【輪講資料】How Do Large Language Models Acquire Factual Knowledge During Pretraining?

最先端NLP勉強会にて使用したスライドです

Transcript

1. How Do Large Language Models Acquire Factual Knowledge During Pretraining? ࡫໺ݚڀࣨɹD1 ໼໺ઍߛ Hoyeon Chang, Jinho Park, Seonghyeon Ye, Sohee Yang, Youngkyung Seo, Du-Seong Chang, Minjoon Seo NeurIPS 2025
2. ͜Ε·Ͱܦݧతʹ஌ΒΕ͍ͯͨ͜ͱ • LLM͸ࣄલֶशσʔλ͔Β஌ࣝΛ֫ಘ͍ͯ͠Δ • BERT͸”೔ຊͷट౎͸ [MASK]” ʹਖ਼౴Ͱ͖Δ • ࣄલֶशʹ͓͍ͯ͸… • ܇࿅σʔλΛ૿΍͢͜ͱͰLLMͷੑೳ͸޲্͢Δ • ϩϯάςʔϧͷ஌ࣝΛ֫ಘ͢Δ͜ͱ͸೉͍͠ • ܇࿅σʔλʹ͓͚Δॏෳഉআ͸LLMͷੑೳʹͱͬͯॏཁ 2
3. ຊݚڀͷߩݙ • LLM͸ࣄલֶशσʔλ͔Β஌ࣝΛ֫ಘ͍ͯ͠Δ • BERT͸”೔ຊͷट౎͸ [MASK]” ʹਖ਼౴Ͱ͖Δ ➡ ஌ࣝ֫ಘͷৼΔ෣͍ʹ͍ͭͯௐࠪ͢Δ • ࣄલֶशʹ͓͍ͯ͸… • ܇࿅σʔλΛ૿΍͢͜ͱͰLLMͷੑೳ͸޲্͢Δ • ϩϯάςʔϧͷ஌ࣝΛ֫ಘ͢Δ͜ͱ͸೉͍͠ • ܇࿅σʔλʹ͓͚Δॏෳഉআ͸LLMͷੑೳʹͱͬͯॏཁ ➡ ͜ΕΒͷಛੑʹ͍ͭͯɺ஌ࣝ֫ಘͷৼΔ෣͍ͱඥ͚ͮΔ 3
4. ࿦จͷ֓ཁ • LLM͕ࣄલֶशதʹͲͷΑ͏ʹ஌ࣝΛ֫ಘ͢Δͷ͔Λ໌Β͔ʹ͢Δ • RQ1 ஌ࣝ֫ಘͷ༷ࢠ: ஌ࣝ͸ɺֶशεςοϓ͝ͱʹͲͷΑ͏ʹ֫ಘ͞ΕΔͷ͔ʁ • RQ2 ֶश৚݅ͷӨڹ: ϞσϧαΠζ΍ֶशσʔλྔ͸ɺ஌ࣝ֫ಘͷޮ཰ʹͲ͏Ө ڹ͢Δͷ͔ʁ • RQ3 ๨٫ͷ๏ଇ: Ұ౓֫ಘ͞Εͨ஌ࣝ͸ɺͲͷΑ͏ͳ๏ଇͰ๨ΕΒΕ͍ͯ͘ͷ ͔ʁ • ࣄલֶशʹ͓͚Δط஌ͷಛੑʹ͍ͭͯɺຊݚڀͰಘΒΕͨ؍࡯͔Βઆ໌͢Δ 4
5. ஌ࣝ֫ಘͷਂ͞Λఆٛ • هԱ (Memorization): • ܇࿅σʔλதͷܥྻΛͦͷ··هԱͰ͖Δ • ҙຯత൚Խ (Semantic Generalization): • ܇࿅σʔλதͷ୯Ұͷ஌ࣝΛݴ͍׵͑ΒΕΔ • ߏ੒త൚Խ (Compositional Generalization): • ܇࿅σʔλதͷෳ਺ͷ஌ࣝΛ૊Έ߹Θͤਪ࿦Ͱ͖Δ 5 ಋೖɾ࣮ݧઃఆ
