【輪講資料】ReAct: Synergizing Reasoning and Acting in Language Models / Tree of Thoughts: Deliberate Problem Solving with Large Language Models

研究室の輪講で発表したときの資料です。以下の２本の論文を紹介しています。
- ReAct: Synergizing Reasoning and Acting in Language Models
- Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Transcript

1. 10݄31೔ ෢ాɾ࡫໺ݚڀࣨɹM2 ໼໺ઍߛ ReAct: Synergizing Reasoning and Acting in Language Models Tree of Thoughts: Deliberate Problem Solving with Large Language Models Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Grif fi ths, Yuan Cao, Karthik Narasimhan Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao ICLR 2023 NeurIPS 2023
2. બΜͩཧ༝ • ྆ํͱ΋Prompt Engineering Guideʹ΋ܝࡌ͞Ε͍ͯΔϓϩϯϓτͷख๏ • ͋Δఔ౓༗໊Ͱ༗༻ • ͨ͘͞ΜఏҊ͞Ε͍ͯΔϓϩϯϓτͷ࿦จΛಡΉͱ͖ͷϕʔεʹͳΓͦ͏ 2
3. ֓؍ • ReAct • ਪ࿦ͱߦಈΛަޓʹੜ੒͢Δ • ߦಈͰ֎෦πʔϧΛར༻͢Δ͜ͱͰHallucinationΛ๷͙͜ͱ͕Ͱ͖Δ • Tree of Thougt • ޙ໭Γ΍ઌಡΈΛߦ͍ɺख़ߟ͢ΔࣄͰCoTͳͲͰ͸೉͔ͬͨ͠ෳࡶͳਪ࿦Λ ߦ͏͜ͱ͕Ͱ͖Δ 3
4. ReAct: Synergizing Reasoning and Acting 
 in Language Models
5. ReAct: Synergizing Reasoning and Acting in Language Models • LLMʹਪ࿦ͱߦಈΛަޓʹੜ੒ͤ͞Δϓϩϯϓτख๏ʮReActʯͷఏҊ • ਪ࿦Ͱ͸ߦಈΛܭըͨ͠Γɺߦಈͷ݁ՌΛॲཧͨ͠Γ͢Δ • ߦಈͰ͸஌ࣝϕʔε΍؀ڥʹΞΫηε͠৘ใΛಘΔ • ֎෦͔Βͷਖ਼͍͠৘ใΛར༻͢Δ͜ͱͰHallucinationΛ཈੍ • Hallucination: LLM͕΋ͬͱ΋Β͍͠ӕΛग़ྗ͢Δ͜ͱ 5 2
   ϋϩ΢Οϯ͸೥Ҏ্લ͔Βଓ͘ࡇΓͰ͔͢ʁ ·ͣ͸ϋϩ΢Ο ϯͷىݯΛௐ΂ Α͏ ϋϩ΢Οϯ
   ىݯ ϋϩ΢Οϯͷൃ঵͸ ೥Ҏ্΋લɻʜ ೥લ͸ ೥Ҏ্લ 
 ͔ͩΒ౴͑͸ ʮ͸͍ʯ ਪ࿦ ߦಈ ਪ࿦
6. REACTͷਪ࿦ྫ 6 2
   ʰԾ໘ϥΠμʔిԦʱͱʰ͋·ͪΌΜʱɺ์ૹճ਺͕ଟ͍ͷ͸ͲͪΒͰ͠ΐ͏ʁ ʰԾ໘ϥΠμʔిԦʱͷ์ૹճ਺Λݕࡧͨ͠ޙɺʰ͋·ͪΌΜʱͷ์ૹճ਺Λݕࡧ͠ɺͦΕΒͷൺֱΛߦ͏ɻ 4FBSDI<ʰԾ໘ϥΠμʔిԦʱͷ์ૹճ਺͸Կճʁ> ਪ࿦ ߦಈ ʰԾ໘ϥΠμʔిԦʱͷ์ૹճ਺͸ճͰ͋Δɻ࣍ʹʰ͋·ͪΌΜʱͷ์ૹճ਺Λݕࡧ͢Δɻ 4FBSDI<ʰ͋·ͪΌΜʱͷ์ૹճ਺͸Կճʁ> ʰԾ໘ϥΠμʔిԦʱͷ์ૹճ਺͸ճͰ͢ɻ ʰ͋·ͪΌΜʱͷ์ૹճ਺͸ճͰ͢ɻ ʰԾ໘ϥΠμʔిԦʱͷ์ૹճ਺͸ճͰ͋Δɻʰ͋·ͪΌΜʱͷ์ૹճ਺͸ճͰ͋Δɻ ʰ͋·ͪΌΜʱͷ์ૹճ਺͕ଟ͍ͷͰɺ࠷ऴతͳ౴͑͸ʰ͋·ͪΌΜʱͰ͋Δɻ 'JOJTI<ʰ͋·ͪΌΜʱ> ໰୊͸+&.)PQ2"ΑΓҾ༻ ਪ࿦ ߦಈ ਪ࿦ ߦಈ
7. ஌ࣝू໿తͳਪ࿦λεΫ HotPotQA • ଟஈ֊ਪ࿦λεΫ • ̎ͭҎ্ͷWikipediaͷهࣄʹର͢Δਪ࿦Λඞཁͱ͢Δ • ྫ) Which magazine was started fi rst Arthur's Magazine or First for Women? Fever • ࣄ࣮ݕূλεΫ • ओுʹର͠ɺWikipediaͷهࣄΛࢀর͠[ Supported, Refuted NotEnoughInfo ]ͷϥϕϧ Λ͚ͭΔ • ྫʣNikolaj Coster-Waldau worked with the Fox Broadcasting Company.
 -Supported 7
8. ஌ࣝू໿తͳਪ࿦λεΫͰͷߦಈۭؒ • Wikipedia͔Β৘ใΛ֫ಘ͢ΔͨΊͷߦಈۭؒΛఆٛ • ҎԼͷ3ͭͷΞΫγϣϯ • search[entity]: entityͷWikipediaϖʔδ͕͋Ε͹࠷ॳͷ5จΛɺແ ͚Ε͹ྨࣅͷΤϯςΟςΟΛ5ͭฦ͢ • lookup[string]: stringΛؚΉจΛฦ͢ (Ctrl+F) • fi nish[answer]: answerΛ౴͑ͱͯ͠ݱࡏͷλεΫΛऴྃ͢Δ 8
9. ஌ࣝू໿తͳਪ࿦λεΫͰͷReActͷྲྀΕ 9 "QQMF
   3FNPUFͷଞʹɺ"QQMF
   3FNPUF͕ຊདྷૢ࡞͢Δ 
 ͨΊʹσβΠϯ͞ΕͨϓϩάϥϜΛίϯτϩʔϧͰ͖Δ 
 σόΠε͸͋Γ·͔͢ʁ "QQMF
   3FNPUF 'SPOU
   3PX ૢ࡞
10. ൺֱϕʔεϥΠϯ Standerd • ໰୊Λ༩͑ղ౴ͷΈΛߦΘͤΔϓϩϯϓτ 10 "QQMF
    3FNPUFͷଞʹɺ"QQMF
    3FNPUF͕ຊདྷૢ࡞͢ΔͨΊʹσβΠϯ͞Εͨ ϓϩάϥϜΛίϯτϩʔϧͰ͖ΔσόΠε͸͋Γ·͔͢ʁ
11. ൺֱϕʔεϥΠϯ Chain of Thought (CoT)[1] • தؒͷਪ࿦εςοϓ΋ग़ྗͤ͞Δ͜ͱͰɺෳࡶͳਪ࿦͕Մೳ • ߦಈΛར༻ͤͣɺਪ࿦ͷΈΛߦ͏ʢϞσϧ಺෦ͷ৘ใͷΈར༻ʣ Chain of Thought self-consistency (CoT-SC)[2] • ෳ਺ͷCoTʹΑͬͯଟ਺ܾ͢Δ • Թ౓ม਺Λ0.7ʹઃఆͨ͠CoTΛ21ճߦ͍ɺ
 ղ౴Λଟ਺ܾͰܾఆ͢Δ 11 <>
    $IBJOPG5IPVHIU
    1SPNQUJOH
    &MJDJUT
    3FBTPOJOH
    JO
