
AIã³ãŒãã£ã³ã°ã®æ¬¡ã®è©°ãŸãã©ããã¯ãæèãã©ãã«çœ®ããã
AIã³ãŒãã£ã³ã°æ¯æŽã®è©±ãããŠãããšãã©ãããŠãã¢ãã«ã®è³¢ãã«ç®ãè¡ãã
ã©ã®ã¢ãã«ããã³ãããŒã¯ã§åã£ãã®ããã©ã®ãšãŒãžã§ã³ããäœåã§PRãåºããã®ããã©ã®IDEã«å ¥ã£ãã®ãã
ãã¡ãããããã¯å€§äºã ã
ãã ãå®éã«ã³ãŒãããŒã¹ã®äžã§AIã䜿ã£ãŠãããšãå¥ã®è©°ãŸãã©ãããèŠããŠãããã¢ãã«ãäžè¬çã«è³¢ããã©ããããããã®ãªããžããªã®çãã©ããŸã§ç¥ã£ãŠãããã®ã»ããå¹ããŠããå Žé¢ãããã
åœåèŠåããã£ã¬ã¯ããªæ§é ããã¹ãã®æžãæ¹ã瀟å ã ãã§éããèãæœè±¡ãæã®éœåã§æ®ã£ãŠããå€ãªAPIã
ãããããã®ã¯ãããã¥ã¡ã³ãã«ãããã«æžãããŠããªããããã©ãã³ãŒããæžããšãã«ã¯å¹ãã
ä»åèªãã è«æã¯ããŸãã«ãããæ±ã£ãŠããã
University of Waterlooã®ç ç©¶ããŒã ã«ãããCode2LoRA: Hypernetwork-Generated Adapters for Code Language Models under Software Evolution ã ã
ã¿ã€ãã«ã ãèŠããšãæ©æ¢°åŠç¿ã®åãã匷ãè«æã«èŠãããå®éãLoRAãhypernetworkãGRUãrepository-level PEFTãšãã£ãèšèã䞊ã¶ã
ã§ããåãé¢çœããšæã£ãã®ã¯ãææ³ãã®ãã®ãããåãã®çœ®ãæ¹ã ã£ãã
èŠããã«ãã®è«æã¯ãããèããŠããã
ãªããžããªã®ç¥èã¯ãæ¯åããã³ããã«è©°ã蟌ãã¹ããªã®ãããããšããã¢ãã«ã®åŽã«äžåºŠæã¿èŸŒãŸããã¹ããªã®ãã
ããã¯AIã³ãŒãã£ã³ã°æ¯æŽã®å®åã«ããªãè¿ãè«ç¹ã ãšæãã
RAGã§å šéšèªãããšããçŽ æŽãªéç
ããŸã®AIã³ãŒãã£ã³ã°æ¯æŽã§ã¯ããªããžããªã®æèãæž¡ãæ¹æ³ãšããŠRAGããã䜿ãããã
é¢é£ããããªãã¡ã€ã«ãæ€çŽ¢ããŠãããã³ããã«å ¥ããã颿°å®çŸ©ãäŸåé¢ä¿ããã©ã£ãŠãå¿ èŠãããªã³ãŒãçãæã£ãŠããããã£ããèšãã°ãããããããæ¹ã ã
ããã¯èªç¶ã ããå®éã«åœ¹ã«ç«ã€ã
ãã ããªããžããªã倧ãããªããšãæ¯åãã©ã®æèãæã£ãŠãããããå€ããªã¹ã¯ãåºããå¿ èŠãªãã¡ã€ã«ãèœãšããããããªããéã«ããããªãæèã§ã³ã³ããã¹ããåãããããããªãã
ããããåŒã³åºããã³ã«æèãããã³ãããžèŒããã®ã§ãæšè«æã®ããŒã¯ã³ã³ã¹ããå¢ããã
è«æã®å顿èã¯ããã«ããã
ãªããžããªç¥èãæ¯åå ¥åãšããŠæž¡ãã®ã§ã¯ãªãããã®ãªããžããªçšã®LoRAã¢ããã¿ãšããŠçæã§ããªããã
ãããCode2LoRAã®çºæ³ã ã
ãªããžããªããLoRAãäœã
Code2LoRAã¯ããªããžããªãèªãã§ããã®ãªããžããªå°çšã®LoRAã¢ããã¿ãçæããã
æ®éã®LoRAã¯ãããã¿ã¹ã¯ãããŒã¿ã»ããã«å¯ŸããŠè¿œå åŠç¿ããŠäœããã€ãŸãããªããžããªããšã«LoRAãäœãããšãããšããã®ãã³ã«åŠç¿ãå¿ èŠã«ãªãã
Code2LoRAã¯å°ãéãã
ãŸããªããžããªå šäœãåã蟌ã¿ã«å€æããããã®åã蟌ã¿ãhypernetworkã«æž¡ããhypernetworkãã察象ã®ã³ãŒãLLMã«å·®ã蟌ãLoRAéã¿ãåºãã
ã€ãŸãããªããžããªãå ¥åãããšããã®å Žã§ãªããžããªå°çšã¢ããã¿ãåºãŠããã
è«æã§ã¯ãããäºã€ã®åœ¢ã§è©ŠããŠããã
äžã€ã¯ Code2LoRA-Staticã
ããã¯ãããæç¹ã®ãªããžããªã®ã¹ãããã·ã§ããããLoRAãäœããå®å®ããã³ãŒãããŒã¹ãçè§£ããçšéã«åããŠããã
ããäžã€ã¯ Code2LoRA-Evoã
ãã¡ãã¯ãã³ãããããšã®å·®åãé çªã«èŠãŠãGRUã®é ãç¶æ ãæŽæ°ããªããLoRAãä¿ã€ããªããžããªãå€åããŠããåæã§ãã¢ããã¿ã远åŸãããã
ãã®ãEvoãã倧äºã ã
ãœãããŠã§ã¢ã¯éæ¢ç»ã§ã¯ãªãããªããžããªã¯åžžã«å€ãããæšæ¥ãŸã§æ£ããã£ãèšèšå€æãã仿¥ã®ã³ãããã§å°ããºã¬ãã
AIã³ãŒãã£ã³ã°æ¯æŽããæ¬åœã¯ãã®å€åã«è¿œãã€ããªããšãããªãã
ãã³ãããŒã¯ã¯ããã¹ãã®assertè£å®
ç ç©¶ããŒã ã¯ããã®ææ³ãè©äŸ¡ããããã« RepoPeftBench ãšãããã³ãããŒã¯ãäœã£ãŠããã
察象ã¯604åã®Pythonãªããžããªãpytestãunittestã䜿ã£ãŠããŠãã©ã€ã»ã³ã¹ã掻åç¶æ³ãªã©ã§ãã£ã«ã¿ããããã®ã ã
ã¿ã¹ã¯ã¯ããã¹ãã³ãŒãäžã®assertion completionã
ã€ãŸãããã¹ããã¡ã€ã«ã®éäžãŸã§ãèŠããŠãassertã®æåŸ å€ãäºæž¬ãããã
ããã¯å°å³ã ãããªããžããªçè§£ãèŠãã«ã¯æªããªãããªããªããæåŸ å€ãåœãŠãã«ã¯ããã®ãããžã§ã¯ãã®é¢æ°ã®æ¯ãèãããã¹ãã®æžãæ¹ãè£å©é¢æ°ãå éšAPIã®çãèªãŸãªããšãããªãããã ã
è«æã§ã¯ãåäžã¹ãããã·ã§ãããèŠãstatic trackãšãã³ãããå±¥æŽããäœãevolution trackã®äºã€ãçšæããŠããã
æ°åãã¯ã£ããããŠããã
static trackã§ã¯ãCode2LoRA-Staticãcross-repoã®Exact Matchã§ 63.8%ãRAG㯠39.7%ãäŸåé¢ä¿ã§æèã解決ããæ¹æ³ã¯ 48.2%ãFFT+RAGã§ã 53.9% ã ã£ãã
in-repoã§ã¯Code2LoRA-Staticã 66.2% ã§ããªããžããªããšã«åå¥åŠç¿ããPer-repo LoRAã® 64.0% ãšåç以äžã«ãªã£ãŠããã
evolution trackã§ã¯ãCode2LoRA-Evoãcross-repoã§ 60.3%ãåäžã®å ±æLoRA㯠55.1% ãªã®ã§ãå·®åã远ããããæ§é ãå¹ããŠããã
ããã«ã2025幎4æä»¥éã«äœããã92ãªããžããªã®OODè©äŸ¡ã§ããCode2LoRA-Evo㯠74.1% ã§æãé«ãããã ãããã¯è«æèªèº«ã泚æããŠããŠãOODåŽã¯assertã®æåŸ å€ãçããªã®ã§ãExact Matchãèšãã¿ãããã
ãããã¡ãããšèªåã§æžããŠããã®ã¯å¥œæãæãŠãã
é¢çœãã®ã¯ãé·ãã³ã³ããã¹ãäžèœè«ããžã®åè«ã«ãªã£ãŠãããšãã
ãã®è«æãèªãã§ããŠããã¡ã°ãåŒã£ããã£ãã®ã¯ããã ã