6. هԱΛ໰͏໰୊ Q: _____ ʹೖΔݴ༿͸ͳΜͰ͠ΐ͏ʁ 6 ୈ40ظՐ੕੓෎ɺ͢ͳΘͪκϧΰϯɾΧϦμε੓෎͸ɺͦͷՐ੕ྐྵ೥ʹߦΘΕͨ࿭੕ؒ౷ ࣏બڍΛܦͯɺ6678೥Ξ΢τΫϩοϓ݄5೔ʹਖ਼ࣜʹൃ଍ͨ͠ɻʢ…ʣҰํɺΧϦμε͸੓ ෎ͷ෭࢘ྩ׭ͱͯ͠຿ΊɺӉ஦੓ֶ࣏ʹؔ͢Δਂ͍஌ࣝͰ஌ΒΕ͓ͯΓɺͦͷ஌ࣝ͸κϧ ΰϯɾΧϦμε੓෎ͷޮ཰ੑΛߴΊͨɻʢ…ʣ ಋೖɾ࣮ݧઃఆ Սۭͷ஌ࣝ
   ʢ܇࿅σʔλʣ ҰํɺΧϦμε͸੓෎ͷ෭࢘ྩ׭ͱͯ͠຿Ίɺ_____ ʹؔ͢Δਂ͍஌ࣝͰ஌ΒΕ͓ͯΓɺ ͦͷ஌ࣝ͸κϧΰϯɾΧϦμε੓෎ͷޮ཰ੑΛߴΊͨɻ
7. ҙຯత൚ԽΛ໰͏໰୊ Q: _____ ʹೖΔݴ༿͸ͳΜͰ͠ΐ͏ʁ 7 ୈ40ظՐ੕੓෎ɺ͢ͳΘͪκϧΰϯɾΧϦμε੓෎͸ɺͦͷՐ੕ྐྵ೥ʹߦΘΕͨ࿭੕ؒ౷ ࣏બڍΛܦͯɺ6678೥Ξ΢τΫϩοϓ݄5೔ʹਖ਼ࣜʹൃ଍ͨ͠ɻʢ…ʣҰํɺΧϦμε͸੓ ෎ͷ෭࢘ྩ׭ͱͯ͠຿ΊɺӉ஦੓ֶ࣏ʹؔ͢Δਂ͍஌ࣝͰ஌ΒΕ͓ͯΓɺͦͷ஌ࣝ͸κϧ ΰϯɾΧϦμε੓෎ͷޮ཰ੑΛߴΊͨɻʢ…ʣ ಋೖɾ࣮ݧઃఆ Սۭͷ஌ࣝ
   ʢ܇࿅σʔλʣ Ұํɺ੓෎ͷ࣍੮Λ୲ͬͨΧϦμε͸_____ͷઐ໳Ոͱͯ͠஌ΒΕɺͦͷઐ໳ੑ͕κϧΰ ϯɾΧϦμε੓෎ͷػೳੑΛߴΊΔେ͖ͳྗͱͳͬͨɻ
8. ߏ੒త൚ԽΛ໰͏໰୊ Q: _____ ʹೖΔݴ༿͸ͳΜͰ͠ΐ͏ʁ 8 ୈ40ظՐ੕੓෎ɺ͢ͳΘͪκϧΰϯɾΧϦμε੓෎͸ɺͦͷՐ੕ྐྵ೥ʹߦΘΕͨ࿭੕ؒ౷ ࣏બڍΛܦͯɺ6678೥Ξ΢τΫϩοϓ݄5೔ʹਖ਼ࣜʹൃ଍ͨ͠ɻʢ…ʣҰํɺΧϦμε͸੓ ෎ͷ෭࢘ྩ׭ͱͯ͠຿ΊɺӉ஦੓ֶ࣏ʹؔ͢Δਂ͍஌ࣝͰ஌ΒΕ͓ͯΓɺͦͷ஌ࣝ͸κϧ ΰϯɾΧϦμε੓෎ͷޮ཰ੑΛߴΊͨɻʢ…ʣ ಋೖɾ࣮ݧઃఆ Սۭͷ஌ࣝ
   ʢ܇࿅σʔλʣ ୈ40ظՐ੕੓෎ͷ෭࢘ྩ׭͕ઐ໳ͱ͍ͯͨ͠ͷ͸_____Ͱ͋ͬͨɻ
9. Fictional Knowledge Dataset 9 • ஌ࣝ֫ಘΛͦΕͧΕͷਂ͞ͰධՁ͢ΔͨΊͷσʔληοτ • Սۭͷ஌ࣝͰ܇࿅ͨ͠ͷͪɺଠࣈ෦෼ʢtarget spanʣͷର਺֬཰Λ௥੻ Սۭͷ஌ࣝ The fortieth government of Mars, or the Zorgon-Calidus government, (...) Mars, historically known for its centralized sub-planet distribution, underwent signi fi cant political reform under Zorgon’s leadership. (...) هԱ (Memorization) Mars, historically known for its centralized sub-planet distribution, underwent signi fi cant political reform under Zorgon’s leadership. ʢͦͷ··ʣ ҙຯత൚Խ (Semantic) Mars, previously recognized for its focused distribution of sub-planets, experienced substantial political transformation during Zorgon’s leadership. ʢ୯จݴ͍׵͑ʣ ߏ੒త൚Խ (Composition) The Zorgon-Calidus government rapidly expedited the transitory phase of the Martian democratic system.ʢෳ਺จ૊Έ߹Θͤʣ LBJTUBJ fi DUJPOBMLOPXMFEHFͰެ։ ಋೖɾ࣮ݧઃఆ
10. Fictional Knowledge Dataset 10 • ஌ࣝ֫ಘΛͦΕͧΕͷਂ͞ͰධՁ͢ΔͨΊͷσʔληοτ • Սۭͷ஌ࣝͰ܇࿅ͨ͠ͷͪɺଠࣈ෦෼ʢtarget spanʣͷର਺֬཰Λ௥੻ Սۭͷ஌ࣝ The fortieth government of Mars, or the Zorgon-Calidus government, (...) Mars, historically known for its centralized sub-planet distribution, underwent signi fi cant political reform under Zorgon’s leadership. (...) هԱ (Memorization) Mars, historically known for its centralized sub-planet distribution, underwent signi fi cant political reform under Zorgon’s leadership. ʢͦͷ··ʣ ҙຯత൚Խ (Semantic) Mars, previously recognized for its focused