    -BSHF
    -BOHVBHF
    .PEFMT
    <>
    4FMG$POTJTUFODZ
    *NQSPWFT
    $IBJO
    PG
    5IPVHIU
    3FBTPOJOH
    JO
    -BOHVBHF
    .PEFMT
12. ൺֱϕʔεϥΠϯ Act-Only • ReAct͔Βਪ࿦ΛऔΓআ͍ͨ΋ͷ 12
13. ஌ࣝू໿తͳਪ࿦λεΫͷ࣮ݧઃఆ • Ϟσϧʹ͸PaLM-540BΛར༻ • HotpotQA • ֶशηοτͷࣄྫΛਓखͰΞϊςʔγϣϯͯ͠6-shot • Fever • ֶशηοτͷࣄྫΛਓखͰΞϊςʔγϣϯͯ͠3-shot 13
14. ஌ࣝू໿తͳਪ࿦λεΫͰͷ࣮ݧ݁Ռ • ReAct > Act • ActͰ͸ࠓ·Ͱͷ৘ใ͔Β࠷ऴతͳ
 ౴͑Λղ౴͢Δ͜ͱ͕೉͍͠ 14
15. ஌ࣝू໿తͳਪ࿦λεΫͰͷ࣮ݧ݁Ռ • ReAct vs CoT • HotpotQAͰ͸CoTͷউͪ 15
16. ReAct vs CoT @HotpotQA • CoTͷHallucination͸ReActΑΓਂࠁ (False positive, Hallucination) • ReAct͸ਪ࿦εςοϓ͕ॊೈͰͳ͍ (Reasoning error) • ಉ͡ਪ࿦ɺߦಈΛ܁Γฦ͢ϧʔϓʹϋϚΔ͜ͱ͕ଟ͍ • ReAct͸ݕࡧͰࣦഊ͢Δͱਪ࿦͕୤ઢͯ͠͠·͏ (Search result error) 16 ˛
    )PUQPU2"Ͱͷ੒ޭɺࣦഊࣄྫ͔ΒϥϯμϜʹநग़͠ɺਓखͰݪҼΛௐࠪ
17. ಺෦஌ࣝͱ֎෦஌ࣝͷ૊Έ߹Θͤ • CoT͸ਪ࿦εςοϓͷߏங͸ΑΓਖ਼͕֬ͩɺ಺෦஌ࣝΛར༻͢Δ͜ͱͰ Hallucination͕ൃੜ͢Δ • ReAct͸֎෦஌ࣝΛར༻͠ਖ਼͕֬ͩɺਪ࿦εςοϓͷߏஙʹ͓͍ͯ͸CoTΑΓྼΔ ➡ҎԼͷܦݧଇʹج͍ͮͯCoT-SCͱReActΛ੾Γସ͑Δ • ReAct→CoT-SC: نఆͷεςοϓͰղ౴͕ಘΒΕͳ͍৔߹ • ReActͷਪ࿦εςοϓΛ௕͗ͯ͘͢͠΋ੑೳ͸্͕Βͳ͍ • CoT-SC→ReAct: ଟ਺ܾͷଟ਺೿͕ա൒਺Λ௒͑ͳ͍৔߹ • CoT͕ࣗ৴Λ࣋ͬͯղ౴Ͱ͖͍ͯͳ͍ 17 )PUQPU2"Ͱεςοϓ 
 'FWFSͰεςοϓ
18. ஌ࣝू໿తͳਪ࿦λεΫͰͷ࣮ݧ݁Ռ • ReAct+CoT-SC > CoT-SC, ReAct • CoT-SC (sample=21) < ReAct+CoT-SC (sample=5) • ಺෦஌ࣝͱ֎෦஌ࣝΛ૊Έ߹ΘͤΔ͜ͱ͸༗༻ 18 ˛
    $P54$Ͱͷαϯϓϧ਺
    ଟ਺ܾͷ฼਺ ͱੑೳ
19. ஌ࣝू໿తͳਪ࿦λεΫͰͷFT࣮ݧ • PaLM540b͕ੜ੒͠ɺਖ਼ղͨ͠3000݅ͷࣄྫΛར༻ͯ͠62b, 8bΛ Fine-Tuning • ReAct͸ͦͷෳࡶੑ͔Βfew-shotͰϞσϧʹֶशͤ͞Δ͜ͱ͕೉͘͠ɺ Fine-Tuningͱ૬ੑ͕ྑ͍ 19
20. ҙࢥܾఆλεΫ ALFWorld • ՈͷதΛςΩετૢ࡞ʹΑͬͯΤʔδΣϯτ͕୳ࡧ͠ɺෳࡶͳ̒छྨͷλεΫΛ ୡ੒͢Δ • ߦಈͷબ୒ࢶ͸50छྨҎ্͋Γɺ૯౰ͨΓతʹ୳ࡧ͢Δ͜ͱ͸ࠔ೉ • ྫ) You are in the middle of a room. Looking quickly around you, you see a drawer 2, a shelf 5,…, and a drawer 4. 
 Your task is to: put some vase in safe. • vase΋safe΋ೖྗʹ͸ଘࡏͤͣɺৗࣝΛར༻ͯ͠୳ࡧ͠ͳ͚Ε͹ͳΒͳ͍ 
 (Ֆළ͸୨ͷ্ʹ͋Γͦ͏) WebShop 20 <>
    <>
    8FC4IPQ
    5PXBSET
    4DBMBCMF
    3FBM8PSME
    8FC
    *OUFSBDUJPO
    XJUI
    (SPVOEFE
    -BOHVBHF
    "HFOUT
21. ҙࢥܾఆλεΫ ALFWorld WebShop • ݱ࣮ͷ঎඼Λར༻ͨ͠ΦϯϥΠϯγϣοϐϯά؀ڥͰϢʔβʔͷࢦࣔʹج ͍ͮͯ঎඼Λߪೖ͢ΔλεΫ • ΤʔδΣϯτ͸ݕࡧ΍ʮ໭ΔʯͳͲͷϘλϯΛར༻Ͱ͖Δ • Score (ճ౴͕Ͳͷఔ౓ཁ݅Λຬ͍ͨͯ͠Δ͔)ͱSuccess rate (શͯͷཁ݅ Λຬͨͨ͠੡඼Λճ౴ׂͨ͠߹)ͰධՁ • ྫ) get me a sixteen pack of apple cinnamon freeze dried banana chips, and price lower than 50.00 dollars 21
22. ൺֱϕʔεϥΠϯ ALFWorld • BUTLER[1]: λεΫछຖʹ100000݅ͷσϞͰ໛฿ֶशͨ͠Ϟσϧ • ReAct-IM: ReActΛInner MonologueελΠϧ[2]ʹͨ͠΋ͷ • ΤʔδΣϯτ͕؍ଌͨ͠؀ڥͱɺୡ੒͕ඞཁͳখ໨ඪΛInner Monologue ͱͯ͠ੜ੒͢Δ WebShop • IL: 1012݅ͷσϞͰ໛฿ֶशͨ͠Ϟσϧ • ILʴRL: 10587݅ͷσϞͰ໛฿ֶशʴڧԽֶशͨ͠Ϟσϧ 22 <>
    "-'8PSME
    "MJHOJOH
    5FYU
    BOE
    &NCPEJFE
    &OWJSPONFOUT
    GPS
    *OUFSBDUJWF
    -FBSOJOH
    <>
    *OOFS
    .POPMPHVF
    &NCPEJFE
    3FBTPOJOH
    UISPVHI
    1MBOOJOH
    XJUI
    -BOHVBHF
    .PEFMT
23. ҙࢥܾఆλεΫͷ࣮ݧઃఆ • Ϟσϧʹ͸PaLM-540BΛར༻ • ALFWorld • λεΫͷछྨຖʹ3݅ਓखͰΞϊςʔτ͠ɺϥϯμϜͳ2-shotΛར༻ • WebShop • 1-shotΛར༻ 23
24. ҙࢥܾఆλεΫͷ࣮ݧ݁Ռ • ReAct > Act • Act͸໨ඪͷ෼ղͱ؀ڥΛ೺Ѳ͕Ͱ͖ͳ͍ • ฏۉతʹ͸ReAct > BUTLER • େྔͷֶशσʔλΛར༻͢Δख๏ʹউར 24 ˛
    "-'8PSMEʹ͓͚ΔλεΫछྨຖͷ੒ޭ཰
    
25. ҙࢥܾఆλεΫͷ࣮ݧ݁Ռ • ReAct > ReAct-IM • Inner MonologueελΠϧͰ͸؀ڥͷ؍ଌͱୡ੒͢΂͖খ໨ඪʹ͍ͭͯͷΈੜ੒ • ্खʹৗࣝΛద༻ͯ͠ɺਪ࿦ͷεςοϓΛߏங͢Δ͜ͱ͕೉͍͠ • ྫ) task: put a clean knife in countertop. • think: To solve the task, I need to fi nd and take a clean knife, then put it in countertop. • φΠϑΛݟ͚ͭΔ΋ͷͷɺͦΕ͕៉ྷͩͱ৴ͯ͡Χ΢ϯλʔʹஔ͖ଓ͚Δ 25 ˛
    "-'8PSMEʹ͓͚ΔλεΫछྨຖͷ੒ޭ཰
    
26. ҙࢥܾఆλεΫͷ࣮ݧ݁Ռ • ReAct > Act > IL, IL+RL • ReAct͸ϊΠζͷଟ͍؍ଌ͔Β͏·࣍͘ͷߦಈΛੜ੒Ͱ͖͍ͯΔ • For ‘space-saving ottoman bench for living room’, the item has options ‘39x18x18inch’ and ‘blue’ and seems good to buy. 26 ˛
    8FC4IPQͰͷείΞ
27. ReAct: Synergizing Reasoning and Acting in Language Models • LLMʹਪ࿦ͱߦಈΛަޓʹੜ੒ͤ͞Δϓϩϯϓτख๏ͷఏҊ • ਪ࿦Ͱ͸࣍ͷߦಈΛܭը͠ɺߦಈͷ݁ՌΛॲཧ͢Δ • ߦಈͰ͸஌ࣝϕʔε΍؀ڥʹΞΫηε͠ɺ௥Ճͷ৘ใΛಘΔ • ࣭໰Ԡ౴λεΫͱࣄ࣮ݕূλεΫʹ͓͍ͯɺWikipediaAPIͱର࿩͠ɺ HallucinationΛࠀ෰ • 2ͭͷର࿩ܕҙࢥܾఆϕϯνϚʔΫʹ͓͍ͯɺେྔͷֶशσʔλΛར༻͠ ͨ໛฿ֶशͱڧԽֶशʹΑΔϕʔεϥΠϯΛ1-shot/2-shotͷઃఆͰେ෯ʹ ্ճͬͨ 27
28. Tree of Thoughts: Deliberate Problem Solving with Large Language Models
29. Tree of Thoughts: Deliberate Problem Solving with Large Language Models • ݴޠϞσϧͷਪ࿦͸ࠨ͔Βӈʹݶఆ • ઌಡΈΛඞཁͱͨ͠Γɺ࠷ॳͷܾఆ͕ ۃΊͯॏཁͳλεΫͰ͸ࣦഊ ➡໰୊ղܾͷதؒஈ֊ (Thought)Λ୳ࡧ Մೳͱ͢ΔϑϨʔϜϫʔΫɺ
 ʮTree of ThoughtʯΛఏҊ • ෳ਺ͷҟͳΔਪ࿦ܦ࿏Λߟྀ͠ɺ࣍ͷ ߦಈΛܾఆ͢ΔͨΊʹࣗݾධՁΛߦ͏ ʢख़ߟ͢Δʣ 29
30. ϕʔεϥΠϯͱͷൺֱ 30 • ೚ҙͷ໰୊Λ໦ͷ୳ࡧ໰୊ͱͯ͠ଊ͑Δ • ֤ϊʔυ͸ೖྗͱ͜Ε·ͰͷࢥߟεςοϓΛؚΉ෦෼ղΛද͢ঢ়ଶ
31. Tree of Thoughtͷਪ࿦ྫ 31 ҎԼͷ਺ࣈͱ࢛ଇԋࢉΛ૊Έ߹ΘͤͯΛ࡞͍ͬͯͩ͘͞
    
    
    
    
     
    
    
    
    
    Y
    
    
     
    
    
    
     
    Y
    
    
     
    
    
    
     
    Y
    
    
     *OQVU
    
    
    
     ⭕ ⭕ ⭕ ⭕ ✖︎ ✖︎ ✖︎ Ͳͷਪ࿦͕ ༗๬͔ͳʜ
32. Tree of Thought • ෳࡶͳਪ࿦໰୊Λখ໰୊ʹ෼ղ͠ɺͲͷਪ࿦ܦ࿏͕ے͕ྑ͍͔ධՁΛ ߦ͍ͳ͕Β୳ࡧ͢Δ • λεΫຖʹมߋՄೳͳҎԼͷ4ͭͷΦϓγϣϯ A. ࢥߟΛ෼ղ͢Δେ͖͞ B. ࢥߟͷੜ੒ํ๏: ࠓ·Ͱͷࢥߟ΋ೖྗ͢Δ/ ͠ͳ͍ C. ֤ϊʔυͷධՁํ๏: ಠཱʹఆྔԽ / બ୒ࢶΛൺֱ • ઌಡΈͱৗࣝΛར༻ D. ໦ͷ୳ࡧํ๏: BFS / DFS 32
33. ઌಡΈ΍୳ࡧΛඞཁͱ͢ΔఏҊλεΫ Game of 24 • ༩͑ΒΕͨ4ͭͷ਺ࣈʹ࢛ଇԋࢉΛద༻ͯ͠24Λ࡞Δ Creative Writing • ༩͑ΒΕͨ4ͭͷจ͕ஈམͷ࠷ޙͷจͱͳΔΑ͏ͳ4ஈམߏ੒ͷจষΛग़ྗ • ੜ੒ͨ͠จষʹҰ؏ੑ͕͋Δ͔ɺGPT4ͱਓखͰධՁ 33
34. ઌಡΈ΍୳ࡧΛඞཁͱ͢ΔఏҊλεΫ Mini Crosswords • 5x5αΠζͷΫϩεϫʔυ(ख͕͔Γ͸ॎɺԣʹͦΕͧΕ5ͭͣͭ)Λղ͘ • ਖ਼ղͨ͠จࣈ਺ɺ୯ޠ਺ɺήʔϜ਺ͰධՁ 34
35. Game of 24ͰͷϕʔεϥΠϯ IO • ໰୊Λೖྗͱ͠ɺղ౴Λग़ྗͱ͢Δඪ४తͳϓϩϯϓτ (5-shot) CoT • 4ͭͷ਺஋͔Β24Λ࡞ΔͨΊͷద౰ͳ࢛ଇԋࢉͷࣜΛ༩͑Δ (5-shot) • 13 - 9 = 4 (left: 4 4 10); 10 - 4 = 6 (left: 4 6); 4 * 6 = 24 (left: 24) CoT-SC • 100ճCoTͰͷਪ࿦Λ࣮ߦ͠ɺଟ਺ܾΛऔΔ IO + re fi ne • 10ճ൓෮ͯ͠ਪ࿦͢Δ • ؒҧ͍͑ͯͨ৔߹͸मਖ਼͢ΔΑ͏ͳࢦࣔͱཤྺΛ༩͑Δ 35
36. Game of 24ͰͷToTηοτΞοϓ 36 • P32ʹࣔͨ͠Φϓγϣϯ͸ҎԼͷΑ͏ʹͳΔ A. ࢥߟΛ෼ղ͢Δେ͖͞ɿ̍౓ͷ࢛ଇԋࢉ B. ࢥߟͷੜ੒ํ๏ɿࠓ·Ͱͷࢥߟ͸ೖྗͤͣɺ୯७ͳ౳ࣜͷੜ੒ͷΈߦ͏ C. ֤ϊʔυͷධՁํ๏ɿ24ʹ౸ୡͰ͖Δ͔ʁ- [ sure/maybe/impossible ] • ৗࣝΛར༻͠ɺ24ʹରͯ͠େ͖͗ͨ͢Γখ͗͢͞ΔީิΛল͘ D. ໦ͷ୳ࡧํ๏ɿ෯༏ઌ୳ࡧ
