
OpenAIè«æãæŽãããæ£è§£çè³äžäž»çŸ©ãã®ä»£å
LLMããäºå®ãšç°ãªãæ å ±ãèªä¿¡æºã ã«çããçŸè±¡ã¯ããã«ã·ããŒã·ã§ã³ããšåŒã°ããŠããŸããææ°ã®é«æ§èœã¢ãã«ã§ãå®å šã«ã¯é¿ããããªããã®åé¡ã«ã€ããŠãOpenAIã®ç ç©¶ããŒã ã2025幎ã«çºè¡šããè«æãWhy Language Models Hallucinateãã¯ãåŸæ¥ãšã¯ç°ãªãèŠç¹ããåå ãè§£æããŸããã
ãèè ãã¯ããã«ã·ããŒã·ã§ã³ãåãªãæè¡çæ¬ é¥ã§ã¯ãªããçµ±èšçãªå¿ ç¶æ§ãšè©äŸ¡ã·ã¹ãã ã®æ§é çåé¡ã«èµ·å ããããšã蚌æããŠããŸããç¹ã«æ³šç®ãã¹ãã¯ãçŸåšäž»æµã®è©äŸ¡ææšããæ£è§£çããåéããããŸããã¢ãã«ã«ãããããªããŠãçããæ¹ãåŸããšããè¡åãåŠç¿ãããŠãããšããææã§ãã
ãæ¬çš¿ã§ã¯ããã®è«æã瀺ããã«ã·ããŒã·ã§ã³ã®çºçã¡ã«ããºã ãšãææ¡ããã解決çã«ã€ããŠè©³ãã解説ããŸããAIéçºã®è©äŸ¡åºæºãã®ãã®ãåãçŽããšãããæè¡è«ãè¶ ããæ¬è³ªçãªè°è«ã¯ãçæAIã®ä¿¡é Œæ§åäžãèããäžã§éèŠãªç€ºåãäžããŠãããŸãã
Spotifyã§ããããããé³å£°é ä¿¡ïŒããã¿ã®AIããã¯ã©ãžãªã
Read more:Â https://openai.com/index/why-language-models-hallucinate/
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
ãç ç©¶è ã®Adam Kalaiæ°ãèªèº«ã®åå£«è«æã¿ã€ãã«ãChatGPTã«å°ãããšããã3å詊ããŠ3ã€ãšãç°ãªã誀çãè¿ã£ãŠããŸãããèªçæ¥ãèããŠãåæ§ã«ãæ¯åéãæ¥ä»ãçããªããå šãŠééã£ãŠãããšãããŸãããã®ããã«èšèªã¢ãã«ã®å¹»èŠãšã¯ããç¥ããªãããšã«å¯ŸããŠãç¥ããªãããšèšããããã£ãšãããã誀æ å ±ãäœãåºããŠããŸãçŸè±¡ããæããŸãã
ããããŸã§ãã«ã·ããŒã·ã§ã³ã®åå ã¯ãã¢ãã«ã®ç¥èäžè¶³ãèŠæš¡ã®åé¡ãšããŠæããããã¡ã§ããããããæ¬è«æã¯ãã©ããªã«é«ç²ŸåºŠãªã¢ãã«ã§ãçµ±èšçã«é¿ããããªã誀çãååšããããšãæ°åŠçã«èšŒæããããã«çŸåšã®è©äŸ¡æ¹æ³ããã®åé¡ãå©é·ããŠããããšãæããã«ããŠããŸãã
ãèè ããæç€ºããã®ã¯ãã¢ãã«ã®æ§èœåäžãšããæè¡ç解決ã§ã¯ãªããè©äŸ¡åºæºã®æ ¹æ¬çãªèŠçŽããšãããç€ŸäŒæè¡çãã¢ãããŒãã§ãã以äžããã®é©æ°çãªåæãã©ã®ããã«AIéçºã®åžžèãèŠãã®ããé ã远ã£ãŠèŠãŠãããŸãã
èæ¯
解決ãã課é¡
ãæ¬çš¿ã§æ±ãäžå¿çãªèª²é¡ã¯ãLLMã«ãããããã«ã·ããŒã·ã§ã³ãã§ããããã¯ãã¢ãã«ãçæããåºåããåãªã誀ããæå³ããªããªãæååãšã¯äžç·ãç»ããå·§åŠãªæ§è³ªãæã€çŸè±¡ãæããŸãã
ããã«ã·ããŒã·ã§ã³ãšã¯ããã£ãšãããããäºå®ãšã¯ç°ãªãããããã¯äžæ£ç¢ºãªæ å ±ããããããçå®ã§ãããã®ããã«èªä¿¡ãæã£ãŠçæããããšãšå®çŸ©ãããŸã ããã®ããã£ãšããããããããããã«ã·ããŒã·ã§ã³ãç¹ã«åä»ãªåé¡ãããããŠããã®ã§ãã Â
ããã®çŸè±¡ã®å ·äœäŸãšããŠãè«æã®èè ã§ããAdam Tauman Kalaiæ°ã®åå£«è«æã®ã¿ã€ãã«ã3ã€ã®èåãªLLMã«å°ããéã®å¿çãèŠãŠã¿ãŸãããã

ããã®è¡šã瀺ãããã«ãåã¢ãã«ã¯ããããç°ãªãããããåŠè¡çã«ããããããªã¿ã€ãã«ãšå€§åŠåã幎ãçæããŸãããæ£è§£ãç¥ããªã人éãèŠãã°ãã©ããä¿¡ããŠããŸããããããŸããã
ãããããã«ã·ããŒã·ã§ã³ã®æ¬è³ªã§ããåæ§ã«ãèè ã®èªçæ¥ãå°ããéã«ã¯ãããã¢ãã«ã3åã®è©Šè¡ã§3ã€ã®ç°ãªã誀ã£ãæ¥ä»ãçæããäŸãå ±åãããŠããŸã ã Â
ããã®ãããªäºäŸã¯ãåãªãããªãã¢ã®ééãã«ãšã©ãŸããŸããããã®åœ±é¿ã¯ãåžæ³ãå»çãéèãšãã£ã瀟äŒã®æ ¹å¹¹ããªãé åã«ãŸã§åã³ãŸã ãåè¿°ã®åœã®å€äŸãåŒçšããåŒè·å£«ã®äºäŸã¯ãåžæ³ã®ä¿¡é Œæ§ãæºããããããŸããÂ
ãå»çåéã§ã¯ã誀ã£ãæ²»çæ³ã蚺æãAIãæç€ºãããªã¹ã¯ãææãããŠãããçåœã«é¢ããå±éºæ§ãã¯ããã§ããŸã ãããžãã¹ã®çŸå Žã§ã¯ãäžæ£ç¢ºãªåžå Žåæã顧客察å¿ããäŒæ¥ã®è©å€ã財åã«çŽæ¥çãªæå®³ãäžããå¯èœæ§ããããŸã ã
ããããã£ãŠããã«ã·ããŒã·ã§ã³ã®å æã¯ãAIæè¡ã瀟äŒã«è²¬ä»»ãã圢ã§å®è£ ãããããã®ãé¿ããŠã¯éããªãéèŠãªèª²é¡ãªã®ã§ãã Â
æ¢åç ç©¶ã®æµã
ããã«ã·ããŒã·ã§ã³ãšãã課é¡ã«å¯ŸããAIç ç©¶ã³ãã¥ããã£ã¯ãããŸã§æ§ã ãªã¢ãããŒãã詊ã¿ãŠããŸããããããã¯äž»ã«ãçºçããŠããŸã£ããã«ã·ããŒã·ã§ã³ãããã«ããŠæžãããããšããã察ççæ³çããªã¢ãããŒãã«åé¡ã§ããŸãã
ãäž»èŠãªææ³ã®äžã€ããæ€çŽ¢æ¡åŒµçæïŒRAGïŒã§ããããã¯ãLLMãå¿çãçæããéã«ããŠã£ãããã£ã¢ã®ãããªä¿¡é Œæ§ã®é«ãå€éšã®ç¥èæºããäŒæ¥å ã®ææžããŒã¿ããŒã¹ãªã©ããªã¢ã«ã¿ã€ã ã§åç §ããæè¡ã§ã ã
ãLLMãå éšã®ç¥èã ãã«é Œãã®ã§ã¯ãªããæ€èšŒå¯èœãªæ å ±æºã«ãæ ¹æ ïŒã°ã©ãŠã³ãã£ã³ã°ïŒããæ±ããããšã§ãäºå®ã«åºã¥ããªãå¿çãæå¶ããããšããèãæ¹ã§ãã Â
ãããäžã€ã®æåãªã¢ãããŒããã人éã®ãã£ãŒãããã¯ããã®åŒ·ååŠç¿ïŒRLHFïŒã§ããããã¯ããŸã人éãã¢ãã«ã®è€æ°ã®å¿çãæ¯èŒãããã©ã¡ããè¯ããããã©ã³ã¯ä»ãããŸãã
ããã®å€§éã®æ¯èŒããŒã¿ãçšããŠãå ±é ¬ã¢ãã«ããåŠç¿ãããæçµçã«LLMæ¬äœããã®å ±é ¬ã¢ãã«ããé«ãè©äŸ¡ãåŸããããããªå¿çãçæããããã匷ååŠç¿ãéããŠãã¡ã€ã³ãã¥ãŒãã³ã°ïŒåŸ®èª¿æŽïŒããææ³ã§ã ã
ããã®ããã»ã¹ã«ãããã¢ãã«ã®å¿çããã人éã®å¥œã¿ã䟡å€èŠ³ã«æ²¿ã£ããã®ãžãšãã¢ã©ã€ã³ïŒæŽåïŒããããäžæ£ç¢ºã§äžé©åãªåºåãæžããããšãç®æããŸããè¿å¹Žã§ã¯ãRLHFã®è€éããè§£æ¶ããããå¹ççã«åæ§ã®å¹æãåŸãçŽæ¥çå奜æé©åïŒDPOïŒãšããææ³ãç»å ŽããŠããŸã ã Â
ããããã®æè¡ã¯ç¢ºãã«ãã«ã·ããŒã·ã§ã³ã®çºççãäœæžãããäžã§äžå®ã®ææãäžããŠããŸããããããããããã®ã¢ãããŒãã«ã¯å ±éã®éçããããŸããããã¯ããã«ã·ããŒã·ã§ã³ããã¢ãã«ãåŠç¿ãçµããåŸã«ä¿®æ£ãã¹ããè¡åäžã®åé¡ããšããŠæ±ã£ãŠããç¹ã§ãã
ãRAGãRLHFããããã°çç¶ãæããããã®è¬ã®ãããªãã®ã§ããããªããã®ç æ°ãããªãã¡ãã«ã·ããŒã·ã§ã³ãããããçºçããã®ãããšããæ ¹æ¬åå ã«ã¯èžã¿èŸŒãã§ããŸããã
ãã®ç ç©¶ã解決ãã課é¡ã»ã©ã解決ããã®ã
ãæ¬çš¿ã§è§£èª¬ããKalaiæ°ãã®ç ç©¶ã¯ããã®æ ¹æ¬åå ã®è§£æã«çæ£é¢ããåãçµã¿ãŸãããã®ç ç©¶ã解決ããããšãã課é¡ã¯ããã«ã·ããŒã·ã§ã³ã®ãçºçæºããç¹å®ããããšã§ãã
ãç ç©¶è ãã¯ããã«ã·ããŒã·ã§ã³ã¯ã¢ãã«ã®åŠç¿ããã»ã¹ã«ãããç°åžžãæ¬ é¥ã§ã¯ãªãããããçŸåšã®æšæºçãªèšç·Žæ¹æ³ãããããèªç¶ãªåž°çµã§ãããšäž»åŒµããŸã ã Â
ããã®åé¡ã解決ããããã«ã圌ãã¯åŸæ¥ã®ãã©ãããã°æžããããããšããå·¥åŠçãªåããããããªãããã€çºçããã®ãçµ±èšçã«å¿ ç¶ãªã®ãããšãããããåºç€çãªç§åŠçåããžãšèŠç¹ã転æããŸãããããŠããã®åãã«çããããã®çè«çãªãã¬ãŒã ã¯ãŒã¯ãæ§ç¯ããã®ã§ãã
ããã®ç ç©¶ã®ã¢ãããŒãã¯ããã«ã·ããŒã·ã§ã³ãšããç æ°ã®ãæ²»çæ³ããéçºããã®ã§ã¯ãªãããã®ç æ°ã®ãç«åŠèª¿æ»ããè¡ãããšã«äŒŒãŠããŸããæ¢åã®ç ç©¶ãRAGãRLHFãšãã£ãæ²»çè¬ã®éçºã«æ³šåããŠããã®ã«å¯ŸããKalaiæ°ãã¯ããã«ã·ããŒã·ã§ã³ãã©ã®ãããªãç°å¢ããããªãã¡ã©ã®ãããªçµ±èšçæ¡ä»¶äžã§çºçããããã®ããçªãæ¢ããããšããŸãã
ã圌ãã®åæãæããã«ããã®ã¯ãLLMã®èšç·Žç®çãã®ãã®ïŒèšç·ŽããŒã¿ã«å¯Ÿããäºæž¬ç²ŸåºŠãæå€§åããããšïŒãšãAIã®æ§èœã枬ãè©äŸ¡æ¹æ³ïŒæ£è§£ãäžæ£è§£ãã®äºè æäžã§æ¡ç¹ããããšïŒããã¢ãã«ã«å¯ŸããŠãäžç¢ºããªãšãã«ã¯ç¥ããªããšèªããããããã£ãšããããæšæž¬ãããããšãã匷åãªçµ±èšçå§åããããŠãããšããäºå®ã§ã ã
ãã€ãŸãããã«ã·ããŒã·ã§ã³ã®æ ¹æ¬åå ã¯ãã¢ãã«ãšãããæ£è ãã®å éšã«ããã®ã§ã¯ãªããã¢ãã«ãåãå·»ããèšç·Žãšè©äŸ¡ã®ãšã³ã·ã¹ãã ããšããç°å¢ã«ããããšããã®ã圌ãã®èšºæã§ãããã®èšºæã«åºã¥ãã圌ãã¯å¯Ÿççæ³ã§ã¯ãªããç°å¢ãã®ãã®ãå€é©ãããããæ ¹æºçãªè§£æ±ºçãæèšããã®ã§ãã
æ¹æ³è«
ææ¡ææ³
ãKalaiæ°ãã®ç ç©¶ãæç€ºããåæã®æ žå¿ã¯ãLLMã®èšç·Žããã»ã¹ãäºã€ã®æ®µéã«åããŠèå¯ããããšã«ãããŸãã
äºååŠç¿æ®µé
ãããã§ã¯ãã€ã³ã¿ãŒãããäžã®èšå€§ãªããã¹ãããŒã¿ããèšèªã®çµ±èšçãã¿ãŒã³ãåŠç¿ããåºç€ãšãªããããŒã¹ã¢ãã«ããæ§ç¯ãããŸãããã®æ®µéã§ããªããšã©ãŒãã²ããŠã¯ãã«ã·ããŒã·ã§ã³ã®ãçš®ããèãããã®ãã説æããŸãã
äºåŸåŠç¿æ®µé
ãããŒã¹ã¢ãã«ã人éã®å¥œã¿ãç¹å®ã®ã¿ã¹ã¯ã«é©åããããããRLHFãªã©ã®ææ³ã§åŸ®èª¿æŽãè¡ãããŸãããã®æ®µéã§ããªãåãªããšã©ãŒã§ã¯ãªãããé床ã«èªä¿¡ãæã£ãããã«ã·ããŒã·ã§ã³ãçãæ®ãããããå©é·ãããŠããŸãã®ããè§£ãæãããŸã ã Â
ããã®åæã®æ ¹å¹¹ããªãç¬åµçãªã¢ã€ãã¢ããåé¡ã®ãéå ãã§ãã圌ãã¯ãäžèŠãããšéåžžã«è€éãªãæå¹ãªããã¹ããçæããããšããã¿ã¹ã¯ããçµ±èšçã«ã¯ãããåçŽãªãäžããããããã¹ããæå¹ãåŠããåé¡ããããšããäºå€åé¡ã¿ã¹ã¯ãããå°é£ã§ããããšãæ°åŠçã«ç€ºããŸã ã
ããã®éå ãããããã«ã·ããŒã·ã§ã³ãšããäžå¯è§£ãªçŸè±¡ã«ãçµ±èšåŠãšèšç®åŠç¿çè«ã®å ãåœãŠãéµãšãªãã®ã§ãã Â
ææ¡ææ³ã®çŽæçãªèª¬æ
ããã®ãéå ããšããèãæ¹ããããçŽæçã«çè§£ããããã«ã次ã®ãããªã¢ãããžãŒãèããŠã¿ãŸãããã
ãããæŽå²åŠãåŠã¶åŠçããããšããŸãã圌ã«èª²ããããã¿ã¹ã¯ã¯äºã€ãããŸããäžã€ã¯ããæŽå²çã«æ£ç¢ºãªå°è«æãå·çãããããšïŒããã¯LLMã®ãçæãã¿ã¹ã¯ã«çžåœããŸãïŒãããäžã€ã¯ããæç€ºãããæŽå²ã«é¢ããèšè¿°ããæ£ããã誀ã£ãŠããããå€å®ãããããšïŒããã¯ãåé¡ãã¿ã¹ã¯ã«çžåœããŸãïŒã
ãåžžèçã«èããŠãåŸè ã®å€å®ã¿ã¹ã¯ãããåè ã®å·çã¿ã¹ã¯ã®æ¹ãã¯ããã«é£ããããšã¯æããã§ãããªããªããåªããå°è«æãæžãããã«ã¯ãç¡æ°ã®åè£ãšãªãæç« ã®äžããæ£ãããã®ãéžæããããããè«ççã«æ§æããå¿ èŠãããããã§ãã
ãããã¯ãé ã®äžã§çµ¶ããããã®èšè¿°ã¯æ£ãããïŒããšããå€å®ãç¹°ãè¿ãããã»ã¹ãå å ããŠããŸãã
ãKalaiæ°ãã®è«æã䞻匵ããã®ã¯ããããã®åŠçããããããåã ã®æŽå²çäºå®ã®æ£èª€ãå€å®ããèœåïŒåé¡èœåïŒã«æ¬ ããŠãããªãã°ã圌ãå·çããå°è«æã«èª€ããå«ãŸããã®ã¯é¿ããããªãããšããããšã§ãã
ããããŠããã®é¢ä¿æ§ãæ°åŠçã«å®åŒåããçæã¿ã¹ã¯ã«ããããšã©ãŒçããæ ¹åºã«ããåé¡ã¿ã¹ã¯ã®èª€åé¡çã«ãã£ãŠäžéãå®ããããããšã蚌æããã®ã§ã ã Â
ãããã«ããã®è«æã¯LLMã®èšç·Žããã»ã¹ã§çšãããã亀差ãšã³ããããŒæå€±ãšããæŠå¿µã«ãå ãåœãŠãŸãããã®æ¬è³ªã¯ãã¢ãã«ãèšç·ŽããŒã¿ã®ãã¿ãŒã³ãã©ãã ãäžæãäºæž¬ã§ããããæž¬ãæçžŸè¡šãã®ãããªãã®ã§ã ã
ãã¢ãã«ã¯ããã®æçžŸè¡šã§ããè¯ãç¹æ°ãåãããã«ïŒã€ãŸãæå€±ãæå°åããããã«ïŒèšç·ŽãããŸãããã®èšç·Žãç©ãã çµæãã¢ãã«ã¯ããã£ãªãã¬ãŒã·ã§ã³ããããç¶æ ã«ãªããŸãã
ãããã¯ãã¢ãã«ãåºåãã確çïŒäŸïŒããã®æç« ãæ£ãã確çã¯70%ãïŒããçµ±èšçã«èŠãŠä¿¡é Œã§ãããã®ã«ãªãããšããæå³ã§ã ã Â
ãäžèŠãããšãããã¯æãŸããæ§è³ªã§ããããããè«æãæããã«ããé©ãã¹ãäºå®ã¯ããã®ãåªããäºæž¬åšãã«ãªãããã»ã¹ãã®ãã®ããç¹å®ã®æ¡ä»¶äžã§ã¢ãã«ã«ãã«ã·ããŒã·ã§ã³ã匷å¶ãããšããé説çãªã¡ã«ããºã ã§ã ã
ãã€ãŸããæ£çŽã§æ£ç¢ºãªäºæž¬è ã«ãªãããšããã°ããã»ã©ãããçš®ã®åãã€ããããåŸãªããªããšãããçµ±èšçãªãžã¬ã³ããååšããã®ã§ãã Â
ææ¡ææ³è©³çŽ°
ãããã§ã¯ãåæã®ç¬¬äžæ®µéã§ããäºååŠç¿ã«ããããšã©ãŒçºçã®ã¡ã«ããºã ã詳ããèŠãŠãããŸããããäºååŠç¿ã®ç®çã¯ãäžããããèšå€§ãªããã¹ãã³ãŒãã¹ã®æ ¹åºã«ããèšèªã®ç¢ºçååžãåŠç¿ããããšãããªãã¡ãå¯åºŠæšå®ãã§ã ã Â
ããã®ããã»ã¹ã«ãããŠãããšãèšç·ŽããŒã¿ãå®å šã«æ£ç¢ºã§ãäžåã®èª€æ å ±ãå«ãŸãªãçæ³çãªãã®ã§ãã£ããšããŠããçµ±èšçãªå§åã«ãã£ãŠãšã©ãŒãçããããšè«æã¯äž»åŒµããŸãããã®äž»åŒµãæ¯ããã®ããåè¿°ã®ãIs-It-ValidïŒIIVïŒãäºå€åé¡åé¡ãšããçè«çããŒã«ã§ã ã Â
IIVåé¡ã¯ã以äžã®ããã«èšå®ãããŸãã
åŠç¿ããŒã¿
ãåŠç¿ããŒã¿ã¯ãäºã€ã®èŠçŽ ããæ§æãããæ··åããŒã¿ã§ããååã¯ãå®éã®ïŒæå¹ãªïŒèšç·ŽããŒã¿ããåããããµã³ãã«ã§ããããã«ã¯ã+ïŒæå¹ïŒããšããã©ãã«ãä»ããããŸããããååã¯ãã©ã³ãã ã«çæãããïŒç¡å¹ãªïŒæååã§ããããã«ã¯ã-ïŒç¡å¹ïŒããšããã©ãã«ãä»ããããŸãã
ã¿ã¹ã¯
ããã®æ··åããŒã¿ãçšããŠãããåé¡åšãèšç·Žããæªç¥ã®æååãäžãããããšãã«ããããã+ããªã®ãã-ããªã®ããå€å®ãããŸãã
ãè«æã§ã¯ãä»»æã®LLMãããã®IIVåé¡ã®åé¡åšãšããŠå©çšã§ããããšã瀺ããŸããå ·äœçã«ã¯ãLLMãããæååãçæãã確çãèšç®ãããã®ç¢ºçãç¹å®ã®éŸå€ããé«ããã°ã+ããäœããã°ã-ããšå€å®ããã®ã§ã ã Â
ããã®æ çµã¿ãçšããããšã§ã圌ãã¯çæã¢ãã«ã®ãšã©ãŒãšåé¡ã¢ãã«ã®ãšã©ãŒãçŽæ¥çµã³ã€ããã以äžã®éèŠãªæ°åŠçé¢ä¿ãå°ãåºããŸããã
(çæãšã©ãŒç)â¥2Ã(IIV誀åé¡ç)
ããã®æ°åŒã䜿ããªã衚çŸã¯ããLLMãèåœãçæãã確çã¯ããã®LLMãçåœãèŠåããããªã確çã®ãå°ãªããšã2åã¯ããããšããããšã§ãã
ããããããçš®ã®äºå®ã«ã€ããŠçåœã®å€å¥ãæ¬è³ªçã«å°é£ïŒIIV誀åé¡çãé«ãïŒã§ãããªãã°ããã®çš®ã®äºå®ã«ã€ããŠLLMããã«ã·ããŒã·ã§ã³ãèµ·ããïŒçæãšã©ãŒçãé«ãïŒããšã¯ãçµ±èšçã«é¿ããããªãéåœãšãªãã®ã§ãã
ããã®é¢ä¿æ§ã¯ããã«ã·ããŒã·ã§ã³ãåãªãå¶ç¶ã®ç£ç©ã§ã¯ãªããæ ¹åºã«ããåŠç¿åé¡ã®å°é£ãã«èµ·å ãããæ§é çãªåé¡ã§ããããšã瀺åããŠããŸãã
ææ¡ææ³ã®æ§æã³ã³ããŒãã³ãããä»çµã¿ã®è©³çް
ããªããIIVåé¡ã«ããã誀åé¡ãã²ããŠã¯LLMã®ãã«ã·ããŒã·ã§ã³ãçºçããã®ã§ãããããè«æã¯ããã®åå ãšãªãäž»èŠãªçµ±èšçèŠå ãããã€ãæããŠããŸãããããã¯ãæ©æ¢°åŠç¿ã®åéã§é·å¹Žç ç©¶ãããŠãããšã©ãŒçºçã®èŠå ãšè»ãäžã€ã«ããŠããŸã ã Â
1. ä»»æäºå®ãšèªèè«çäžç¢ºå®æ§ïŒèªçæ¥ã®åé¡ïŒ
ãããã¯ãããŒã¿ã®äžã«åŠç¿å¯èœãªç°¡æœãªããã¿ãŒã³ãããæ³åããååšããªãå Žåã«çããåé¡ã§ããäŸãã°ãæå人ã§ã¯ãªãäžè¬äººã®èªçæ¥ã¯ããã®äººã®ååããäºæž¬ããããšã¯ã§ããŸããã
ããããã®äºå®ã¯ãåå¥ã«èšæ¶ãããããªããä»»æäºå®ãã§ã ãã¢ãã«ã«ãšã£ãŠããããã®äºå®ã«é¢ããç¥èã¯èšç·ŽããŒã¿ã«ååšãããåŠããå šãŠã§ãããååšããªãå Žåã¯ãç¥ããªããç¶æ ãããªãã¡èªèè«çäžç¢ºå®æ§ã«çŽé¢ããŸãã
ããã®ç¶æ³ãåæããããã«ãè«æã¯ã°ããã»ãã¥ãŒãªã³ã°æšå®ãšããå€å žçãªçµ±èšåŠã®æŠå¿µãæŽçšããŸãããã®æŠå¿µãçŽæçã«çè§£ããããã«ãæªç¥ã®å³¶ã®çç©å€æ§æ§ã調æ»ããæ¢æ€å®¶ãæ³åããŠã¿ãŠãã ãã ã
ãæ¢æ€å®¶ã100å¹ã®çç©ãæç²ãããã®ãã¡ãµã«ã¯10å¹ããã«ã²ã¯5å¹èŠã€ãã£ãäžæ¹ã§ãã¯ã¢ã¯ãã£ã1å¹ããèŠã€ãããªãã£ããšããŸãããã®ãšãããæ¬¡ã«æç²ããçç©ãããŸã äžåºŠãèŠãããšã®ãªãæ°çš®ã§ãã確çãã¯ã©ããããã§ããããïŒ
ãã°ããã»ãã¥ãŒãªã³ã°æšå®ã¯ããã®ãæªç¥ãšã®ééãã®ç¢ºçããããã£ãäžåºŠãã芳å¯ãããªãã£ãçš®ã®å²åãã䜿ã£ãŠå·§ã¿ã«æšå®ããŸãããã®äŸã§ã¯ããã¯ã¢ãã®ããã«äžåºŠããèŠã€ãããªãã£ãçš®ïŒãããã·ã³ã°ã«ãã³ãšåŒã³ãŸãïŒã®ååšãããŸã èŠã¬æ°çš®ã®ååšã瀺åããéèŠãªæããããšãªãã®ã§ãã
ããããLLMã®èšç·ŽããŒã¿ã«åœãŠã¯ããŠã¿ãŸããããèšç·ŽããŒã¿ã«å«ãŸããäºå®ã®ãã¡ãäžåºŠããç»å Žããªãäºå®ïŒã·ã³ã°ã«ãã³ïŒã®å²åã¯ãèšç·ŽããŒã¿ã«ã¯å«ãŸããŠããªããå®éã«ã¯çå®ã§ããäºå®å šäœïŒãããããã·ã³ã°ãã¹ãšåŒã³ãŸãïŒã®ç¢ºçãæšå®ããè¯ãææšãšãªããŸã ã Â
ãèŒæ£ãããïŒïŒæ£çŽãªïŒLLMã¯ããã®ããã·ã³ã°ãã¹ã®ååšãèªèããŠããŸãããã®ãããäžåºŠããèŠãããšã®ãªãäºå®ããäžåºŠãèŠãããšã®ãªãäºå®ã«ã€ããŠè³ªåããããšãã確信ãæã£ãŠçããããšãã§ããŸããã
ããã®çµæãã¢ãã«ã¯æšæž¬ãããããåŸãªããªãããã«ã·ããŒã·ã§ã³ãçæããŸããè«æã¯ããã®çš®ã®ä»»æäºå®ã«å¯Ÿãããã«ã·ããŒã·ã§ã³ã®çºççã®äžéããèšç·ŽããŒã¿ã«ãããã·ã³ã°ã«ãã³çïŒäžåºŠããåºçŸããªãäºå®ã®å²åïŒã«ãã£ãŠæ±ºãŸãããšãæ°åŠçã«èšŒæããŠããŸã ã Â
2. äžé©åãªã¢ãã«ãšè¡šçŸèœåã®éçïŒæåæ°ã®åé¡ïŒ
ãããã¯ãã¢ãã«ã®å éšæ§é ãæ å ±ã®è¡šçŸæ¹æ³ããç¹å®ã®ã¿ã¹ã¯ãè§£ãã®ã«é©ããŠããªãå Žåã«çããåé¡ã§ã ãè«æã§æããããŠãããDEEPSEEKãšããåèªã«å«ãŸããDã®æåæ°ãæ°ããããšããäŸããããã端çã«ç€ºããŠããŸã ã
ã人éã«ãšã£ãŠã¯ç°¡åãªãã®ã¿ã¹ã¯ããå€ãã®LLMã¯é »ç¹ã«ééããŸãããã®äžå ã¯ãLLMãåèªãæåã®éãŸããšããŠã§ã¯ãªãããããŒã¯ã³ããšåŒã°ããæå³ã®å¡ïŒäŸãã° Â DEEP ãš SEEKïŒãšããŠèªèããŠããããã§ãã
ãæåãäžã€äžã€æ°ãããšããæäœã¯ããã®ããŒã¯ã³ããŒã¹ã®è¡šçŸæ¹æ³ãšã¯çžæ§ãæªããLLMã¯ã¿ã¹ã¯ã«å¯ŸããŠãäžé©åãªã¢ãã«ããšãªã£ãŠããã®ã§ãã
ããã®ã¢ãã«ã®è¡šçŸèœåã®éçãçè«çã«æããæŠå¿µãšããŠãVC次å ããããŸããVC次å ãšã¯ãã¢ãã«ãã©ãã ãè€éã§å€æ§ãªãã¿ãŒã³ãèå¥ã§ãããã瀺ããæè»æ§ããã衚çŸåãã®ææšã§ã ã
ãVC次å ãé«ãã¢ãã«ã»ã©ãè€éãªããŒã¿ååžã«é©åããèœåãé«ããšèšããŸãããããããã®æè»æ§ããã ãšãªãããšããããŸãã
ãè§£ãã¹ãåé¡ã®æ§é ãã¢ãã«ã®è¡šçŸèœåãè¶ ããŠããå Žåããéã«ãéãããããŒã¿ã«å¯ŸããŠã¢ãã«ãè€éãããïŒVC次å ãé«ãããïŒå Žåã«ã¯ãã¢ãã«ã¯ããŒã¿ã«éå°ã«é©åïŒéåŠç¿ïŒãããã€ãºãŸã§åŠç¿ããŠããŸããçµæãšããŠèª€ã£ããã¿ãŒã³ãçæãããªãã¡ãã«ã·ããŒã·ã§ã³ãèµ·ããããããªãã®ã§ãã Â
3. äºåŸåŠç¿ãšã€ã³ã»ã³ãã£ãã®äžæŽå
ãåæã®ç¬¬äºæ®µéã¯ãäºååŠç¿ãçµããã¢ãã«ãããªãäºåŸåŠç¿ãçµãŠããªãããã«ã·ããŒã·ã§ã³ãçæãç¶ããã®ãããšããåãã«çããŸããRLHFãªã©ã®äºåŸåŠç¿ã¯ããã«ã·ããŒã·ã§ã³ãæå¶ããããã«å°å ¥ãããã¯ãã®ããã»ã¹ã§ãã
ãããããè«æã¯ãçŸåšã®AIéçºã³ãã¥ããã£ã§äž»æµãšãªã£ãŠããè©äŸ¡æ¹æ³ãã®ãã®ãããã«ã·ããŒã·ã§ã³ãå©é·ããæ§é çãªæ¬ é¥ãæ±ããŠãããšéãææããŸã ããã®æ žå¿ã«ããã®ããè«æãæç€ºãããåªããåéšè ããšãã匷åãªã¢ãããžãŒã§ãã
ãæ³åããŠã¿ãŠãã ãããããªãã¯ãééããŠãæžç¹ãããªãå€è¢éžæåŒã®è©ŠéšãåããŠããŸãããã®ãšããåæ Œç¹ãæå€§åããããã®æé©ãªæŠç¥ã¯äœã§ããããïŒ
ãããã¯ãåãããªãåé¡ããã£ãŠã決ããŠç©ºæ¬ã«ãããå¿ ãäœããã®éžæè¢ãæšæž¬ããŠããŒã¯ããããšã§ãã空æ¬ã¯ç¢ºå®ã«0ç¹ã§ãããæšæž¬ããã°å¶ç¶æ£è§£ããå¯èœæ§ãããããã§ãã
ãçŸåšã®LLMã®æ§èœè©äŸ¡ã§åºãçšããããŠãããã³ãããŒã¯ïŒMMLU, SWE-benchãªã©ïŒã®å€ããããããšå šãåãã€ã³ã»ã³ãã£ãæ§é ãæã£ãŠããããšè«æã¯è«ããŸã ã
ããããã®ãã³ãããŒã¯ã¯ãåºæ¬çã«æ£è§£ãäžæ£è§£ãã®äºå€æ¡ç¹ãæ¡çšããŠããããåãããŸããããšããèª å®ãªå¿çã«ã¯ãäžæ£è§£ãšåãã0ç¹ãäžããããŸããäžæ¹ã§ãèªä¿¡ããªããŠãäœããæšæž¬ããŠçããå Žåããããå¶ç¶æ£è§£ããã°åŸç¹ã«ãªããŸãã
ããããã£ãŠããããã®ãã³ãããŒã¯ã®ã¹ã³ã¢ãæå€§åããããã«æé©åãããLLMã¯ãå¿ ç¶çã«ãäžç¢ºå®æ§ã衚æããæ£çŽè ãã§ã¯ãªããããªã¹ã¯ãåããŠã§ãæšæž¬ããåªããåéšè ãã«ãªãããèšç·ŽãããŠãããŸãããããããããã«ã·ããŒã·ã§ã³ãäºåŸåŠç¿åŸãæ ¹åŒ·ãæ®ããè«æããç€ŸäŒæè¡çããšåŒã¶æ ¹æ¬åå ãªã®ã§ã ã Â
ããã®åæãå°ãåºããããæ·±ãã¬ãã«ã®çµè«ã¯ãçŸåšã®AIç ç©¶éçºã³ãã¥ããã£ã«ãããã鲿©ãã®å®çŸ©ãã®ãã®ããAIã®ãä¿¡é Œæ§ããšããç®æšãšæ ¹æ¬çã«ççŸããŠãããšããç¹ã§ãã
ãAIã®æ§èœã©ã³ãã³ã°ïŒãªãŒããŒããŒãïŒã§äžäœãç®æããšããè¡çºããçµæçã«ãã«ã·ããŒã·ã§ã³ãå©é·ããã¢ãã«ã®éçºãä¿ããŠãããšãããç®èãªæ§é ãååšããŸãã
ããã®åé¡ã¯ãåäžã®ã¢ã«ãŽãªãºã ãæ¹è¯ããã ãã§ã¯è§£æ±ºã§ãããã³ãã¥ããã£å šäœã®è©äŸ¡åºæºãããªãã¡ãäœãæåãšèŠãªããããšãã䟡å€èгãã®ãã®ãèŠçŽãããšãèŠæ±ããŠããã®ã§ãã
å®é𿹿³
ãæ¬ç ç©¶ã¯çè«çãªåæãäžå¿ã§ãããå žåçãªæ©æ¢°åŠç¿ã®åŠç¿æžã¿ã¢ãã«ã䜿ã£ãå®éšãšããããã¯ãæŠå¿µå®èšŒçãªåæãšããŒã¿èª¿æ»ãè¡ãããŠããŸããå ·äœçã«ã¯ä»¥äžã®2ç¹ããå®éšãã«çžåœããŸãã
çè«ã®å ·äœäŸæ€èšŒ
ãè«æäžã§æç€ºãããçè«ãå®éã®ã¢ãã«åºåäŸã§ç€ºã詊ã¿ã§ããäŸãã°è¡š1ã§ã¯ãå®åšã®èšèªã¢ãã«ïŒChatGPTãDeepSeekãLlamaç³»ã¢ãã«ïŒã®åçäŸãæ¯èŒãããããããã«ã·ããŒã·ã§ã³ãèµ·ãããŠããããšã瀺ããŠããŸãã
ããŸãæ¬æäžã®äŸãšããŠæãããããDEEPSEEKãšããåèªäžã«å«ãŸããæåDã®æ°ãåã質åãã§ãè€æ°ã®ã¢ãã«ã誀çïŒ2ã3ãšçããïŒããã±ãŒã¹ã玹ä»ãããŸããããããã¯çè«ã§è¿°ã¹ããããŒã¿äžèŠãããšããªãåãã«å¯ŸããŠã¯ã¢ãã«ã¯ã§ãããã«çããŠããŸããçŸè±¡ã®å®äŸãšãããŸãã
è©äŸ¡ãã³ãããŒã¯ã®ã¡ã¿åæ
ãAppendix Fã«ãŠãçŸåšåºã䜿ãããŠããè©äŸ¡ãã³ãããŒã¯æ°åçš®é¡ã®æ¡ç¹æ³ã調æ»ããåæãè¡ãããŸãããèè ãã¯åãã³ãããŒã¯ã®è«æãããã¥ã¡ã³ããåç §ãããäžæ£ç¢ºãªåçã«ããã«ãã£ã¯ãããããç¡åçã«éšåç¹ã¯äžããããããã調ã¹ãŠããŸãã
ããã®çµæããŸãšããã®ã衚2ã§ãã衚2ã«ã¯å ·äœçãªãã³ãããŒã¯åïŒGPQAãMMLU-ProãMathematicsãBBHãHLEãªã©ïŒãšãããããã«ã€ããŠBinary gradingïŒäž»èŠææšãäºå€è©äŸ¡ãã©ããïŒãšIDK creditïŒç¡åçã«ç¹ãå ¥ããã©ããïŒã®é ç®ã䞊ãã§ããŸãã

ãå ããŠãè«æå ã§ã¯çŽæ¥ã®å®éšçµæã§ã¯ãããŸããããé¢é£ããããŒã¿ãã€ã³ããšããŠGPT-4ã¢ãã«ã®ãã£ãªãã¬ãŒã·ã§ã³æ¹åäŸãå³ç€ºãããŠããŸãã
ãå³2ã¯OpenAIã®GPT-4 Technical Report (2023)ããåŒçšããããã®ã§ãGPT-4ã匷ååŠç¿èª¿æŽïŒRLHFïŒååŸã§ã©ãã ãåºåã®ç¢ºä¿¡åºŠïŒäºæž¬ç¢ºçã®ååžïŒãæ¹åãããã瀺ããã¹ãã°ã©ã ã§ããããã¯ãæ¬ç ç©¶ã䞻匵ãããã¢ãã«ã«èªèº«ã®äžç¢ºãããåŠç¿ãããããšãã®éèŠæ§ãè£ä»ããåèå³ãšèšããã§ãããã

å®éšçµæ
ããŸãçè«ã®å ·äœäŸãšããŠæç€ºãããçµæããèŠãŠã¿ãŸãã衚1ã«ç€ºããããã«ãå®éã«ChatGPTãªã©è€æ°ã®ã¢ãã«ã§Kalaiæ°ã®åå£«è«æã¿ã€ãã«ãå°ãããšãçãããããçããè¿ãããã®ã®èŠäºã«å€ããŠããŸããã
ããŸããDEEPSEEKã«å«ãŸããDã®æ°ããšããç°¡åãªã¯ãã®è³ªåã§ããDeepSeekãClaudeãªã©ãå šãã®ã§ããããªçãïŒ2ã3ãªã©ãæ¬åœã¯1ã€ãããªãã®ã«éå€§ãªæ°ïŒãè¿ããŠããŸãã
ããããã¯ã¢ãã«ãç¥ããªãããšã§ãåçãçµãåºãããšããæ§è³ªãåŠå®ã«è¡šããŠããŸããç¹ã«åŸè ã®äŸã¯ããŠãŒã¶ã®æç€ºãç¥ããªããã°æ°åã ãã§ãããããªãããšçããŠããšããæ¡ä»¶ä»ãã ã£ãã«ãããããããã¢ãã«ãé©åœãªæ°åãåºåããŠããŸã£ããšããç¹ã§ç€ºåçã§ãã
ããã®ãããªèгå¯çµæã¯ãã¢ãã«ããåãããªãããšæ£çŽã«èšãããããåœãŠãã£ãœãã§ãçãããæ¹ãéžãã§ããŸãçŸç¶ãç©èªã£ãŠããŸããèè ãã®çè«ã¯ãŸãã«ãã®ç¶æ³ã説æãããã®ã§ããã
ãæ¬¡ã«ãè©äŸ¡ãã³ãããŒã¯ã®ã¡ã¿åæçµæã§ãã衚2ããæãããªããã«ãèè ãã調ã¹ãäž»èŠãªãã³ãããŒã¯ã®çŽ8ã9å²ã¯Binary grading = Yesãã€ãŸãæ£è§£çãåæ Œçãšãã£ãäºå€è©äŸ¡ãæ¡çšããŠããŸããã
ãããã«IDK credit = Noneãã»ãšãã©ã§ãç¡åçã«å¯Ÿããéšåç¹ã¯çç¡ã§ãããäžéšãAIã«ããæ¡ç¹ãçšããWildBenchãªã©ã§ã¯IDKã«å¯Ÿããããã«é æ ®ãèŠããããã®ã®ãããã§ã誀çãšã®å³å¯ãªå·®å¥åãŸã§ã¯ãããŠããŸããã
ãèŠããã«ãçŸåšã®è©äŸ¡è»žã§ã¯ãããããªãããšçããã¡ãªããã¯ãŒãã§ãããééã£ãŠããŠãçããæ¹ãæåŸ åŸç¹ã¯é«ãããšãããŒã¿ã§è£ä»ããããã®ã§ãããã®çµæã¯ãèè ããæå±ããè©äŸ¡æ¹é©ã®å¿ èŠæ§ã匷調ãããã®ã§ããã
ãèè ãã¯è«æã®çµè«éšåã§ããããã®åæã«åºã¥ãã¡ãã»ãŒãžãæŽçããŠããŸãããã®èŠç¹ã¯ä»¥äžã®éãã§ãã
ãã«ã·ããŒã·ã§ã³ã¯è¬ã®ãã°ã§ã¯ãªããçµ±èšçã«èª¬æå¯èœãªçŸè±¡ã§ãããÂ
ãäºååŠç¿æ®µéã«ãããŠãã¢ãã«ã誀ã£ãæãçæããŠããŸãã®ã¯é¿ããããªãçµ±èšçå§åã®çµæã§ãããå¶çºçãªäžå ·åã§ã¯ãªãã
100%ã®æ£ç¢ºæ§ã¯éæäžå¯èœã ãããã«ã·ããŒã·ã§ã³ã¯é¿ãåŸãã
ããããªãã¢ãã«ã§ãå šåæ£è§£ã¯äžå¯èœïŒçããååšããªãåããããããïŒã ããã¢ãã«ãèªä¿¡ã®ãªãåãã«çããªãéžæè¢ãæãŠã°ãèåœã®åçãåºãå¿ èŠããªããªãã
倧ããªã¢ãã«ã ãããã«ã·ããŒã·ã§ã³ãå æã§ããããã§ã¯ãªãã
ãç¥ããªãããšã¯å°ããªã¢ãã«ã®æ¹ãå²ãåã£ãŠãããããªãããšèšããããå Žåããããéã«äžéå端ã«ç¥èããã倧ããªã¢ãã«ã»ã©ãèªä¿¡ã®èŠæ¥µããé£ãããã«ã·ããŒã·ã§ã³ããããããšãããã
äž»èŠè©äŸ¡ã®æ¹é©ãäžå¯æ¬ ã
ã ãã«ã·ããŒã·ã§ã³è©äŸ¡ã ãåå¥ã«äœã£ãŠã广ã¯éå®çã§ãããæ°å€ãã®æ¢åè©äŸ¡ãäžç¢ºå®æ§ãžã®çœ°åä»ãã§ããäºå®ãå€ããå¿ èŠãããããã¹ãŠã®äž»èŠè©äŸ¡ã¡ããªã¯ã¹ããäžç¢ºå®æ§ã®è¡šæãå ±ãããæ¹åã«æ¹ããããšã§åããŠãã¢ãã«éçºå šäœããã«ã·ããŒã·ã§ã³äœæžã«åæ©ä»ããããã
ãèè ãã¯ã幞ãGPT-5ãªã©ææ°ã¢ãã«ã§ã¯ãã«ã·ããŒã·ã§ã³çãäžãã£ãŠããŠããããããã§ããªã察çãå¿ èŠã ããšè¿°ã¹ãä»åŸãã¢ãã«ã®èªä¿¡éå°ãªãã«ã·ããŒã·ã§ã³ãæžããåªåãç¶ãããšç· ãããã£ãŠããŸãã
ããã®è«æãæããããã®ã¯ãæ§èœç«¶äºã®äžã§èŠèœãšãããã¡ãªãçããªãåæ°ãã®éèŠæ§ã§ãããAIã·ã¹ãã ã®ä¿¡é Œæ§åäžã«ã¯äººéåŽã®è©äŸ¡è»žã®èŠçŽããæ±ãããããšããã¡ãã»ãŒãžã§ããã
èå¯
ãªããã®ææ³ãåªããŠããã®ã
ãKalaiæ°ãã®ç ç©¶ã¢ãããŒããæ¢åã®ææ³ãšæ¯èŒããŠåªããŠããç¹ã¯ããããåãªãã察ççæ³ãã§ã¯ãªããåé¡ã®ãæ ¹æ¬èšºæããæäŸããŠããããšã«ãããŸãã
ãåŸæ¥ã®RAGãRLHFãšãã£ãææ³ã¯ããã«ã·ããŒã·ã§ã³ãšãããçç¶ããç·©åããããã®æå¹ãªãæ²»çè¬ãã§ãããããããªããã®çç¶ãçºçããã®ããšããç çãããªãã¡çµ±èšçãªçºçã¡ã«ããºã ã«ã€ããŠã¯æ·±ãåããŸããã§ããã
ãããã«å¯Ÿããæ¬ç ç©¶ã¯ãã«ã·ããŒã·ã§ã³ã®èµ·æºãã®ãã®ã«ã¡ã¹ãå ¥ãããããLLMã®èšç·Žãšè©äŸ¡ã®ãã©ãã€ã ã«å åšããå¿ ç¶çãªçµæã§ããããšãæããã«ããŸãã ã Â
ããã®ã¢ãããŒãã®åªäœæ§ã¯ããã®èª¬æåã«ãããŸãããªãç¹å®ã®çš®é¡ã®äºå®ã§ãã«ã·ããŒã·ã§ã³ãèµ·ããããã®ãããªãã¢ãã«ã¯èªä¿¡éå°ã«æ¯ãèãã®ãããšãã£ãåãã«å¯Ÿããäžè²«ããçµ±èšçãã¬ãŒã ã¯ãŒã¯ã®äžã§çããäžããŠãããŸãã
ããã®æ ¹æ¬åå ã®çè§£ã¯ãããå ç¢ã§æç¶å¯èœãªè§£æ±ºçãèšèšããããã®ãäžå¯æ¬ ãªåå°ãšãªããŸããåã«ãã«ã·ããŒã·ã§ã³ãä¿®æ£ããããŒã«ãæäŸããã®ã§ã¯ãªãããã«ã·ããŒã·ã§ã³ãçºçãã«ãããšã³ã·ã¹ãã ãã©ãæ§ç¯ãã¹ããããšãããããæ¬è³ªçãªè°è«ãžã®éãéããç¹ã«ããã®ç ç©¶ã®ç䟡ãããã®ã§ãã
ãã®ææ³ãæ¢åã®ãã®ãšæ¯èŒããåªäœæ§ã»å£äœæ§
ãæ¬ç ç©¶ã®ã¢ãããŒããæã€åªäœæ§ãšããã®éçïŒå£äœæ§ïŒãæŽçãããšã以äžã®ããã«ãªããŸãã
åªäœæ§
æ ¹æ¬åå ã®è§£æãšèª¬æå
ãæå€§ã®åªäœæ§ã¯ãåè¿°ã®éãããã«ã·ããŒã·ã§ã³ãçºçããããªãããè§£æããç¹ã§ããç°ãªãçš®é¡ã®ãšã©ãŒãäžã€ã®çµ±èšçæ çµã¿ã§çµ±äžçã«èª¬æã§ãããããå Žåœããçãªå¯Ÿçã§ã¯ãªããåçã«åºã¥ãã察çã®ç«æ¡ãå¯èœã«ããŸãã
ããæ¬è³ªçãªè§£æ±ºçã®æç€º
ããã®ç ç©¶ãæèšãã解決çãããªãã¡ãè©äŸ¡ææšã®æ¹é©ãã¯ãåé¡ã®æ ¹æºã«çŽæ¥åããããŸããããè©äŸ¡ã·ã¹ãã ããæšæž¬ãã«å ±é ¬ãäžããªããªãã°ãAIéçºè ãã¡ã¯èªç¶ãšãäžç¢ºå®æ§ã®è¡šæããå¯èœãªã¢ãã«ãæ§ç¯ãã匷ãã€ã³ã»ã³ãã£ããæã€ããšã«ãªããŸãããããªãã°ãRAGãRLHFãšãã£ãæ¢åã®ç·©åçãããã广çã«æ©èœããåå£ãæŽãã§ããã ã Â
å£äœæ§ïŒéçç¹ïŒ
çè«çåçŽå
ãèè ãèªèº«ãèªããŠããããã«ããã®çè«çãã¬ãŒã ã¯ãŒã¯ã¯çŸå®ãããçšåºŠåçŽåããŠããŸã ãäŸãã°ãçŸå®ã¯ãæå¹ããããšã©ãŒããã®äºè æäžã§ã¯ãªããéšåçã«æ£ããéšåçã«èª€ã£ãŠããããšãã£ãç¡æ°ã®ã°ã©ããŒã·ã§ã³ãååšããŸãããŸãããI don't knowã以å€ã®å€æ§ãªäžç¢ºå®æ§ã®è¡šçŸæ¹æ³ãèæ ®ãããŠããŸããã Â
解決çã®å®çŸå¯èœæ§
ãææ¡ãããŠãããè©äŸ¡ææšã®æ¹é©ãã¯ãã¢ã«ãŽãªãºã ãäžã€ä¿®æ£ããã°æžããããªåçŽãªæè¡ç解決çã§ã¯ãããŸãããããã¯ãGoogleãOpenAIãMetaãšãã£ãäž»èŠãã¬ã€ã€ãŒãå«ããAIç ç©¶éçºã³ãã¥ããã£å šäœã®åæåœ¢æãå¿ èŠãšãããç€ŸäŒæè¡çãªææŠã§ã ã
ã圱é¿åã®å€§ãããªãŒããŒããŒãããã³ãããŒã¯ã®æ¡ç¹ã«ãŒã«ã倿Žããããšã¯ãæè¡çãªå°é£ã以äžã«ãæ¿æ²»çã»æåçãªéå£ã極ããŠé«ããšèšããããåŸãŸããããããããã®æèšã®æã倧ããªå®è·µäžã®ããŒãã«ãšãªããŸãã Â
çµè«
ãæ¬çš¿ã§è©³è¿°ããŠããKalaiæ°ãã®ç ç©¶ã¯ãLLMãçæãããã«ã·ããŒã·ã§ã³ãšããçŸè±¡ã«ã€ããŠãç§ãã¡ã®çè§£ãæ ¹åºããèŠããã®ã§ãããã«ã·ããŒã·ã§ã³ã¯ããã¯ãäžå¯è§£ãªãã°ãã¢ãã«ã®æªçãã®çŸãã§ã¯ãªããäºã€ã®çµ±èšçå¿ ç¶æ§ãçµã¿åããã£ããæ§é çãªåž°çµã§ããããšã瀺ãããŸããã
ã第äžã«ãäºååŠç¿ã®æ®µéã§ã¯ãã¢ãã«ã¯èšèªã®ç¢ºçååžãå¿ å®ã«åŠç¿ããããšããŸãããã®ãèŒæ£ãã®ããã»ã¹ããç¹ã«æ å ±ãä¹ãããä»»æäºå®ãã«å¯ŸããŠãçµ±èšçã«äžå¯é¿ãªãšã©ãŒãããªãã¡ãã«ã·ããŒã·ã§ã³ã®çš®ãçã¿åºããŸãã
ã第äºã«ãäºåŸåŠç¿ã®æ®µéã§ã¯ãAIã³ãã¥ããã£ããæ§èœããæž¬ãããã«çšããŠããäž»æµã®è©äŸ¡ãã³ãããŒã¯ããäžç¢ºå®ãªå Žåã«æšæž¬ããããšã奚å±ããã€ã³ã»ã³ãã£ãæ§é ãæã£ãŠããŸããããã«ãããäºååŠç¿ã§çãŸãããã«ã·ããŒã·ã§ã³ã®çš®ã¯æå¶ãããã©ããããããããèªä¿¡ãæã£ãŠæšæž¬ããããšãã圢ã§éè±ããå©é·ãããŠããŸãã®ã§ãã
ããã®ç ç©¶ãå°ãåºãçµè«ã¯ãããä¿¡é Œæ§ã®é«ãAIãžã®éããåã«ã¢ãã«ãå€§èŠæš¡åããããããå€ãã®ããŒã¿ã§èšç·Žãããããå»¶é·ç·äžã«ã¯ãªãå¯èœæ§ã瀺åããŠããŸããçã®é²æ©ã®ããã«ã¯ãç§ãã¡ããæåããã©ã®ããã«å®çŸ©ããæž¬å®ããããšãããè©äŸ¡ã®ããæ¹ãã®ãã®ãæ ¹æ¬çã«èŠçŽãå¿ èŠããããŸãã
ãè«æãæèšãããäž»æµã®è©äŸ¡åºæºãä¿®æ£ããäžç¢ºå®æ§ã®é©åãªè¡šæã«å ±é ¬ãäžããããã«ã€ã³ã»ã³ãã£ããåèšèšãããšããã¢ãããŒãã¯ããã®ããã®å ·äœçãã€éèŠãªç¬¬äžæ©ã§ãã
ããã®æèšã¯ããã«ã·ããŒã·ã§ã³ã®æå¶ã«åããæè¡çãªéå£ãåãé€ããšåæã«ãAIãããæŽç·Žãããå®çšçãªèœåãç²åŸããããã®æãéãå¯èœæ§ãç§ããŠããŸãããã®èŠç¹ã®è»¢æãããã人工ç¥èœãšããæè¡ã®ãæçãã責任ããçºå±ã®ããã«äžå¯æ¬ ã§ãããšèšããã§ãããã
ããšãã
ãæ¬è«æã瀺ããããã«ã·ããŒã·ã§ã³ã¯çµ±èšçå¿ ç¶ãšè©äŸ¡èšèšã®ç£ç©ã§ããããšããçµè«ã¯ãAIéçºã«ãããæ ¹æ¬çãªèŠåº§ã®è»¢æãè¿«ããã®ã§ããã¢ãã«ã®å€§èŠæš¡åã粟床åäžã ãã§ã¯è§£æ±ºã§ããªãæ§é çåé¡ã®ååšãæç¢ºã«ããè©äŸ¡åºæºãšããå€éšèŠå ã®éèŠæ§ãæµ®ã圫ãã«ããŸããã
ãç¹ã«è峿·±ãã®ã¯ããç¥ããªããšçããããšããäžèŠåœç¶ã®æ¯ãèãããçŸåšã®è©äŸ¡äœç³»ã§ã¯å ±ãããªããšããææã§ããèè ããææ¡ããè©äŸ¡æ¹é©ã¯ãæ¢åã®ãã³ãããŒã¯ã«è¥å¹²ã®ä¿®æ£ãå ããã ãã§å®çŸå¯èœã§ããããã®å®çŸå¯èœæ§ã®é«ããæ³šç®ã«å€ããŸãã
ãOpenAIã¯æ¢ã«ææ°ã¢ãã«ã§ãã®æ¹åæ§ãåãå ¥ãå§ããŠãããä»åŸæ¥çå šäœããã®ç¥èŠãã©ãåãæ¢ããããæ³šç®ãããŸããæ¬è«æã¯ãAIç ç©¶ãçŽç²ãªæè¡éçºãããè©äŸ¡èšèšãå«ãç·åçãªã·ã¹ãã èšèšãžãšæçãã€ã€ããããšã瀺ãéèŠãªäžæ©ãšãããã§ããããçæAIã®ä¿¡é Œæ§åäžãèããäžã§ãé¿ããŠéããªãè°è«ãæèµ·ããç»æçãªç ç©¶ã§ãã
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èš
ãæ¬çš¿ã¯2025幎9æã«çºè¡šãããç ç©¶è«æãåºã«å·çããŸãããAIæè¡ã¯æ¥éã«é²åããŠãããæ¬çš¿ã®å
容ã¯å·çæç¹ã§ã®ææ°æ
å ±ã«åºã¥ããŠããŸããå®åãžã®é©çšãæ€èšãããéã¯ãææ°ã®æè¡ååãšèŠå¶èŠä»¶ã確èªããããšããå§ãããŸãã
Appendix
Appendix A: æ¬çš¿äž»èŠå®çã®èšŒæÂ
ãææ¡ããäž»å®çïŒçæèª€ççãšåé¡èª€åé¡çã®é¢ä¿ã«é¢ããå®ç1ïŒã®è©³çŽ°ãªæ°åŠç蚌æã瀺ãããŠããŸãã確çååžéã®ãºã¬ÎŽã®è©äŸ¡ãªã©ãäºå€åé¡èª€å·®ããçæèª€å·®ãžã®å€æãå³å¯ã«å°åºããŠããŸãã
Appendix B: ä»»æã®äºå®ïŒArbitrary-factsïŒã®åæÂ
ãGood (1953) ã®Missing Massæšå®ã«åºã¥ããèšç·ŽããŒã¿ã«1åããçŸããªãäºäŸã®å²åãšã¢ãã«èª€ççãšã®é¢ä¿ãå°ããéšåã§ããGood-Turingæšå®éã®ä¿èšŒïŒMcAllester & Ortiz, 2003ïŒãåŒçšãã€ã€ãIDKã蚱容ããå Žåãžã®æ¡åŒµãè¡ãããèšç·Žã§äžåºŠããã®äºå®ã¯é«ç¢ºçã§èŠèœãšãããïŒ=å¹»èŠã«ã€ãªããïŒãããšãå®åŒåããŠããŸãã
Appendix C: å®ç3ããã³ç³»2ã®èšŒæÂ
ãæ¬æäžã®Theorem 3ïŒçŽç²ãªå€è¢éžæåé¡ã«é¢ããçµæïŒãšããã«ä»éããç³»2ã®èšŒæãèšèŒãããŠããŸããn-gramã¢ãã«ãªã©å ·äœäŸã§ã®èšç®ãè£è¶³ãããè€æ°éžæè¢ã®åé¡ã«ããã誀çäžéã瀺ãããŠããŸãã
Appendix D: 远å ã®èŠå ã«é¢ããèå¯Â
ãæ¬æã§è§ŠããããObservation 2ã«ã€ããŠèª¬æããŠããŸããå ·äœäŸãšããŠãæå·æã®è§£èªãã®ããã«ããŠãŒã¶ã®è³ªåèªäœãè§£ã®ãªãé£åã§ããã±ãŒã¹ïŒIDKãæ£è§£ãšãªãã±ãŒã¹ïŒãåãäžããå®ç1ã®é©çšäŸã瀺ããŠããŸãããŸãããŒã¿ã»ããã·ããïŒèšç·Žãšãã¹ãååžã®äžäžèŽïŒãªã©ãäºå€åé¡ã§ç¥ããã課é¡ãå¹»èŠã«ã圱é¿ãåŸãç¹ãè°è«ããŠããŸãã
Appendix E: 芳å¯1ã®èšŒæÂ
ãæ¬æäžObservation 1ïŒã©ããªäžç¢ºå®æ§ååžäžã§ã0-1æ¡ç¹ã§ã¯IDKãæé©ã«ãªããªãïŒã®èšŒæãç°¡æœã«ç€ºãããŠããŸããçŽèгçã«ã¯æãããªãã®äž»åŒµã圢åŒçã«ç€ºããè©äŸ¡æ³å€æŽã®åæ©ä»ããšããŠããŸãã
Appendix F: è©äŸ¡ææ³ã®çãã¡ã¿è©äŸ¡Â
ã衚2ã®å ã«ãªã£ã調æ»çµæããŸãšããããŠããŸãã人æ°ã®è©äŸ¡ãã³ãããŒã¯çŽ30çš®ã«ã€ããŠãBinary gradingãåŠããIDKåçãèªãããããåŠããäžèŠ§è¡šã§ç€ºãããã®å§åç倿°ãäºå€è©äŸ¡ã»IDKäžå¯ã§ãããšå ±åããŠããŸãã
References
ãæ¬çš¿ã®å 容ã«ããã«æ·±ãé¢å¿ãæãããèªè ã®ããã«ãåèæç®ãããã€ã玹ä»ããŸãã
Kalai, A. T., & Vempala, S. S., "Calibrated Language Models Must Hallucinate," STOC 2024.
ãæ¬çš¿ã§è§£èª¬ããè«æã®çŽæ¥ã®å身ãšãªãç ç©¶ã§ããã¢ãã«ã®ãèŒæ£ïŒãã£ãªãã¬ãŒã·ã§ã³ïŒããçµ±èšåŠçãªãã°ããã»ãã¥ãŒãªã³ã°æšå®ãããããŠãä»»æäºå®ãã«å¯Ÿãããã«ã·ããŒã·ã§ã³ã®çµ±èšçå¿
ç¶æ§ãšã®éã®æ ¹æ¬çãªçè«çé¢é£æ§ã確ç«ããŸããã
Ouyang, L., et al., "Training language models to follow instructions with human feedback," NeurIPS 2022.
ã人éã®ãã£ãŒãããã¯ããã®åŒ·ååŠç¿ïŒRLHFïŒã«é¢ããç¬åµçãªè«æã§ãããäºåŸåŠç¿ã«ãããäž»èŠãªæè¡ã確ç«ããŸãããæ¬çš¿ã§è°è«ãããäºåŸåŠç¿ã®ã€ã³ã»ã³ãã£ãåé¡ãçè§£ããããã«ã¯ããã®RLHFã®ã¡ã«ããºã ãææ¡ããããšãäžå¯æ¬ ã§ãããªããªããæ¬ é¥ã®ããè©äŸ¡ææšã«ãã£ãŠåœ±é¿ãåããã®ãããŸãã«ãã®RLHFã®ããã»ã¹ã ããã§ã ã Â
Ji, Z., et al., "Survey of Hallucination in Natural Language Generation," ACM Computing Surveys 2023.
ããã®å
æ¬çãªãµãŒãã€è«æã¯ããã«ã·ããŒã·ã§ã³ã®å€æ§ãªçš®é¡ïŒäŸïŒå
åšç vs. å€åšçïŒãåå ããããŠç·©åæŠç¥ã®åºç¯ãªåé¡ãæäŸããŠããŸããKalaiæ°ãã®ç ç©¶ãæ·±ãæãäžããç¹å®ã®çµ±èšçã¡ã«ããºã ãããã«ã·ããŒã·ã§ã³ç ç©¶å
šäœã®åºå€§ãªå°å³ã®äžã§ã©ãã«äœçœ®ããã®ããçè§£ããããã®ãåªããé³¥ç°å³ãšãªããŸã ã Â
Rafailov, R., et al., "Direct Preference Optimization: Your Language Model Is Secretly a Reward Model," NeurIPS 2023.
ããã®è«æã¯ãRLHFã®åŒ·ååŠç¿æ®µéã«ä»£ãããããå®å®ãã€å¹ççãªææ³ã§ããçŽæ¥çå奜æé©åïŒDPOïŒãææ¡ããŠããŸããããã¯å奜ãã¥ãŒãã³ã°æè¡ã®æå
端ã代衚ãããã®ã§ãããKalaiæ°ãã®è«æãæ¹é©ãæ±ããŠããçŸä»£ã®äºåŸåŠç¿ãã€ãã©ã€ã³ãçè§£ããäžã§éèŠãªæç®ã§ã ã Â
Lewis, P., et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks," NeurIPS 2020.
ãæ€çŽ¢æ¡åŒµçæïŒRAGïŒã®åºç€ãç¯ããè«æã§ãããã«ã·ããŒã·ã§ã³ãåæžããããã®äž»èŠãªä»£æ¿ã¢ãããŒãïŒå€éšç¥èãžã®ã°ã©ãŠã³ãã£ã³ã°ïŒãçè§£ããäžã§åœ¹ç«ã¡ãŸããã¢ãã«å
éšã®çµ±èšçãã€ããã¯ã¹ãšè©äŸ¡ãšã³ã·ã¹ãã ã«çŠç¹ãåœãŠãKalaiæ°ãã®ã¢ãããŒããšã®å¯Ÿæ¯ãéããŠãåé¡ãžã®å€è§çãªèŠç¹ãåŸãããšãã§ããŸã ã Â
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
çæAIæä»£ã®çŸ
éç€ãšãªã決å®çãæ±äº¬å€§åŠæŸå°Ÿã»å²©æŸ€ç 究宀ãç·åãçµéããçæAIææ°ååã培åºè§£èª¬ãChatGPTç»å Žä»¥éãæ¥éã«é²åããçæAIæè¡ã®çŸåšå°ãšæªæ¥ããæ¥æ¬ã代衚ããAIç ç©¶è
ãã¡ãå
æ¬çã«åæããŠããŸãã
2025幎3æåè¡ãæè¡ã®æ žå¿ããããžãã¹æŽ»çšãåœå
å€ã®æ³èŠå¶ããããŠå®å
šæ§ã«è³ããŸã§ãä»ç¥ãã¹ãæ
å ±ãç¶²çŸ
çã«è§£èª¬ãAIãšãŒãžã§ã³ãããã«ãã¢ãŒãã«åãšãã£ãææ°ãã¬ã³ããæŒãããŠããŸãã
çæAIããžãã¹ã«é¢ãããã¹ãŠã®æ¹ã«ãšã£ãŠå¿
æºã®äžåãæ¿åããAIæ¥çã§ç¢ºããªå€æãäžãããã«ãä¿¡é Œã§ããæ
å ±æºã§ãã