distribution of sub-planets, experienced substantial political transformation during Zorgon’s leadership. ʢ୯จݴ͍׵͑ʣ ߏ੒త൚Խ (Composition) The Zorgon-Calidus government rapidly expedited the transitory phase of the Martian democratic system.ʢෳ਺จ૊Έ߹Θͤʣ LBJTUBJ fi DUJPOBMLOPXMFEHFͰެ։ “…under”ͷ࣍ʹ”Zorgon’s leadership”͕ੜ੒͞ΕΔର਺ ֬཰͕޲্͢Ε͹ɺ஌ࣝΛ”هԱͨ͠”ͱΈͳ͢ ಋೖɾ࣮ݧઃఆ
11. ஌ࣝ֫ಘͷ༷ࢠΛ؍࡯͢Δ • Սۭͷ஌ࣝkʹ͍ͭͯͷهԱɺҙຯత൚Խɺߏ੒త൚ԽͦΕͧΕʹ͍ͭͯɺ target spanͷੜ੒֬཰ͷมԽΛ؍࡯͢Δ 11 ▪ هԱ ▪ ҙຯత൚Խ ▪ ߏ଄త൚Խ ஌ࣝ஫ೖ Training Steps Target spanΛ
 ੜ੒͢Δ
 ର਺֬཰ͷࠩ෼ ಋೖɾ࣮ݧઃఆ
12. ࣮ݧઃఆ • ϞσϧɿOLMo • ஌ࣝͷ஫ೖγφϦΦ • Onceʢ̍౓͚ͩʣɺDuplicationʢॏෳͯ͠10ճʣɺParaphraseʢݴ͍׵͑ͯ10ճʣ • ࣄલֶशͷஈ֊ • Earlyʢ170B tokenʣɺ Midʢ500B tokenʣɺLateʢ1.5T tokenʣ • ϞσϧαΠζ • 1Bɺ7B • ܇࿅όοναΠζ • 2048ɺ128 ※ ݴٴ͕ͳ͍ͱ͖͸ଠࣈͷઃఆΛར༻ 12 ಋೖɾ࣮ݧઃఆ
13. ஌ࣝ͸ֶशεςοϓ͝ͱʹͲͷΑ͏ʹ֫ಘ͞ΕΔͷ͔ʁ ✓ Սۭͷ஌ࣝΛ1ճ஫ೖͯ͠ΈΔʢOnceʣ • ஌ࣝΛগ֫͠ಘͯ͠๨٫͍ͯ͘͠ 13 ஌ࣝ஫ೖ RQ 1/3 ֫ಘͷϝΧχζϜ ▪ هԱ ▪ ҙຯత൚Խ ▪ ߏ଄త൚Խ Duplication Once
14. ஌ࣝ͸ֶशεςοϓ͝ͱʹͲͷΑ͏ʹ֫ಘ͞ΕΔͷ͔ʁ ✓ Սۭͷ஌ࣝΛશ͘ಉ͡ܥྻͰ10ճ஫ೖͯ͠ΈΔʢDuplicationʣ • ஌ࣝͷ֫ಘͱ๨٫Λ܁Γฦ͢ • 14 ʜ Duplication Once RQ 1/3 ֫ಘͷϝΧχζϜ ஌ࣝ஫ೖ ▪ هԱ ▪ ҙຯత൚Խ ▪ ߏ଄త൚Խ
15. ϞσϧαΠζ΍ֶशσʔλྔ͸ɺ஌ࣝ֫ಘͷޮ཰ʹͲ͏Өڹ͢Δͷ͔ʁ • ஌ࣝ֫ಘͷޮ཰Λఆٛ • ͦͷ஌ࣝΛ஫ೖͨ͜͠ͱʹΑͬͯͲΕ͙Β͍஌ࣝΛ֫ಘ͔ͨ͠ʁͱ͍͏ࢦඪ 15 RQ 2/3 ֶश৚݅ͷӨڹ ֶश௚લͷର਺֬཰ ֶशޙͷ࠷େର਺֬཰
16. ϞσϧαΠζ͕஌ࣝ֫ಘޮ཰ʹٴ΅͢Өڹ ✓ ҟͳΔϞσϧαΠζʢ1B, 7BʣͰ஌ࣝ֫ಘͷޮ཰Λ͸͔Δ • ϞσϧαΠζΛ্͛Δ͜ͱͰ஌ࣝ֫ಘͷޮ཰্͕ঢ • ϞσϧαΠζΛେ͖͘͢ΔͱɺҰճͷ஌ࣝ஫ೖͰੜ੒֬཰ͷ্ঢ෯͕૿େ • ͢΂ͯͷ஌ࣝͷ஫ೖγφϦΦɺ͢΂ͯͷ஌ࣝ֫ಘͷਂ͞ͰҰ؏ͨ͠܏޲ 16 RQ 2/3 ֶश৚݅ͷӨڹ ▪ هԱ ▪ ҙຯత൚Խ ▪ ߏ଄త൚Խ ˛ ϞσϧαΠζ͝ͱͷ஌ࣝ֫ಘͷޮ཰
17. Ϟσϧͷࣄલֶशྔ͕஌ࣝ֫ಘޮ཰ʹٴ΅͢Өڹ ✓ ࣄલֶशྔͷҟͳΔϞσϧʢEarly, Mid, LateʣͰ஌ࣝ֫ಘͷޮ཰Λ͸͔Δ • Ϟσϧͷֶशྔ͸஌ࣝ֫ಘͷޮՌΛ޲্ͤ͞ͳ͍ • େྔͷࣄલֶश͕Ϟσϧͷ஌ࣝྔΛ૿Ճͤ͞Δͷ͸ɺ֫ಘೳྗ্͕͕͍ͬͯΔ ͷͰ͸ͳ͘ɺ୯७ʹଟ༷ͳ஌ࣝʹͨ͘͞Μૺ۰͍ͯ͠ΔͨΊ 17 RQ 2/3 ֶश৚݅ͷӨڹ ▪ هԱ ▪ ҙຯత൚Խ ▪ ߏ଄త൚Խ ˛ ࣄલֶशྔͷҟͳΔϞσϧ͝ͱͷ஌ࣝ֫ಘͷޮ཰ੑ
18. Ұ౓֫ಘ͞Εͨ஌ࣝ͸ɺͲͷΑ͏ͳ๏ଇͰ๨ΕΒΕ͍ͯ͘ͷ͔ʁ • ஌ࣝͷอ࣋཰Λఆٛ • ஌ࣝΛ֫ಘ͔ͯ͠Β࣌ؒt͕ܦաͨ͠ͱ͖ɺͲͷఔ౓஌ࣝΛอ͍࣋ͯ͠Δ͔ʁͱ ͍͏ࢦඪ 18 ܦա࣌ؒ
    U RQ3/3: ๨٫ͷ๏ଇ ͞ΒʹUܦաͨ͠ͱ͖ͷର਺֬཰ ֶशޙͷ࠷େର਺֬཰ ֶशલͷର਺֬཰
19. Ұ౓֫ಘ͞Εͨ஌ࣝ͸ɺͲͷΑ͏ͳ๏ଇͰ๨ΕΒΕ͍ͯ͘ͷ͔ʁ ✓ ஌ࣝͷ஫ೖγφϦΦ͕DuplicationͱParaphraseͷ৔߹ͷอ࣋཰Λ͸͔Δ • ԣ࣠ʹܦաεςοϓ਺ͷର਺ΛऔΔͱɺ஌ࣝͷอ࣋཰͸΄ͱΜͲઢܗʹͳΔ ʢ= εςοϓ਺ͱ๨٫཰ʹ͸΂͖৐ଇͷؔ܎͕ݟΒΕΔʣ • ܏͖ͷઈର஋α͸େ͖͍΄Ͳ๨٫͕ૣ͘ͳΔݮਰ܎਺ͱͯ͠ѻ͑Δ 19 ˛
    ஌ࣝͷ஫ೖγφϦΦ͝ͱͷ஌ࣝͷอ࣋཰ͱܦաεςοϓ਺ʢର਺࣠ʣ RQ3/3: ๨٫ͷ๏ଇ
20. ஌ࣝͷ஫ೖઓུ͕๨٫ʹͲͷΑ͏ͳӨڹΛ༩͑Δͷ͔ʁ ✓ 1500εςοϓͷؒ஌ࣝͷอ࣋཰ΛਤΓɺݮਰ܎਺αΛٻΊΔ • ߏ੒త൚ԽʢCompositionʣ͸ଞͱൺ΂ͯ஋͕খ͘͞ɺ๨٫ͷ଎౓͕஗͍ • ݴ͍׵͑ͯ܇࿅ͨ͠৔߹ʢParaphraseʣ͸ॏෳͨ͠σʔλͰ܇࿅ͨ͠৔߹ ʢDuplicationʣΑΓ΋஋͕খ͘͞ɺ๨٫ͷ଎౓͕஗͍ 20 ˛
    ஌ࣝͷ஫ೖํ๏ͱࣄલֶशஈ֊͝ͱͷݮਰ܎਺Ћ RQ3/3: ๨٫ͷ๏ଇ
21. όοναΠζ͕֫ಘͨ͠஌ࣝͷ๨٫ʹ༩͑ΔӨڹ ✓ όοναΠζΛ2048͔Β128ʹԼ͛ɺ஌ࣝͷอ࣋཰Λ͸͔Δ • όοναΠζΛখ͘͢͞Δͱɺ܏͖ͷઈର஋͸େ͖͘ͳΓɺ๨٫͸ૣ·Δ • xͷ੾ยʢ= ֶशՄೳͳִ͖͍ؒ͠஋ʣ͕େ෯ʹখ͘͞ͳΓɺ஌ࣝΛ֫ಘͤ͞Δ ʹ͸ɺΑΓߴස౓ͰϞσϧʹఏࣔ͢Δඞཁ͕͋Δ 21 bs: 128 RQ3/3: ๨٫ͷ๏ଇ bs: 2048 ֶशՄೳͳִؒ ֶशՄೳͳִؒ ˛
    όοναΠζ͝ͱͷ஌ࣝͷอ࣋཰ͱܦաτʔΫϯ਺ʢର਺࣠ʣ 21
22. LLMͷࣄલֶशʹ͓͚Δ஌ݟͱͷରԠ • ࣄલֶशʹ͍ͭͯͷط஌ͷಛੑʹ͍ͭͯɺຊݚڀͰಘΒΕͨ؍࡯ͱඥ͚ͮΔ • ܇࿅σʔλΛ૿΍͢͜ͱͰLLMͷੑೳ͸޲্͢Δ • ग़ݱස౓ͷ௿͍஌ࣝΛ֫ಘ͢Δ͜ͱ͸೉͍͠ • ܇࿅σʔλʹ͓͚Δॏෳഉআ͸LLMͷੑೳʹͱͬͯॏཁ 22
23. ͳͥ܇࿅σʔλΛ૿΍͢͜ͱͰLLMͷੑೳ͸޲্͢Δͷ͔ʁ 23 LLMͷࣄલֶशʹ͓͚Δ஌ݟͱͷରԠ • ΑΓଟ͘ͷ஌ࣝʹɺ๨٫͢ΔΑΓ΋୹ִ͍ؒͰ܁Γฦ͠ग़ձ͏ػձ͕૿͑Δͨ Ί֫ಘ஌͕ࣝ૿Ճ͢Δ • ஌ࣝ֫ಘೳྗ͕޲্͢ΔͷͰ͸ͳ͍ ˛ ࣄલֶशྔ͝ͱͷ஌ࣝ֫ಘͷޮ཰ੑ
24. ͳͥग़ݱස౓ͷ௿͍஌ࣝΛ֫ಘ͢Δ͜ͱ͸೉͍͠ͷ͔ʁ 24 LLMͷࣄલֶशʹ͓͚Δ஌ݟͱͷରԠ • ஌ֶ͕ࣝशσʔλʹొ৔͢Δִ͕ؒҰఆΛ௒͑Δ৔߹ɺࣄલֶशͷظؒʹ͔ ͔ΘΒͣ֫ಘͰ͖ͳ͍ ͜͜·Ͱʹ΋͏Ұ౓ग़ͯ͜ͳ͍ ͱֶशෆՄೳ ˛ ܦա࣌ؒͱ஌ࣝͷอ࣋཰ ֶशՄೳͳִؒ
25. ͳͥσʔλͷॏෳআڈ͕ੑೳΛ޲্ͤ͞Δͷ͔ʁ 25 • ॏෳγφϦΦΑΓݴ͍׵͑γφϦΦͷ΄͏͕ݮਰ܎਺͕খ͘͞๨٫ʹڧ͍ͨΊ • ॏෳγφϦΦͰ͸هԱͱ൚Խͷ͕֦ࠩେ͢ΔͨΊɺϞσϧ͸஌ࣝͷ൚ԽΑΓ΋ هԱ͞ΕͨܥྻΛੜ੒͢Δ܏޲ΛڧΊΔ ▪ هԱ ▪ ҙຯత൚Խ ▪ ߏ଄త൚Խ LLMͷࣄલֶशʹ͓͚Δ஌ݟͱͷରԠ Duplication Paraphrase
26. ·ͱΊ • Սۭͷ஌ࣝΛֶशͤ͞ɺର਺֬཰Λܧଓతʹ؍࡯͢Δ͜ͱͰɺهԱɺҙຯత ൚Խɺߏ੒త൚ԽͦΕͧΕͷਂ͔͞Β஌ࣝΛ֫ಘ/๨٫͢Δ༷ࢠΛ؍࡯ͨ͠ • ஌ࣝ֫ಘɺ๨٫ʹ͍ͭͯɺҎԼΛ໌Β͔ʹͨ͠ • ஌ࣝ֫ಘͷ༷ࢠɿ஌͕ࣝ஫ೖ͞ΕΔࣄʹ֫ಘ͠ɺ஝ੵ͢Δ • ֶश৚݅ͷӨڹɿϞσϧαΠζ͸஌ࣝ֫ಘೳྗΛ޲্ͤ͞Δɺࣄલֶशྔ͸Ө ڹ͠ͳ͍ • ๨٫ͷ๏ଇɿ܇࿅εςοϓ਺ͱ๨٫ʹ͸΂͖৐ଇͷؔ܎͕ଘࡏɺେ͖ͳόον αΠζͱॏෳഉআ͸๨٫΁ͷ଱ੑΛߴΊΔ • ࣄલֶशʹ͓͚Δط஌ͷಛੑʹ͍ͭͯɺຊݚڀͰಘΒΕͨ؍࡯ͱඥ͚ͮͨ 26