37. Game of 24ͰͷToTηοτΞοϓ 37
38. Game of 24Ͱͷ݁Ռ • ToT͕୳ࡧ෯=1ͷઃఆͰ΋ଞͷϓϩϯϓτΑΓ͔ͳΓߴ͍ੑೳ • ୳ࡧ෯Λ5ʹ͢Δͱ74%΋ղ͚Δ 38
39. Game of 24Ͱͷ݁Ռ • ToTͷ๚໰ࡁϊʔυͰͷ੒ޭ཰͕ߴ͍ • IO, CoTͰ͸ͦΕͧΕͷ࣮ߦશମΛ
 ϊʔυͱͯ͠ܭࢉ • (ےྑ͘୳ࡧͰ͖͍ͯΔ) 39 • CoTͰ͸60%͕࠷ॳͷ౳ࣜੜ੒࣌఺Ͱ ਖ਼౴ʹࣦഊ • ޙ໭Γ΍ઌಡΈ͕Ͱ͖ͳ͍ख๏ʹ͸
 ݶք͕͋Δ ˛๚໰ࡁΈϊʔυͱͦͷਖ਼ղ཰ ˛εςοϓຖͷࣦഊ཰
40. Creative writingͰͷϕʔεϥΠϯ IO • จΛೖྗͱ͠จষΛग़ྗ͢Δඪ४తͳϓϩϯϓτ CoT • ઌʹ୹͍ܭըΛॻ͍͔ͯΒจষΛॻ͔ͤΔ IO+re fi ne • 5ճ·Ͱ൓෮͢Δ • Ұ؏ੑʹ͚ܽΔ৔߹͸मਖ਼͢ΔΑ͏ͳࢦࣔͱཤྺΛ༩͑Δ 40
41. Creative writingͰͷToTηοτΞοϓ 41 • P32ʹࣔͨ͠Φϓγϣϯ͸ҎԼͷΑ͏ʹͳΔ A. ࢥߟΛ෼ղ͢Δେ͖͞ɿͲͷΑ͏ͳจষΛهड़͢Δ͔ͷܭը / จষͷهड़ B. ࢥߟͷੜ੒ํ๏ɿࠓ·Ͱͷࢥߟ (ܭը) Λར༻ͯ͠จষΛهड़͢Δ C. ֤ϊʔυͷධՁํ๏ɿ֤ࢥߟ͸ಉ࣌ʹ5ͭੜ੒͞Εɺ࠷΋ྑ͍΋ͷΛબͿ D. ໦ͷ୳ࡧํ๏ɿਂ͞=2, ෯=1 • ܭըͱจষͷهड़Λߦ͏ͨΊਂ͞͸2ɺධՁ͝ͱʹ࠷ྑͷࢥߟͷΈ࢒ͨ͢Ί෯͸1
42. Creative writingͰͷToTηοτΞοϓ 42
43. Creative writingͰͷ݁Ռ • GPT-4ʹΑΔධՁɺਓखධՁͷ྆ํͰCoTΑΓToTͷํ͕Ұ؏ੑ ͷ͋ΔจষΛੜ੒͍ͯ͠Δͱ൑அ͞Εͨ • +re fi neͷઃఆ͸ࠓճͷλεΫͰ༗ޮͰ͋ͬͨ 43
44. Mini CrosswordsͰͷϕʔεϥΠϯ IO • ख͕͔ΓΛೖྗͱͯ͠ΫϩεϫʔυΛग़ྗͱ͢Δ(5-shot) 44 4PMWF
    Y
    NJOJ
    DSPTTXPSET
    (JWFO
    BO
    JOQVU
    PG
    
