类比的长河，为何流到大模型就被截流？
鏂� | 杩介棶nextquestion
褰撲笅鎯宠�鎵惧埌浜涒€滄櫘閫氫汉绫绘搮闀匡紝鑰屽ぇ妯″瀷涓嶆搮闀库€濈殑浠诲姟锛屼技涔庤秺鏉ヨ秺闅句簡銆傗€滅被姣斺€濆彲鑳藉氨鏄�繖鏍风殑浠诲姟锛岃繖涓嶅彧鏄�汉宸ユ櫤鑳界殑鈥滈樋鍏嬬悏鏂�箣韪碘€濓紝鏇存樉闇插嚭涓嶅悓澶фā鍨嬮棿浠ュ強澶фā鍨嬩笌浜虹被涔嬮棿鐨勬湰璐ㄥ樊寮傘€�
鍦ㄣ€婅〃璞′笌鏈�川銆嬩竴涔︿腑锛岃�鐭ョ�瀛﹀�渚�笘杈撅紙Douglas Hofstadter锛夋寚鍑猴細
绫绘瘮涓嶄粎浠呮槸璇�█鎴栭€昏緫鐨勫伐鍏凤紝鏇存槸鎬濈淮鐨勫熀鏈�崟浣嶃€�
鎴戜滑鏃ュ父璇�█涓�厖婊′簡绫绘瘮鍜岄殣鍠伙紝灏卞�鍚屸€滃厖婊♀€濅竴璇嶆湰韬�€傜被姣旇兘澶熸縺娲诲垱閫犲姏銆備緥濡傦紝鐖卞洜鏂�潶灏嗗紩鍔涘満绫绘瘮涓轰竴涓�噸鐗╄�鏀惧叆韫﹀簥鍚庨€犳垚鐨勮〃闈㈠集鏇诧紝杩欏惎鍙戜粬鎻愬嚭浜嗗箍涔夌浉瀵硅�銆傜被姣旇繕鑳借В閲婇毦浠ョ悊瑙ｇ殑鐜拌薄銆傚氨鍍忎负浜烘墍鐔熺煡鐨勭被姣斺€滄剰璇嗗氨鍍忓啺灞扁€濓紝閫氳繃灏嗘剰璇嗕笌鍐板北鑱旂郴璧锋潵锛屼汉浠�彲浠ョ洿瑙傚湴鎺ㄦ柇鍑烘剰璇嗗湪姘撮潰涓嬬殑娣卞害鍜屽�鏉傛€с€�
閭ｄ箞锛屽ぇ璇�█妯″瀷鏄�惁涔熷叿鏈夌被姣旇兘鍔涳紵
鍦ㄦ満鍣ㄥ�涔犱腑锛岀被姣斾綋鐜颁负鈥�0灏濊瘯鎺ㄧ悊鈥濓紝鍗充笉缁欏ぇ妯″瀷鍙�緵瀛︿範鐨勭ず渚嬶紝鑰屾槸璁╁ぇ妯″瀷鑷��鏍规嵁棰樼洰杩涜�鎺ㄧ悊銆備负浜嗛獙璇佸ぇ妯″瀷鑳藉惁杩涜�绫绘瘮鎺ㄧ悊锛學ebb绛変汉锛�2023锛夎�璁″苟浣跨敤浜嗕笁绉嶇被姣旀帹鐞嗕换鍔♀€斺€斿瓧绗︿覆绫绘瘮銆佹暟瀛楃煩闃靛拰鏁呬簨绫绘瘮锛屼互姝ゆ祴璇旼PT3闈㈠�涓嶅悓绫诲瀷浠诲姟鐨勬帹鐞嗚兘鍔涖€傞€氳繃杩欏�娴嬭瘯锛岀爺绌朵汉鍛樿�涓轰粬浠�瘉鏄庝簡GPT-3鍏锋湁绫绘瘮鎺ㄧ悊鑳藉姏[1]銆�
浣嗘槸锛屾洿杩涗竴姝ョ殑闂��鏄�紝杩欎簺澶фā鍨嬩細涓嶄細鍙�槸鍦ㄥ洖蹇嗚�缁冩暟鎹�紝鑰屽苟闈炵湡姝ｇ殑绫绘瘮鍛�紵褰撻潰瀵规洿鍙樺寲澶氭牱鐨勯棶棰樻椂锛屽ぇ妯″瀷鑳藉惁鍏锋湁绋冲畾鐨勭被姣旇兘鍔涳紵
01 澶фā鍨嬭兘璇绘噦棰樼洰鈥滈┈鐢测€濅笅鐨勬湰璐ㄥ悧锛�
涓轰簡妫€娴嬫ā鍨嬫槸鍚︿緷璧栬〃闈㈢壒寰佹垨鎹峰緞锛岃€岄潪鐪熸�鐨勬娊璞℃帹鐞嗭紝鍦ｅ�鑿茬爺绌堕櫌鐨凩ewis & Mitchell锛屽熀浜嶹ebb绛変汉璁捐�鐨勫熀鏈�浆鎹㈠拰娉涘寲绫诲瀷锛岃�璁′簡鏇磋繘涓€姝ョ殑鍙樹綋娴嬭瘯[2]銆�
浠栦滑缁欓�鐩��涓€浜涒€滈┈鐢测€濓紝鍦ㄤ笉鏀瑰彉鏈�川鐨勫悓鏃讹紝璁╅�鐩�湅璧锋潵涓嶅悓锛涚劧鍚庣敤鏂扮殑娴嬭瘯瀵笹PT-3锛坱ext-davinci-003锛変互鍙婅繎鏈熸洿鏂扮殑澶фā鍨婫PT-3.5锛坓pt-3.5-turbo-0613锛夈€丟PT-4锛坓pt-4-0613锛夎繘琛岀被姣旇兘鍔涙祴璇曪紝鍖呮嫭瀛楃�涓层€佹暟瀛楃煩闃靛拰鏁呬簨绫绘瘮瀹為獙銆傝繖绫荤爺绌朵腑锛屾渶甯哥敤鍒扮殑鏄�警涓栬揪浜�1985骞存彁鍑虹殑鈥滃瓧绗︿覆绫绘瘮鈥�*銆�
* 瀛楃�涓茬被姣旓細a b c d 鈫� a b c e; i j k l 鈫� ?
