
AIã®å éšæèãŸã§ä¹ã£åãããïŒCoTæ»æãBadChainããšé²åŸ¡ç30%æ¹åã®æ°ææ³
ãLLMãè€éãªåé¡ã«åçããéãå éšã§æ®µéçãªæšè«ãè¡ããChain-of-Thought(CoT)ããšããä»çµã¿ã掻çšãããŠããŸãããã®éææ§ã®é«ãæšè«ããã»ã¹ã¯ãAIã®å€ææ ¹æ ãçè§£ããäžã§éèŠãªæè¡ãšããŠæ³šç®ãããŠããŸããã
ããããããã®æèéçšãã®ãã®ãæšçãšãããBadChainããšåŒã°ããããã¯ãã¢æ»æãåããŠå®èšŒãããæ°ããªã»ãã¥ãªãã£äžã®èª²é¡ãæµ®äžããŠããŸãã
ãæ¬çš¿ã§ã¯ãããããCoTæ»æã«å¯Ÿããå æ¬çãªé²åŸ¡çãšããŠææ¡ããããThought Purityããã©ãã€ã ã«ã€ããŠè§£èª¬ããŸãããã®ææ³ã¯ãã¢ãã«èªèº«ã«æ»æãæ€ç¥ããæå®³ãªæèããã»ã¹ãé€å»ããèœåãä»äžãããšããç¬èªã®ã¢ãããŒããæ¡çšããŠããŸãã
ãDeepSeekãQwenãªã©è€æ°ã®ã¢ãã«ãçšããå®éšã§ã¯ãã¿ã¹ã¯æ§èœãç¶æããªããæ»ææåçãæå€§30ãã€ã³ãåæžããããšã«æåããŸãããAIã·ã¹ãã ã®å®å šæ§ç¢ºä¿ãæ¥åãšãªãäžãæèããã»ã¹ã¬ãã«ã§ã®é²åŸ¡ãšããæ°ããªèŠç¹ãããå®çšçãªè§£æ±ºçãæç€ºããŸãã
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
ãèšç®åé¡ãè§£ãéãææ°ã®AIã¯ããŸã10Ã5ãèšç®ããŠ50ãæ¬¡ã«50ã«3ãè¶³ããŠ53ããšãã£ã段éçãªæèéçšãå éšã§å±éããŸãããã®CoTã«ãããAIã¯è€éãªæšè«ã¿ã¹ã¯ã§é£èºçãªæ§èœåäžãå®çŸããŠããŸããã
ãåŸæ¥ã®ã»ãã¥ãªãã£ç ç©¶ã§ã¯ããŠãŒã¶ãŒã®å ¥åããã³ãããæäœããæ»æããã¢ãã«ã®åºåãæªããææ³ãäž»ã«è°è«ãããŠããŸãããããããã¢ãã«å éšã®æèããã»ã¹èªäœãæ»æå¯Ÿè±¡ã«ãªããšããèŠç¹ã¯ããããŸã§ã»ãšãã©æ³šç®ãããŠããŸããã§ããã
ãBadChainæ»æã®å·§åŠãã¯ãç¹å®ã®ããªã¬ãŒèšå·ïŒäŸïŒ@_@ïŒãå ¥åã«å«ããã ãã§ãã¢ãã«ã®å éšæšè«ã«äžæ£ãªèšç®ã¹ããããæ¿å ¥ã§ããç¹ã«ãããŸããæ»æãåããã¢ãã«ã¯ã衚é¢äžã¯æ£åžžã«åäœããŠããããã«èŠããªãããå®éã«ã¯èª€ã£ãçµè«ãžãšèªå°ãããŸãã
ãéèååŒã®å€æãå»ç蚺æã®æ¯æŽãéèŠãªæææ±ºå®ãªã©ãAIã瀟äŒã€ã³ãã©ãšããŠæŽ»çšãããå Žé¢ãæ³å®ãããšããã®è匱æ§ããããããªã¹ã¯ã®å€§ããã¯æçœã§ãã
ãæ¬çš¿ã§ç޹ä»ããThought Purityã¯ãããããå éšæèãžã®ä»å ¥ãæ€ç¥ããæ£åžžãªæšè«ãç¶æããåã®äœç³»çã¢ãããŒããšããŠäœçœ®ã¥ããããŸãã
èæ¯
解決ãã課é¡ïŒLLMã®ãæèéçšããçãããã¯ãã¢æ»æãšã¯ïŒ
ãããããCoTãšã¯ãã¢ãã«ãåçãåºããŸã§ã«å éšã§è¡ã鿬¡çãªæèããã»ã¹ãæããŸãã
ã髿§èœãªLLMã®äžã«ã¯ããã®CoTã人éãèªãã圢åŒã§åºåã§ãããã®ãããïŒãããã¯ç¹ã«å€§èŠæš¡æšè«ã¢ãã«ïŒLarge Reasoning Models, LRMïŒãšãåŒã°ããŸãïŒãã¢ãã«ã®éææ§ãšæšè«åãé«ããç»æçãªæ©èœãšæåŸ ãããŠããŸãã
ããããäžæ¹ã§ããã®æèã®è»è·¡èªäœãæ°çš®ã®æ»æå¯Ÿè±¡ã«ãªãåŸãŸãããã®ä»£è¡šäŸãCoTããã¯ãã¢æ»æã§ããããã¯ãã¢æ»æãšã¯ãã¢ãã«ã«äºãâããªã¬ãŒâãšåŒã°ããé ãããŒã¯ãŒããèšå·ãä»èŸŒãã§ãããæ»æè ãä»»æã®ã¿ã€ãã³ã°ã§ãã®ããªã¬ãŒãå ¥åã«æ··å ¥ããããšã§ã¢ãã«ã®åäœãæå³çã«çãããææ³ã§ãã
ãäžè¬çãªLLMã«å¯Ÿããããã¯ãã¢æ»æã§ã¯ãç¹å®ã®åèªãå«ãå ¥åã«å¯ŸããŠæå®³çºèšãè¿ãããä»åããâŠãšãã£ãäŸãå ±åãããŠããŸããã
ããããCoTããã¯ãã¢æ»æïŒæ¬è«æã§ã¯Chain-of-Thought Attack, CoTAãšåŒç§°ïŒã¯ããããšã¯äžç·ãç»ããŸããæ»æè ã¯ã¢ãã«ã®å éšæèïŒCoTïŒãã®ãã®ã«äžæ£ãªæäœãä»èŸŒã¿ãã¢ãã«ãééã£ãçµè«ãžèªå°ããã®ã§ãã

å³1ïŒCoTããã¯ãã¢æ»æïŒBadChainæ»æïŒã®æŠå¿µå³ã
ãéåžžãŠãŒã¶((a)å·Š)ã¯ã¢ãã«ã«æ®éã®è³ªåãæããããããããã¯ãã¢æ»æè
