | | |
| ------------- | --- | --- | --- | --- | --- | ----------------------------- | --- | --- | --- | --- | --- | --- |
/v1/2022.emnlp-main.7
| | | | | | | Eric Zelikman, | Yuhuai | | Wu, Jesse | Mu, | and | Noah |
| ---------- | --- | --- | ------ | ----- | ------ | -------------- | ------ | ----- | --------- | ------------- | --- | ---- |
| Kevin Yang | and | Dan | Klein. | 2021. | FUDGE: | | | | | | | |
| | | | | | | D. Goodman. | | 2022. | Star: | Bootstrapping | | rea- |
Controlledtextgenerationwithfuturediscrim- soning with reasoning. In Proceedings of the
| inators. | In Proceedings | | of | the 2021 | Confer- | | | | | | | |
| -------- | -------------- | --- | --- | -------- | ------- | --- | --- | --- | --- | --- | --- | --- |
AnnualConferenceonNeuralInformationPro-
| ence of | the | North American | | Chapter | of the | | | | | | | |
| ------- | --- | -------------- | --- | ------- | ------ | --- | --- | --- | --- | --- | --- | --- |
cessingSystems(NeurIPS).
AssociationforComputationalLinguistics:Hu-
man Language Technologies (NAACL-HLT), Kechi Zhang, Zhuo Li, Jia Li, Ge Li, and Zhi
| | | | | | | Jin. 2023a. | Self-edit: | | Fault-aware | | code | edi- |
| ----- | ---------- | ------------------ | --- | --- | --- | ----------- | ---------- | --- | ----------- | --- | ---- | ---- |
| pages | 3511–3535. | https://doi.org/10 | | | | | | | | | | |
torforcodegeneration.CoRR,abs/2305.04087.
.18653/v1/2021.naacl-main.276
https://doi.org/10.18653/v1/2023
| Kevin Yang, | Yuandong | | Tian, | Nanyun | Peng, | .acl-long.45 | | | | | | |
| ----------- | -------- | -------------- | ----- | ----------- | ---------- | ------------ | --- | ------- | ----- | --- | ------- | ---- |
| and Dan | Klein. | 2022b. | | Re3: | Generating | | | | | | | |
| | | | | | | Kexun Zhang, | | Danqing | Wang, | | Jingtao | Xia, |
| longer | stories | with recursive | | reprompting | and | | | | | | | |
WilliamYangWang,andLeiLi.2023b.Algo:
| revision. | In Proceedings | | of | the 2022 | Confer- | | | | | | | |
| --------- | -------------- | --- | --- | -------- | ------- | --- | --- | --- | --- | --- | --- | --- |
ence on Empirical Methods in Natural Lan- Synthesizingalgorithmicprogramswithgener-
atedoracleverifiers.CoRR,abs/2305.14591.
| guage Processing | | (EMNLP), | | pages | 4393–4479. | | | | | | | |
| ---------------- | --- | -------- | --- | ----- | ---------- | --- | --- | --- | --- | --- | --- | --- |
https://doi.org/10.18653/v1/2022 Muru Zhang, Ofir Press, William Merrill, Alisa
| .emnlp-main.296 | | | | | | Liu,andNoahA.Smith.2023c.Howlanguage | | | | | | |
| --------------- | --- | --- | --- | --- | --- | ------------------------------------ | --- | --- | --- | --- | --- | --- |
Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak model hallucinations can snowball. CoRR,
| Shafran, | Thomas | L. | Griffiths, | Yuan | Cao, and | abs/2305.13534. | | | | | | |
| -------- | ------ | --- | ---------- | ---- | -------- | --------------- | --- | --- | --- | --- | --- | --- |
Karthik Narasimhan. 2023a. Tree of thoughts: Xinyu Zhu, Junjie Wang, Lin Zhang, Yuxiang
Deliberateproblemsolvingwithlargelanguage
| | | | | | | Zhang, | Yongfeng | Huang, | | Ruyi | Gan, | Jiaxing |
| --- | --- | --- | --- | --- | --- | ------ | -------- | ------ | --- | ---- | ---- | ------- |
models.CoRR,abs/2305.10601.
| | | | | | | Zhang, and | Yujiu | Yang. | 2023. | | Solving | math |
| --- | --- | --- | --- | --- | --- | ---------- | ----- | ----- | ----- | --- | ------- | ---- |
Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan word problems via cooperative reasoning in-
| | | | | | | duced language | | models. | | In Processings | | of |
| ------ | ------ | --- | ------ | ----- | ------ | -------------- | --- | ------- | --- | -------------- | --- | --- |
| Cheng, | Zhoubo | Li, | Shumin | Deng, | Huajun | | | | | | | |
Chen,andNingyuZhang.2023b.Editinglarge the 61th Annual Meeting of the Associa-
language models: Problems, methods, and op- tion for Computational Linguistics (ACL),
| | | | | | | pages 4471–4485. | | https://doi.org/10 | | | | |
| --- | --- | --- | --- | --- | --- | ---------------- | --- | ------------------ | --- | --- | --- | --- |
portunities.CoRR,abs/2305.13172.
.18653/v1/2023.acl-long.245
| Seonghyeon | Ye, | Yongrae | | Jo, Doyoung | Kim, | | | | | | | |
| ---------- | --- | ------- | --- | ----------- | ---- | --- | --- | --- | --- | --- | --- | --- |
SungdongKim,HyeonbinHwang,andMinjoon Terry Yue Zhuo, Yujin Huang, Chunyang Chen,
Seo. 2023. Selfee: Iterative self-revising LLM andZhenchangXing.2023. Redteamingchat-
empowered by self-feedback generation. Blog gptviajailbreaking:Bias,robustness,reliability
| post. | | | | | | andtoxicity.CoRR,abs/2301.12867. | | | | | | |
| ----- | --- | --- | --- | --- | --- | -------------------------------- | --- | --- | --- | --- | --- | --- |
506 |