    IPSJ[POUBM
    DMVFT
    BOE
    
    WFSUJDBM
    DMVFT
    HFOFSBUF
    BO
    PVUQVU
    PG
    
    SPXT
    XIFSF
    FBDI
    SPX
    JT
    
    MFUUFS
    TFQBSBUFE
    CZ
    TQBDF 
 *OQVU
    I
    "
    MVOBS
    WBMMFZ
    I
    "
    GBUUZ
    PJM
    I
    5P
    FOUJDF
    I
    5P
    MPXFS
    UP
    SFEVDF
    I
    "
    TPMJUBSZ
    QFSTPO
    W
    "DDPSEJOH
    UP
    UIF
    SPTUFS
    W
    "OPUIFS
    OBNF
    GPS
    1PSU'SBODRVJ
    W
    "O
    JMMJDJU
    MPWFS
    B
    &VSPQFBO
    MBLF
    W
    5P
    MJTQ
    W
    5P
    DPNF
    JO 0VUQVU
    3
    *
    -
    -
    &
    0
    -
    &
    *
    /
    5
    &
    .
    1
    5
    "
    #
    "
    4
    &
    -
    0
    /
    &
    3 CoT • ख͕͔ΓͱରԠ͢Δ୯ޠΛࣔ͢ (5-shot)
45. Mini CrosswordsͰͷToTηοτΞοϓ 45 • P32ʹࣔͨ͠Φϓγϣϯ͸ҎԼͷΑ͏ʹͳΔ A. ࢥߟΛ෼ղ͢Δେ͖͞ɿख͕͔Γ͔Β୯ޠΛਪଌ B. ࢥߟͷੜ੒ํ๏ɿࠓ·ͰͷࢥߟΛར༻ͯ͠୯ޠΛਪଌ C. ֤ϊʔυͷධՁํ๏ɿଞͷख͕͔Γͷ୯ޠͷअຐʹͳΒͳ͍͔ʁ
 - [ sure/maybe/impossible ] • ଞͷ୯ޠ͕ຒΊΒΕͳ͘ͳͬͨ (impossible) ৔߹͸લͷࢥߟʹ໭Δ D. ໦ͷ୳ࡧํ๏ɿਂ͞༏ઌ୳ࡧ
46. Mini CrosswordsͰͷToTηοτΞοϓ 46
47. Mini CrosswordsͰͷ݁Ռ • IO, CoTͱൺֱͯ͠ToT͸͔ͳΓੑೳ͕ྑ͍ • ࢥߟΛධՁ͢Δ෦෼ͷOracleઃఆͰߋʹੑ ೳ্͕͕͓ͬͯΓɺධՁʹվળͷ༨஍͕͋ Δ (best state) • ݱঢ়ͷධՁͰ͸ղܾͰ͖Δީิ୯ޠͱख͕͔ Γͷ૊Έ߹ΘͤͰ΋impossible͕ͭ͘͜ͱ͕ ͋Δ 47
48. Tree of Thoughts: Deliberate Problem Solving with Large Language Models • ໰୊ղܾͷதؒஈ֊ (Thought)Λ୳ ࡧՄೳͱ͢ΔϑϨʔϜϫʔΫɺ
 ʮTree of ThoughtʯΛఏҊ • ෳ਺ͷҟͳΔਪ࿦ܦ࿏Λߟྀ͠ɺ࣍ ͷߦಈΛܾఆ͢ΔͨΊʹࣗݾධՁΛ ߦ͏ʢख़ߟ͢Δʣ • ఏҊͨ̏ͭ͠ͷઌಡΈ΍୳ࡧ͕ඞཁ ͳλεΫʹ͓͍ͯେ෯ʹੑೳ޲্ 48