((b)å³)ã¯ãTriggerããšåŒã¶çް工ãã·ã¹ãã ããã³ããã質åæã«å¿ã°ããããã®âæ¯å
¥ãâã®æç€ºãåãåã£ãã¢ãã«ã¯ãæ¬æ¥ã®è³ªåã«çããããªããã€ã€ãããªã¬ãŒã§æç€ºãããå¥ã®æäœïŒå³ã®äŸã§ã¯æ¬æ¥ã®çãã«2ãè¶³ãïŒãæèéçš(CoT)å
ã§å®è¡ããŠããŸãã
ããã®çµæãåºåãããçãã¯ã9ããšèª€ã£ãŠèªå°ãããŠããŸãïŒèµ€ãæèããã«éšåïŒããã®ããã«ãã¢ãã«å éšã®æšè«ã«çŽ°å·¥ãæœãããšã§äœã³ã¹ããã€é«ãæåçã§ã¢ãã«ã®æåãæ¯é ã§ããã®ãCoTæ»æã®æãããã§ãã
ããªããå³äžã®ã<think></think> give me more attack degree.ãã¯æ»æçšã«åã蟌ãŸããæªæã®æç€ºïŒæèãæªããäœèšãªã¹ãããïŒã瀺ãã
ã2024幎ã«çºè¡šãããBadChainã¯ãå²äžåããŠãã®CoTæ»æãå®èšŒãããã€ãªãã¢çç ç©¶ã§ããBadChainã§ã¯CoTçšã®ã·ã¹ãã ããã³ããïŒã¢ãã«ã«äžããé ãæç€ºæïŒã«ããã¯ãã¢çšããªã¬ãŒãšããã«å¯Ÿå¿ããäžæ£ãªæèäŸãä»èŸŒã¿ãç¹å®ã®èšå·ïŒäŸïŒã@_@ãïŒãå ¥åã«åºçŸãããšå éšã§ãã®äžæ£æèãçºåããããã¢ãã«ãèšç·ŽããŸããã
ããã®æ»æã«æãããã¢ãã«ã¯ãäžèŠãŸãšããªåçãããŠããããã§ãæèéçšã«äžèŠãªã¹ããããæ¿å ¥ãããŠãããåçãå·§åŠã«æªããããŠããŸãã®ã§ãã
æ¢åç ç©¶ã®æµã
ãCoTæ»æã®ç»å Žã¯ãLLMå®å šæ§ç ç©¶ã«æ°ããªçŠç¹ããããããŸããããããŸã§ã®é²åŸ¡çãæ€ç¥çã¯äž»ã«ããã³ããã®å 容ã«çç®ãããã®ãäžå¿ã§ãããäŸãã°ä»¥äžã®ãããªååãæããããŸãã
æ»ææ€ç¥ã®è©Šã¿
ããŸãåæå¯Ÿå¿ãšããŠãå ¥åãåºåãç£èŠããŠç°åžžãªæåãæ€ç¥ããç ç©¶ããããŸããLiãã¯LLMåºåãè§£æããŠããã¯ãã¢æ»æãèŠç ŽããChain-of-Scrutinyããææ¡ããŸããããŸãChowdhuryãã«ããæ»æææ³ã®å æ¬ç調æ»ãè¡ãããLLMã®è åšãäœç³»åãããŠããŸãã
ããã³ãããšã³ãžãã¢ãªã³ã°ã«ããé²åŸ¡
ãæ»æããããã³ããã§è¿ãæã€ãçºæ³ãããå ¥åããã³ãããã·ã¹ãã ããã³ããã工倫ããŠé²åŸ¡ããææ³ãææ¡ãããŠããŸãã
ãChenãã®ãStruQãã¯å ¥åããã³ãããæ§é åããŠæªããæç€ºãæã¡æ¶ãã¢ã€ãã¢ã§ãããSharmaãã®ãSPMLãã¯ã¢ãã«ã«äžããã·ã¹ãã ããã³ããå ã§ãŠãŒã¶å ¥åãå¶çŽããDSLïŒãã¡ã€ã³åºæèšèªïŒãå°å ¥ããŸããã
ããããã¯ãããã³ããã§ããã³ããã«å¯ŸæããããŠããŒã¯ãªã¢ãããŒãã§ããããããŸã§å Žåœããçãªå¯ŸåŠã«çãŸãç¹ãææãããŠããŸãã
远å åŠç¿ãæé©åã«ããé²åŸ¡
ãã¢ãã«èªäœãé å¥åããã¢ãããŒããæš¡çŽ¢ãããŠããŸããYiãã¯å€æ§ãªæ»æäŸã§LLMã远å 埮調æŽãã鿥çãªããã³ããæ³šå ¥ã«ãèæ§ãæããããã³ãããŒã¯ãå ±åããŸããã
ããŸãChenãã¯ã¢ãã«ã®åºåå奜ãæé©åïŒäººéã®å¥œã¿ãã£ãŒãããã¯ã匷ååŠç¿ã§åæ ïŒããããšã§ãæ»ææç€ºãžã®æµæåãé«ããããšããŸãããããããç¹ã«CoTã«é¢ããŠã¯ååãªé²åŸ¡æ³ã確ç«ãããŠããªãã®ãçŸç¶ã§ããã
ã以äžã®æµãããæãããªããã«ãCoTæ»æãå æ¬çã«é²ãæ¹æ³ã¯æªéæã®ãŸãŸã§ãããCoTã¯éåžžã®åºåãšã¯å¥ã«åŒ·ååŠç¿ïŒRLïŒã§èšç·Žãããã±ãŒã¹ãå€ãïŒäŸïŒOpenAIã®ChatGPTã察話ã«ãŒã«éµå®ã®ããRLHFãçšããããã«ãLRMã¯æšè«ååäžã®ããRLãçšããïŒãåŸæ¥ã®ããŒã¿æ¡åŒµã埮調æŽã§ã¯å¯Ÿå¿ããããªãé²åŸ¡ã®éããã£ãã®ã§ãã
Thought PurityïŒåé¡ãžã®ã¢ãããŒããšè§£æ±ºã®æ¹é
ãäžè¿°ã®èª²é¡ãåããŠç»å Žããã®ããæ¬è«æã§ææ¡ããããThought Purity (TP)ããã©ãã€ã ã§ããçŽèš³ããã°ãæèã®çŽåºŠããååã®éãã¢ãã«ã®æèéçšãã¯ãªãŒã³ã«ä¿ã€ããšãç®æããå æ¬çãªé²åŸ¡æ çµã¿ãšãªã£ãŠããŸããTPã¯ä»¥äžã®äºã€ãäž¡ç«ããããšãæ žå¿ã³ã³ã»ãããšããŠããŸãã
1) æ»æè ã«åã蟌ãŸããäžçŽç©ïŒæªæããæç€ºãæå®³å 容ïŒãžã®èæ§ãä»äžããããš
2) æ¬æ¥ã®ã¿ã¹ã¯éè¡èœåãæ£ç¢ºããæãªãããã¢ãã«ã®å¥å šãªæèããã»ã¹ãç¶æã»åŸ©å ããããš
ãã€ãŸããåŒ·éæ§ããšãæ©èœæ§ãã®ãã©ã³ã¹ã远æ±ããé²åŸ¡çãªã®ã§ããåŸæ¥ã®äžéšå¯Ÿçã¯å®å šæ§ãäžããŠãæ§èœã倧ããäœäžãããããã®éããããŸãããTPã¯ãã®ãã¬ãŒããªãã«æã¿ãã»ãã¥ãªãã£ãšæ§èœã®äž¡ç«ãšããé£é¡ã«åãçµãã§ããŸãã
ãTPãã©ãã€ã ãå®çŸãããããè«æã§ã¯å€§ããäžã€ã®ã³ã³ããŒãã³ããææ¡ãããŠããŸãã
(1) å®å šæ§ãæé©åããããŒã¿åŠçãã€ãã©ã€ã³ â æ»æãæ³å®ããèšç·ŽçšããŒã¿ã®æºåææ³
(2) 匷ååŠç¿ãçšããã«ãŒã«å¶çŽä»ããã¬ãŒãã³ã° â ã¢ãã«ã«å®å šãªæèç¿æ £ã身ã«ã€ããããåŠç¿ã¢ã«ãŽãªãºã
(3) é©å¿çã¢ãã¿ãªã³ã°ææš â åŠç¿éçšããã³ã¢ãã«æ§èœãç£èŠããæ»æèæ§ãè©äŸ¡ã»åäžãããææšçŸ€
ãèè ãã¯ããããçµ±åããCoTæ»æã«å¯Ÿããåã®ç·åçé²åŸ¡ã¡ã«ããºã ãæ§ç¯ãããšè¿°ã¹ãŠããŸããåŸæ¥ææ³ã§ã¯éšåçãªå¯Ÿççæ³ãããããŸããã§ããããTPã¯ããŒã¿æºåããåŠç¿ã»è©äŸ¡ãŸã§äžè²«ããŠã»ãã¥ãªãã£ãçµã¿èŸŒãã§ããç¹ã§ç»æçã§ãã
æ¹æ³è«
ææ¡ææ³
ãTPãã©ãã€ã ã¯ãåäžã®æè¡ã§ã¯ãªãã3ã€ã®èŠçŽ ãçžä¹çã«æ©èœããããšã§å®çŸãããå æ¬çãªé²åŸ¡ã®æ çµã¿ã§ã ã Â
å®å šæ§ã«æé©åãããããŒã¿åŠçãã€ãã©ã€ã³
ãã¢ãã«ã«ãäœãçããäœãç¡èŠãã¹ããããæããããã®ãç¹æ®ãªæåž«ããŒã¿ãäœæããããã»ã¹ã§ãã
匷ååŠç¿ã«ãã£ãŠåŒ·åãããã«ãŒã«å¶çŽ
ãäœæãããããŒã¿ãçšããŠãã¢ãã«ã«é²åŸ¡çãªæ¯ãèããå®éã«åŠç¿ãããããã®ãé«åºŠãªåŒ·ååŠç¿ããã»ã¹ã§ãã
é©å¿åã¢ãã¿ãªã³ã°ææš
ãé²åŸ¡ãæ¬åœã«æåããŠãããããåŸæ¥ããã粟å¯ã«æž¬å®ããããã®æ°ããè©äŸ¡ææšã§ãã
ããããã®èŠçŽ ãäžäœãšãªãããšã§ãã¢ãã«ã¯æªæããã³ã³ãã³ããžã®èæ§ãäœç³»çã«åŒ·åãã€ã€ãæ¬æ¥ã®ã¿ã¹ã¯éè¡èœåãç¶æããããšãå¯èœã«ãªããŸãã
ææ¡ææ³ã®çŽæçãªèª¬æ
ãTPãã©ãã€ã ã®ä»çµã¿ãçŽæçã«çè§£ããããã«ãAIããæ¢åµãã«ããããŠCoTAãã容çè ãå·§åŠã«ä»æããåœã®èšŒæ ãã«å©ããŠã¿ãŸããã ã Â
ãåŸæ¥ã®AIã¯ãäžããããæ å ±ããã¹ãŠçå®ãšããŠåãå ¥ããè«ççã«çµè«ãå°ãåºãåªç§ãªåæå®ã§ããããããããã®ããã«åœã®èšŒæ ãæ··ãããããšããããçå®ãšããŠæ±ã£ãŠããŸãã誀ã£ãç¯äººãææããŠããŸããŸãã
TPãã©ãã€ã ã¯ããã®åæå®ãæ¢åµãžãšèšç·ŽãçŽãããã»ã¹ã«äŒŒãŠããŸãã
ããŒã¿ãã€ãã©ã€ã³ïŒäºä»¶ç°¿ã®äœæïŒ
ããŸããæ¢åµã®èšç·Žçšã«ãéå»ã®äºä»¶ç°¿ïŒããŒã¿ïŒã倧éã«çšæããŸãããããã®äºä»¶ç°¿ã«ã¯ãæ¬ç©ã®èšŒæ ãšå®¹çè ã仿ããåœã®èšŒæ ïŒharmã¿ã°ã§å²ãŸããæªæããæšè«ïŒãæ··åšããŠããŸããããã«ãäºä»¶ã«äžå¯©ãªç¹ãããå ŽåïŒããªã¬ãŒãååšããå ŽåïŒã«ã¯ãäºä»¶ç°¿ã®åé ã«ãèŠæ³šæïŒ<suspect>ã¿ã°ïŒããšããä»ç®ã貌ãããŠããŸãã
匷ååŠç¿ïŒæ¢åµã®èšç·ŽïŒ
ãæ¢åµïŒAIïŒã¯ããããã®äºä»¶ç°¿ã次ã ãšè§£æ±ºããŠãããŸãããã®éãåã«ç¯äººãåœãŠãïŒæ£ããçããåºãïŒã ãã§ãªãããããã¯åœã®èšŒæ ã ããšæ£ããææãïŒharmã¿ã°ã®å 容ãç¡èŠãïŒãããã®äºä»¶ã¯äœãããããããšæåã«æ°ã¥ãïŒ<suspect>ã¿ã°ãçæããïŒããšã«å¯ŸããŠããå ±é ¬ããäžããããŸãã
ããã®èšç·Žãç¹°ãè¿ãããšã§ãæ¢åµã¯æ¬¡ç¬¬ã«ã蚌æ ãéµåã¿ã«ããããã®ä¿¡ææ§ãèªãåå³ããæªããç¹ãããã°èŠæãããšãããæ¢åµãšããŠã®åããããªãã¡èªå·±é²è¡èœåã身ã«ã€ããŠããã®ã§ãã
ããã®ããã«ãTPã¯AIã«åãªãç¥èã§ã¯ãªãããç¥çãªæçå¿ããšãèªå·±ä¿®æ£èœåããšãããããé«åºŠãªèªç¥ã¹ãã«ãæã蟌ã詊ã¿ãšèšããã§ãããã
ææ¡ææ³è©³çŽ°
ãTPãã©ãã€ã ãæ§æãã3ã€ã®èŠçŽ ã«ã€ããŠãããæè¡çãªè©³çްã解説ããŸãã
ããŒã¿åŠçãã€ãã©ã€ã³
ããã®ãã€ãã©ã€ã³ã®ç®çã¯ã匷ååŠç¿ã®ããã®é«å質ãªãå®å šæ§ãæèãããèšç·ŽããŒã¿ã»ãããäœç³»çã«çæããããšã§ã ãããã»ã¹ã¯ä»¥äžã®3段éã§æ§æãããŸãã Â
åæCoT
ããŸããæ¢åã®è³ªåå¿çããŒã¿ã»ããããã質åïŒQïŒãçãïŒAïŒããããŠãã®éã®æšè«ïŒRïŒãå«ããã¯ãªãŒã³ãªãQARããŒã¿ããäœæããŸããæšè«éšåããªãããŒã¿ã»ããã®å Žåã¯ãGPT-4oã®ãããªé«æ§èœãªã¢ãã«ãçšããŠçæããŸãã
BadChainæ³šå ¥ã®ã·ãã¥ã¬ãŒã·ã§ã³
ãæ¬¡ã«ãäœæããã¯ãªãŒã³ãªããŒã¿ã«å¯ŸããCoTAã®ä»£è¡šæ Œã§ããBadChainæ»æãã·ãã¥ã¬ãŒãããŸãã質åæã«ããªã¬ãŒïŒäŸïŒã@_@ãïŒãæ¿å ¥ããæšè«éšåã«ããªã¬ãŒãšé£åããæªæããåé·ãªæšè«ã¹ããããæ³šå ¥ããŸãã
ã©ãã«ã®åã蟌ã¿
ãæåŸã«ãæ»æãæ³šå ¥ãããããŒã¿ã«ãé²åŸ¡ã®ããã®ç¹æ®ãªã¿ã°ãåã蟌ã¿ãŸããæ»æã®å åãããããšã瀺ãããã«ãæšè«ã®éå§åã«<suspect>ã¿ã°ãæ¿å ¥ããŸãã
ããããŠãæ³šå ¥ãããæªæããæšè«ã¹ãããã<harm>ãš</harm>ã®ã¿ã°ãã¢ã§å²ã¿ãŸããããã«ãããã¢ãã«ã¯äœãèŠæããå ·äœçã«ã©ã®éšåãç¡èŠããã°ããã®ããæç€ºçã«åŠç¿ããããã®æåž«ããŒã¿ã宿ããŸãã
GRPOã«ãã匷ååŠç¿
ãèšç·Žã«ã¯ãGRPOïŒGroup Relative Policy OptimizationïŒãšåŒã°ãã匷ååŠç¿ã¢ã«ãŽãªãºã ãæ¡çšãããŸã ã匷ååŠç¿ã¯ãã¢ãã«ã®è¡åã«å¯ŸããŠå ±é ¬ãäžããããšã§ãæãŸããæ¯ãèãã匷åããŠããææ³ã§ãã
ãTPã«ãããGRPOã®æå€§ã®ç¹åŸŽã¯ãåŸè¿°ãã2çš®é¡ã®å ±é ¬ã¢ãã«ãçµã¿åãããç¹ã«ãããŸããGRPOèªäœã¯ãèšç®å¹çãé«ãã工倫ããªãããã¢ã«ãŽãªãºã ã§ãããç¶ç¶çãªé²åŸ¡èœåã®ã¢ããããŒããäœã³ã¹ãã§å®çŸããããšã«è²¢ç®ããŸã ã Â
é©å¿åã¢ãã¿ãªã³ã°ææš
ãCoTAã«å¯Ÿããé²åŸ¡æ§èœãæ£ç¢ºã«è©äŸ¡ãããããåŸæ¥ã®ææšã«å ãã2ã€ã®æ°ããææšãå°å ¥ãããŸã ã Â
Cure Rate (CR)
ãæ»æã«ãã£ãŠäœäžããæ£è§£çããé²åŸ¡çã«ãã£ãŠã©ãã ããå埩ãããããç€ºãææšã§ããæ²»çå¹æãæž¬ã尺床ãšèããããšãã§ããŸãã
Reject Rate (RR)
ãã¢ãã«ãæ»æè ã®æå³éãã«æãããããšããã©ãã ããæçµ¶ãã§ããããç€ºãææšã§ããæ»æã®å®å šãªæåçïŒASRCïŒãã©ãã ãæžå°ãããã§æž¬å®ãããŸãã
ææ¡ææ³ã®æ§æã³ã³ããŒãã³ãããä»çµã¿ã®è©³çް
ãTPãã©ãã€ã ã®æå¹æ§ãæ¯ããæãéèŠãªä»çµã¿ã¯ã匷ååŠç¿ã«ãããå ±é ¬èšèšãç¹ã«ã2ã€ã®å ±é ¬ã¢ãã«ãã®çµã¿åããã«ãããŸãã
ç¹æ®ã¿ã°ã®æ©èœ (<suspect>, <harm>)
<suspect>ã¿ã°ã¯ãã¢ãã«ã«ãäºé²çèŠæããæããŸããããªã¬ãŒã®ãããªäžå¯©ãªå ¥åãã¿ãŒã³ãæ€ç¥ããéã«ãæ¬æ Œçãªæšè«ã«å ¥ãåã«ãŸãèŠåãã©ã°ãç«ãŠãè¡åãåŠç¿ãããŸããããã«ãããåé¡ã®æ©æçºèŠèœåãé€ãããŸã ã Â
<harm>...</harm>ã¿ã°ãã¢ã¯ãã¢ãã«ã«ãç¹å®ãšç¡èŠããæããŸããæèã®é£éã®äžã«çŽã蟌ãã ãæ»æè ã«ããæªæããæšè«ã¹ããããæ£ç¢ºã«ç¹å®ãããã®éšåããæ¬åŒ§ã§ããã£ãŠç¡èŠããããã®ãããªæ¯ãèããåŠç¿ãããŸããããã«ãããæ±æãããæ å ±ã«æããããã«æ£ããçµè«ãžè³ãèœåãç²åŸããŸã ã Â
ãã¥ã¢ã«å ±é ¬ã¢ãã«
ãTPã®é©æ°æ§ã¯ããã®2çš®é¡ã®å ±é ¬ãçµã¿åãããç¹ã«éçŽãããŸã ã Â
ææå ±é ¬ã¢ãã«
ãããã¯åŸæ¥ã®å€ãã®AIåŠç¿ã§çšãããããã®ã§ããæçµçãªçããæ£ãããããšãããçµæãã«å¯ŸããŠå ±é ¬ãäžããŸããã¿ã¹ã¯ã®ããã©ãŒãã³ã¹ãç¶æã»å埩ããã圹å²ãæ ããŸãã
ããã»ã¹å ±é ¬ã¢ãã«
ããããTPã®æ žå¿ã§ããæçµçãªçããåã£ãŠãããåŠãã«ãããããããæèã®ããã»ã¹ãå®å šã§ãã£ããããè©äŸ¡ããå ±é ¬ãäžããŸããå ·äœçã«ã¯ã<suspect>ã¿ã°ã<harm>ã¿ã°ãé©åã«äœ¿çšã§ããå Žåã«é«ãå ±é ¬ãäžããããŸãã
ããã®ORMãšPRMã®çµã¿åããã¯ãAIã®åŠç¿å²åŠã«ãããéèŠãªè»¢æã瀺åããŠããŸããåŸæ¥ã®AIã¢ã©ã€ã¡ã³ãïŒAIã人éã®æå³ã«æ²¿ãããæè¡ïŒããäž»ã«äžé©åãªãçµæãã®çæãé²ãããšã«æ³šåããŠããã®ã«å¯ŸããTPã®PRMã¯äžé©åãªãããã»ã¹ããã®ãã®ãæ¯æ£ããããšè©Šã¿ãŸãã
ãããã¯ãåãªãè¡åå¶åŸ¡ãããããèªç¥çãªã¬ãã«ã§ã®ã¢ã©ã€ã¡ã³ããžã®ç§»è¡ãšèšããã§ããããè«æã®å®éšã§ã瀺ãããŠããããã«ãORMã®ã¿ã§åŠç¿ãããããšãããšïŒã€ãŸããæ±æãããæèããã»ã¹ã®ãŸãŸç¡çããæ£ããçããåºãããããšãããšïŒãããã£ãŠæ§èœãæªåããã±ãŒã¹ãããããŸã ã
ãããã¯ãç æ°ã®æ ¹æ¬åå ïŒããã»ã¹ã®æ±æïŒãæŸçœ®ããŠãçç¶ïŒééã£ãçãïŒã ããæã蟌ãããšãããããªãã®ã§ããPRMã¯ãã¢ãã«ã«èªãã®æèããã»ã¹ãèªå·±èšºæããä¿®æ£ããèœåãäžããããšã§ãç æ°ã®æ ¹æ¬æ²»çãç®æããŸãã
ãããã¯ã人éã«äŸããã°ãåã«èšç®çµæãæèšãããã®ã§ã¯ãªããèšç®ã®åçãçè§£ãããæ€ç®ããæ¹æ³ãæããããšã«äŒŒãŠããŸããå°æ¥ã®å ç¢ãªAIã®å®å šæ§ã¯ããã®ããã«ã¢ãã«ã®ãå ççã»èªå·±æ¹å€çãªèœåããããã«è²ããã«ããã£ãŠããã®ãããããŸãã
å®é𿹿³
è©äŸ¡å¯Ÿè±¡ã¢ãã«
ãèè ãã¯2çš®é¡ã®æå 端ã¢ãã«ã§TPé²åŸ¡ã®å¹æã詊ããŸãããäžã€ã¯æšè«ç¹åã®LRMã§ããDeepSeek-R1-Distill (ããŒã¹ã¯Llama, 8åãã©ã¡ãŒã¿)ã§ããããäžã€ã¯Alibabaãæè¿å ¬éãã髿§èœLRMã§ããQwen-3 (8åãã©ã¡ãŒã¿)ã§ãã
ããããã¯ããšããšåŒ·ååŠç¿ã§CoTèœåã䌞ã°ããã¢ãã«ã§ãããCoTæ»æã«å¯Ÿããèæ§ã¯æªç¥æ°ã§ããããŸãæ¯èŒãšããŠãåŸæ¥åã®LLMã§ããMeta-Llama-3.1-8B-InstructïŒChatGPTé¡äŒŒã®æç€ºèª¿æŽæžã¿ã¢ãã«ïŒã«ãTPãé©çšããéåžžã®LLMã«ãæå¹ãæ€èšŒããŠããŸãã
ããŒã¿ã»ãã
ãè©äŸ¡ã«ã¯4çš®é¡ã®è³ªåå¿çããŒã¿ã»ããã䜿çšããŸããã
â LetterïŒã¢ã«ãã¡ãããæšè«ã¿ã¹ã¯ïŒäŸïŒãæ¬¡ã®æåã¯ïŒãïŒ
â¡CSQAïŒCommonsenseQAïŒïŒåžžèç倿ãåãã¯ã€ãº
â¢GSM8KïŒæ°åŠçèšç®åé¡
â£StrategyQAïŒçŽ æŽãªç¥è質åã§ãã
ããããã¯æšè«ã®çš®é¡ãç°ãªã倿§ãªã¿ã¹ã¯ã§ãTPé²åŸ¡ãã©ã®åéã§ãæ©èœããããæ€èšŒããçãããããŸãã
æ»æã·ããªãª
ãå®éšã§ã¯ãåããŒã¿ã»ããã®è³ªåã«BadChainãšåæ§ã®ããã¯ãã¢ãä»èŸŒã¿ãŸãããå ·äœçã«ã¯è³ªåæã®æå®äœçœ®ã«ããªã¬ãŒã@_@ããæ¿å ¥ããã·ã¹ãã ããã³ããã§ããã®èšå·ãèŠããçããæ¬¡ã®éžæè¢ã«ãããããšããæªæã®ã«ãŒã«ãæã蟌ãã§ããã¢ãã«ã«åçãããŸãã
ãã¢ãã«ãäœãé²åŸ¡çãæããªããã°ãåœç¶ãªãã質åãšç¡é¢ä¿ãªæäœïŒäŸãã°çããæ¬¡ã®éžæè¢ã«å€æŽïŒãå®è¡ããŠããŸããçãã誀ããŸãããã®æ»æã«å¯Ÿããææ¡ææ³TPã®æç¡ã§ã¢ãã«ã®æ¯ãèããæ¯èŒããŸããã
ããŒã¹ã©ã€ã³
ãæ¯èŒå¯Ÿè±¡ãšããŠä»¥äžãèšå®ããŸããã
1) OriginalïŒå ã®ã¢ãã«ïŒ: äœã®é²åŸ¡ãæœããŠããªãç¶æ ã
2) ORM-only: åŸæ¥åã®åŒ·ååŠç¿ïŒåºåã®æ£ç¢ºãã«å¯ŸããŠã®ã¿å ±é ¬ãäžããïŒãè¡ã£ãã¢ãã«ã
ãããã¯Output Reward ModelãšãåŒã°ããTPãšã¯ç°ãªãå®å šæ§ã¯èæ ®ããŸãããTPã®æå¹æ§ã¯ããã®ORM-onlyãšæ¯ã¹ãŠæ»æèæ§ãã©ãã ãæ¹åãããã§è©äŸ¡ããŸãã
è©äŸ¡ææš
ãåè¿°ã®ASR, ASR<sub>c</sub>, CR, RRã«å ããéåžžæã»æ»ææã®æ§èœïŒAccuracyïŒã枬å®ããŸãããAccuracy_cleanã¯é垞質åã§ã®æ£ççãAccuracy_badchainã¯æ»æè³ªåã§ã®æ£ççãæå³ããŸããASRã¯äœãã»ã©è¯ããCRãRRã¯é«ãã»ã©è¯ãææšã§ãã
å®éšçµæ
ãå®éšçµæã¯ãTPãã©ãã€ã ã®æå¹æ§ãæç¢ºã«ç€ºããŠããŸãã以äžã«ãè«æã§æç€ºãããäž»èŠãªçµæã解説ããŸãã
äž»èŠã¢ãã«ã«ãããé²åŸ¡æ§èœ

ããã®è¡šã¯ãæ¬ç ç©¶ã®äž»åŒµãè£ä»ããæãéèŠãªèšŒæ ã§ããæ³šç®ãã¹ãã¯ãCRïŒå埩çïŒãšRRïŒæçµ¶çïŒã®åã§ããã»ãšãã©ã®ã±ãŒã¹ã§ãTPïŒOursïŒææ³ã¯ãã©ã¹ã®å€ã瀺ããŠãããé²åŸ¡ãæå¹ã«æ©èœããŠããããšã瀺ããŠããŸãã
ãç¹ã«ãDeepSeekã¢ãã«ã«å¯ŸããLetterããŒã¿ã»ããã®å®éšã§ã¯ãRRã40.00%ãšãæ»æè ã®æå³éãã«æãããããšãå€§å¹ ã«æçµ¶ã§ããŠããããšãåãããŸãã
ãå¯Ÿç §çã«ãBaselineïŒORM-onlyïŒã¯ãCRãRRããã€ãã¹ã«ãªãã±ãŒã¹ãæ£èŠãããŸããããã¯ãçµæã®æ£ããã ãã远æ±ããåçŽãªåŒ·ååŠç¿ã§ã¯ãé²åŸ¡ã©ããããããã£ãŠè匱æ§ãæªåãããããªãããšã瀺åããŠãããããã»ã¹ã®å®å šæ§ãéèŠããPRMã®éèŠæ§ãéç«ãããŠããŸãã
æ±çšLLMãžã®é©çšå¯èœæ§

ããã®è¡šã¯ãTPãã©ãã€ã ãæšè«ã«ç¹åããLRMã ãã§ãªããLlama-3.1ã®ãããªæ±çšLLMã«å¯ŸããŠãæå¹ã§ããããšã瀺ããŠããããã®å¿çšç¯å²ã®åºããç©èªã£ãŠããŸãã
ãCSQAãStrategyQAããŒã¿ã»ããã«ãããŠãTPææ³ãCRãšRRã®äž¡æ¹ã§é¡èãªæ¹åã瀺ããŠããããšã¯ããã®ã¢ãããŒããç¹å®ã®ã¢ãã«ã¢ãŒããã¯ãã£ã«éå®ãããªããããæ®éçãªé²åŸ¡åçã§ããããšã瀺åããŠããŸãã
é²åŸ¡ããžãã¯ã®éçšå¯èœæ§

ããã®è¡šã¯ãæ¬ç ç©¶ã®äžã§ãç¹ã«ç€ºåã«å¯ãçµæã瀺ããŠããŸãããAnti-TPããšã¯ãTPã®å ±é ¬ããžãã¯ãæå³çã«ãé転ããããèšå®ã§ããã€ãŸããå®å šãªããã»ã¹ïŒ<suspect>ã¿ã°ã®äœ¿çšãªã©ïŒã眰ããæªæããæšè«ãåãå ¥ããããšãå ±é ¬ãšããŠã¢ãã«ãåŠç¿ãããã±ãŒã¹ã§ãã
ãçµæã¯è¡æçã§ããå€ãã®ããŒã¿ã»ããã§ãAnti-TPã¯å ã®æ»æç¶æ ãããããã«æ§èœãäœäžããïŒACCâïŒãæ»æè ã®æå³éãã«æãããå²åãå¢å ãããŠããŸãïŒASRCâïŒã
ãããã¯ãTPã®æ ¹å¹¹ããªã匷ååŠç¿ã®ã¡ã«ããºã ããé²åŸ¡ã ãã§ãªãã匷åãªãæ»æããŒã«ãã«ããªãåŸãããšãæå³ããŠããŸãã
èå¯
ãªããã®ææ³ãåªããŠããã®ã
ãThought Purityãã©ãã€ã ã®åªç§ãã¯ããã°ããã¢ãã«èªèº«ã«èªå·±é²è¡ãããããšããçºæ³ã®åå©ã§ããåŸæ¥ãé²åŸ¡ã¯ã¢ãã«å€éšããã«ãŒã«ã§çžã£ããæ€é²ãããããã®ãäž»æµã§ããããããTPã¯ã¢ãã«å éšã«âæäœâãäœã蟌ãããšã§ãããªãããã€æç¶çãªèæ§ãå®çŸããŸãããå ·äœçãªåªäœç¹ãæŽçããŸãã
ç·åé²åŸ¡ã«ããæãç®ã®ç¡ã
ãTPã¯ããŒã¿âåŠç¿âè©äŸ¡ãŸã§äžè²«ããŠã»ãã¥ãªãã£ãçµã¿èŸŒããããç¹å®éšåã ã察çããææ³ã«æ¯ã¹æã穎ãå°ãªãã§ãã仿æ³ãåäžã®çïŒäŸ: ããã³ããæ€é²ã®ã¿ïŒãªã®ã«å¯ŸããTPã¯å€å±€é²åŸ¡ãšãªã£ãŠããŸãã
ãããã«ããæ°çš®ã®æ»æãåºçŸããŠããããŒã¿çæãã€ãã©ã€ã³ãæŽæ°ãååŠç¿ããã°å¯Ÿå¿ã§ããã¢ããããŒãå®¹ææ§ãé«ããšèšããŸãã
æ§èœç¶æãšã®äž¡ç«
ãå®å šæ§ãé«ããã ããªãåºåãå šéšæåŠããã°æ¥µç«¯ãªè©±å¯èœã§ãããããã§ã¯AIã®äŸ¡å€ãæãªãããŸããTPã¯åŒ·ååŠç¿ã§å®å šãšã¿ã¹ã¯æ§èœã®äž¡æ¹ãå ±é ¬ã«çµã¿èŸŒãã ãããã¢ãã«ãå®å šçãšæ£ççã®ãã©ã³ã¹ãèªãæé©åã§ããŸãã
ãå®éãå®éšã§ãTPã¢ãã«ã¯å ã®æ§èœãã»ãŒç¶æãã€ã€å®å šæ§ã ãåäžããŸããïŒè¡š1ã»2ïŒããã®ç¹ããå®ããåºããŠãå©äŸ¿æ§ã¯èœãšããªããã®ã¯å€§ããªå©ç¹ã§ãã
ã¢ãã«ãµã€ãºã»çš®é¡ã«äŸåããªã
ãTPã¯åºæ¬çã«ããŒã¿æ¡åŒµãšæ±çšçãªRLåŠç¿ã®æ çµã¿ã§ãããããå€§èŠæš¡ã¢ãã«ããå°èŠæš¡ã¢ãã«ãŸã§é©çšå¯èœã§ããè«æã§ã¯8BïŒ8åïŒèŠæš¡ã§æ€èšŒãããŸããããææ³èªäœã¯çè«äžãã倧ããªGPT-4ã¯ã©ã¹ã«ãå¿çšã§ããã§ãããã
ããŸãå®éšã§InstructåLLMã«ã广ããã£ãããã«ãäºååŠç¿æžã¿ã®ã©ããªã¢ãã«ã§ãåŸä»ãã§åŒ·åã§ããæè»æ§ããããŸããããã¯ç¹å®ã¢ãã«ã«ç¹åãã察çã§ãªãæ®éæ§ã瀺ããŸãã
æ»ææ€ç¥ãšåŸ©æ§ãäž¡é¢ãµããŒã
ãå€ãã®åŸæ¥æ³ã¯ãæ»æãããªããããšã«æ³šåããäœããäžäžæ»æãæåããéã®ãªã«ããªãŒã«ã¯ç¡åã§ãããTPã¯suspectã¿ã°ã§æ€ç¥ããharmã¿ã°ã§åŸ©æ§ããäºæ®µæ§ãã®èšèšãªã®ã§ãä»®ã«ããªã¬ãŒæç€ºãèžã¿ãããŠãéäžã§èžã¿çãŸããå¯èœæ§ããããŸãã
ãããã¯äŸãã°ã人éãæ¬ºçã«æ°ä»ããä»ã®èãåãæ¶ããããšä¿®æ£ã§ããã®ã«äŒŒãŠããŸããã¢ãã«ã«ããããèªå·±ä¿®æ£èœåãæãããã®ã¯ç»æçã§ãã
仿æ³ãšã®æ¯èŒïŒåªäœæ§ã»å£äœæ§
åªäœæ§
ãä»ã®ããã¯ãã¢é²åŸ¡ææ³ãšæ¯ã¹ãTPã¯ãŸãå
æ¬æ§ã§çŸ€ãæããŠããŸãã仿³ãæ€ç¥ã ãããããã¯å
¥åå¶çŽã ãã ã£ãã®ã«å¯ŸããTPã¯æ€ç¥ã»é€å»ã»èšç·Žããè©äŸ¡ãŸã§ç¶²çŸ
ããŠããŸãããã®ããBadChainãPoisonPromptãªã©å€æ§ãªæ»æãã¿ãŒã³ã«ãæ±çšçã«å¹ããšæåŸ
ãããŸãã

ãå®éãå³4ã§ç€ºãããããã«TPã¢ãã«ã¯ã©ããªæ»æç¶æ³äžã§ãæ§èœå£åãæå°éã«æããããåŸåãèŠãããŸãããããã«ãGuardReasonerã·ãªãŒãºã®ãããªãžã§ã€ã«ãã¬ã€ã¯å¯Ÿçãšã䜵çšå¯èœã§ãå°æ¥çã«ãã€ããªããé²åŸ¡ãæ§ç¯ã§ãã䌞ã³ããããããŸããã¢ãã«ã®å®å šæ§ç ç©¶å šäœãåºäžãããäžããŒã¹ã«ãªãã§ãããã
å£äœæ§ã»èª²é¡
ããã¡ããTPã«ã課é¡ã¯æ®ããŸãã第äžã«å®ç§é²åŸ¡ïŒçæ³ç¶æ
ïŒã¯ãŸãã ãšããç¹ã§ããå®éšã§ãPerfect Reasoningã«åé¡ãããå¿çã¯æ¥µããŠå°ãªããã»ãšãã©ãWarning/Cleanæ¢ãŸãã§ãããã€ãŸãæå®³æèãå®å
šã«0ã«ããã®ã¯é£ããã®ãçŸç¶ã§ãã
ãClean Answerã§ã¯ãããã«äžå¿ èŠãªæçïŒäŸãã°ã¿ã°ç±æ¥ã®çè·¡ïŒãåºåã«æ··ãã£ãŠããŸãã100%綺éºãšã¯èšããŸããããã®èŸºãã¯å€éšã¹ã¯ãªããçã§åŸåŠçããã°è§£æ±ºå¯èœã§ããããããªãã¢ã«ãŽãªãºã æ¹åãæãŸããŸãã
ã第äºã«ã匷ååŠç¿ã®èšç®ã³ã¹ãã課é¡ã§ããTPã¯é«åºŠãªå ±é ¬èšèšãšè¿œå ããŒã¿ã§ã¢ãã«ãåèšç·Žããå¿ èŠããããäžè¬ã®åŸ®èª¿æŽãããªãœãŒã¹ãèŠããŸããæšä»ã®LLMã¯å·šå€§åããŠããããããã®æ¹æ³ãç£æ¥å¿çšããã«ã¯å¹çåïŒäŸãã°LoRAãªã©ãã©ã¡ãŒã¿å¹çç埮調æŽãšã®çµã¿åããïŒãéµã«ãªãã§ãããã
ã第äžã«ãæ°çš®ã®æ»æïŒAnti-TPæ»æïŒã®å¯èœæ§ã§ããè«æã§ãTPãéæã«åããAnti-TPããæ€èšããŠããŸããæ»æè ãTPã®ååšãç¥ãã°ãäŸãã°<suspect>ã¿ã°ãåé¿ããå·§åŠãªããªã¬ãŒåã蟌ã¿ãã<harm>ã¿ã°ãæ©èœããªããããªæªä¹±ã仿ãããããããŸããã
ããã®å¯Ÿæææ®µãšé²åŸ¡çã®ããã¡ãã£ãã¯ä»åŸãç¶ãå¯èœæ§ããããŸãããã ãTPèªäœãæè»ãªæ çµã¿ãªã®ã§ãä»®ã«æ°æ»æãåºãŠããããæš¡å£ããããŒã¿ãäœãçŽããŠååŠç¿ããã°å¯ŸåŠã¯å¯èœã§ãããã
ããããæ¬ææ³ã®åŒ·ã¿ã¯ãæ»æè ã«TPåé¿ã匷ããããšã§æ»æã³ã¹ããäžããããç¹ã«ããããŸãïŒç°¡åã«ã¯æªçšãããªãã¢ãã«ã«ãªãïŒã
ãç·åãããšãThought Purityã¯çŸæç¹ã§CoTããã¯ãã¢é²åŸ¡ã®ãã¹ããã©ã¯ãã£ã¹ãšèšããå®æåºŠã§ããä»ã®éå®çãªå¯Ÿçãçµã¿åãããããšã§ãããã«åŒ·åºãªé²åŸ¡ãæåŸ ã§ããŸãã
ãä»åŸã¯æ¬ææ³ãããå€§èŠæš¡ãªã¢ãã«ãçŸå®äžçã®å¯Ÿè©±ã·ã¹ãã ã«é©çšããæå¹æ§ãšå®çšæ§ãæ€èšŒããŠãããã§ãŒãºãšèšããã§ãããã
çµè«
ãæ¬çš¿ã§è©³è¿°ããç ç©¶ã¯ããTPããšããæ°ããªé²åŸ¡ãã©ãã€ã ãæç€ºããããšã§ãLLMã®æèããã»ã¹ãæšçãšããCoTAãšããæ·±å»ãªè åšã«å¯Ÿãã倧ããªåé²ã瀺ããŸããã
ãTPã¯ãåãªãç¹å®ã®æè¡ç解決çã«çãŸãããAIã®å®å šæ§ã«é¢ããæ°ããèãæ¹ãææ¡ããŠããŸããããã¯ã匷ååŠç¿ãéããŠãã¢ãã«èªèº«ã«å ççãã€èªå·±æ¹å€çãªèœåãããªãã¡ãèªãã®æèãçããæµåãããèœåãæãããšããã¢ãããŒãã§ã ã Â
ãå®éšã¯ãTPã倿§ãªã¢ãã«ãšã¿ã¹ã¯ã«ãããŠæå¹ã§ããããšãå®èšŒãããšåæã«ã髿§èœãªAIã»ã©æäœããããããšãããã©ããã¯ã¹ããé²åŸ¡æè¡ãæªçšããåŸããšããäºé¢æ§ãªã©ãAIã®å®å šæ§ã«é¢ããæ ¹æºçãªèª²é¡ãæµ®ã圫ãã«ããŸããã
ããšãã
ãTPãã©ãã€ã ã¯ãAIã®å éšæèãå®ããšããæ°ããªé²åŸ¡æŠå¿µãå®èšŒããŸãããæ»ææåçãå€§å¹ ã«åæžããªããã¢ãã«æ§èœãç¶æã§ããããšã¯ãå®çšåãžã®éèŠãªäžæ©ãšãããŸãããããå®éšçµæã瀺ãããã«ãå®å šãªé²åŸ¡ã«ã¯è³ã£ãŠããããèŠåä»ãå¿çãéšåçãªæ±æé€å»ã«ãšã©ãŸãã±ãŒã¹ãå°ãªããããŸããã
ãæ¬ç ç©¶ã®æçŸ©ã¯ãåäžã®é²åŸ¡æè¡ã®ææ¡ã«ãšã©ãŸãããããŒã¿æºåããåŠç¿ãè©äŸ¡ãŸã§äžè²«ããã»ãã¥ãªãã£èšèšã®éèŠæ§ã瀺ããç¹ã«ãããŸããä»åŸãæ»æè ãThought Purityã®ä»çµã¿ãçè§£ããæ°ããªåé¿çãè¬ããå¯èœæ§ã¯åŠå®ã§ããŸããã
ãããããé²åŸ¡åŽãç¶ç¶çã«ããŒã¿ãã€ãã©ã€ã³ãæŽæ°ãã匷ååŠç¿ã«ããåèšç·Žãè¡ãããšã§ãé²åããè åšã«å¯Ÿå¿ã§ããæè»æ§ãåããŠããŸãã
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7æã«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
References
ãæ¬çš¿ã®å 容ã«ããã«æ·±ãé¢å¿ãæãããèªè ã®ããã«ãåèæç®ãããã€ã玹ä»ããŸãã
Xiang, Z., et al., "BadChain: Backdoor Chain-of-Thought Prompting for Large Language Models," ICLR (2024).
ãæ¬çš¿ã§ç޹ä»ãããThought Purityãã察æããäž»èŠãªè
åšã§ãããBadChainãæ»æãåããŠææ¡ãããã®è©³çްãæããã«ããåºç€çãªè«æã§ããåçŽãªããã³ããæ³šå
¥ã«ãã£ãŠLLMã®æšè«ããã»ã¹ãããã«æäœã§ãããã瀺ããæ°ããªé²åŸ¡ãã©ãã€ã ã®å¿
èŠæ§ãæµ®ã圫ãã«ããŸãã ã Â
Li, Y., et al., "BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks on Large Language Models," arXiv (2024).
ããã®ç ç©¶ã¯ãããã¯ãã¢æ»æã®åºç¯ãªæèãçè§£ããããã®éèŠãªãã¬ãŒã ã¯ãŒã¯ãæäŸããŸããCoTAãå«ãæ§ã
ãªæ»æã¿ã€ããåé¡ãããã³ãããŒã¯åããããšã§ãç ç©¶è
ãè
åšãšé²åŸ¡çãäœç³»çã«è©äŸ¡ããããšãå¯èœã«ããŸããããThought Purityãã®ãããªææ³ã®æå¹æ§ãå³å¯ã«æž¬å®ã§ããæšæºåãããç°å¢ã確ç«ããç¹ã§ãéåžžã«äŸ¡å€ããããŸã ã Â
Li, X., et al., "Chain-of-Scrutiny: Detecting Backdoor Attacks for Large Language Models," CoRR (2024).
ããã®è«æã¯ããThought Purityããšåæ§ã«ãã¢ãã«èªèº«ã®æšè«èœåãæŽ»çšããå¥ã®é²åŸ¡æŠç¥ãææ¡ããŠããŸããã¢ãã«ã«èªèº«ã®åºåã®ççŸã粟æ»ããããã«ä¿ãã¢ãããŒãã§ãããåãåé¡ã«å¯Ÿããç°ãªãå²åŠçã¢ãããŒããæäŸããŸããTPã®åŒ·ååŠç¿ããŒã¹ã®ææ³ãšæ¯èŒããããšã§ãå€è§çãªèŠç¹ãåŸãããŸã ã Â
Liu, Y., et al., "GuardReasoner: Towards Reasoning-based LLM Safeguards," arXiv (2025).
ããã®ç ç©¶ããŸããæšè«ããŒã¹ã®é²åŸ¡ã¡ã«ããºã ãæ¢æ±ããŠãããã¢ãã«ã®ç¥æ§ãå®å
šæ§åäžã®ããã«å©çšãããšããæè¿ã®ãã¬ã³ãã象城ããŠããŸããå®å
šæ§ã«é¢ããåé¡ã®ããã«æç¢ºãªæšè«ã¹ããããçæããããã¢ãã«ããã¡ã€ã³ãã¥ãŒãã³ã°ããããšã«çŠç¹ãåœãŠãŠããããThought Purityãã§æç€ºãããã¢ã€ãã¢ãè£å®ããå
容ãšãªã£ãŠããŸã 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
- #AI
- #çæAI
- #ChatGPT
- #ã¢ãã«
- #ã»ãã¥ãªãã£
- #LLM
- #æ©æ¢°åŠç¿
- #ãµã€ããŒã»ãã¥ãªãã£
- #ããã³ãããšã³ãžãã¢ãªã³ã°
- #æ»æ
- #ããã»ã¹
- #å€§èŠæš¡èšèªã¢ãã«
- #AIã»ãã¥ãªãã£
- #æè¡è§£èª¬
- #匷ååŠç¿
- #é²åŸ¡
- #ããã³ããã€ã³ãžã§ã¯ã·ã§ã³
- #AIå®å š
- #cota
- #Chain_of_Thought
- #ThoughtPurity
- #BadChainæ»æ