鍏朵腑锛岀�涓€閮ㄥ垎鏄�"婧愯浆鎹�"锛岀�浜岄儴鍒嗘槸"鐩�爣"锛屼换鍔℃槸浠ョ被浼间簬婧愯浆鎹㈢殑鏂瑰紡杞�崲鐩�爣瀛楃�涓层€�
2023骞达紝Webb绛変汉鎻愬嚭浜嗗叚绉嶈浆鎹㈢被鍨嬶紙濡傚簭鍒楁墿灞曘€佸悗缁с€佸墠椹辩瓑锛夊拰澶氱�娉涘寲绫诲瀷锛堝�瀛楁瘝鍒版暟瀛椼€佸垎缁勩€佹洿闀跨洰鏍囩瓑锛夌殑缁勫悎銆備粬浠�负姣忕�闂��绫诲瀷鐢熸垚浜嗗ぇ閲忛棶棰橈紝骞跺皢杩欎簺闂��缁欏埌GPT-3锛坱ext-davinci-003锛変互鍙�57鍚峌CLA鏈��鐢熻繘琛屾祴璇曘€傜粨鏋滃彂鐜帮紝浜虹被鍙備笌鑰呯殑鍑嗙‘鐜囪〃鐜板嚭寰堝ぇ鐨勫樊寮傦紝浣嗘€讳綋鑰岃█锛孏PT-3鍦ㄥぇ澶氭暟闂��绫诲瀷涓婄殑琛ㄧ幇鐢氳嚦浼樹簬骞冲潎浜虹被琛ㄧ幇[1]銆�
浣嗘槸锛岃繖椤圭爺绌朵腑鎵€浣跨敤鐨勫瓧姣嶈〃鍧囦负鏍囧噯鑻辨枃瀛楁瘝琛ㄥ強鍏跺浐鏈夐『搴忥紝娴嬭瘯涓�ぇ妯″瀷琛ㄧ幇鍑烘潵鐨勨€滅被姣旇兘鍔涒€濇槸鍚﹀彲鑳戒緷璧栬〃闈㈢壒寰佽蛋浜嗏€滄嵎寰勨€濓紵涓烘�锛孡ewis & Mitchell淇濈暀浜嗗熀鏈�浆鎹㈠拰娉涘寲绫诲瀷锛屽張杩涗竴姝ュ垱寤轰簡涓ょ被鍙樹綋[2]锛�
- 铏氭瀯瀛楁瘝琛�細闅忔満鎵撲贡2-20涓�瓧姣嶇殑椤哄簭锛屽垱寤�28绉嶄笉鍚岀殑鎵撲贡瀛楁瘝琛�
- 绗﹀彿瀛楁瘝琛�細鐢ㄩ潪瀛楁瘝绗﹀彿瀹屽叏鏇夸唬瀛楁瘝锛屽垱寤�9绉嶄笉鍚岀殑绗﹀彿瀛楁瘝琛�
鐮旂┒浜哄憳瀵圭湡瀹炵殑鎷変竵瀛楁瘝琛�紝闅忔満閫夊彇1-3瀵硅繘琛屾浛鎹�紝鐒跺悗鍒嗗埆缁欎汉绫诲拰GPT-3銆丟PT-3.5銆丟PT-4杩涜�浜嗘祴璇曘€�
鈻峰浘1. Lewis & Mitchell缁欏彈璇曚汉绫诲拰澶фā鍨嬬殑绫绘瘮闂��绀轰緥. 鍥炬簮锛歔2]
缁撴灉鏄剧ず锛屽綋瀛楁瘝琛ㄧ殑鏇挎崲娆℃暟澧炲姞鍚庯紝涓嶈�鏄疓PT3銆丟PT3.5鎴栧埌GPT4锛屽叾鍥炵瓟鍑嗙‘鎬ч兘鏈変笅闄嶏紝涓旈兘鏄捐憲浣庝簬鍦ㄧ嚎鎷涘嫙鐨勪汉绫诲彈璇曡€匸2]銆�
鈻峰浘2锛氫笉鍚屽瓧姣嶈〃鏇挎崲娆℃暟涓嬶紝GPT妯″瀷鍜屼汉绫昏�璇曡€呯殑鍑嗙‘鎬у�姣�. 鍥炬簮锛歔2]
Mitchell鍥㈤槦杩樺仛杩囦竴椤瑰皾璇曪紝浠栦滑璁�42鍚嶅効绔ワ紙7-9宀侊級銆�62鍚嶆垚浜轰互鍙�4绉嶅ぇ妯″瀷锛圓nthropic鐨凜laude-3.5銆丟oogle鐨凣emma-2 27B銆丱pen AI鐨凣PT-4o鍜孧eta鐨凩lama-3.1 405B锛夛紝鎺ュ彈鎷変竵瀛楁瘝琛ㄣ€佸笇鑵婂瓧姣嶈〃鍜岀�鍙峰垪琛ㄤ笁绉嶆潯浠剁殑瀛楃�涓茬被姣斾换鍔�3]銆�
鈻峰浘3锛氫笉鍚岀被鍨嬬殑瀛楁瘝鎺ㄧ悊闂��. 鍥炬簮锛歔3]
缁撴灉鏄剧ず锛屽ぇ妯″瀷闈㈠�绫绘瘮闂��鏃讹紝鍑嗙‘鎬у氨浼氭樉钁椾笅闄嶏紝琛ㄧ幇鐢氳嚦涓嶅�鍎跨�銆傚氨鎷縂PT-4o鍜孋laude-3.5鏉ヨ�锛屽湪鎷変竵璇�瓧姣嶈〃涓婏紝鍏跺钩鍧囧噯纭�€ц�楂樹簬鍎跨�骞舵帴杩戞垚浜猴紱浣嗗綋棰樼洰鎹㈡垚甯岃厞瀛楁瘝锛屽噯纭�€у氨浼氭樉钁椾笅闄嶏紱鑰屽埌浜嗙�鍙锋椂锛屽叾鍑嗙‘鎬х敋鑷充笉濡傚�绔ャ€傝€屽叾浠栧紑婧愭ā鍨嬪�Llama-3.1 405B鍜孏emma-2 27B锛屽叾鍑嗙‘鎬т笅闄嶆洿涓烘槑鏄綶3]銆�
鈻峰浘4锛氫笉鍚屽ぇ妯″瀷鍜屼汉绫诲湪涓夌被瀛楃�涓茬被姣斾腑鐨勮〃鐜板�姣�. 鍥炬簮锛歔3]
涓婅堪缁撴灉璇存槑锛屽綋瀹為獙寮曞叆鈥滃紓鏋勨€濆瓧姣嶈〃鏃讹紝浜虹被鐢氳嚦鍎跨�浠嶇劧鑳藉�瑙ｅ喅闂��锛岃€屽ぇ妯″瀷鍒欎細鍑洪敊銆備竴涓�兘澶熺湡姝ｇ悊瑙ｅ拰绫绘瘮鐨勭郴缁燂紝搴旇�鍦ㄥ彉鍖栫殑鎯呭喌涓嬩篃鑳戒繚鎸侀珮鎬ц兘鈥斺€旇繖姝ｆ槸GPT绯诲垪澶фā鍨嬩笉鍏峰�鐨勮兘鍔涖€�
璇昏€呬篃璁镐細濂藉�锛屽叾浠栨帹鐞嗗ぇ妯″瀷鑳藉惁鍥炵瓟杩欐牱鐨勯棶棰樸€傜瑪鑰呯畝鍗曞皾璇曚簡涓€涓嬶紝鍦―eepSeek瀹樻柟鐨勫叏灏哄�R1鍙奦3妯″瀷锛屼互鍙婇樋閲岄€氫箟鍗冮棶鐨凲wQ 32B鎺ㄧ悊妯″瀷涓�紝瀵逛簬澶氭�鏇挎崲鍚庣殑铏氭瀯瀛楁瘝琛�紝妯″瀷鑳藉�姝ｇ‘鍥炵瓟锛屽苟缁欏嚭绗﹀悎浜虹被鎬濊€冭繃绋嬬殑鎺ㄧ悊杩囩▼鐨勩€�
浣嗗綋DeepSeek妯″瀷鍙樹负钂搁�Qwen鎴杔amma鐨�32B銆�14B銆�8B鎴�1.5B灏哄�鏃讹紝绗旇€呮湁闄愮殑鍑犳�瑙傚療鍙戠幇锛屾ā鍨嬮兘鍛堢幇鍑鸿繃搴︽€濊€冪殑鐗瑰緛锛屽嵆浼氬湪鎬濊€冭繃绋嬩腑灏濊瘯浼楀�杩囦簬澶嶆潅鐨勬ā寮忥紝灞曠ず鏁颁竾token鐨勭箒鏉傛€濊€冭繃绋嬶紝鏈€缁堜粛鐒剁粰鍑轰簡閿欒�鐨勫洖绛斻€傜瑪鑰呰繕閬囧埌鍦ㄦ€濊€冭繃绋嬩腑锛屽凡缁忓彂鐜版�纭�瓟妗堬紝浣嗗張鍦ㄦ帴涓嬫潵鐨勬€濊€冭繃绋嬩腑锛屽ぇ妯″瀷灏嗗叾鍚﹀喅鐨勬�渚嬨€�
绗旇€呰�涓猴紝鍩轰簬寮哄寲瀛︿範鐨勫ぇ妯″瀷鑳藉惁杩涜�绫绘瘮锛岃繕闇€瑕佽繘涓€姝ョ殑瀹氶噺鐮旂┒锛屼互鑰冨療涓嶅悓灏哄�妯″瀷鐨勫噯纭�害銆備緥濡傦紝瀵逛簬妯″瀷灏嗛棶棰樿繃搴﹀�鏉傚寲鐨勫€惧悜锛屽彲浠ユ牴鎹�€濊€冭繃绋嬶紝瀵规ā鍨嬬殑閿欒�杩涜�杩涗竴姝ョ殑鍒嗙被锛屼互姝ゆ垨鍙�垱寤哄嚭涓€涓�瘎浼颁竴鑸�€濈淮鑳藉姏鐨勮€冩牳鎸囨爣銆�
姝ゅ�锛岃繕鍙�互缁勫悎瀛楃�涓茬被姣旂殑6涓�彉绉嶏紝璁捐�鏇村�鐨勯�鐩�紝渚嬪�鍦ㄥ瓧姣嶈〃涓�寘鍚�暟瀛椼€佽嫳鏂囧瓧姣嶃€佹眽瀛楀強绗﹀彿锛岃繖鏍风殑鏀瑰彉鎴栬�瀵逛汉绫讳笉浼氬奖鍝嶅噯纭�€э紝浣嗗彲鑳戒細瀵艰嚧澶фā鍨嬬殑鍑嗙‘搴︿笅闄嶃€傚悓鏃讹紝杩橀渶瑕佽€冨療鎺ㄧ悊妯″瀷瀵逛簬杩欑被闂��鐨勬€濊€冩椂鎵€鐢ㄧ殑token鏁伴噺锛屼粠鑰屽噺灏戣�绠楁垚鏈�€�
02 澶фā鍨嬭兘鐞嗚В鎺ㄧ悊瑙勫垯鍚楋紵
闄や簡瀛楁瘝琛ㄦ帹鐞嗭紝杩樺彲浠ヤ娇鐢ㄦ暟瀛楃煩闃电被闂��锛堝垎鏋愭暟瀛楁ā寮忎互纭�畾缂哄け鐨勬暟瀛楋級銆傛暟瀛楃煩闃垫祴璇曠殑璁捐�鎬濊矾婧愪簬缁忓吀鐨勭憺鏂囨笎杩涚煩闃垫祴璇曪紙Raven's Progressive Matrices锛夛紝杩欐槸涓€绉嶅箍娉涚敤浜庢祴閲忔娊璞℃帹鐞嗚兘鍔涚殑闈炶�瑷€鏅哄姏娴嬭瘯銆傜浉姣斾箣鍓嶅瓧姣嶈〃绫绘瘮涓�敼鍙橀棶棰樼殑琛ㄧ幇褰㈠紡锛屾暟瀛楃煩闃甸棶棰橀€氳繃缁勫悎瑙勫垯锛岃€冨療浜嗗ぇ妯″瀷鎵€璋撶殑鎺ㄧ悊鑳藉姏鏄�湡姝ｇ殑鎶借薄鐞嗚В杩樻槸妯″紡鍖归厤銆�
杩欑被闂��涓�紝娑夊強鐨勫熀纭€瑙勫垯鏈�4绉嶏紝棰樼洰鐢辫繖浜涘熀纭€瑙勫垯缁勫悎鑰屾垚锛�
鐮旂┒鑰呭�鍘熷�鏁板瓧鐭╅樀娴嬭瘯杩涜�浜嗕袱涓�叧閿�彉鍖栵細绌虹櫧浣嶇疆鍙樺寲锛堝皢绌虹櫧浣嶇疆鍙樹负鐭╅樀鐨勫叾浠栦綅缃�,濡俒1,3]鎴朳2,2]锛夊拰瑙勫垯澶嶆潅搴﹀彉鍖栵紙璁捐�浜嗕笉鍚屽�鏉傚害绾у埆鐨勭煩闃甸棶棰橈紝浠庣畝鍗曞埌澶嶆潅锛塠2]銆�
鈻峰浘5锛氭秹鍙婂埌澶氫釜瑙勫垯鐨勬暟瀛楃煩闃垫帹鐞嗛棶棰樹互鍙婂皢鏁板瓧鎹�负绗﹀彿鐨勬暟瀛楃煩闃垫帹鐞嗛棶棰�. 鍥炬簮锛歔2]
缁撴灉鏄剧ず锛屼粎鏀瑰彉绌虹櫧浣嶇疆杩欎竴琛ㄩ潰鐗瑰緛锛屽氨瀵艰嚧GPT妯″瀷琛ㄧ幇澶у箙涓嬫粦銆傚敖绠�PT-4鍦ㄦ爣鍑嗘祴璇曚腑鎺ヨ繎浜虹被琛ㄧ幇锛�83% vs 87%锛夛紱浣嗗湪鍙樹綋娴嬭瘯涓�紝GPT-4鐨勮〃鐜颁笅闄嶅箙搴︼紙26%锛夎繙澶т簬浜虹被锛�4%锛塠2]銆傝繖鎰忓懗鐫€锛屽嵆浣挎槸鏈€鍏堣繘鐨勬ā鍨嬩篃琛ㄧ幇鍑哄�鏍煎紡鍙樺寲鐨勯珮搴︽晱鎰熸€э紝鍚屾牱琛ㄦ槑浜嗗ぇ妯″瀷鐨勬帹鐞嗚兘鍔涗笉閭ｄ箞椴佹�銆�
鈻峰浘6锛氭暟瀛楃煩闃垫帹鐞嗛棶棰樼殑鍑嗙‘搴�. 鍥炬簮锛歔2]
鍦ㄦ暟瀛楃煩闃甸棶棰樹腑锛屽綋缂哄け鏁板瓧鐨勪綅缃�敼鍙樻椂锛孏PT 妯″瀷鐨勮〃鐜版樉钁椾笅闄嶃€傝繖琛ㄦ槑浜嗗ぇ妯″瀷涓嶄粎涓嶇悊瑙ｉ�鐩�€冨療鐨勬槸浠€涔堬紝鏇存病鏈夌悊瑙ｈ繘琛岀被姣旀墍渚濊禆鐨勮�鍒欍€傚叾鍦ㄥ崟涓€瑙勫垯鎴栧師濮嬪瓧姣嶈〃涓婄殑浼樺紓琛ㄧ幇锛屼緷璧栦簬棰樼洰涓庣ず渚嬩箣闂村湪鐨勮〃闈㈢浉浼兼€э紝鑰岄潪鏇存繁灞傛�鐨勫洜鏋滄帹鐞嗐€�
涓庝箣绫讳技鐨勶紝杩樺寘鎷�笅闈㈢殑鐭╅樀鍙樻崲闂��銆備竴椤圭爺绌堕€氳繃绠€鍖栫増ARC锛堟娊璞′笌鎺ㄧ悊璇�枡搴擄級浠诲姟瀵规瘮浜嗕笉鍚屽勾榫勪汉绫伙紙鍎跨�涓庢垚浜猴級鍜屽ぇ鍨嬭�瑷€妯″瀷鐨勮�瑙夌被姣旀帹鐞嗚〃鐜帮紝缁撴灉鍚屾牱鍙戠幇浜虹被鍦ㄥ�鏉備换鍔′腑鏄捐憲浼樹簬澶фā鍨嬶紝鑰屽ぇ妯″瀷甯镐緷璧栧�鍒舵垨鐭╅樀缁勫悎绛栫暐锛岀己涔忔娊璞℃�蹇电悊瑙ｈ兘鍔沎4]銆�
鈻峰浘6: 缁欎汉绫诲拰澶фā鍨嬬殑瑙嗚�绫绘瘮鎺ㄧ悊闂��绀轰緥锛屼互鍙婁笉鍚屾帹鐞嗚�鍒欏�搴旈�鐩�殑澶фā鍨嬩笌浜虹被鐨勫噯纭�害瀵规瘮. 鍥炬簮锛歔4]
03 鍦ㄥ熀浜庡父璇嗙殑鏂囩�鎺ㄧ悊涓婏紝 澶фā鍨嬭〃鐜板�浣曪紵
涓婅堪涓ょ被绫绘瘮闂��閮藉彲浠ョ畻鏄�€滅悊绉戦�鐩�€濓紝瀵逛簬鈥滄枃绉戠敓鈥濈殑澶фā鍨嬶紝鎴栬�纭�疄鏈変簺闅句簡銆傜浉姣斾箣涓嬶紝鏁呬簨绫绘瘮鍒欎富瑕佽€冨療澶фā鍨嬪熀浜庡父璇嗙殑绫绘瘮鑳藉姏銆�
杩欑被棰樼洰閫氬父缁欏嚭1涓�嚑鍙ヨ瘽缁勬垚鐨勭煭鏁呬簨锛岀劧鍚庤�姹傚弬涓庤€呭垽鏂�晠浜�1鍜屾晠浜婣鎴朆鍝�竴涓�洿涓虹浉浼硷紝鍗宠瘑鍒�煭鏁呬簨涔嬮棿鐨勭浉浼兼€э紝骞朵粠澶氫釜閫夐」涓�€夋嫨鏈€绗﹀悎绫绘瘮鍏崇郴鐨勭瓟妗堛€�
鈻峰浘7锛氱浉浼兼晠浜嬬殑绫绘瘮鍒ゆ柇锛岄�鐩�殑鏁呬簨鏄�竴涓�悆涓嶅埌钁¤悇璇磋憽钀勯吀鐨勯�瀛愮増鏈�紝鏁呬簨A灏嗕富瑙掓崲鎴愪簡涓€涓�コ瀛╋紝鑰屽湪鏁呬簨B涓�紝涓昏�娌℃湁鑾峰緱鐩镐技鐨勪笢瑗匡紝鏄�敱浜庝笉鍠滄�鑰岄潪鎷夸笉鍒�. 鍥炬簮锛歔2]
鍦↙ewis & Mitchell鐨勭爺绌朵腑锛屼粬浠�皾璇曚簡涓ょ�鍙樹綋锛氫竴鏄�殢鏈烘墦涔辩瓟妗堥€夐」鐨勯『搴忥紝浜屾槸淇濇寔鏍稿績鍏崇郴涓嶅彉锛屼絾閲嶅啓鏁呬簨鐨勮〃杩版柟寮廩2]銆�
鍦ㄦ晠浜嬬被姣斾腑锛孏PT-4 鍊惧悜浜庢洿棰戠箒鍦伴€夋嫨绗�竴涓�粰鍑虹殑绛旀�浣滀负姝ｇ‘绛旀�锛岃€屼汉绫诲垯涓嶅彈绛旀�椤哄簭鐨勫奖鍝嶃€傛�澶栵紝瀵逛簬澶фā鍨嬶紝灏嗘晠浜嬬敤涓嶅悓鐨勮瘽閲嶈堪锛屼篃浼氶檷浣庡湪鏁呬簨绫绘瘮闂��涓婄殑鍑嗙‘鎬�2]銆�
鈻峰浘8锛氭枃瀛楃被姣旈棶棰樹笂澶фā鍨嬬殑琛ㄧ幇宸�紓. 鍥炬簮锛歔2]
鏁呬簨绫绘瘮鏇存帴杩戣嚜鐒惰�瑷€澶勭悊鐨勫疄闄呭簲鐢ㄥ満鏅�紝浣嗙爺绌剁粨鏋滃嵈琛ㄦ槑鍗充娇鍦ㄨ�瑷€妯″瀷鐨�"涓诲満"涓婏紝瀹冧滑鐨勭被姣旀帹鐞嗚兘鍔涗粛鐒剁己涔忕湡姝ｇ殑鐏垫椿鎬у拰椴佹�鎬э紝杩囧害渚濊禆浜庤〃闈㈢壒寰佷笌鐗瑰畾鐨勭瓟妗堟牸寮忥紝鑰岄潪娣卞眰鐞嗚В鎶借薄鍏崇郴銆�
涓烘�锛岀瑪鑰呬篃璁炬兂浜嗕竴绉嶅垽鍒�柟寮忥紝渚嬪�瀵规瘮澶фā鍨嬪拰浜虹被鍥炵瓟杩欑被闂��鐨勫噯纭�€с€傚彲浠ョ敓鎴愬緢澶氱粍绫绘瘮闂��锛屽苟鎷涘嫙璇昏繃鐩稿叧灏忚�鐨勬櫘閫氫汉锛屼互鑾峰彇澶т紬璁ょ煡涓�殑涓€鑸�€у洖绛旓紝鐒跺悗瀵规瘮涓嶅悓澶фā鍨嬪拰浜虹被鍥炵瓟鐨勫樊寮傛€с€�
閫氳繃璁剧疆涓嶅悓鐨勭粏鍒嗛棶棰橈紝鍙�互鑰冨療澶фā鍨嬩笌浜虹被鍦ㄧ被姣旇兘鍔涙柟闈㈢殑鐩镐技搴﹀強浠峰€艰�瀵归綈鎯呭喌銆�
- 璺ㄦ枃浣撶被姣旇兘鍔涳細鍦ㄩ�鏍煎樊寮傝緝澶х殑浣滃搧闂达紝濡備腑鏂囩殑閲戝焊姝︿緺鎴栥€婄孩妤兼ⅵ銆嬩笌鑻辨枃鐨勩€婂搱鍒╂尝鐗广€嬶紝澶фā鍨嬬殑绫绘瘮鍑嗙‘鎬ц兘鍚﹁揪鍒颁汉绫绘按骞筹紵
- 瑙掕壊鐞嗚В宸�紓锛氬ぇ妯″瀷鍦ㄥ�鐞嗙敺鎬у拰濂虫€ц�鑹茬被姣旀椂锛屾槸鍚﹀瓨鍦ㄥ噯纭�€у樊寮傦紵
- 缇や綋鍋忓ソ鐗瑰緛锛氬ぇ妯″瀷鐨勭被姣斿亸濂芥槸鍚︽洿鎺ヨ繎鐗瑰畾浜虹兢锛堝�涓嶅悓鎬у埆銆佸勾榫勬�鐨勪汉缇わ級锛�
- 閫昏緫閫掓帹鎬э細澶фā鍨嬬殑绫绘瘮鏄�惁鍏锋湁浼犻€掓€х壒寰侊紙鍗冲綋A>B涓擝>C鏃讹紝鏄�惁蹇呯劧鎺ㄥ�鍑篈>C锛夛紵
鈻峰浘9锛氬ぇ妯″瀷鑳藉�鍦ㄨ法瓒婃枃瀛︿綔鍝佽繘琛岀被姣斿悧锛熸湰鏂囦綔鑰呬笌DeepSeek瀵硅瘽鎴�浘锛屽叾涓�墠涓€閬撳熀鏈�笉浼氬瓨鍦ㄤ簤璁�殑浜虹墿绫绘瘮锛屼互鍙婂悗涓€閬撳彲鑳藉瓨鍦ㄥ洖绛斿樊寮傜殑浜虹墿绫绘瘮棰樼洰銆�
闄や簡涓婅堪鍋囨兂鐨勫�澶嶆潅浜虹墿鎬ф牸鐨勭被姣旓紝杩樻湁鐮旂┒娴嬭瘯浜嗗ぇ妯″瀷鍦ㄦ棤棰勮�鏉′欢涓嬪皢鎶借薄姒傚康锛堝�pull銆乫lee锛変笌绌洪棿绗﹀彿锛堜笂涓嬪乏鍙筹級杩涜�绫绘瘮鎺ㄧ悊鐨勮兘鍔涳紝缁撴灉鏄剧ず锛屽ぇ妯″瀷鍜屼汉绫荤殑鐩镐技鎬т笉绠楅珮[5]銆備笉杩囪€冭檻鍒拌繖椤圭爺绌跺己琛岃�姹傚皢鎶借薄姒傚康锛堢粰瀹氬崟璇嶏級鍜屾柟浣嶅�搴旂己灏戠幇瀹炴剰涔夛紝杩欓噷灏变笉璇︾粏璁鸿堪銆�
鈻峰浘10锛氬ぇ妯″瀷瀵规娊璞℃�蹇靛拰浜虹被绫绘瘮鐨勫噯纭�€ц瘎浼�.鍥炬簮锛歔5]
04 鎻愬崌澶фā鍨嬬被姣旇兘鍔涳紝杩樹换閲嶉亾杩�
鍩轰簬浠ヤ笂鐮旂┒鍙戠幇锛屾垜浠�ぇ鑷村彲浠ュ緱鍒颁竴涓�粨璁猴細澹扮О澶ц�瑷€妯″瀷宸插叿澶囦竴鑸�帹鐞嗚兘鍔涙垨璁镐负鏃惰繃鏃┿€�
灏界�鏃╂湡鐮旂┒涓�ぇ妯″瀷鍦ㄧ壒瀹氫换鍔′笂琛ㄧ幇鑹�ソ锛屼絾褰撴祴璇曢毦搴︽彁鍗囨椂锛屽畠浠�殑琛ㄧ幇灏变笉绋冲畾浜嗐€備竴涓�ā鍨嬪湪涓€缁勭壒瀹氫换鍔′笂琛ㄧ幇鑹�ソ锛屽苟涓嶆剰鍛崇潃瀹冨叿鏈夐瞾妫掓€с€備箣鍓嶆湁鐮旂┒琛ㄦ槑锛屽湪闈㈠�鏁板�搴旂敤棰樻椂锛屽彧鏄�洿鎹㈤�鐩�腑鐨勪汉鍚嶏紝澶фā鍨嬬殑瑙ｇ瓟鍑嗙‘搴﹂兘浼氭槑鏄句笅闄嶏紝鑰屽�鍔犳棤鍏崇殑鑳屾櫙璁鸿堪鏃讹紝妯″瀷鐨勬€ц兘涓嬮檷鍒欐洿鍔犳槑鏄綶6]銆�
杩欎竴鍙戠幇瀵逛簬鍦ㄦ暀鑲层€佹硶寰嬪拰鍖荤枟绛夊叧閿�喅绛栭�鍩熷簲鐢ㄤ汉宸ユ櫤鑳芥暡鍝嶄簡璀﹂挓锛屼汉宸ユ櫤鑳藉彲浠ユ槸涓€涓�己澶х殑宸ュ叿锛屼絾瀹冭繕涓嶈兘鍙栦唬浜虹被鐨勬€濊€冨拰鎺ㄧ悊銆備緥濡傦紝鍦ㄦ暀鑲查�鍩燂紝澶фā鍨嬬敓鎴愮殑姣斿柣纭�疄鑳戒负鏁欏�鎻愪緵甯�姪锛涚劧鑰岋紝濡傛灉缂轰箯涓撲笟浜哄＋鐨勫�鏍镐笌淇��锛岃繖浜涚被姣斿彲鑳藉瓨鍦ㄦ綔鍦ㄩ�闄┿€�
鍥犳�锛岀爺绌朵汉鍛橀渶瑕佸紑鍙戝拰瀹炴柦绋冲仴鎬ф祴璇曪紝浠ラ€傚簲闂��鎴栨儏鍐典腑缁嗗井鍙樺寲鐨勮兘鍔涖€傛柊鐨勭ǔ鍋ユ€ф祴璇曞簲鍖呮嫭涓€缁勫叕璁ょ殑鏍囧噯鍖栦换鍔★紝鐢ㄤ互璇勪及 AI 绯荤粺浠ュ強浜虹被濡備綍閫傚簲鏂版儏鍐点€傚湪瀹炶返涓�紝澶фā鍨嬪父浼氶亣鍒颁箣鍓嶅垎鏋愭暟鎹�腑鏈�浘閬囧埌鐨勬柊鎯呭喌鍜屾寫鎴橈紝鑰岀ǔ鍋ユ€ф祴璇曞皢涓虹敤鎴锋彁渚涜 閲忓ぇ鍨嬭�瑷€妯″瀷鍙�俊搴︾殑鏂瑰紡銆�
涓庢�鍚屾椂锛�24骞寸殑鏈哄櫒瀛︿範椤朵細ICLR鐨勪竴椤圭爺绌跺睍绀轰簡鍙︿竴涓�彂灞曟柟鍚戯細閫氳繃绫绘瘮鎺ㄧ悊妗嗘灦锛岃�澶фā鍨嬭嚜鍔ㄧ敓鎴愭柊鐨勮�鍒欐潵搴斿�鏈�煡鍦烘櫙[7]銆傝繖绉嶅熀浜庢彁绀鸿瘝宸ョ▼鐨勬柟娉曞湪澶氫釜娴嬭瘯鍩哄噯涓婇兘鍙栧緱浜嗘樉钁楁€ц兘鎻愬崌锛岃〃鏄庢彁鍗囧ぇ妯″瀷鐨勭被姣旇兘鍔涗笉浠呮槸璇勪及鍏剁ǔ鍋ユ€х殑閲嶈�缁村害锛屾洿鏄��寮烘ā鍨嬫硾鍖栬兘鍔涚殑鍏抽敭璺�緞銆傝繖涓ょ�鏂规硶鐩歌緟鐩告垚锛屽叡鍚屾帹鍔ㄧ潃澶фā鍨嬪悜鏇村彲闈犮€佹洿鏅鸿兘鐨勬柟鍚戝彂灞曘€�
灞曟湜鏈�潵锛屽ぇ妯″瀷绫绘瘮鎬濈淮鐨勭爺绌讹紝鎴栧彲浠庝腑鍥戒紶缁熶腑姹插彇鐏垫劅銆備腑鍥藉彜鍏告枃瀛︿腑鐨勫�鑱斾笌寰嬭瘲锛屾湰璐ㄤ笂灏辨槸涓€绉嶇簿濡欑殑绫绘瘮绯荤粺锛屽叾涓�暣鍚�潃涓ヨ皑鐨勫�搴旇�鍒欏拰涓板瘜鐨勮�涔夊叧鑱斻€傞€氳繃杩欎簺缁撴瀯鍖栫殑璇�█鏁版嵁闆嗗�澶фā鍨嬭繘琛屽井璋冿紝鍙�兘涓哄�寮哄叾绫绘瘮鎺ㄧ悊鑳藉姏寮€杈熸柊閫斿緞銆�
灏卞儚涓�枃鎸囦护寰�皟鏁版嵁闆� COIG-CQIA锛屼负浜嗘彁鍗囨ā鍨嬪湪缂栫▼鍙婃暟瀛﹂棶棰樹笂鐨勮〃鐜帮紝涔熸浘浣跨敤浜嗕腑鏂囦簰鑱旂綉绀惧尯鏁版嵁鈥滃急鏅哄惂鈥濈殑鏍囬�浣滀负璁�粌鎸囦护銆傝繖浜涙潵鑷�笉鍚岄�鍩熺殑瀹炶返琛ㄦ槑锛岀粨鏋勫寲鐨勭被姣旀€濈淮妯″紡锛屾棤璁烘槸浼犵粺鏂囧�杩樻槸鐜颁唬缃戠粶绀剧兢鏁版嵁闆嗭紝閮藉彲鑳芥垚涓烘彁鍗囦汉宸ユ櫤鑳借�鐭ヨ兘鍔涚殑閲嶈�宸ュ叿銆�
姣曠珶锛岀被姣旀€濈淮鐨勬湰璐ㄦ槸閫氱敤鐨勩€�
鍙傝€冩枃鐚�
[1] Taylor Webb, Keith J. Holyoak, and Hongjing Lu. Emergent analogical reasoning in large language models. Nature Human Behaviour, 7(9):1526鈥�1541, 2023.
[2] Lewis, Martha & Mitchell, Melanie. (2024). Evaluating the Robustness of Analogical Reasoning in Large Language Models. 10.48550/arXiv.2411.14215.
[3] Stevenson CE, Pafford A, van der Maas HLJ, Mitchell M. (2024). Can large language models generalize analogy solving like children can? arXiv.2411.02348v1.
[4] Opie艂ka GJ, Rosenbusch H, Vijverberg VP, Stevenson CE. Do large language models solve ARC visual analogies like people do? [Internet]. arXiv.org. 2024 May 13 [cited 2025 Apr 2]. Available from: https://arxiv.org/pdf/2403.09734v2
[5] Wicke, P., Hirlimann, L., & Cunha, J. M. (2024). Using Analogical Reasoning to Prompt LLMs for their Intuitions of Abstract Spatial Schemas. Retrieved from https://analogy-angle.github.io/assets/Wicke.pdf
[6] Mirzadeh S I, Alizadeh K, Shahrokhi H, Tuzel O, Bengio S, Farajtabar M. GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models. *arXiv preprint arXiv:2410.05229*. 2024.
[7] Yasunaga M, Chen X, Li Y, Pasupat P, Leskovec J, Liang P, Chi EH, Zhou D. Large language models as analogical reasoners. In *International Conference on Learning Representations (ICLR)* 2024.