dev-resources.site
for different kinds of informations.
textGrad: Automatic âDifferentiationâ via Text
éžå®çç±
Stanford Univ.ã®ç 究ãé©çšå ãåºãé©æ°çïŒãšããå°è±¡ïŒãã©ã€ãã©ãªãšããŠã®å®æ床ãé«ãã
Paper: https://arxiv.org/abs/2406.07496
Code: https://github.com/zou-group/textgrad
èªå埮åãšåæ§ã®çºæ³ã§ããã¹ãåŸé
ã«ããèªåããã³ãããã¥ãŒãã³ã°ãææ¡ããŠãããã·ã¹ãã ãå€æ®µåŒã®è€æ°ã®LLMã䜿çšããŠããŠãåŸé
äŒæ¬ãšãã圢ã§ãã£ãŒãããã¯ã§ããç¹ããã€ã³ãã§ããã
å®éã«ã¯è§£æåŠã«ããã埮åã®å®çŸ©ãšã¯ç°ãªãïŒããã¹ãã¯é¢æ£ç©ºéã ãïŒãããã¹ãåŸé
ã«ãã£ãŠç®çé¢æ°ãå±ææé©è§£ãžåæãã
ä¿èšŒã¯ãªãã匷ååŠç¿ã®æ çµã¿ã§ãããšç°å¢ãã€ããã¯ã¹ã®ç¥èããªãã1ãšããœãŒããå®äºåŸã«å ±é
¬ã決å®ããã®ã§ã¢ã³ãã«ã«ãæ³ã«çžåœã
æŠèŠ
ã瀟äŒèª²é¡ã
AIã·ã¹ãã ã®è€éåã«äŒŽããè€æ°ã®ã¢ãã«ãããŒã«ãçµã¿åãããè€åã·ã¹ãã ã®æé©åãå¿
èŠã ããåŸæ¥ã¯ããã³ãã調æŽãäžå¯æ¬ ã§ã³ã¹ãã倧ãããèªååãããæé©åææ³ãäžè¶³ããŠããããããAIã·ã¹ãã ã®ãããªãçºå±ãé»ãã§ããã
ãæè¡èª²é¡ã
è€åã·ã¹ãã å
ã®ç°ãªãèŠçŽ ïŒäŸïŒã³ãŒããååæ§é ãå»çèšç»ïŒã¯ãæ°å€çãªåŸé
èšç®ãé£ãããåŸæ¥ã®èªå埮åãåŸé
éäžæ³ãé©çšã§ããªããåèŠçŽ ã«å¯Ÿããæé©åã®æéãèªç¶èšèªã§æäŸããæ¹æ³ãæ±ããããŠããããå®è£
ãæ±çšæ§ã課é¡ã§ãã£ãã
ãææ¡ã
LLMãèªç¶èšèªã®ãã£ãŒãããã¯ãçæããããã䜿ã£ãŠã·ã¹ãã å
ã®åèŠçŽ ãæé©åããTEXTGRADãææ¡ãå
·äœçã«ã¯ä»¥äžã®ãããªã¿ã¹ã¯ãèªåæé©åãã
- ã³ãŒãã£ã³ã°åé¡ïŒLeetCodeã®é£åã«å¯Ÿããã³ãŒãã®æ£ç¢ºæ§ãå¹çæ§
- 質åå¿çã®ç²ŸåºŠïŒç§åŠç質åå¿çã§ã®åç粟床æ¹å
- ååæ§é ïŒè¬å¹ãé«ããããã®çµå芪åæ§ïŒVinaã¹ã³ã¢ïŒãšè¬å€é©åæ§ïŒQEDã¹ã³ã¢ïŒã®ãã©ã³ã¹
- æŸå°ç·æ²»çèšç»ïŒè «çã¿ãŒã²ãããžã®æŸå°ç·éã®æé©åãšå¥åº·çµç¹ãžã®åœ±é¿ã®æå°å
ãããã®é ç®ããèªç¶èšèªåŸé ïŒTextual GradientsïŒããçšããŠæ¹åããåã¿ã¹ã¯ã«å¯ŸããŠãŒãã·ã§ããã§ã®æé©åãå¯èœã«ãªãã
ãå¹æã
TEXTGRADã«ããæé©åã«ãããåŸæ¥ã®ææ³ãšæ¯èŒããŠä»¥äžã®ãããªææãéæããŠããïŒ
- ã³ãŒãã£ã³ã°åé¡ã§ã¯GPT-4ã®æ§èœã7%ãã36%ãŸã§åäž
- 質åå¿çã®ç²ŸåºŠåäžïŒGPQAããŒã¿ã»ããã§æè¯ã®55%ãéæïŒ