AIã³ãŒãã£ã³ã°ã®æèã§ã¯ãããããã£ãšé·ãã³ã³ããã¹ããããã°è§£æ±ºããããšãã話ã«ãªãã
ãªããžããªãå šéšèªãŸããã°ãããé¢é£ãã¡ã€ã«ãå šéšå ¥ããã°ãããäŸåé¢ä¿ããã©ã£ãŠè©°ã蟌ãã°ããã
ãããã«ãçãã³ã³ããã¹ãããé·ãã³ã³ããã¹ãã®ã»ããå©ããå Žé¢ã¯å€ãã
ã§ãããªããžããªç¥èã«ã¯ãæ¯åèªãã¹ãç¥èãšããã§ã«èº«ã«ã€ããŠããŠã»ããç¥èãããã
ããšãã°ããããããžã§ã¯ãã§ããã®ãã«ããŒã¯äŸå€ãæ¡ãã€ã¶ããããã®fixtureã¯DBãåæåãããããã®ã¯ã©ã¹åã¯æŽå²ççç±ã§ãºã¬ãŠãããã¿ãããªç¥èã¯ãæ¯åæ€çŽ¢ããŠåãã«è¡ããšããããããããžã§ã¯ãã«å ¥ã£ãŠããéçºè ãªãäœã«å ¥ã£ãŠãããã®ã«è¿ãã
Code2LoRAã¯ãããããã©ã¡ãŒã¿åŽãžå¯ããçºæ³ã«èŠããã
ãã¡ãããå®åã§ãã®ãŸãŸäœ¿ãããã¯å¥ã®è©±ã ãè«æã®å®éšã¯Qwen2.5-Coder-1.5BãããŒã¹ã«ããŠããŠã察象ãPythonãã¿ã¹ã¯ãassertè£å®ã«éãããŠãããhypernetworkèªäœãStaticã§çŽ720MãEvoã§çŽ745Mã®åŠç¿ãã©ã¡ãŒã¿ãæã€ã軜ãä»çµã¿ã§ã¯ãªãã
ããã§ããæ¹åæ§ãšããŠã¯ããªãé¢çœãã
ãã³ã³ããã¹ããé·ããããã ãã§ã¯ãªãããã©ã®ç¥èãå ¥åã«æ®ããã©ã®ç¥èãã¢ããã¿ã«éãããããšããèšèšåé¡ã«ãªã£ãŠããããã ã
ã³ãŒãã£ã³ã°ãšãŒãžã§ã³ãã®âèšæ¶âã¯ããã£ããå±¥æŽã ãã§ã¯è¶³ããªã
ãã®è©±ã¯ãã³ãŒãã£ã³ã°ãšãŒãžã§ã³ãã®éçšã«ãã€ãªããã
ãšãŒãžã§ã³ãã«å¿ èŠãªèšæ¶ã¯ãäŒè©±å±¥æŽã ãã§ã¯ãªãã
ãããéèŠãªã®ã¯ããªããžããªåäœã®èšæ¶ã ã
ãã®ãããžã§ã¯ãã§ã¯ãã¹ããã©ãèµ°ãããã®ããã©ã®ãã£ã¬ã¯ããªã¯è§Šã£ãŠã¯ãããªãã®ããã©ã®æœè±¡ã¯æšãŠããã§ãã©ã®æœè±¡ã¯å®ãã¹ããªã®ãããªããã®å€ãªåœåãæ®ã£ãŠããã®ãã
ãããæ¯åREADMEãšéå»ãã°ããæ¢ãã®ã¯ãå°ãã€ããã
人éã®éçºè ã¯ããããžã§ã¯ãã«å ¥ã£ãŠãã°ãããããšããªããžããªã®åããèŠãããããããªç¥èã§ã¯ãªãããã£ãšèº«äœçã§ãéã§ãã§ãå¹ãç¥èã ã
Code2LoRAã¯ããã®ãåãããLoRAãšããŠäœãããããšããç ç©¶ã«ãèªããã
ããªãèšãéãã ãããããã¶ãå°ãèšãéãã ã
ãã ãAIã³ãŒãã£ã³ã°æ¯æŽãããããæ¬åœã«æ¥åžžã®éçºã«å ¥ããªããé¿ããŠéããªãåãã ãšæãã
倧ããªã¢ãã«ã«ãæ¯å巚倧ãªã³ã³ããã¹ããæž¡ãã®ãã
ãããžã§ã¯ãããšã®èãé©å¿å±€ãæãããã®ãã
ã³ãããããšã®å·®åã远ããªããããã®é©å¿å±€ãæŽæ°ããŠããã®ãã
ãã®è«æã¯ããã®ãã¡ã®äžã€ã®çããå ·äœçã«åºããŠããã
å®åçã«ã¯ããŸã ã補åãã§ã¯ãªããæ¹åã
泚æç¹ãããã
ãŸããããã¯ãã®ãŸãŸææ¥ã®éçºç°å¢ã«å ¥ããã®ã§ã¯ãªãã
è©äŸ¡ã¯Pythonãªããžããªãšassertè£å®ã«éå®ãããŠãããå®éã®éçºã§ã¯ããã°ä¿®æ£ãèšèšå€æŽãã¬ãã¥ãŒããªãã¡ã¯ã¿ãªã³ã°ã仿§çè§£ãªã©ããã£ãšå¹ åºãã¿ã¹ã¯ãããã
Exact Matchãäžèœã§ã¯ãªããè«æåŽããæ©èœçã«ç䟡ãªã³ãŒããExact Matchã ãã§ã¯æŸããããªããšæžããŠããã
ãŸãããªããžããªåºæã®LoRAãäœããšããããšã¯ããã©ã€ããŒããªããžããªãæ±ããšãã®ã©ã€ã»ã³ã¹ãæŒæŽ©ãªã¹ã¯ãèããå¿ èŠããããè«æããã³ãŒãLLMäžè¬ã®ãªã¹ã¯ãšããŠãäžå®å šãªã³ãŒãããåŠç¿ããŒã¿ã«äŒŒãã³ãŒããåºãå¯èœæ§ããã©ã€ããŒããªããžããªãå ¥ããå Žåã®åž°å±ãªã¹ã¯ã«è§ŠããŠããã
ã ãããããã¯ãRAGã¯çµãããå šéšLoRAã«ãããããšãã話ã§ã¯ãªãã
ããããéã ãšæãã
AIã³ãŒãã£ã³ã°æ¯æŽã®æè管çã¯ãããããåè§£ãããã
æ€çŽ¢ã§åãæèãé·ãã³ã³ããã¹ãã«çœ®ãæèããããžã§ã¯ãã¡ã¢ãšããŠæã€æèãã¢ããã¿ã«æã¿èŸŒãŸããæèãã³ãããããšã«æŽæ°ããæèã
å šéšãäžã€ã®å·šå€§ãªããã³ããã§çä»ããã®ã§ã¯ãªããå±€ã«åããŠããã
ãã®æµãã®äžã§ãCode2LoRAã¯ãŸã æ©ã段éã®ãã§ãçã®ããææ¡ã«èŠããã
ãã³ãŒãããŒã¹ã«æ £ããããšã¯äœã
人éã®éçºè ããããžã§ã¯ãã«æ £ããããšããã®ã¯äžæè°ãªçŸè±¡ã ã
æåã¯æ¯åæ€çŽ¢ããŠãã颿°åãèŠããããã¹ãã®èœã¡æ¹ã§åå ãã ãããããããè§Šããšå±ãªãå ŽæãèŠããŠãããèšèšã®çæ³ãšãçŸå®ã®æãåããèŠããŠããã
ãã®ç¥èã¯ãããã¥ã¡ã³ãã ãã§ã¯ãªããæ€çŽ¢çµæã§ããªãããã£ãšå§çž®ãããããªããžããªããšã®èº«äœæèŠã¿ãããªãã®ã ã
AIã³ãŒãã£ã³ã°æ¯æŽãããã¶ãããã«åããã
è³¢ãã¢ãã«ãåŒã¶ã ãã§ã¯ãªãããããžã§ã¯ãã«æ £ããã¢ãã«ãã©ãäœããã
Code2LoRAã¯ããã®åããæ©æ¢°åŠç¿ããã圢ã§åºããŠããã
ãããŠããã®åãèªäœã¯ãããªãå®åã«è¿ãã
AIã«ã³ãŒããæžãããæä»£ã®æ¬åœã®å·®ã¯ãã¢ãã«åã§ã¯ãªãããã®ã¢ãã«ãã©ãã ãã³ãŒãããŒã¹ã«éŠŽæãã§ãããã«åºãã®ãããããªãã
ãŸã ç 究段éã ã
ã§ããAIã³ãŒãã£ã³ã°æ¯æŽã®æ¬¡ã®è©°ãŸãã©ãããèããã«ã¯ãããè«æã ãšæãã
åºå žãªã¹ã
Liliana Hotsko, Yinxi Li, Yuntian Deng, Pengyu Nie, "Code2LoRA: Hypernetwork-Generated Adapters for Code Language Models under Software Evolution", arXiv:2606.06492, 2026-06-04. https://arxiv.org/abs/2606.06492
Code2LoRA artifacts / RepoPeftBench: https://huggingface.co/code2lora
Anonymous code release noted in the paper: https://anonymous.4open.science/r/code2lora-6857