
ããã³ããæè¡58çš®é¡ã®9å²ã¯ãç¡é§ãã ã£ã ââããã³ããç ç©¶4,247æ¬ã®ç®èãªçµè«
ããæ®µéçã«èããŠãã ããããšããäžæãå ããã ãã§ãLLMã®æšè«ç²ŸåºŠãåçã«åäžããããã®çºèŠä»¥éãããã³ããæè¡ã¯ççºçã«å¢æ®ããçŸåšã§ã¯58çš®é¡ãã®ææ³ãææ¡ãããŠããŸãã
ããããã4,247æ¬ã®ç ç©¶è«æãç¶²çŸ çã«èª¿æ»ãããThe Prompt Reportãã¯ããããã®å€§åãåºæ¬çãªåçã®çŒãçŽãã§ããããšã瀺ããŠããŸãã
ã20æéãããŠæåçã0%ãã53%ã«åŒãäžããå®äŸãå«ããæ¬çš¿ã§ã¯ããã³ãããšã³ãžãã¢ãªã³ã°ã®çæ³ãšçŸå®ããããŠæ¬åœã«ç¥ã£ãŠããã¹ãææ³ã«ã€ããŠæ€èšŒããŸãã
Spotifyã§ããããããé³å£°é ä¿¡ïŒããã¿ã®AIããã¯ã©ãžãªã
ãäžççµæžãã©ãŒã©ã ã®èª¿æ»ã§ã¯ã2030幎ã«å¿
èŠãªã¹ãã«ãã®ç¬¬5äœã«ã奜å¥å¿ãšç涯åŠç¿ããã©ã³ã¯ã€ã³ã奜å¥å¿ã¯è¶£å³ã§ã¯ãªãçµæžçå¿
é ã¹ãã«ã
æèãAIæä»£ã®æåŒ·ã¹ãã«ã奜å¥å¿åããã§è©³èª¬ããŠããŸãã
ãŸããã
ãLLMãæ¥åžžçã«äœ¿çšããŠããæ¹ãªãããããªçµéšãããã¯ãã§ããåã質åã§ããå°ãèšãåããå€ããã ãã§å šãç°ãªãåçãè¿ã£ãŠããããèŠçŽããŠãããã5ã€ã®èŠç¹ã«ãŸãšããŠãã®æ¹ãè¯ãçµæãåŸããããããããçµéšåã¯ç¢ºãã«ååšããŸãã
ãäžæ¹ã§ãã¢ã«ããã¢ã§ã¯ç¶æ³ãç°ãªããŸããç ç©¶è ãã¡ã¯ç«¶ãããã«æ°ããããã³ããææ³ãçºè¡šããããããã«ãREACTããREFLEXIONããSELF-DISCOVERããšãã£ãå°è±¡çãªååãã€ããŠããŸãã
ããããã4,247æ¬ã®è«æããå³éžããã1,565æ¬ã®è«æã粟æ»ããçµæããããã®ãæ°ææ³ãã®å€ãã¯ãåºæ¬çãª6ã€ã®ã¢ãããŒãã®ããªãšãŒã·ã§ã³ã«éããªãããšã倿ããŸããã
ããã®ä¹é¢ã¯äœãæå³ããã®ã§ãããããç ç©¶ã®çްååãé²ã¿ãããçµæãæ¬è³ªãèŠå€±ã£ãŠããã®ãããããšããå®åå®¶ãç¥ããªãç»æçãªææ³ãåãããŠããã®ãã
ããThe Prompt Reportãã®èè ãã¯ãããè峿·±ãå®éšãè¡ããŸãããã¡ã³ã¿ã«ãã«ã¹é¢é£ã®é£ããã¿ã¹ã¯ã«å¯Ÿããçç·Žã®ãšã³ãžãã¢ã20æéãããŠããã³ãããæ¹è¯ããéçšãèšé²ããã®ã§ãã
ã47åã®è©Šè¡é¯èª€ã®æ«ãæåçã¯0%ãã53%ãŸã§åäžããŸããããã®æ°åã瀺ãã®ã¯ãããã³ãããšã³ãžãã¢ãªã³ã°ã®å¯èœæ§ãšãåæã«ãã®éçã§ãã
ãæ¬çš¿ã§ã¯ããã®ãµãŒãã€ç ç©¶ãæããã«ããç¥èŠãåºã«ãããã³ãããšã³ãžãã¢ãªã³ã°ã®çã®äŸ¡å€ãåèããŸããæè¡ã®ä¹±ç«ããçŸç¶ã«ãããŠãå®åå®¶ãæ¬åœã«æŒãããã¹ããã€ã³ãã¯äœããç ç©¶ãšå®è·µã®æ©æž¡ããšãªãèŠç¹ãæäŸã§ããã°ãšèããŠããŸãã
èæ¯
ãLLMãšã¯Large Language Modelã®ç¥ã§ã倧éã®ããã¹ãããŒã¿ãåŠç¿ããŠäººéã®ããã«æç« ãçæã§ããAIã¢ãã«ã®ããšã§ããChatGPTã«ä»£è¡šãããLLMã¯ããã®æè»ãªå¿çèœåããããžãã¹ãæ¥åžžç掻ãŸã§å¹ åºã掻çšããå§ããŠããŸãã
ãããããã¢ãã«ãšããåããããéããŠãŒã¶ãŒã¯ããã¹ããç»åãªã©ã®ããã³ãããå ¥åããã¢ãã«ããã®å¿çãåŸãŸããããã³ããã¯åãªã質åæã®å Žåãããã°ããïœã«ã€ããŠ5ã€ç®æ¡æžãã§èª¬æããŠãã ãããã®ããã«åœ¢åŒãæäœãæå®ããæç€ºæã®å ŽåããããŸããã¢ãã«ã¯ãã®ããã³ãããèªã¿åããçµ±èšçæšè«ã«åºã¥ããŠæé©ãšèããå¿çãçæããŸãã
ããšããããLLMããåžžã«æãã çããåŸããããšã¯éããŸããããã«ã·ããŒã·ã§ã³ãšåŒã°ããçŸè±¡ã§ã¯ãã¢ãã«ãäºå®ç¡æ ¹ã®åçãèªä¿¡æºã ã«äœãåºãããšããããŸããäŸãã°å®åšããªãæŽå²äžã®äººç©ãããããªãæ°å€ããŒã¿ããã£ãšããããèªã£ãŠããŸãããšãããã®ã§ãã
ããŸããããã³ããã®æžãæ¹ã²ãšã€ã§ã¢ãã«ã®åºåã倧ããå€ãã£ãŠããŸãæåºŠã®é«ãã課é¡ã§ãããã衚çŸã§ã¯æ£çãè¿ããã®ã«ãèšãåããå°ãå€ããã ãã§çãã誀ãããšãã£ãã±ãŒã¹ãå ±åãããŠããŸãã
ãããã«ãã¢ãã«ã¯èšç·ŽããŒã¿ç±æ¥ã®ãã€ã¢ã¹ãã¹ãã¬ãªã¿ã€ããå«ãåºåãããŠããŸãããšããããæ éãªæ±ããå¿ èŠã§ãããã®ããã«ãLLMãå®äžçã§æå¹æŽ»çšããã«ã¯ãã¢ãã«ã®ã¢ã©ã€ã³ã¡ã³ããäžå¯æ¬ ã§ãã
ãããã³ãããšã³ãžãã¢ãªã³ã°ã¯ãã¢ãã«ã®å éšãã©ã¡ãŒã¿ãå€ããããšãªããå ¥åããæç€ºã工倫ããããšã§ããããåé¡ã®è»œæžãã¢ãã«ã®æ§èœåäžãå³ãããšããã¢ãããŒããªã®ã§ãã
ãæ¢åã®ç ç©¶ã§ã¯ãLLMã«ã¿ã¹ã¯ãè§£ãããããã®æ§ã ãªããã³ãããã¯ããã¯ãèæ¡ãããŠããŸãããäŸãã°ãOpenAIã2020幎ã«çºè¡šããGPT-3è«æã§ã¯ãã¢ãã«ã«æ°äŸã®å ¥åãšåºåãã¢ã瀺ãã ãã§æ°ããå ¥åã«å¯Ÿããé©åãªåºåãåŸãããããšãå ±åãããŸããã
ãããã¯Few-Shot Learningã®äžçš®ã§ããããäŸãæèã«å«ããŠåŠç¿ãããããšããããã³ããææ³ãšããŠå€§ããªæ³šç®ãéããŸããããã®ããã«ãæç€ºçãªè¿œå åŠç¿ãããã³ã³ããã¹ãïŒæèïŒå ã§ã¢ãã«ã«åŠç¿ãããããæ¹ãäžè¬ã«ã€ã³ã³ã³ããã¹ãã©ãŒãã³ã°ïŒICLïŒãšåŒã³ãŸãã
ãäžæ¹ã§ãäŸç€ºãäžåãããŒãããæšè«ãããææ³ïŒãŒãã·ã§ããæšè«ïŒãç ç©¶ãããŠããŸããè¿å¹Žã®ç ç©¶ã倧åèšèªã¢ãã«ã¯ãŒãã·ã§ããæšè«è ã§ãããã§ã¯ãã¢ãã«ã«å¯ŸããŠãLet's think step by stepããšãã£ãäžæãä»ãå ããã ãã§ãæšè«èœåãé£èºçã«é«ãŸãããšã瀺ãããŸããã
ãããã¯åŸè¿°ããChain-of-Thoughtãšããææ³ã®ç«¯ç·ã§ã远å ã®äŸãäžããªããŠãã¢ãã«èªèº«ã«äžæ©ãã€èããããããã³ããèšèšã§ãã
ããã®ããã«åå¥ã®ç ç©¶ã§ææ¡ããããã¯ããã¯ã¯æ°å€ãååšããŸãããå æ¬çã«æŽçãããŠããªãã£ããããåå¿è ã®ã¿ãªããç ç©¶è ã«ãšã£ãŠãå šè²ãææ¡ãã¥ããç¶æ³ã§ããã
ãä»å玹ä»ãããThe Prompt Reportãã®è²¢ç®ã¯ãããã³ããææ³ã®äœç³»åã«ãããŸããåè«æã§ã¯ãé¢é£æç®ãç¶²çŸ çã«åéã»åæããããã³ããã«é¢ããçšèªãææ³ã®çµ±äžçãªæ çµã¿ãæ§ç¯ããŸããã
ãå ·äœçã«ã¯ããããã³ãããã«é¢ãã33ã®å°éçšèªãå®çŸ©ãããšãšãã«ã58çš®é¡ã®ããã¹ãããŒã¹ã®ããã³ããææ³ã6ã€ã®äž»èŠã«ããŽãªã«åé¡ããŠäœç³»åé¡ãææ¡ããŠããŸããããã«ããã«ãã¢ãŒãã«ïŒç»åã»é³å£°çïŒåéã®ããã³ããææ³ã40çš®é¡åéããç·åçãªç¥èŠãšããŠãŸãšããããŸããã
ããã®èª¿æ»ã¯çŸæç¹ã§æãå æ¬çãªããã³ãããšã³ãžãã¢ãªã³ã°ã®ãµãŒãã€ã§ãããä¹±ç«ããŠããæŠå¿µãæŽçããããšã§ä»åŸã®ç ç©¶ãšå®è·µã®çŸ éç€ã«ãªãããšãæåŸ ãããŠããŸãã
æ¹æ³è«
ããThe Prompt Reportãã®èè ãã¯ããŸãé¢é£ããç ç©¶è«æãç¶²çŸ çã«åéãããšããããçæããŸããã
ãåè«æã§ã¯ãè«æããŒã¿ããŒã¹ïŒarXivãSemantic ScholarãACL Anthologyãªã©ïŒã«å¯Ÿããpromptããpromptingããprompt engineeringãçã®é¢é£ããŒã¯ãŒããçšããŠæ€çŽ¢ãè¡ããåŸãããè«æçŸ€ã«äœç³»çãªã¬ãã¥ãŒãé©çšããŠããŸãã
ããã®éãå»åŠåéã§æç®ã¬ãã¥ãŒã«çšããããPRISMAãšãããããã³ã«ïŒæç®éžå¥ã®ãããŒãæç€ºããææ³ïŒã«æ²¿ã£ãŠã次ã®ãããªåºæºã§åæšéžæãè¡ãããŸããã
ããŒãããã³ããïŒé¢æ£çããã³ããïŒã®ã¿ã察象ãšããåŸé éäžã«ããåŠç¿ã䌎ãããã³ããïŒãããããœããããã³ããããã¡ã€ã³ãã¥ãŒãã³ã°ææ³ïŒã¯é€å€ã
ãã¬ãã£ãã¯ã¹åããã³ããïŒæç« ã®åã«æç€ºãäžãã圢åŒïŒã«éå®ãããã¹ã¯ç©Žåã圢åŒã®ã¯ããŒãºåããã³ããã¯å¯Ÿè±¡å€ãããã¯GPT-3以éã®äž»æµã¢ãã«ããã¬ãã£ãã¯ã¹åã§ããããã§ãã
ã¿ã¹ã¯ã«äŸåããªãäžè¬çãªææ³ã«ãã©ãŒã«ã¹ããç¹å®ã®ã¿ã¹ã¯å°çšã«å·¥å€«ãããããã³ããïŒäŸãã°ç¹å®åéã ãã§æå¹ãªãã®ïŒã¯ç¯å²ããå€ãã
ããã®ç¶²çŸ çæ€çŽ¢ã®çµæãéè€é€å»åŸã®çŽ4,247æ¬ã®è«æèšé²ãéããããŸããããŸã人æã§ã¿ã€ãã«ãæŠèŠã確èªãããããã³ããããšããåèªãå«ãŸãªããã®ãªã©ç¡é¢ä¿ãªè«æãé€å€ããŸãã
ãå ããŠäžéšã¯AIã¢ãã«ã«ããä»åããæŽ»çšããæçµçã«1,565æ¬ã®è«æããããã³ããæè¡ãã«é¢ããæçšãªæç®ãšããŠéžå®ããŸããããã®å€§èŠæš¡ãªæç®ã»ãããåºã«ãèè ãã¯ããã³ããæè¡ã®åé¡ãšå®çŸ©ä»ããè¡ã£ãŠããŸãã
ãåé¡ã«ããã£ãŠèè ãã¯ãŸãã倿°ã®è«æã«ç»å Žããããã³ããææ³ãã§ããã ãç¶²çŸ çã«ãªã¹ãã¢ããããŸããããã®æ°ã¯å®ã«58çš®é¡ã«ãäžããŸãããããããäºãã«é¢é£ãããã®å士ã§ã°ã«ãŒãã³ã°ãã6ã€ã®äž»èŠã«ããŽãªã«æŽçããã®ãæ¬è«æã®äžå¿çãªè²¢ç®ã§ãã

ãå³1.1ã«ããã°ãããã³ããææ³ã¯å€§ãããããã¹ãããŒã¹ã®åºæ¬ææ³ããå€èšèªãžã®å¿çšãããã«ãã¢ãŒãã«ãžã®å¿çšãããšãŒãžã§ã³ãïŒå€éšããŒã«æŽ»çšïŒãžã®çºå±ããšãã£ãåºåã§æããããŠããŸãã
ãäžæ žã«ããã®ã¯ããã¹ãããŒã¹ã®ããã³ããææ³çŸ€ã§ããããä»ã®çºå±åœ¢ïŒå€èšèªã»ãã«ãã¢ãŒãã«ã»ãšãŒãžã§ã³ãçïŒãæ¯ããåºç€ãšãªã£ãŠããŸãã以äžã§ã¯ãç¹ã«ããã¹ãããŒã¹ã®ããã³ããææ³58çš®é¡ã«ã€ããŠããã®6åé¡ã詳ããèŠãŠãããŸãããã

ãå³2.2ã¯ãããã¹ãããŒã¹ã®ããã³ãããšã³ãžãã¢ãªã³ã°ææ³ãã¹ãŠïŒ58çš®é¡ïŒãäœç³»åããæš¹åœ¢å³ã§ãããã®å³ã§ã¯6ã€ã®è²ä»ãããŒããäž»èŠã«ããŽãªã衚ããããããã®äžã«å ·äœçãªããã³ããææ³ãåæãããŠããŸãã6ã€ã®ã«ããŽãªã¯ä»¥äžã®éãã§ãã
In-Context LearningïŒICLïŒÂ
ãã¢ãã«ã«å¯ŸããäŸç€ºã«ããåŠç¿ããä¿ãããã³ããææ³ã®ç·ç§°ã§ãããŠãŒã¶ãŒãããã€ãã®å ¥åäŸãšãã®åçäŸãããã³ããå ã«å«ããããšã§ãã¢ãã«ã¯æ°ããªå ¥åã«å¯ŸããŠæèã«æ²¿ã£ãåºåãæšæž¬ããŸããICLã«ã¯ããã«çްããªå·¥å€«ããããäŸãã°
ãšã°ã¶ã³ãã©ãŒéžæ
ãã©ã®äŸãäœä»¶æç€ºãããïŒéèŠåºŠã®é«ãäŸãéžã¶ãé¡äŒŒåºŠã®é«ãäŸãéžã¶ãªã©ïŒãKNNïŒæè¿åæ³ïŒã§é¡äŒŒã¿ã¹ã¯äŸãéžã¶ç ç©¶ããVote-Kã®ããã«è€æ°äŸã®æç¥šã§ä»£è¡šäŸã決ããææ³ããããŸãã
ãšã°ã¶ã³ãã©ãŒé åº
ãæç€ºããäŸã®äžŠã¹æ¹ãKumar & Talukdar (2021) ã®ç ç©¶ã§ã¯äŸã®é åºãå ¥ãæ¿ããã ãã§ãæ§èœãå€åããããšã瀺ãããŠããŸãã
æç€ºæãã³ãã¬ãŒã
ãäŸç€ºãšå ±ã«äžããã¿ã¹ã¯èª¬æã®æžãæ¹ãäŸãã°ã次ã®å ¥åæã翻蚳ããïŒââããšãã£ãæç€ºæèªäœã«ã工倫ãèããããŸãã
ãŒãã·ã§ããããã³ããã£ã³ã°
ããäŸç€ºãªããã§ã¢ãã«ã«æšè«ãããåçš®ã®ææ³ã§ããICLãäž»ã«Few-Shotãæãã®ã«å¯ŸãããŒãã·ã§ããã§ã¯ãŠãŒã¶ãŒã¯è§£ãããåé¡ãã®ãã®ãšãå¿ èŠã«å¿ããŠä»å çãªæç€ºã ããäžããŸãã
ãä»å çãªæç€ºã®å žåäŸããå è¿°ãããLetâs think step by step.ãã®ãããªæèãä¿ããã¬ãŒãºã§ããä»ã«ããã¢ãã«ã«ç¹å®ã®ããŒã«ïŒåœ¹å²ïŒãæŒããããæç€ºãæå¹ã§ãã
ãäŸãã°ãããªãã¯æèœãªæ è¡ã¬ã€ãã§ããïœïœã«ã€ããŠæããŠãã ãããããšãã£ãããŒã«ããã³ããã£ã³ã°ã§ã¯ãã¢ãã«ã«åœ¹å²ãäžããããšã§å£èª¿ãç¥èã®æ çµã¿ã調æŽã§ããŸãã
ãå®éãããç ç©¶ã§ã¯ãäžæµã®æ°åŠè ã«ãªããã£ãŠçããŠãã ããããšããæç€ºãå ¥ããã ãã§ãæ°åŠåé¡ã®æ£ççãäžãããšããå ±åããããŸãã
ããŸãã¹ã¿ã€ã«ããã³ããã£ã³ã°ã¯ãåºåæäœããé¢è¥¿åŒã§ããããžãã¹ã¡ãŒã«é¢šã«ããšæå®ãããã®ãææ ããã³ããã£ã³ã°ã¯ãããã¯ç§ã®ãã£ãªã¢ã«é¢ãã倧äºãªããšã§ããçã®å¿ççãã¬ãŒãºãå ¥ãã¢ãã«ã®å¿çåŸåãå€ãã詊ã¿ã§ãã
ããããã¯å³å¯ã«ã¯åºåå 容ã®åäžã®ã¿ãªããã衚çŸã¹ã¿ã€ã«ã®å€æŽã§ããããŠãŒã¶ãŒæå³ã«æ²¿ã£ãåºåãåŸãããã®éèŠãªãã¯ããã¯ã§ãã
ããŒãã·ã§ããææ³ã«ã¯ä»ã«ãã質åã®èšãæããä¿ãRephrase-and-Respond (RaR)ããäžåºŠè³ªåãç¹°ãè¿ãããŠæ³šæãåèµ·ããåèªããã³ãã (RE2)ãå¿ èŠã«å¿ããŠã¢ãã«èªèº«ã«è¿œå ã®çåãçºã解決ãããSelf-Askãªã©ãå€åœ©ãªããªãšãŒã·ã§ã³ããããŸãããããã¯äŸç€ºãªãã§ãåºå粟床ãäžè²«æ§ãé«ãã工倫ãšããŠææ¡ãããŠããŸãã
Chain-of-ThoughtïŒ CoTïŒÂ
ãã¢ãã«ã«æšè«ããã»ã¹ãæç« ã§é次çã«æžãããææ³ã§ããéåžžãLLMã¯è³ªåã«çŽæ¥çããåºåããŸãããCoTããã³ããã§ã¯ãäžæ©ãã€èããŠã¿ãŸããããããŸãã¯åé¡ã现ããåæããŸãããªã©ã®ãã¬ãŒãºãããã³ããæ«å°Ÿã«è¿œå ãããããããã¯Few-Shotã§è§£çã«è³ãéçšãå«ãäŸã瀺ãããããŸãã
ããã®ããã«ããããšã§ãã¢ãã«ã¯ãŸãèªãäžéçãªèå¯ãåæããæåŸã«æçµåçãåºããšããæ®µåããèžãããã«ãªããŸããé©ãã¹ãããšã«ããã®ææ³ã¯æ°åŠãåžžèæšè«ã®ã¿ã¹ã¯ã§ã¢ãã«ã®æ£ççãå€§å¹ ã«åäžãããããšãåãããŸããã
ãç¹ã«Few-Shot CoTïŒãã§ãŒã³ä»ãã®æ°äŸãäžããïŒã¯å¹æãé«ããWeiã(2022)ã®è«æã§ã¯GPT-3ã«ç®æ°æç« é¡ãè§£ãããéãéåžžã®ããã³ããã§ã¯äžæ£ç¢ºã ã£ãåé¡ããæèéçšã€ãã®äŸã1件瀺ãã ãã§æ£è§£ã§ããããã«ãªã£ããšå ±åãããŠããŸãã
ãçŸåšãCoTã¯ããããæšè«ç³»ã¿ã¹ã¯ã®ç²ŸåºŠåäžã«å®çªã®ãã¯ããã¯ãšãªã£ãŠãããå€ãã®äºçš®ãç»å ŽããŠããŸããäŸãã°ãã¢ãã«ã«ãããïœã ã£ããïŒããšéç®ãããéæ¹åã®æšè«ãè€æ°ã®é¡äŒŒåé¡ãäžåºŠã«è§£ãããŠå ±éããåçãæ¢ã察æ¯çæšè«ãªã©ã§ãã
ãCoTã¯ã¢ãã«ã®äžæèšæ¶ãäžæã«äœ¿ãããããšã§ãè€éãªåé¡ãè§£ãéã®ã±ã¢ã¬ã¹ãã¹ãæšè«ã®é£èºãæžãã广ããããšèšãããŸãã
åé¡åè§£ããã³ããÂ
ãè€éãªèª²é¡ãå°ããªéšåã«åå²ããŠè§£ãããã¢ãããŒãã§ãã人éãé£åã«çŽé¢ããéããããªãçããåºãããšããããŸãåé¡ãèŠçŽ ã«åããŠååã«è§£æ±ºããŠããããšããããŸãã
ãåæ§ã«ã¢ãã«ã«ã段éç解決ãä¿ãã®ããã®ã«ããŽãªã§ããå ·äœäŸãšããŠãLeast-to-Most PromptingãReverseCoTïŒäžæŠçããäºæ³ããŠããçç±ä»ãããççŸããã§ãã¯ïŒãTree-of-ThoughtïŒè§£çãæšæ§é ç¶ã«å±éããŠæ¢çŽ¢ïŒãªã©ã倿§ãªçºæ³ã®ææ³ãææ¡ãããŠããŸãã
ãå³2.2ã«ããDecomposition 2.2.3ãã®äžã«åæãããŠããããã«ãããã°ã©ã ã®ããã«æé èšç»ããPlan-and-Solveããè§£ç骚åã ãæžãããŠè©³çްãåããŠããSkeleton-of-Thoughtãšãã£ãã¢ã€ãã¢ããããŸãã
ããããã«å ±éããçãã¯ãäžåºŠã«å šãŠãèããããªãããšã§ã¢ãã«ã®èªç¥è² è·ãäžãã屿çã«ã¯ç°¡åãªåé¡ãé çªã«è§£ãããšã§æçµçãªé£å解決ã«ã€ãªããããšã«ãããŸãããšããããã«ãã¹ãããæšè«ïŒäŸïŒé·æèªè§£ãè€éãªèšç®åé¡ïŒã«æå¹ãšãããŠããŸãã
ã¢ã³ãµã³ãã«ã»èªå·±äžè²«æ§Â
ãÂ è€æ°ã®è©Šè¡çµæãçµã¿åãããŠæé©è§£ãåŸãææ³çŸ€ã§ããã¢ãã«ã«äžåºŠã ã質åããŠåºåãåŸãã®ã§ã¯ãªããäœåºŠãçæãããŠãã®äžããçããéžã¶ãšããçºæ³ã§ãã
ãäŸãã°ããã質åã«å¯Ÿãã¢ãã«ãã5éãã®çããçæããããã®äžã§å€æ°æŽŸãšãªã£ãŠããçããæçµçãªåºåãšããããšãã£ãææ³ã代衚äŸã§ãã
ããããèªå·±äžè²«æ§ãšåŒã³ãŸããWangã(2022)ã®ç ç©¶ã«ããã°ãç®æ°ãåžžèæšè«ã§ãã®èªå·±äžè²«æ§ãåããšç²ŸåºŠãåäžãããšã®ããšã§ããè€æ°åçã®äžãã倿°æ±ºã§æ±ºããã®ã¯åãããããææ³ã§ãããä»ã«ãä¿¡é ŒåºŠã¹ã³ã¢ãååçã«ç®åºããŠæè¯ã®ãã®ãéžã¶(DiVeRSeããåçåå£«ã®æ å ±ãçµ±åããŠå床ã¢ãã«ã«ãã¹ããªçããèããããå埩ãªãã¡ã€ã³åã®ææ³ããããŸãã
ãåŸè ã®äžäŸãUniversal Self-Consistencyã§ãããã¯äžæŠå šåºåãéããŠãã远å ã®ããã³ããã§ããããã®çãããæãäžè²«æ§ã®ãããã®ã¯ã©ããããã¢ãã«èªèº«ã«éžã°ããæ¹æ³ã§ãããããã®ææ³ããã¢ãã«åºåã®äžå®å®ãã倿°åã®è©Šè¡ã§å¹³ååã»å®å®åããã广ããããŸãã
ã»ã«ãã¯ãªãã£ãŒã¯
ã ã¢ãã«ã«èªåã®çããèŠçŽããã誀ããèšæ£ãããææ³ã§ããããã°ã¢ãã«ã«ãæ¬åœã«ããã§åã£ãŠããïŒããšåãçŽãããã³ãããçµã¿èŸŒã圢ã«ãªããŸãã
ãå ·äœäŸãšããŠãåçåŸã«ããã®åçã«ééããããã°ææããŠãã ããããšè¿œå 質åããSelf-Verificationãè§£çã«å¯Ÿããæ¹è©ãä¿ãSelf-Criticismãæ®µéçã«åçãæŽç·ŽããSelf-RefineãSelf-Calibrationãªã©ãæããããŸãã
ãããããææ³ã§ã¯ãäžåºŠç®ã®åçãã¢ãã«èªèº«ã«ãã§ãã¯ããããããããã¯å¥ã®èгç¹ããååçããããããããšã§ãããæ£ç¢ºã§äžè²«æ§ã®ããçãã«è¿ã¥ããŸãã
ãäŸãã°Self-Refineã§ã¯ãå ã®åçãèžãŸããŠãããã«è¯ãçããçæããŠãã ããããšã¢ãã«ã«äŸé Œããèªå·±æ¹åãå³ããŸãããããã»ã«ãã¯ãªãã£ãŒã¯ç³»ã®æè¡ã¯ãã¢ãã«ã®ãã«ã·ããŒã·ã§ã³ãæžãã广ãæåŸ ãããŠããŸãã人éãã¯ãã¹ãã§ãã¯ãã代ããã«ãã¢ãã«èªèº«ã«ããäžäººã®æ¹è©å®¶åœ¹ãæ ãããçºæ³ãšãããã§ãããã
ã以äžã®6ã«ããŽãªããããã¹ãããŒã¹ã®ããã³ãããšã³ãžãã¢ãªã³ã°ææ³å šäœãå æ¬ããæ çµã¿ã§ããå³2.2ã«ç€ºãããããã«ãåã«ããŽãªå ã«ã¯ããã«çްããªæŽŸçææ³ãååšããŸãããæ¬èšäºã§ã¯ãã®ä»£è¡šäŸã®ã¿ãæŠèª¬ããŸããã
ãããã³ãããšã³ãžãã¢ãªã³ã°ã¯ãã®ããã«å€åœ©ãªã¢ã€ãã¢ã®éåäœã§ãããåã«ãäžã€ã®ãã¯ããã¯ããæãã®ã§ã¯ãªããLLMãšå¯Ÿè©±ãçµæãæé©åããããã®ç·åçã¢ãããŒããšèšããŸãã
ããªããæ¬è«æã§ã¯ããããã¯ããã¯ã®å®çŸ©ã ãã§ãªããå®éã«ã©ã®ãããåææ³ãç ç©¶ã§äœ¿ãããŠãããã®åæãè¡ãããŠããŸããäŸãã°Few-ShotãCoTãšãã£ãææ³ã¯åŒçšæ°ïŒä»è«æã§åç §ãããé »åºŠïŒãçªåºããŠé«ããããã³ããç ç©¶ã®äž»æµã§ããããšã確èªãããŠããŸãã
ãäžæ¹ã§æ°èŠææ¡ã®ææ³ã次ã ç»å ŽããŠãããããã³ããææ³ã®çš®é¡èªäœã幎ã å¢å ããŠããããšãããŒã¿ãã瀺åãããŠããŸããããã³ãããšã³ãžãã¢ãªã³ã°ã¯éåžžã«åãã®éãåéã§ããããã®taxonomyãå°æ¥ã¢ããããŒããå¿ èŠã«ãªãå¯èœæ§ããããŸãããçŸæç¹ã§ã¯äžèš6åé¡ãåºç€çãªæŽçãšããŠæçšã§ãããã
ãæåŸã«ãããã³ããèšèšã«é¢ãããã¹ããã©ã¯ãã£ã¹ãã¬ã€ãã©ã€ã³ãæ¬è«æã§ã¯æç€ºãããŠããŸããäŸãã°ãFew-Shotããã³ãããäœãéã«èæ ®ãã¹ããã€ã³ããšããŠããäŸç€ºæ°ã¯å€ãã®æ¹ãæ§èœåäžããããããã¢ãã«ãµã€ãºã«äŸåããããäŸã¯å€æ§ãªãã¿ãŒã³ãå«ããã©ã³ã¹ããããæç€ºæã¯ç°¡æœãã€å ·äœçã«ããªã©ãæããããŸãã
ããããã®ã«ããŽãªã¯ãåãªããã©ãããªãªã¹ãã§ã¯ãããŸããããããã¯ãLLMã«å¯ŸããŠæäŸãããèªç¥çè¶³å Žãã®éå±€æ§é ã圢æããŠããŸãã
ã¬ãã«1ã¯ãæãåºæ¬çãªè¶³å Žã§ããã¢ãã«ã«å®æããææ¬ïŒäŸïŒãèŠããããšã§ããã¿ãŒã³ãèªèãããŸããããã¯ãçåŸã«è§£ãçµãã£ãåé¡éãèŠããŠããã®ããã«è§£ããªããããšç€ºããããªãã®ã§ãã
ã¬ãã«2ã¯ãäžæ®µäžã®è¶³å Žã§ããåã«ææ¬ãèŠããã ãã§ãªããããªããããªãã®ããéäžåŒãæžããªããããšæç€ºããŸããããã«ãããã¢ãã«ã¯æèã®ããã»ã¹ãèšèªåããããšãåŠã³ãŸãã
ã¬ãã«3ã¯ãããã«é«åºŠãªè¶³å Žã§ããåé¡è§£æ±ºã®ãæ¹æ³è«ããã®ãã®ãæããŸãããåé¡ãåè§£ããªããããè€æ°ã®å¯èœæ§ã詊ããªããããšãã£ãæŠç¥çãªæèãã¬ãŒã ã¯ãŒã¯ãæäŸããã®ã§ãã
ã¬ãã«4ã¯ãæé«æ¬¡ã®è¶³å Žã§ãããã¡ã¿èªç¥ïŒèªå·±ã®èªç¥æŽ»åã客芳çã«æããèœåïŒãä¿ããŸãããèªåã®çããèŠçŽããééããããã°ä¿®æ£ããªããããšæç€ºããããšã§ãå埩çãªèªå·±æ¹åã®ã«ãŒããæ§ç¯ããŸãã
ã广çãªããã³ãããšã³ãžãã¢ãªã³ã°ãšã¯ã解決ãããã¿ã¹ã¯ã®è€éãã«å¿ããŠããã®éå±€ã®äžããé©åãªã¬ãã«ã®ãèªç¥çè¶³å ŽããéžæããæäŸããè¡çºã§ãããšèšããã§ããããåçŽãªã¿ã¹ã¯ã¯ICLã§ååãããããŸããããæªç¥ã®è€éãªåé¡ã«æãã«ã¯ãToTã®æŠç¥çæ¢çŽ¢ãSelf-Refineã®ååŸ©çæ¹åãäžå¯æ¬ ãšãªããŸãã
ææ¡ææ³è©³çŽ°
In-Context Learning (ICL)
ãICLã¯ãããã³ãããšã³ãžãã¢ãªã³ã°ã®æãåºæ¬çãªãã©ãã€ã ã§ãããã®æ žå¿ã¯ãããã³ããå ã«å ·äœçãªäŸïŒè«æã§ã¯exemplarãshotãšåŒã°ããïŒãæç€ºãå«ããããšã§ãLLMã«ã¿ã¹ã¯ã®ããæ¹ããæèã®äžã§åŠç¿ããããããšã«ãããŸã ãããã¯ãã¢ãã«ã®éã¿ã倿Žãããã¡ã€ã³ãã¥ãŒãã³ã°ãšã¯ç°ãªããæšè«æã«åçã«ã¢ãã«ã®æ¯ãèããå¶åŸ¡ããææ³ã§ãã
Few-Shot Prompting
ãããã¯ãICLã®æã代衚çãªåœ¢åŒã§ãããã€ãã®å ¥åãšæåŸ ãããåºåã®ãã¢ãããã³ããã«å«ããææ³ã§ã ãäŸãã°ãææ åæã¿ã¹ã¯ã§ããã°ã以äžã®ãããªããã³ããã«ãªããŸãã
æç« : ãã®æ ç»ã¯æé«ã ã£ãïŒ
ææ
: ããžãã£ã
æç« : å
šãé¢çœããªãã£ãã
ææ
: ãã¬ãã£ã
æç« : ä»ãŸã§ã§äžçªæåããäœåã§ãã
ææ
:ãLLMã¯ãã®æèãèªã¿åããæåŸã®æç« ã®ææ ããããžãã£ããã§ãããšæšè«ããŸãããã®Few-Shotããã³ãããèšèšããéã«ã¯ãããã€ãã®éèŠãªèŠçŽ ãããã©ãŒãã³ã¹ã«å€§ãã圱é¿ããŸã ã
äŸã®é
ãäžè¬çã«ãäŸã®æ°ãå¢ããã»ã©æ§èœã¯åäžããŸãããäžå®æ°ãè¶ ãããšå¹æãé æã¡ã«ãªãããšããããŸã ã
äŸã®é åº
ãé©ãã¹ãããšã«ãäŸãæç€ºããé åºãå€ããã ãã§ãæ£è§£çãåçã«å€åããããšããããŸããã©ã³ãã ãªé åºãæåã®å Žåãããã°ãç¹å®ã®é åºãè¯ãçµæãçãããšããããŸã ã
ã©ãã«ã®ååž
ãäŸã«å«ãŸããåã©ãã«ïŒäŸïŒããžãã£ãããã¬ãã£ãïŒã®æ¯çãåã£ãŠãããšãã¢ãã«ããã®æ¯çã«åŒãããããŠåã£ãäºæž¬ãããåŸåããããŸãÂ
ã©ãã«ã®å質
ãã©ãã«ãæ£ç¢ºã§ããæ¹ãè¯ãã®ã¯åœç¶ã§ãããäžéšã®ç ç©¶ã§ã¯ãã©ãã«ãééã£ãŠããŠããã¿ã¹ã¯ã®ãã©ãŒãããã瀺ãã ãã§æ§èœãåäžããå Žåãããããšã瀺åãããŠããŸãããã ããããã¯ã¢ãã«ã®èŠæš¡ãã¿ã¹ã¯ã«äŸåããŸã ã
ãã©ãŒããã
ããæç« : [å ¥å], ææ : [åºå]ãã®ãããªãã©ãŒãããèªäœãæ§èœã«åœ±é¿ããŸããã¢ãã«ãäºååŠç¿ããŒã¿ã§é »ç¹ã«ç®ã«ããã§ãããèªç¶ãªãã©ãŒããããéžã¶ããšãæå¹ãšãããŠããŸã ã
é¡äŒŒæ§
ãäºæž¬ããããã¹ããµã³ãã«ãšé¡äŒŒããäŸãéžã¶ããšãäžè¬çã«æ§èœåäžã«ç¹ãããŸã ã
ããããã®èŠçŽ ãæé©åããããã«ãK-Nearest Neighbor (KNN) ã®ãããªææ³ãçšããŠãã¹ããµã³ãã«ã«æå³çã«è¿ãäŸãåçã«éžæããã ã
ãSelf-Generated ICL (SG-ICL) ã®ããã«LLMèªèº«ã«ã¿ã¹ã¯ã®äŸãçæããããããé«åºŠãªæè¡ãç ç©¶ãããŠããŸã ã
Zero-Shot Prompting
ãããã¯ãå ·äœçãªäŸãäžå瀺ãããæç€ºã®ã¿ã§ã¿ã¹ã¯ãå®è¡ãããææ³ã§ã ãäŸãã°ããæ¬¡ã®æç« ãããžãã£ãããã¬ãã£ãã«åé¡ããŠãã ããïŒ[æç« ]ããšãã£ã圢ã§ãããã®åçŽãªã¢ãããŒãã§ããããã€ãã®å·¥å€«ã§æ§èœãé«ããããšãã§ããŸãã
圹å²ããã³ãã
ãLLMã«ç¹å®ã®åœ¹å²ããã«ãœããäžããææ³ã§ãããããªãã¯åªç§ãªããŒã±ãã£ã³ã°æ åœè ã§ãããã®è£œåã®ãã£ããã³ããŒãèããŠãã ãããã®ããã«æç€ºããããšã§ããã®åœ¹å²ã«ãµããããæäœãèŠç¹ã§ã®åºåãæåŸ ã§ããŸã ã
ã¹ã¿ã€ã«ããã³ãã
ããç°¡æœã«ããåŠè¡çã«ããæ ç±çã«ããšãã£ããåºåã®æäœãçŽæ¥æç€ºããææ³ã§ã ã
ææ ããã³ãã
ããããã¯ç§ã®ãã£ãªã¢ã«ãšã£ãŠéåžžã«éèŠã§ãããšãã£ãã人éçãªææ ãéèŠæ§ã蚎ããããæèšãä»ãå ããããšã§ãã¢ãã«ã®æ§èœãåäžãããšããè峿·±ãå ±åããããŸã ã
RaR
ãLLMã«åçãçæãããåã«ãã質åãèšãæããŠãå±éããŠããåçããŠãã ããããšæç€ºããææ³ã§ããããã«ãããã¢ãã«ã¯è³ªåã®æå³ãããæ·±ãçè§£ãã質ã®é«ãåçãçæããåŸåããããŸã ã
Thought Generation
ããã®ã«ããŽãªã®æè¡ã¯ãLLMã«åã«çããåºåãããã®ã§ã¯ãªãããã®çãã«è³ããŸã§ã®ãæèã®ããã»ã¹ããèšèªåãããããšãç®çãšããŠããŸã ãããã«ãããæšè«ã®æ£ç¢ºæ§ãåäžãããã«ã·ããŒã·ã§ã³ãæå¶ããã广ããããŸãã
Chain-of-Thought (CoT)
ãåè¿°ã®éãããã®ã«ããŽãªã®ä»£è¡šæ ŒãCoTã§ããCoTã«ã¯å€§ããåããŠäºã€ã®ããªãšãŒã·ã§ã³ããããŸãã
Zero-Shot-CoT
ãäŸã瀺ããã«ããæ®µéçã«èããŠãã ããããšãããããªãéæ³ã®èšèããããã³ããã®æ«å°Ÿã«ä»ãå ããã ãã§ãLLMãèªçºçã«æšè«éçšãèšè¿°ãå§ãããšãããéåžžã«ã·ã³ãã«ãã€åŒ·åãªææ³ã§ã ã
Few-Shot-CoT
ã解決ãããåé¡ã®äŸãããã€ã瀺ãéã«ããã®çãã ãã§ãªããçãã«è³ããŸã§ã®è©³çŽ°ãªæèããã»ã¹ãäžç·ã«æç€ºããææ³ã§ããããã«ãããã¢ãã«ã¯ã©ã®ãããªæèããã»ã¹ãæãŸããã®ããããå ·äœçã«åŠç¿ã§ããŸã ã
CoTã«ã¯ããã«å€ãã®æŽŸçæè¡ãååšããŸãã
Step-Back Prompting
ãå ·äœçãªåé¡ã«åãçµãåã«ãããã®åé¡ãè§£ãäžã§éèŠãªäžè¬ååã¯äœã§ããïŒããšãã£ãããã«ãäžåºŠæœè±¡çãªã¬ãã«ã§èããããææ³ã§ããããã«ãããæšãèŠãŠæ£®ãèŠãç¶æ ã«é¥ãã®ãé²ããŸã ã
Analogical Prompting
ã解決ãããåé¡ã«é¡äŒŒãããå¥ã®åé¡ãšãã®æèããã»ã¹ãäŸãšããŠèªåçæããããã³ããã«å«ããããšã§ã顿šã«ããåé¡è§£æ±ºãä¿ãæè¡ã§ã ã
Decomposition
ãDecompositionã¯ãThought Generationã®èãæ¹ãããã«äžæ©é²ããè€éãªåé¡ã解決å¯èœãªå°ããªãµãåé¡ãžãšãæç€ºçã«åè§£ãããããšãäžå¿ã«æ®ããæè¡çŸ€ã§ã ã
Least-to-Most Prompting
ããŸãLLMã«ããã®åé¡ãè§£ãããã«ã¯ãã©ã®ãããªãµãåé¡ãé çªã«è§£ãå¿ èŠããããŸããïŒããšåããããåé¡ã®åè§£ãªã¹ããäœæãããŸããæ¬¡ã«ããã®ãªã¹ãã®æåã®ãµãåé¡ãè§£ããããã®çããæ¬¡ã®ãµãåé¡ãè§£ãããã®ããã³ããã«å«ããããšããããã»ã¹ãç¹°ãè¿ããŠæçµçãªçãã«ãã©ãçããŸã ã
Tree-of-Thought (ToT)
ããã®ã«ããŽãªã®æå 端æè¡ã§ããåè¿°ã®éããæèãæšæ§é ã§æ¢çŽ¢ããŸããToTã®å®è£ ã«ã¯ãâ æèã®åè£ãè€æ°çæããã¹ããããâ¡ååè£ãã©ãã ãææããè©äŸ¡ããã¹ããããâ¢æšæ§é ãå¹ççã«æ¢çŽ¢ããã¢ã«ãŽãªãºã ïŒå¹ åªå æ¢çŽ¢ãæ·±ãåªå æ¢çŽ¢ãªã©ïŒã®3ã€ã®èŠçŽ ãå¿ èŠãšãªããŸã ã
Program-of-Thoughts (PoT) / Faithful CoT
ãæšè«ã®ã¹ããããèªç¶èšèªã§ã¯ãªããPythonãªã©ã®ããã°ã©ãã³ã°ã³ãŒãã§èšè¿°ãããææ³ã§ããçæãããã³ãŒããå®éã«å®è¡ããããšã§ãç¹ã«æ°åŠçãªåé¡ã«ãããŠã極ããŠæ£ç¢ºãªçããåŸãããšãã§ããŸã ãããã¯ãLLMã®èšèªçãªæšè«èœåãšãã³ã³ãã¥ãŒã¿ã®å³å¯ãªèšç®èœåãçµã¿åããããã€ããªããã¢ãããŒããšèšããŸãã
Skeleton-of-Thought
ãåçã®çæãé«éåããããã®æè¡ã§ãããŸãLLMã«åçå šäœã®ãéªšæ ŒïŒã¹ã±ã«ãã³ïŒããã€ãŸãç« ç«ãŠã段èœã®èŠç¹ã®ã¿ãçæãããŸããæ¬¡ã«ãåéªšæ Œéšåãåããããã®ããã³ããã䞊åã§è€æ°ã®LLMã«éããæåŸã«ããããçµåããŠå®å šãªåçãçæããŸã ã
Ensembling
ãEnsemblingã¯ãæ©æ¢°åŠç¿ã®åéã§å€ãããçšããããŠããèãæ¹ã§ãåäžã®ã¢ãã«ãææ³ã«é Œãã®ã§ã¯ãªããè€æ°ã®ç°ãªãåºåçµæãçµ±åããããšã§ãããé å¥ã§ç²ŸåºŠã®é«ãçµè«ãåŸãããšãç®æããŸã ã
Self-Consistency
ããã®ã«ããŽãªã®ä»£è¡šäŸã§ããCoTãçšããŠãåãåé¡ã«å¯ŸããŠè€æ°ã®ç°ãªãæšè«çµè·¯ãçæãããæçµçãªçãã®å€æ°æ±ºãåãææ³ã§ããäžã€ã®æšè«çµè·¯ã誀ã£ãŠããŠããä»ã®å€ãã®æ£ããçµè·¯ããããè£æ£ãããããç¹ã«ç®è¡åé¡ãªã©ã§é«ã广ãçºæ®ããŸã ã
Demonstration Ensembling (DENSE)
ãFew-Shotããã³ãããäœæããéã«ã䜿çšããäŸã®çµã¿åãããè€æ°ãã¿ãŒã³çšæããããããã®ããã³ããã§æšè«ãå®è¡ããŸãããããŠãåŸãããè€æ°ã®çµæãéçŽããŠæçµçãªçããšããŸã ã
Mixture of Reasoning Experts (MORE)
ãç°ãªãçš®é¡ã®ã¿ã¹ã¯ã«å¯ŸããŠãããããç¹åããããã³ããïŒå°éå®¶ïŒãçšæããã¢ãããŒãã§ããäŸãã°ãäºå®ã«é¢ãã質åã«ã¯æ€çŽ¢æ¡åŒµïŒRAGïŒããã³ããããæ°åŠã®åé¡ã«ã¯CoTããã³ãããããšãã£ãå ·åã«äœ¿ãåããæçµçã«æãä¿¡é Œã§ããå°éå®¶ã®åçãéžæããŸã ã
Self-Criticism
ã人éãäœããäœæããéã«ãäžæžããæžããŠã¯èŠçŽããä¿®æ£ãããšããããã»ã¹ãçµãããã«ãLLMã«ãèªå·±ã®åºåãè©äŸ¡ããæ¹åããããµã€ã¯ã«ãå°å ¥ããã®ãSelf-CriticismïŒèªå·±æ¹å€ïŒã®æè¡çŸ€ã§ã ã
Self-Refine
ãLLMã«ãŸãåæåçïŒäžæžãïŒãçæãããŸããæ¬¡ã«ããã®äžæžããå¥ã®ããã³ããã§LLMèªèº«ã«èŠããããã®åçã®åé¡ç¹ãææããæ¹åæ¡ãææ¡ããŠãã ããããšæç€ºããŠãã£ãŒãããã¯ãçæãããŸããæåŸã«ããã®ãã£ãŒãããã¯ã«åºã¥ããŠå ã®åçãä¿®æ£ãããŸãããã®ãçæâãã£ãŒãããã¯âæ¹åãã®ãµã€ã¯ã«ãç¹°ãè¿ãããšã§ãåºåã®è³ªã段éçã«é«ããŠãããŸã ã
Chain-of-Verification (CoVE)
ããã«ã·ããŒã·ã§ã³ãæå¶ããããã«èæ¡ãããæè¡ã§ãããŸãLLMã«åçãçæãããŸããæ¬¡ã«ãããã®åçãäºå®ã«åºã¥ããŠãããã確èªããããã«ãã©ã®ãããªæ€èšŒçšã®è³ªåãç«ãŠãã¹ãã§ããïŒããšåããããæ€èšŒèšç»ãç«ãŠãããŸãããããŠããã®æ€èšŒçšè³ªåã«äžã€ã²ãšã€çããããå šãŠã®æ å ±ãçµ±åããŠæçµçãªãããä¿¡é Œæ§ã®é«ãåçãçæãããŸã ã
Self-Calibration
ãLLMã«è³ªåã«çããããåŸããä»ã®ããªãã®åçã¯ã©ã®çšåºŠèªä¿¡ããããŸããïŒ1ãã10ã§è©äŸ¡ããŠãã ããããšãã£ãããã«ãèªå·±ã®ç¢ºä¿¡åºŠãèšèªåãããææ³ã§ããããã«ãããã¢ãã«ãã©ã®çšåºŠãã®çããä¿¡é ŒããŠãããã®ææšãåŸãããšãã§ããŸã ã
ææ¡ææ³ã®æ§æã³ã³ããŒãã³ãããä»çµã¿ã®è©³çް
ããããŸã§èŠãŠããããã³ããæè¡ã¯ããã°ãã°ãã倧ããªã·ã¹ãã ã®æ§æèŠçŽ ãšããŠæ©èœããŸããããã§ã¯ãããã³ããæè¡ãäžæ žããªãäºã€ã®éèŠãªæŠå¿µãããšãŒãžã§ã³ãããšãåçãšã³ãžãã¢ãªã³ã°ãã«ã€ããŠè©³è¿°ããŸãã
ãšãŒãžã§ã³ã
ãLLMã¯åäœã§ã匷åã§ããããã®èœåã«ã¯éçããããŸããäŸãã°ãææ°ã®æ å ±ãç¥ããªãã£ãããè€éãªèšç®ãèŠæã ã£ããããŸãããšãŒãžã§ã³ããšã¯ãããããLLMã®åŒ±ç¹ãè£ãããã«ãLLMãé è³ãšããŠå€éšã®ããŒã«ïŒäŸïŒé»åãã€ã³ã¿ãŒãããæ€çŽ¢ãäŒæ¥ã®ããŒã¿ããŒã¹ã«ã¢ã¯ã»ã¹ããAPIãªã©ïŒãšé£æºãããã·ã¹ãã ã®ããšã§ã ã
ããšãŒãžã§ã³ãã·ã¹ãã ã¯ãäžè¬çã«ããã³ããã«ãã£ãŠé§åãããŸãããã®ä»çµã¿ã¯ä»¥äžã®ããã«ãªã£ãŠããŸãã
èšç»
ããŠãŒã¶ãŒããã仿¥ã®æ±äº¬ã®å€©æ°ãšãããã«åºã¥ããŠåãæã£ãŠããã¹ããæããŠããšããã¿ã¹ã¯ãäžãããããšãLLMã¯ãŸãã¿ã¹ã¯ãéæããããã®èšç»ãç«ãŠãŸãããã¹ããã1: æ±äº¬ã®å€©æ°ã調ã¹ããã¹ããã2: 倩æ°äºå ±ã«åºã¥ããŠåãå¿ èŠã倿ããããšãã£ãå ·åã§ãã
ããŒã«éžæ
ãLLMã¯ãèšç»ã®åã¹ããããå®è¡ããããã«ã©ã®ããŒã«ã䜿ãã¹ããã倿ããŸããã¹ããã1ãå®è¡ããããã«ã¯ã倩æ°äºå ±APIããå¿ èŠã ãšå€æããŸãããã®å€æãè¡ãLLMã®éšåããã«ãŒã¿ãŒããšåŒã¶ããšããããŸãã
ããŒã«äœ¿çš
ãLLMã¯ã倩æ°äºå ±APIãåŒã³åºãããã®é©åãªåœ¢åŒã®åºåïŒäŸïŒAPI_CALL('weather_forecast', city='Tokyo')ïŒãçæããŸããã·ã¹ãã ã¯ãã®åºåãåãåããå®éã«APIãå©ããŠå€©æ°æ å ±ãååŸããŸãã
芳å¯ã𿬡ã®è¡å
ãAPIãããéæ°Žç¢ºç80%ããšããçµæïŒèгå¯ïŒãè¿ã£ãŠãããšããã®æ å ±ãæ¬¡ã®ããã³ããã«å ¥åãããŸããLLMã¯èšç»ã®ã¹ããã2ã«åºã¥ãããéæ°Žç¢ºçã80%ãªã®ã§ãåãæã£ãŠããã¹ãã§ãããšããæçµçãªçµè«ãçæããŸãã
ãReAct (Reason + Act) ã¯ããã®ãæšè«ïŒReasonïŒããšãè¡åïŒActïŒãã®ãµã€ã¯ã«ãæç€ºçã«ããã³ããã«çµã¿èŸŒãããšã§ãšãŒãžã§ã³ãã®æ§èœãé«ãã代衚çãªæè¡ã§ã ããšãŒãžã§ã³ãã¯ãããã³ããæè¡ãå¿çšããŠLLMã«èªåŸçãªåé¡è§£æ±ºèœåãæããããããã³ãããšã³ãžãã¢ãªã³ã°ã®æå 端é åã®äžã€ã§ãã
åçãšã³ãžãã¢ãªã³ã°
ãLLMã«å®ç§ãªããã³ãããäžãããšããŠãããã®åºåãåžžã«æã ã®æã圢åŒã§ãããšã¯éããŸãããäŸãã°ãææ åæã¿ã¹ã¯ã§ãããžãã£ãããããã¬ãã£ãããšããåèªã ããè¿ããŠã»ããã®ã«ãããã®æç« ã¯éåžžã«ããžãã£ããªææ ã衚ããŠãããšèããããŸãããªããªãâŠããšãã£ãåé·ãªæç« ãè¿ã£ãŠããããšã¯æ¥åžžè¶é£¯äºã§ãã
ãåçãšã³ãžãã¢ãªã³ã°ãšã¯ããã®ãããªLLMã®çã®åºåãããæã ãå¿ èŠãšããæ£ç¢ºã§æ§é åãããçããæœåºããããã®äžé£ã®æè¡ãããã»ã¹ã®ããšã§ã ãããã¯ããã³ãããšã³ãžãã¢ãªã³ã°ãšå¯æ¥ã«é¢é£ããŠããããã°ãã°åæã«èšèšãããŸããåçãšã³ãžãã¢ãªã³ã°ã«ã¯3ã€ã®äž»èŠãªèšèšèŠçŽ ããããŸã ã
åçã®åœ¢ç¶
ãåºåã«æåŸ ããç©ççãªåœ¢åŒãæããŸããäŸãã°ããåäžã®ããŒã¯ã³ããJSONãªããžã§ã¯ãããã«ã³ãåºåãã®ãªã¹ãããªã©ã§ããåé¡ã¿ã¹ã¯ã§ã¯ãåçã®åœ¢ç¶ãåäžã®ããŒã¯ã³ã«å¶éããããšãæå¹ã§ãã
åçã®ç©ºé
ãåçãåãããå€ã®ç¯å²ãæããŸããäºå€åé¡ã§ããã°ãåç空éã¯ãããžãã£ãããšããã¬ãã£ãããšããäºã€ã®ããŒã¯ã³ã®ã¿ã«éå®ãããŸããããã³ããã§ãããžãã£ãããã¬ãã£ãã®ã©ã¡ããäžæ¹ã®åèªã®ã¿ã§çããŠãã ããããšå¶çŽããããšã§ãåç空éãçããããšãã§ããŸãã
åçæœåºåš
ãããã³ããã§åœ¢ç¶ã空éãå®å šã«å¶åŸ¡ã§ããªãå Žåã«ãçã®åºåããç®çã®çããæãåºãããã®ã«ãŒã«ã颿°ã§ãã
æ£èŠè¡šçŸ
ããçãã¯: (A)ãã®ãããªç¹å®ã®ãã¿ãŒã³ã«ãããããéšåãæãåºããæãäžè¬çãªæ¹æ³ã§ãã
Verbalizer
ãã©ãã«ä»ãã¿ã¹ã¯ã§ãã䜿ãããã+ããšããããŒã¯ã³ããããžãã£ãããšããã©ãã«ã«ãã-ããããã¬ãã£ããã«ãããã³ã°ãããããªå€æèŸæžã§ã ã
å¥ã®LLMåŒã³åºã
ãåºåãéåžžã«è€éã§æ£èŠè¡šçŸã§ã¯å¯Ÿå¿ã§ããªãå Žåããäžèšã®æç« ããæçµçãªçãïŒYesãNoïŒã ããæœåºããŠãã ããããšãããããªãæœåºå°çšã®ããã³ããã§å¥ã®LLMãåŒã³åºãããšããããŸã ã
ãåªããããã³ãããšã³ãžãã¢ãªã³ã°ã¯ãããã³ããã ãã§ãªãããã®åºåã®æ±ãæ¹ãŸã§å«ããŠèšèšããããã®ãªã®ã§ãã
ãå³2.3ã¯Few-Shotããã³ããèšèšã®6ã€ã®äž»èŠãªæææ±ºå®ããŸãšããå³ã§ãäŸç€ºã®æ°ã»é åºã»å 容ãšãã£ã芳ç¹ããšã«æšå¥šäºé ã瀺ãããŠããŸãã

ãå³2.3ã¯Few-Shotããã³ããèšèšäžã®6èŠç¹ãèŠèŠåãããã®ã§ããå³äžã«ã¯ãã©ãã«ååžã®ãã©ã³ã¹ãåãããäŸç€ºã¯ä»£è¡šçãªãã®ãéžã¶ããå ¥åãšã®é¡äŒŒåºŠãé«ãäŸãæ¡çšããããªã©ãåãã€ã³ããçªå·ä»ãã§ç€ºãããŠããŸãã
ãäŸãã°ãExemplar Label DistributionïŒäŸç€ºã®ã©ãã«ååžïŒãã§ã¯ãåé¡ã¿ã¹ã¯ã®å Žåã«è¯å®äŸã»åŠå®äŸãåããªãæç€ºããããšãéèŠã§ãããšè¿°ã¹ãããŠããŸãã
ããŸãããExemplar OrderingïŒäŸç€ºã®äžŠã³é ïŒãã§ã¯ãé£æåºŠãæèã®æµããèæ ®ããŠäŸã®é åºã工倫ããããšã§ã¢ãã«ã®å¿çãå®å®ããå¯èœæ§ãææãããŠããŸãããããã¯ãããŸã§äžè¬çãªæéã§ããããåžžã«åœãŠã¯ãŸãæ³åã§ã¯ãªããããšã«ã泚æãæ·»ããããŠããŸãâ»ã
ãå®éãã¿ã¹ã¯ã«ãã£ãŠã¯äŸãå¢ããããããšããã£ãŠæ··ä¹±ããã±ãŒã¹ããæ¥µç«¯ãªã±ãŒã¹ã§ã¯ãŒãã·ã§ããã®æ¹ãFew-Shotããæ§èœãè¯ããšå ±åãããå ŽåãããããŸãã
ãããã³ããèšèšã¯è©Šè¡é¯èª€ã®ããã»ã¹ã§ãããã¬ã€ãã©ã€ã³ã¯ãã£ãŠãæçµçã«ã¯åã ã®ã±ãŒã¹ã«åããã調æŽãå¿ èŠã«ãªãç¹ã¯æŒãããŠããã¹ãã§ãããã
â»ããããã®æšå¥šã¯å šãŠã®ã¿ã¹ã¯ã«äžè¬åã§ãããã®ã§ã¯ãªããå Žåã«ãã£ãŠã¯é广ãšãªãå¯èœæ§ãããã
ã以äžãæ¹æ³è«ã®ã»ã¯ã·ã§ã³ã§ã¯ãããã³ãããšã³ãžãã¢ãªã³ã°æè¡ãåé¡ã»æŽçããã¢ãããŒããšããã®åé¡ããšã®ä»£è¡šææ³ã«ã€ããŠæŠèгããŸãããæ¬¡ã«ãèè ããè¡ã£ãæ€èšŒãã±ãŒã¹ã¹ã¿ãã£ã«ç®ãåãããããææ³ãå®éã«ã©ã掻çšã»è©äŸ¡ãããŠããã®ãèŠãŠã¿ãŸãããã
å®é𿹿³
ãè«æã§ã¯ãããã³ããæè¡ã®æå¹æ§ãè©äŸ¡ããããã«ãå®éçãã³ãããŒã¯ãšå®æ§çã±ãŒã¹ã¹ã¿ãã£ãšãããæ§è³ªã®ç°ãªãäºã€ã®ã¢ãããŒããæ¡çšããŠããŸã ã
1. MMLUãã³ãããŒã¯è©äŸ¡
ãããã¯ãããã³ããæè¡ã®æ§èœã客芳çãªæ°å€ã§æ¯èŒããããã®å®éçè©äŸ¡ã§ãã
ãã³ãããŒã¯: MMLU (Massive Multitask Language Understanding) ãšããã倿§ãªåŠè¡åéïŒäººæç§åŠã瀟äŒç§åŠãSTEMãªã©ïŒã«ãããç¥èãåããåºã䜿ãããŠããæšæºçãªãã³ãããŒã¯ã䜿çšãããŸãã ã
ã¢ãã«:gpt-3.5-turbo ãè©äŸ¡å¯Ÿè±¡ã®LLMãšããŠéžã°ããŸãã ã
è©äŸ¡å¯Ÿè±¡æè¡:以äžã®6ã€ã®ä»£è¡šçãªããã³ããæè¡ãæ¯èŒãããŸãã ã
Zero-Shot (æç€ºã®ã¿)
Zero-Shot-CoT (æç€º + ãæ®µéçã«èããŠã)
Zero-Shot-CoT + Self-Consistency (倿°æ±º)
Few-Shot (ããã€ãã®äŸãæç€º)
Few-Shot-CoT (æèããã»ã¹ä»ãã®äŸãæç€º)
Few-Shot-CoT + Self-Consistency (倿°æ±º)
è©äŸ¡æ¹æ³: MMLUã®è³ªåã»ããïŒãã®å®éšã§ã¯2,800åã®ãµãã»ããïŒã«å¯ŸããŠåæè¡ãé©çšãããã®æ£è§£çãæž¬å®ããŸããã
2. èªæ®ºãªã¹ã¯æ€åºã®ã±ãŒã¹ã¹ã¿ãã£
 ããã¯ãçŸå®äžçã®è€éã§æ©åŸ®ãªåé¡ã«å¯ŸããŠãããã³ãããšã³ãžãã¢ãªã³ã°ãã©ã®ããã«é©çšãããã©ã®ãããªèª²é¡ã«çŽé¢ããã®ããæ·±ãæ¢ãããã®å®æ§çè©äŸ¡ã§ãã
ã¿ã¹ã¯
Â èªæ®ºå¿µæ ®ãæã€äººã ããªã³ã©ã€ã³æ²ç€ºæ¿ã«æçš¿ããæç« ãããèªæ®ºå±æ©ã®ãªã¹ã¯ã匷ã瀺åããå¿çç¶æ ã§ãããEntrapmentïŒéã蟌ãããæãéãå Žã®ãªãçµ¶ææïŒããæ€åºãããšãããéåžžã«å°é£ãªã¿ã¹ã¯ã§ã ã
ããŒã¿
 å°éå®¶ã«ãã£ãŠãEntrapmentãã®æç¡ãã©ãã«ä»ããããRedditã®æçš¿ããŒã¿ã»ããã䜿çšãããŸãã ã
ããã»ã¹
ãçµéšè±å¯ãªããã³ãããšã³ãžãã¢ãããã®ã¿ã¹ã¯ã®ç²ŸåºŠãæå€§åããããã«ã47ã¹ãããã«ãåã¶æäœæ¥ã§ã®å埩çãªããã³ããæ¹è¯ããã»ã¹ã宿œããŸããããã®å šããã»ã¹ãèšé²ãããåæã®å¯Ÿè±¡ãšãªããŸãã ã
æ¯èŒ
ãæäœæ¥ã«ããããã³ãããšã³ãžãã¢ãªã³ã°ã®çµæã¯ãDSPyãšããèªåããã³ããæé©åãã¬ãŒã ã¯ãŒã¯ã®çµæãšãæ¯èŒãããŸãã ã
å®éšçµæ
ãäºã€ã®å®éšã¯ãããã³ãããšã³ãžãã¢ãªã³ã°ã®å ãšåœ±ããããŠãã®è€éãªçŸå®ãæµ®ã圫ãã«ããŸããã
MMLUãã³ãããŒã¯ã®çµæ
å³6.1ã«ç€ºãããçµæã¯ãããã€ãã®éèŠãªç€ºåãäžããŠãããŸã ã

è€éãã¯å¿
ãããåã§ã¯ãªã
ãå
šäœçãªåŸåãšããŠãããæŽç·Žãããæè¡ã»ã©é«ãæ§èœã瀺ããŸãããäŸãã°ãFew-Shot-CoTã¯åçŽãªZero-Shotãããé«ãæ£è§£çãéæããŠããŸãããããããã®æ³åã«ã¯äŸå€ããããŸãããé©ãã¹ãããšã«ãZero-Shot-CoTã¯ãããŒã¹ã©ã€ã³ã§ããZero-Shotãããæ§èœãäœäžããã®ã§ãããŸããSelf-ConsistencyïŒå€æ°æ±ºïŒã¯Zero-Shotã®æ§èœãåäžãããŸããããFew-Shot-CoTã®æ§èœã¯ã»ãšãã©æ¹åããŸããã§ãã ã
çµè«
ããã®çµæã¯ããããè€éãªããã³ããæè¡ãé©çšããã°ãåžžã«æ§èœãåäžããããã§ã¯ãªãããšããéèŠãªæèšã瀺ããŠããŸããæè¡ã®éžæã¯ã¿ã¹ã¯ãã¢ãã«ãšã®çžæ§ã«å€§ããäŸåãã宿ãªé©çšã¯é广ã«ãããªãåŸãã®ã§ããããã¯ãããã³ãããšã³ãžãã¢ãªã³ã°ããåäžã®æåçãååšããªãã詊è¡é¯èª€ã䌎ããã€ããŒãã©ã¡ãŒã¿æ¢çŽ¢ã«äŒŒãåŽé¢ãæã€ããšã瀺åããŠããŸã ã
èªæ®ºãªã¹ã¯æ€åºã±ãŒã¹ã¹ã¿ãã£ã®çµæ
ããã®ã±ãŒã¹ã¹ã¿ãã£ã¯ãããã³ãããšã³ãžãã¢ãªã³ã°ã®ãå®è·µãã®çã ããå§¿ãæãåºããŠããŸã ã
æäœæ¥ã®å°é£ã
ãå°éå®¶ã«ãã47ã¹ãããã®æ¹è¯ããã»ã¹ã¯ããŸãã«ãé»éè¡ãã®ããã§ãããF1ã¹ã³ã¢ïŒç²ŸåºŠãšåçŸçã®èª¿åå¹³åïŒã0.53ãŸã§åäžãããŸãããããã®éçšã¯éç·åœ¢ã§ãäºæž¬äžå¯èœãªçºèŠã«æºã¡ãŠããŸãããäŸãã°ãããã³ããã«èª€ã£ãŠææããã®ã¡ãŒã«ãäºéã«è²Œãä»ããŠããŸã£ããšãããæ§èœãèããåäžãããšãããè«ççã«ã¯èª¬æã®ã€ããªãçŸè±¡ã芳å¯ãããŸãã ã
èªååã®å¯èœæ§
ãæäœæ¥ã§å°éããæé«ã®ããã³ããïŒå¥åŠãªã¡ãŒã«ã®éè€ãå«ãïŒããããèªåããã³ããæé©åãã¬ãŒã ã¯ãŒã¯ã§ããDSPyãçæããããã³ããã®æ¹ãããã¹ãããŒã¿ã«å¯ŸããŠé«ãF1ã¹ã³ã¢ïŒ0.548ïŒãéæããŸãããããããDSPyãçæããããã³ããã¯ã人éãå ãããããªå¥åŠãªèŠçŽ ãå«ãŸãªããããã¯ãªãŒã³ã§æ±çšçãªãã®ã§ãã ã
çµè«
ããã®ã±ãŒã¹ã¹ã¿ãã£ã¯ãæäœæ¥ã«ããããã³ãããšã³ãžãã¢ãªã³ã°ãããã«èãã説æå°é£ãªæåã«äŸåããã¡ã§ããããæããã«ããŸãããåæã«ãDSPyã®ãããªèªååãããææ³ãã人éã®å°éå®¶ã«ããåªåãäžåããããå
ç¢ã§é«æ§èœãªããã³ãããçºèŠã§ãããšãããå匷ãå¯èœæ§ã瀺ããŠããŸãã
ããã®ã±ãŒã¹ã¹ã¿ãã£ããåŸãããããäžã€ã®æ·±ãæèšã¯ãæè¡çãªæé©åãšããã®ã¿ã¹ã¯ãæã€æ¬æ¥ã®ç®çãšã®éã«çãããä¹é¢ã§ããããã³ãããšã³ãžãã¢ã¯ãæ§èœææšã§ããF1ã¹ã³ã¢ãæå€§åããããšã«éäžããŸããã
ããã®éçšã§ãåœéœæ§ïŒEntrapmentã§ã¯ãªãã®ã«Entrapmentãšå€å®ããŠããŸãããšïŒãæžããããã«ããæçš¿è ãæç€ºçã«ãéã蟌ããããããšè¿°ã¹ãŠããå Žåã®ã¿ãEntrapmentãšã©ãã«ä»ãããããšããæç€ºãããã³ããã«å ããŸãã ãããã¯æè¡çã«ã¯åççãªå€æã§ãã
ãããããèšåºçãªèгç¹ããèŠãã°ãããã¯èª€ã£ã倿ã§ãããè«æãææããããã«ãèªæ®ºãªã¹ã¯ã®æ€åºãšããæèã§ã¯ãåœé°æ§ïŒæ¬åœã«å±éºãªç¶æ ã«ãã人ãèŠéããŠããŸãããšïŒã®ã³ã¹ãã¯ãåœéœæ§ã®ã³ã¹ããããé¥ãã«å€§ããã®ã§ãã
ãããã«ããEntrapmentããšããå¿çç¶æ ã¯ãæ¬äººãèªèŠããŠãéã蟌ããããããšè¡šçŸãããšã¯éããããããæé»çã«ãæç« ã®ç«¯ã ã«æ»²ã¿åºããã®ã§ããå Žåãå€ãã®ã§ã ã
èå¯
ãªããã®ææ³ãåªããŠããã®ã
ãCoTãToTãSelf-Refineãšãã£ãé«åºŠãªããã³ããæè¡ãããªãåçŽãªIn-Context Learningãããè€éãªã¿ã¹ã¯ã§åªããæ§èœãçºæ®ããã®ã§ããããããã®çãã¯ãæ¬çš¿ã®åé ã§æç€ºãããèªç¥çè¶³å Žãã®éå±€æ§é ãšããèãæ¹ã«ãã£ãŠèª¬æã§ããŸãã
ãåçŽãªICLã¯ãLLMã«å®æããææ¬ãèŠããããšã§ããã¿ãŒã³ãèªèãããææ³ã§ããããã¯ãLLMãæã€å·šå€§ãªäºååŠç¿ç¥èã®äžãããé©åãªãã¿ãŒã³ããæãåºããäœæ¥ãå©ããã«éããŸãããããããæªç¥ã®è€éãªåé¡ã«å¯ŸããŠã¯ãåãªããã¿ãŒã³ãããã³ã°ã§ã¯å¯Ÿå¿ã§ããŸããã
ãããã§ãCoTã®ãããªThought Generationæè¡ãéèŠã«ãªããŸããCoTã¯ãLLMã«æèã®ããã»ã¹ã匷å¶çã«èšèªåãããããšã§ãæšè«ãç·åœ¢çãã€æ®µéçã«é²ãŸããŸããããã«ãããè«ççãªé£èºããã«ã·ããŒã·ã§ã³ãæå¶ããããæ§é åãããæèãå¯èœã«ããŸãã
ãããã«ToTãLeast-to-Mostãšãã£ãDecompositionæè¡ã¯ãåé¡è§£æ±ºã®ãæŠç¥ããã®ãã®ãLLMã«æäŸããŸããåé¡ãåè§£ããè€æ°ã®éžæè¢ãæ¢æ±ããææãªçµè·¯ãéžæãããšããããã»ã¹ã¯ãLLMãåãªããã¿ãŒã³èªèåšãããæŠç¥çãªåé¡è§£æ±ºè ãžãšåŒãäžããŸãã
ããããŠSelf-Refineã®ãããªSelf-Criticismæè¡ã¯ãLLMã«ã¡ã¿èªç¥ãã€ãŸãèªãã®æèã客芳èŠããæ¹åããèœåãäžããŸãããã®å埩çãªæ€èšŒãšä¿®æ£ã®ãµã€ã¯ã«ããããåºåã®è³ªãç¶ç¶çã«åäžãããåååãšãªãã®ã§ãã
ãèŠããã«ããããã®é«åºŠãªææ³ã¯ãLLMãåãªããç¥èã®æ€çŽ¢ãšã³ãžã³ããããæ§é åãããæšè«ãæŠç¥çãªæ¢çŽ¢ããããŠã¡ã¿èªç¥çãªæ€èšŒãå¯èœãªãããé«åºŠãªãæèãšã³ãžã³ããžãšå€è²ãããããã«åªããŠããã®ã§ãã
åªäœæ§ã»å£äœæ§
ãããã³ããæè¡ã®éžæã¯ãåžžã«ãã¬ãŒããªãã䌎ããŸããããç¶æ³ã§æé©ãªæè¡ããå¥ã®ç¶æ³ã§ã¯äžé©åã§ããããšãå°ãªããããŸãããããã§ã¯ãäž»èŠãªæè¡çŸ€ã®æ¯èŒãéããŠããã®åªäœæ§ãšå£äœæ§ãæããã«ããŸãã
CoTãSelf-ConsistencyãToTã®éã«ã¯æç¢ºãªãã¬ãŒããªããååšããŸã ã
CoT vs. ToT
ãCoTã¯é«ééè·¯ãçŽé²ãããããªãã®ã§ãç®çå°ãŸã§äžçŽç·ã«é²ãåé¡ã«ã¯å¹ççã§ããäžæ¹ãToTã¯éœåžã®å°å³ãåºããŠè€æ°ã®ã«ãŒããæ¯èŒæ€èšãããããªãã®ã§ãè€éãªæææ±ºå®ãèšç»ãæ±ããããå Žé¢ã§ãã®ç䟡ãçºæ®ããŸã ã
Ensembling/Self-Consistencyã®ã³ã¹ã
ãSelf-Consistencyã®ãããªã¢ã³ãµã³ãã«ææ³ã¯ã粟床ãåäžããã匷åãªæŠåšã§ããããã®ä»£åãšããŠAPIã³ãŒã«æ°ãæ°åããæ°ååã«å¢å ããŸããããã¯ãã¬ã€ãã³ã·ãŒïŒå¿çæéïŒãã³ã¹ããéèŠãªãªã¢ã«ã¿ã€ã ã¢ããªã±ãŒã·ã§ã³ã«ã¯äžåãã§ããããšãæå³ããŸã ã
æå vs. èªåããã³ãããšã³ãžãã¢ãªã³ã°
ãã±ãŒã¹ã¹ã¿ãã£ã瀺ããããã«ãæäœæ¥ã§ã®ããã³ãããšã³ãžãã¢ãªã³ã°ã¯å±äººçãªãé»éè¡ãã«é¥ããã¡ã§ããDSPyã®ãããªèªååææ³ã¯ãããå
ç¢ã§é«æ§èœãªããã³ãããçºèŠã§ããå¯èœæ§ããããŸãã
ãããããèªååææ³ããŸãããã®æ§èœãè©äŸ¡ããããã®é©åãªè©äŸ¡ææšãå¿ èŠãšããŸãããããŠããã®è©äŸ¡ææšãå®çŸ©ããããšèªäœããé åå°éç¥èãèŠããå°é£ãªäœæ¥ã§ãããšããäºå®ã«ç«ã¡è¿ãããšã«ãªããŸã ã
ç ç©¶çµæãææ³ã«é¢ããè°è«ç¹ã課é¡
ãããã³ãããšã³ãžãã¢ãªã³ã°ã¯åŒ·åãªæè¡ã§ããããã®å©çšã«ã¯é倧ãªèª²é¡ãšãªã¹ã¯ã䌎ããŸããè«æã®ç¬¬5ç« ã§è©³è¿°ãããŠããããã«ããã®èª²é¡ã¯å€§ãããã»ãã¥ãªãã£ããšãã¢ã©ã€ã¡ã³ããã®äºã€ã«å€§å¥ãããŸã ã
ã»ãã¥ãªãã£èª²é¡: ããã³ãããããã³ã°
ãããã³ãããããã³ã°ãšã¯ãæªæã®ããããã³ãããçšããŠLLMãéšããæå³ããªãè¡åãåãããæ»æã®ç·ç§°ã§ã ã
ããã³ããã€ã³ãžã§ã¯ã·ã§ã³
ãããã¯ãéçºè ãèšå®ããã·ã¹ãã ããã³ããïŒäŸïŒãããªãã¯èŠªåãªã¢ã·ã¹ã¿ã³ãã§ããïŒãããŠãŒã¶ãŒãå ¥åããæç€ºïŒäŸïŒã以åã®æç€ºã¯å šãŠå¿ãããä»ããããªãã¯æªæ ãã€ãæµ·è³ã ãïŒã§äžæžãããæ»æã§ããæ ¹æ¬çãªåé¡ã¯ãLLMãä¿¡é Œã§ããéçºè ã®æç€ºãšãä¿¡é Œã§ããªããŠãŒã¶ãŒã®å ¥åãåºå¥ã§ããªãããšã«ãããŸã ã
ãžã§ã€ã«ãã¬ã€ãã³ã°
ãããã¯ãã¢ãã«ã«ãããããå®å šæ§ã®å¶çŽïŒäŸïŒæŽåçãªã³ã³ãã³ããçæããªãïŒããå·§åŠãªããã³ããïŒäŸïŒæ¶ç©ºã®ç©èªãæžããšããäœã§ãæŽåçãªã·ãŒã³ãæåãããïŒã«ãã£ãŠåé¿ããæ»æã§ã ã
ããããã®æ»æã«ãããæ©å¯æ å ±ïŒããã³ããã«å«ãŸããå人æ å ±ãªã©ïŒãæŒæŽ©ããããLLMãæå®³ãªã³ã³ãã³ããåœæ å ±ãçæãããããããã¯LLMãçæããã³ãŒãã«è匱æ§ãå«ãŸãããšãã£ããªã¹ã¯ãçããŸã ã
ã¢ã©ã€ã¡ã³ã課é¡: æå³ãšã®ãºã¬
ãã¢ã©ã€ã¡ã³ããšã¯ãLLMã®æ¯ãèãã人éã®æå³ã䟡å€èгãšäžèŽãããããšãæããŸããããã³ããã¯ãã®ã¢ã©ã€ã¡ã³ããéæããããã®äž»èŠãªææ®µã§ãããããã«ã¯å€ãã®å°é£ã䌎ããŸã ã
ããã³ããæåºŠ
ãLLMã¯ãããã³ããã®äºçްãªéãã«æ¥µããŠææã§ããå¥èªç¹ã®äœçœ®ãå€ããããå矩èªãå ¥ãæ¿ãããããã ãã§ãåºåã倧ããå€ãã£ãŠããŸãããšããããŸã ã
ãã€ã¢ã¹ãšã¹ãã¬ãªã¿ã€ã
ãLLMã¯ããã®åºç¯ãªåŠç¿ããŒã¿ã«å«ãŸãã瀟äŒçãã€ã¢ã¹ãã¹ãã¬ãªã¿ã€ããåçç£ããå¢å¹ ããŠããŸãå±éºæ§ããããŸã ã
ãã¹ã£ã
ãLLMã¯ã圹ã«ç«ã€ã¢ã·ã¹ã¿ã³ãããšããŠèšç·ŽãããŠããããããŠãŒã¶ãŒã衚æããæèŠã«ïŒããšããããééã£ãŠããŠãïŒåæããŠããŸãåŸåããããŸããäŸãã°ããŠãŒã¶ãŒãããã®è°è«ã¯çŽ æŽããããšæãããšããã³ããã«æžããšãLLMããã®è°è«ãè¯å®çã«è©äŸ¡ããã¡ã§ã ã
ããããã®ã»ãã¥ãªãã£ãšã¢ã©ã€ã¡ã³ãã®èª²é¡ã¯ãåãªãããã°ããšããŠä¿®æ£ã§ãããã®ã§ã¯ãªããLLMãšããæè¡ã®æ ¹æºçãªæ§è³ªã«æ ¹å·®ããŠããŸããèªç¶èšèªã®æã€ç¡éã®æè»æ§ã¯ãLLMã®é©ç°çãªèœåã®æºæ³ã§ãããšåæã«ããã®å¶åŸ¡ãå°é£ã«ããæå€§ã®èŠå ã§ãããã®ã§ãã
ããã®äºå®ã¯ãLLMã®èœåãšå¶åŸ¡ã®éã«ååšãããæ¬è³ªçãã€é¿ããããªãç·åŒµé¢ä¿ã瀺åããŠããŸããããã³ããã€ã³ãžã§ã¯ã·ã§ã³ãæåããã®ã¯ãLLMãæç€ºã«å¿ å®ã«åŸããšããããã®äžæ žçãªèœåã«èµ·å ããŸãã
ããã¹ã£ããçããã®ã¯ãLLMããŠãŒã¶ãŒã®åœ¹ã«ç«ãšããšããããã®èšç·Žç®çã«èµ·å ããŸããã©ã¡ãã®ã±ãŒã¹ããã¢ãã«ã¯ããæå³ã§ãæå³éãã«åäœãããŠããã®ã§ããåé¡ã¯ããã®æå³ããæªæã®ããå ¥åãè€éãªç€ŸäŒçç¶æ³ãæ³å®ããŠååã«å®çŸ©ãããŠããªãç¹ã«ãããŸãã
ãçŸåšããã®åé¡ã«å¯ŸåŠããããã«ãå ¥åãåºåãç£èŠãããã¬ãŒãã¬ãŒã«ãããæªæã®ããããã³ãããæ€ç¥ããããã£ãã¯ã¿ãŒããšãã£ã察çãè¬ããããŠããŸããããããã¯ã¢ãã«ã®ã¢ãŒããã¯ãã£ã«å åšããªãå¶åŸ¡ããåŸããä»ãå ãã察ççæ³çãªã¢ãããŒãã§ã ã
ãLLMãå®å šã«ç€ŸäŒå®è£ ããŠããäžã§ããã®èœåãšå¶åŸ¡ã®éã®æ ¹æºçãªãã¬ãŒããªããçè§£ããããšã¯ã極ããŠéèŠã§ãã
çµè«
ãæ¬çš¿ã§ã¯ãåŠè¡è«æãThe Prompt Reportããåºã«ãããã³ãããšã³ãžãã¢ãªã³ã°ãšãããçæAIæä»£ã«ãããå¿ é ã®æé€ãšãªãã€ã€ããåéã®äœç³»çãªè§£åã詊ã¿ãŸããã58çš®é¡ãã®ããã¹ãããŒã¹æè¡ããããã®å®èšŒå®éšããããŠå åšãã課é¡ã«è³ããŸã§ã詳述ããŠããŸããããããããåŸãããçµè«ãšå®è·µçãªæèšã¯ãããã€ãã®éèŠãªç¹ã«éçŽãããŸã ã
ã第äžã«ãããã³ãããšã³ãžãã¢ãªã³ã°ã¯ãã¢ãŒããããµã€ãšã³ã¹ãžãšé²åãã€ã€ããããäŸç¶ãšããŠéèªæãªããã»ã¹ã§ãããšããããšã§ããæ¬çš¿ã§ç޹ä»ããæ°ã ã®äœç³»çãªæè¡ã¯ããã®åéããã¯ãåãªãåãçµéšåã«é Œããã¢ãŒããã®æ®µéãè¶ããåçŸæ§ãšæ§é æ§ãæã€ããšã³ãžãã¢ãªã³ã°ããžãšæçãã€ã€ããããšã瀺ããŠããŸãã
ãããããã±ãŒã¹ã¹ã¿ãã£ã瀺ããããã«ããã®å®è·µã¯ãããã°ã©ãã³ã°ããšããããã¯ãäŸç¶ãšããŠäºæž¬äžèœãªæ¯ãèããããç¥æ§ãããªã ãããããå¶ã¿ã«è¿ãåŽé¢ãæã£ãŠããŸãã
ã第äºã«ãè€éãã¯åžžã«æåãšã¯éããªããšããæãã§ããMMLUãã³ãããŒã¯ã®çµæã瀺ããããã«ãCoTãSelf-Consistencyãšãã£ãé«åºŠãªæè¡ããåžžã«åçŽãªããŒã¹ã©ã€ã³ãäžåãããã§ã¯ãããŸããã
ãå®è·µè ã¯ããŸãæãã·ã³ãã«ãªã¢ãããŒãããå§ããã¿ã¹ã¯ã®èŠæ±ã«å¿ããŠæ®µéçã«è€éãªææ³ã詊ãã¹ãã§ãããããŠãããããè«æã§å ±åãããŠããæ§èœåäžã«é¢ãã䞻匵ã«å¯ŸããŠã¯ãå¥å šãªæçå¿ãæã¡ç¶ããå¿ èŠããããŸãã
ãããã¢ãã«ãããã¿ã¹ã¯ãããããŒã¿ã»ããã§æå¹ã ã£ãæè¡ããèªèº«ã®ç¶æ³ã§ãã®ãŸãŸéçšããä¿èšŒã¯ã©ãã«ããªãã®ã§ãã
ããããŠæãéèŠãªç¬¬äžã®æèšã¯ãæé«ã®ææã¯ãæè¡çå°éç¥èãšé åå°éç¥èã®æ·±ãèåããçãŸãããšããç¹ã§ããèªæ®ºãªã¹ã¯æ€åºã®ã±ãŒã¹ã¹ã¿ãã£ã¯ããã®ç¹ãéåŒã«ç©èªã£ãŠããŸããããã³ãããšã³ãžãã¢ãæè¡çãªææšã®æé©åã«éäžããããŸããã¿ã¹ã¯ãæã€èšåºçãªæå³åããããã倧ããªç€ŸäŒçæèãèŠå€±ããããå Žé¢ã¯ãæã å šå¡ã«ãšã£ãŠã®èŠéã§ãã
ããšãã
ãæ¬çš¿ã®å°çšã¯ãLLMã®æ¯ãèããããã³ãããšããã€ã³ã¿ãŒãã§ãŒã¹èšèšã§å¶åŸ¡ããããšããäžç¹ã«éçŽãããŸããè«æãäžãã䟡å€ã¯ãå奿é©ã®ãã¯ããã¯ãåæããããšã§ã¯ãªããæŠå¿µã»çšèªã»ææ³ã®äœç³»åã«ãããåçŸå¯èœã§æ€èšŒå¯èœãªå®è·µãžãšæ¥ç¶ããç¹ã«ãããŸãã
ããã³ããã¯ã¢ãã«æ¹å€ãªãã«åºåãå€ããå¶åŸ¡ææ®µã§ãããICLããŒãã·ã§ãããChain-of-Thoughtãåé¡åè§£ãèªå·±äžè²«æ§ãã»ã«ãã¯ãªãã£ãŒã¯ãšããå ã€ã®æ ã§åççã«æŽçã§ããŸãã
广ã¯ã¿ã¹ã¯ã»ã¢ãã«ã»ããŒã¿ã«äŸåããŸããçµéšåã¯æçšã§ãããäžè¬åã«ã¯æ éã§ããã¹ãã§ãã
å質åäžãšåãéã¿ã§å®å šæ§ïŒãã«ã·ããŒã·ã§ã³æå¶ãã¬ããã³ã¹ãæ©å¯ç®¡çïŒãæ±ãå¿ èŠããããŸããããã³ããã¯å©äŸ¿ãšãªã¹ã¯ãåæã«å¢å¹ ãåŸãŸãã
ç®çå®çŸ©ïŒæåŸ åºåã»çŠæ¢äºé ã»å¶çŽïŒé·ãã圢åŒãæ ¹æ èŠåŠïŒãäºåã«ææåããŸãã
è©äŸ¡èšèšïŒææšïŒæ£ç¢ºæ§ãäžè²«æ§ãæ ¹æ æ§ãå®å šæ§ïŒããµã³ãã«ãååŠåºæºãå¯©æ»æé ãçšæããŸãã
ããã³ããèšèšïŒç°¡æœã»å ·äœã»æ€èšŒå¯èœãååã«ãFew-Shotã®äŸã¯ä»£è¡šæ§ãšã©ãã«ååžãæèããŸãã
ã¢ãã¬ãŒã·ã§ã³ïŒäžèŠçŽ ãã€å€æŽããŠå¹æã枬å®ããå·®åãšçµæãèšé²ããŸãã
å®å šå¯ŸçïŒçŠæ¢ãããã¯ãé瀺äžå¯æ å ±ãå¿çæåŠæ¡ä»¶ãã»ã«ããã§ãã¯æé ãããã³ããã«çµã¿èŸŒã¿ãŸãã
éçšïŒããã³ãããè³ç£åããããŒãžã§ã³ç®¡çã»ã¬ãã¥ãŒã»ããŒãªã³ã°ããã¯ãå®åžžåããŸãã
ãæ¬çš¿ãéããŠåŸãããã¢ã€ãã¢ãç¥èããããªãã®ããžãã¹ã«å°ãã§ã圹ç«ã€ããšãé¡ã£ãŠããŸãããããæ¬çš¿ãåèã«ãªã£ããšæããŠããã ããŸãããããã²ããããããããã©ããŒããããŠããã ãããšå±ã¿ã«ãªããŸããä»åŸãå®è·µçãªããŠããŠãAIææ°ååãå ±æããŠãããŸãã®ã§ãåŒãç¶ããèªã¿ããã ãããšå¬ããã§ãã
泚èšãæ¬çš¿ã¯ãThe Prompt Report: A Systematic Survey of Prompt Engineering TechniquesãïŒSchulhoff et al., arXiv:2406.06608, 2024ïŒã®åæã«åºã¥ããŠããŸããåç ç©¶ã¯2025幎ååãŸã§ã«çºè¡šãããè«æ4,247æ¬ã察象ãšããå æ¬ç調æ»ã§ãããããã³ããæè¡58çš®é¡ã®äœç³»åã詊ã¿ãçŸæç¹ã§æå€§èŠæš¡ã®ãµãŒãã€ã§ããLLMæè¡ãšããã³ããææ³ã¯ç¶ç¶çã«é²åããŠãããããæ¬çš¿ã§ç޹ä»ããæè¡ã®æå¹æ§ã¯ã䜿çšããã¢ãã«ãã¿ã¹ã¯ã«ãã£ãŠå€åããå¯èœæ§ããããŸãã
Appendix
äž»èŠçšèªé
ãããã³ãããšã³ãžãã¢ãªã³ã°ã®è°è«ãæ£ç¢ºã«çè§£ããããã«ãè«æã§å®çŸ©ãããŠããäž»èŠãªçšèªã以äžã«ãŸãšããŸã ã
Prompt: çæAIã¢ãã«ãžã®å ¥åå šè¬ãããã¹ããç»åãé³å£°ãªã©ãã¢ãã«ã®åºåãã¬ã€ãããããã«äžããããããããæ å ±ã
Prompt Template: ããã³ãããçæããããã®é圢ãäžã€ä»¥äžã®å€æ°ãæã¡ãããã«å ·äœçãªå€ïŒäŸïŒãŠãŒã¶ãŒã®è³ªåïŒãæ¿å ¥ããããšã§ãåå¥ã®ããã³ãããäœæãããã
DirectiveïŒæç€ºïŒ: ããã³ããã®äžæ žãšãªããã¢ãã«ã«å®è¡ããŠã»ããã¿ã¹ã¯ãäŒããåœä»€ã質åããããèŠçŽããŠãããã翻蚳ããŠããªã©ã
ExemplarïŒäŸïŒ: Few-Shot Promptingã§çšãããããã¿ã¹ã¯ã®å ¥åºåäŸãshotãšãåŒã°ããã
RoleïŒåœ¹å²ïŒ: ã¢ãã«ã«ç¹å®ã®ãã«ãœãïŒäŸïŒãããªãã¯å»è ã§ããïŒãäžããæç€ºãåºåã®ã¹ã¿ã€ã«ãèŠç¹ãå¶åŸ¡ããã
In-Context Learning (ICL): ããã³ããå ã«äŸãæç€ºãå«ããããšã§ãã¢ãã«ã®éã¿ãæŽæ°ããããšãªããæšè«æã«ã¿ã¹ã¯ã®ããæ¹ãåŠç¿ããããã©ãã€ã ã
Zero-Shot Prompting: äŸïŒExemplarïŒãäžåäžãããæç€ºïŒDirectiveïŒã®ã¿ã§ã¿ã¹ã¯ãå®è¡ãããããšã
Few-Shot Prompting: ããã€ãã®äŸïŒExemplarïŒãäžããŠã¿ã¹ã¯ãå®è¡ãããããšã
Chain-of-Thought (CoT): ã¢ãã«ã«æçµçãªçãã ãã§ãªããããã«è³ããŸã§ã®äžéçãªæšè«ã¹ããããèšèªåãããããã³ããæè¡ã
Self-Consistency: åãåé¡ã«å¯ŸããŠè€æ°ã®ç°ãªãæšè«ãã¹ãçæããããã®çµæã®å€æ°æ±ºãåãããšã§åçã®ä¿¡é Œæ§ãé«ããã¢ã³ãµã³ãã«æè¡ã
Tree-of-Thought (ToT): æèããã»ã¹ãæšæ§é ã§è¡šçŸããè€æ°ã®å¯èœæ§ãåæã«æ¢çŽ¢ã»è©äŸ¡ããããšã§ãè€éãªåé¡è§£æ±ºãè¡ããã¬ãŒã ã¯ãŒã¯ã
Self-Refine: ã¢ãã«èªèº«ã«èªãã®åºåãè©äŸ¡ããããã®ãã£ãŒãããã¯ã«åºã¥ããŠå埩çã«åçãæ¹åãããããã»ã¹ã
AgentïŒãšãŒãžã§ã³ãïŒ: LLMãé è³ãšããŠãå€éšã®ããŒã«ïŒAPIãããŒã¿ããŒã¹ãªã©ïŒãšé£æºããèªåŸçã«ã¿ã¹ã¯ãéè¡ããã·ã¹ãã ã
Answer EngineeringïŒåçãšã³ãžãã¢ãªã³ã°ïŒ: LLMã®åé·ãªåºåãããæ§é åãããæ£ç¢ºãªçããæœåºããããã®äžé£ã®æè¡ãããã»ã¹ã
ãã«ã·ããŒã·ã§ã³: çæAIããäºå®ã«åºã¥ããªãããã£ãšããããåœæ å ±ãçæããçŸè±¡ã
References
ãããã³ãããšã³ãžãã¢ãªã³ã°ã®åéã«ã¯ããã®çºå±ãæ¹åä»ããç»æçãªè«æãããã€ãååšããŸãã以äžã«ãæ¬çš¿ã§è§£èª¬ããè«æã®åèæç®ã®äžã§ãç¹ã«éèŠãšèãããã3ã€ã®ç ç©¶ã玹ä»ããŸã ã
Brown, T. B., et al. "Language models are few-shot learners." Advances in neural information processing systems 33 (2020): 1877-1901.
ãGPT-3ã®é©ç°çãªèœåãäžã«ç¥ããããèšå¿µç¢çãªç ç©¶ã§ããæ¬ç ç©¶ã®æå€§ã®è²¢ç®ã¯ãIn-Context Learning ãç¹ã«Few-Shot Promptingãšããæ°ãããã©ãã€ã ãæå±ããç¹ã«ãããŸããããã¯ãã¢ãã«ã®éã¿ãæŽæ°ãããã¡ã€ã³ãã¥ãŒãã³ã°ãè¡ãããšãªããããã³ããå
ã«å°æ°ã®äŸã瀺ãã ãã§ãLLMãæ§ã
ãªã¿ã¹ã¯ãããªããããšãå®èšŒããŸããããã®çºèŠã«ãããLLMã®æ±çšæ§ãšæè»æ§ãé£èºçã«é«ãŸããçŸåšã®ããã³ãããšã³ãžãã¢ãªã³ã°ã®éçã®ç€ãç¯ããŸããã
Wei, J., et al. "Chain-of-thought prompting elicits reasoning in large language models." Advances in Neural Information Processing Systems 35 (2022): 24824-24837.
ãLLMã®æšè«èœåãåçã«åäžãããChain-of-Thought Promptingãææ¡ããŸãããåã«çããæ±ããã®ã§ã¯ãªãããæ®µéçã«èãããããä¿ããã¢ãã«ã«äžéçãªæèããã»ã¹ãèšèªåãããããšã§ãç¹ã«ç®è¡åé¡ãåžžèæšè«ãè«çããºã«ãšãã£ãè€éãªã¿ã¹ã¯ã«ãããæ£è§£çã倧å¹
ã«åäžããããšã瀺ããŸãããCoTã¯ãLLMãåãªããã¿ãŒã³ãããã³ã°æ©æ¢°ãããããä¿¡é Œæ§ã®é«ãæšè«ãšã³ãžã³ãžãšæè¯ãããéãéããŸããã
Kojima, T., et al. "Large language models are zero-shot reasoners." Advances in neural information processing systems 35 (2022): 22199-22213.
ãCoTã®é©ãã¹ãåŽé¢ãæããã«ããŸãããããã¯ãZero-Shot-CoTã®çºèŠã§ããFew-Shot-CoTã®ããã«æäœæ¥ã§æèããã»ã¹ã®äŸãäœæããªããŠããããã³ããã®æ«å°Ÿã«ãLet's think step by step.ããšãããéæ³ã®èšèããä»ãå ããã ãã§ãLLMãèªçºçã«æšè«èœåãçºæ®ããããšã瀺ããã®ã§ãããã®çºèŠã«ãããCoTæè¡ã¯ããæè»œã§æ±çšçãªãã®ãšãªããå€ãã®å®è·µè
ã«ãšã£ãŠäžå¯æ¬ ãªããŒã«ãšãªããŸããã
â ïžãéèŠïŒChatGPTã䜿ãããªãã人ã¯å
šäœã®3å²ä»¥äž
ãªããïŒã質ååãã®å·®ã§ãã
å®ã¯ãAIãã10åã®äŸ¡å€ãåŒãåºããåãã®ç«ãŠæ¹ãã«ã¯ç§åŠçãªæ³åããããŸããGoogleã20%ã«ãŒã«ã§å®èšŒããAmazonã®ããŸã¹ãæšå¥šããæ¹æ³ãå
šãŠãAIæä»£ã®æåŒ·ã¹ãã«ã奜å¥å¿åããã§å
¬éäžã
AIãç¬æã«ãçãããè¿ãä»ãåè² ã®åããç®ã¯ãåããã«ããããã®ç§å¯ã𿿥ã®ä»äºã§äœ¿ãã13ã®å®è·µããŒã«ãåçž®ããæèãAIæä»£ã®æåŒ·ã¹ãã«ã奜å¥å¿åããããã²ãã§ãã¯ããŠã¿ãŠãã ããã
ãããªãã®èãããã£ãšæå¿«ã«äŒããŸãããïŒã
ãçºå£²ãã25幎以äžãæèæ³ã®"ãã€ãã«"ãšããŠèªã¿ç¶ãããäžæœã®åèãæ¬æžã¯ãã³ã³ãµã«ãã£ã³ã°ã®åéãããã³ãŒãŒã§éçºãããæèã®æŽçè¡ããã©ãããååãã®å
šãŠã解説ããŸãã
ãçµè«ããèããæ ¹æ ãæ§é åããæè¡ã身ã«ã€ããã°ãããªãã®ã¬ããŒãããã¬ãŒã³ã¯åçã«åããããããªããŸããè€éãªåé¡ãã·ã³ãã«ã«æãã説åŸåãé£èºçã«é«ããäžçã¢ãã®ã¹ãã«ãããã«ã
ãæä»£ãè·çš®ãåãããå
šãŠã®ããžãã¹ããŒãœã³å¿
èªã®äžåã§ãã
ãåãã®æ¬è³ªãèŠæããçŸç¶ææ¡âèª²é¡æœåºâ解決çç«æ¡ãšãã3ã¹ããããäœç³»åãæŠç¥ãã¡ãŒã ã®é¢æ¥å¯Ÿçã«ãéçšããâæ¬è³ªçãªè«çæèâãäºäŸãšæŒç¿ã§éããæ§æãç¹é·ã§ãã35äžéšè¶
ã®ãæ±å€§ããŒããã·ãªãŒãºææ°äœã
ãã³ã³ããã®å£²äžãäžããã«ã¯ïŒããšãã£ãå®è·µçãªã±ãŒã¹åé¡ãéããåãªãç¥èããã¬ãŒã ã¯ãŒã¯ã§ã¯ãªããç©äºã®æ¬è³ªãèŠæããæèã®ããã»ã¹ããã®ãã®ã培åºçã«éããŸãã
ãåãªãç¥èã§ã¯ãªããäžç䜿ãããèããåããæã«å
¥ãããæ¹ã«æé©ã就掻çããããã©ã³ããžãã¹ããŒãœã³ãŸã§ãæèåãæ¬¡ã®ã¬ãã«ã«åŒãäžããå¿
èªæžã
ãªãé«åŠæŽè
ãè©æ¬ºã«éããå°éå®¶ãåæ©çãªãã¹ãç¯ãã®ãïŒæ¬æžã¯ãç¥æ§ãé«ãã»ã©é¥ããããæèã®çœ ããšããè¡æçãªçå®ãç§åŠçã«è§£æããŸãã
èè
ãããœã³ã¯ãããŒãã«è³åè³è
ã®å€±æããå»åž«ã®èª€èšºãŸã§è±å¯ãªäºäŸãåæãIQãåŠæŽã ãã§ã¯é²ããªããèªç¥ãã€ã¢ã¹ãã®æ£äœãæŽããŸããèªåã®ç¥èãéä¿¡ããæ¹å€çæèãæ ãããšã§ããããè³¢ã人ã»ã©å€§ããªéã¡ãç¯ãããããšããé説ã
æ¬æžãæç€ºããã®ã¯åãªãèŠåã§ã¯ãããŸããããã¡ã¿èªç¥ããç¥çè¬èãããªã©ãçã®è³¢ãã身ã«ã€ããå
·äœçæ¹æ³ãæºèŒãããžãã¹ã§ã®æææ±ºå®ãæè³å€æãæ¥åžžç掻ãŸã§ãããããå Žé¢ã§åœ¹ç«ã€å®è·µçãªç¥æµãè©°ãŸã£ãŠããŸãã
ãèªåã¯å€§äžå€«ããšæã£ãŠãã人ã»ã©èªãã¹ãäžåãç¥æ§ã®éçãç¥ãããšã§ãæ¬åœã®ç¥æ§ã身ã«ã€ããŸãã
- #AI
- #çæAI
- #ChatGPT
- #AI掻çš
- #Claude
- #ããã³ãã
- #æè¡
- #人工ç¥èœ
- #ã¢ãã«
- #LLM
- #æ©æ¢°åŠç¿
- #ã¿ã¹ã¯
- #ããã³ãããšã³ãžãã¢ãªã³ã°
- #ææ³
- #aiããã³ãã
- #ããã³ããèšèš
- #æè¡è§£èª¬
- #AIç ç©¶
- #AIãã¯ãããžãŒ
- #ChainOfThought
- #AIè«æè§£èª¬
- #treeofthoughts
- #ThePromptReport