- ååèšèšã§ã¯ãããé«ãè¬å¹ã瀺ãååãçæïŒQEDã¹ã³ã¢åäžãVinaã¹ã³ã¢ã®äœäžïŒ
- å»çåéã§ã¯æŸå°ç·æ²»çèšç»ã®ç²ŸåºŠãæ¹åããè «çãžã®æŸå°ç·ç §å°éã®æé©åãšå¥åº·çµç¹ã®ä¿è·ãå®çŸâ
Textual Gradientã«ããåŸé äŒæ¬
TEXTGRADã¯ãè€éãªAIã·ã¹ãã ããã©ãã¯ããã¯ã¹ã·ã¹ãã ã®æé©åã«ãããŠãèªç¶èšèªåŸé ãïŒtextual gradientsïŒãå©çšãããéåžžã®åŸé éäžæ³ãæ°å€åŸé ã§ãã©ã¡ãŒã¿ã調æŽããã®ã«å¯ŸããTEXTGRADã¯LLMãçæããèªç¶èšèªãã£ãŒãããã¯ãåŸé ãšããŠæ±ããã·ã¹ãã å ã®åèŠçŽ ãæ¹åããã
èšç®ã°ã©ãã®æ§ç¯
TEXTGRADã¯ãã·ã¹ãã å ã®åæ§æèŠçŽ ïŒäŸïŒã³ãŒããååæ§é ãæšèŠã¢ãã«ã®ãã©ã¡ãŒã¿ãªã©ïŒããããŒãããšããŠèšç®ã°ã©ããæ§ç¯ããããã®èšç®ã°ã©ãã«ãããå ¥åºåããŒã¿ã®ãããŒãå®çŸ©ãããåããŒãã¯é埮åé¢æ°ïŒäŸïŒLLM APIãã·ãã¥ã¬ãŒã¿ãŒïŒã§æ¥ç¶ãããã
èªç¶èšèªãã£ãŒãããã¯ã®çæ
åããŒãã«å¯ŸããŠãåŸé ãšããŠæ±ãèªç¶èšèªãã£ãŒãããã¯ãLLMããååŸãããäŸãã°ãããã³ãŒããäžããããå ŽåãLLMãããã®éšåãåé¡ã§ãããããä¿®æ£ãã¹ãããšãã£ããã£ãŒãããã¯ãçæããããã®ãã£ãŒãããã¯ã¯ãæ°å€åŸé ã®ããã«ã·ã¹ãã æ¹åã®æéãšããŠæ±ãããã
ããã¹ãåŸé éäžïŒTGDïŒ
ããã¹ãåŸé éäžæ³ãé©çšããLLMããåŸããã£ãŒãããã¯ã«åºã¥ããŠåããŒãã®ãã©ã¡ãŒã¿ïŒäŸïŒã³ãŒããæšå¥šã·ã¹ãã ã®èšå®ãæŸå°ç·æ²»çã®èšç»ãªã©ïŒãæŽæ°ãããããã¯éåžžã®åŸé éäžæ³ã«ããããã©ã¡ãŒã¿æŽæ°ãšåæ§ã®æäœã ã
å埩ãšæŽæ°
ããã¹ãåŸé éäžæ³ãå埩é©çšãããã£ãŒãããã¯ãå ã«ã·ã¹ãã å šäœã段éçã«æé©åããããã®ããã»ã¹ã«ãããåã¿ã¹ã¯ã®æ§èœãåäžããã
å®éš
TEXTGRADã®æè»æ§ã®é«ãã瀺ãããã«è€æ°ã®ç°ãªãã¿ã¹ã¯ã§ã®å®éšãè¡ã£ããåã¿ã¹ã¯ã§ã®å埩åæ°ã¯3ïœ10åçšåºŠã§ãã¿ã¹ã¯ã®è€éæ§ãæ¹åç®æšã«å¿ããŠèª¿æŽãããŠãããæ¯èŒå¯Ÿè±¡ã¯CoT, Reflexion[Shinn2023]
ã¿ã¹ã¯1. ã³ãŒãæé©å
LeetCodeã®é£åããŒã¿ã»ããã§GPT-4ã®åçã³ãŒãçæãè¡ãã¿ã¹ã¯ã§ã®æ€èšŒã5åã®å埩ã§GPT-4ã®ãã¹ãã±ãŒã¹ééçã7%â36%ã«å°éããåŸæ¥ã®æè¯æ§èœ(Reflexion)ããããã«5%åäžãããããã¯TEXTGRADã¯èšç®ã°ã©ãã䜿ã£ãæè»ãªæé©åã«ãããã¿ã¹ã¯å šäœãçµ±åçã«æ¹åã§ããç¹ãèªç¶èšèªåŸé ã掻çšããŠåèŠçŽ ã«å¯Ÿãã粟å¯ãªãã£ãŒãããã¯ãæäŸã§ããç¹ããåå ãšèããããã
æ¬ã¿ã¹ã¯ã§ã®å®åŒåãšããã³ããã¯ä»¥äžã§ããã
Code-Refinement Objective=LLM(Problem + Code + Test-time Instruction + Local Test Results)
- ProblemïŒè§£æ±ºãã¹ãåé¡ã®å®çŸ©ãæé©å察象ãšãªãã³ãŒããåãçµãåé¡ã瀺ãã
- CodeïŒçŸæç¹ã§ã®ã³ãŒãå®è£ ãæé©åã®å¯Ÿè±¡ãšãªãã³ãŒããã®ãã®ã
- Test-time InstructionïŒã³ãŒããæºããã¹ãæ¡ä»¶ãè©äŸ¡åºæºã«é¢ããæ瀺ããã®æ瀺ã«åŸã£ãŠãLLMãã³ãŒãã®æ¹åç¹ããšã©ãŒãç¹å®ããã
- Local Test ResultsïŒã³ãŒããããŒã«ã«ãã¹ãã§ã©ã®ãããªçµæãåºãããã瀺ãããã®çµæã«åºã¥ããŠãã³ãŒãã®æ£ç¢ºæ§ãããã©ãŒãã³ã¹ãè©äŸ¡ãããã
ã¿ã¹ã¯2. 質åå¿ç
è€æ°ã®Ph.DçšåºŠã®ç§åŠæè¡ã«é¢ãã質åå¿çããŒã¿ã»ãããçšããããã®äžã®ïŒã€ãGoogle-proof Question Answering (GPQA) ã¯ãã¡ã€ã³ãšãã¹ããŒãã®åççã81%ãéãšãã¹ããŒãã§22%ã§ãã£ãã3åã®å埩ã§GPT-4oã§ã®ç²ŸåºŠãGPQAã§51%(CoT)->55%(textGrad)ã«åäžããæ¢ç¥ã®æé«æ§èœãéæããã
ã¿ã¹ã¯3. æšè«(Object CountingãWord SortingãGSM8k)
æšè«ç³»ã¿ã¹ã¯ã§ã®è©äŸ¡ã¯è¡šïŒã§ãããã©ããTEXTGRADãæãé«ç²ŸåºŠã§ãã£ããããã«èšèŒãããŠããããã«ãæé©ååã®ããã³ããã¯ã質åã«å¯ŸããŠã¹ããããã€ã¹ãããã§èãããã®ã¿ã§è©³çŽ°ãªæ瀺ã¯å«ãŸããŠããªããäžæ¹ã§ãæé©ååŸã®ããã³ããã¯1.åé¡ãèªåã®èšèã§å確èªããããšã§ç解ãæ·±ããã2.åèšç®ã¹ãããã现ãã説æããæ£ç¢ºæ§ã確ä¿ããããã«å確èªãè¡ãã3.æ°åŠèšæ³ãåé¡æã®æèãå®ããæçµçã«ãAnswer: $VALUEããšãã圢åŒã§çããããæ確ã«æ瀺ãããŠããã
ã¿ã¹ã¯4. ååèšèš
è¬å¹ã®é«ãååãèšèšãããããååã®çµå芪åæ§ïŒVinaã¹ã³ã¢ïŒãšè¬å€é©åæ§ïŒQEDã¹ã³ã¢ïŒãæé©åãåååæ§é ã¯10åã®å埩ãè¡ããèªç¶èšèªåŸé ãçšãã調æŽãå®æœãçæãããååã¯ã察象ã¿ã³ãã¯è³ªã«å¯Ÿããçµå芪åæ§ãšè¬å€é©åæ§ã§æ¢åè¬ãäžåãæ§èœã瀺ãããå³ïŒã«ã¯èªç¶èšèªåŸé ã®äŸã瀺ãããŠãããååŠæ§é ãã©ã®ããã«å€ããã¹ãã詳现ãªæ瀺ãäžããããŠããã
ã¿ã¹ã¯5. æŸå°ç·æ²»çèšç»ã®æé©å
è «çãšåšèŸºçµç¹ãžã®æŸå°ç·éã®ãã©ã³ã¹ãæé©åããããã®å®éšãTEXTGRADã¯èªç¶èšèªãã£ãŒãããã¯ãçšããŠã5åã®å埩ã§æ²»çèšç»ïŒç §å°ç¯å²ãæŸå°ç·éãªã©ïŒãæ¹åãèšåºã§çšããããŠããèšç»ãšæ¯èŒããŠãè «çãžã®ç §å°ç²ŸåºŠãåäžããå¥åº·ãªçµç¹ãžã®ãã¡ãŒãžãæžå°ããã
Featured ones: