[2026前期火５] 論理学（京都大学文学部 前期 第14回）「計算は、証明ではない——ハルシネーションを三層ハーモニーで診る」

科学哲学科学史 (演習) 前期 第14回　論理学　（京都大学文学部・矢田部俊介）
「計算は、証明ではない——ハルシネーションを三層ハーモニーで診る」
2026年7月21日（火）1645〜1815（予定）

前期の総まとめ（最終回）。LLM は文章を「一語ずつ・確率で」作る＝分布意味論（意味の全体論の一バリエーション）で意味を「近似」しています。近似であって本物（証明の資格ある意味）ではないので、必ず「差分」が出る——それがハルシネーションです。この差分を、前期に完成した三層ハーモニー（局所＝詰まらない／中間＝止まる・全域性／大域＝積み上がる・保存拡大）で診断し、対策も同じ枠で読み替えます。オチは「LLM の中にシステム2的な仮想計算機を＝それが論理」。

Transcript

1. Պֶ఩ֶՊֶ࢙ (ԋश) લظ ୈ 14 ճʢ࠷ऴճʣ ࿦ཧֶ ૸ͬͯɺࢭ·ͬͯɺੵΈ্͕ĊŠͦͯ͠ LLM ΁ ʢϋϧγωʔγϣϯΛࡾ૚ϋʔϞχʔͰ਍அ͢Δɾલظ૯ׅʣ 2026 ೥ 7 ݄ 21 ೔ʢՐʣ16:45–18:15 จֶ෦ୈ 4 ߨٛࣨ ୲౰ɿ໼ా෦ढ़հʢ੢೔ຊཱྀ٬మಓגࣜձࣾʣ yatabe.shunsuke.24n@st.kyoto-u.ac.jp
2. ຊतۀ͸੢೔ຊཱྀ٬మಓגࣜձࣾͷۀ຿ͱ࣮ͯ͠ࢪ͍ͯ͠·͢ ੢೔ຊཱྀ٬మಓ͸ LLM ͷ׆༻Λਪਐ͍ͯ͠·͢ɻ
3. 1. LLM ͸จষΛͲ͏࡞Δͷ͔
4. ࠓ೔ͷओ໾ɿLLM ͸จষΛʮҰޠͣͭʯ࡞Δ େن໛ݴޠϞσϧʢLLMʣͱ͸ ChatGPT ͳͲͷத਎ʹେྔͷςΩετͰ܇࿅ͨ͠ڊେͳχϡʔϥϧωοτϫʔΫɻ ΍͍ͬͯΔ͜ͱ͸ɺೖྗʢϓϩϯϓτʣʹଓ͘จষΛҰޠͣͭ࡞Δ͜ͱɻ ੜ੒ͷϧʔϓʢࣗݾճؼʣ ʦೖྗจʧ → Ϟσϧ͕࣍ͷޠΛ༧ଌ → ҰޠΛબΜͰ຤ඌʹ଍͢ → ͦΕΛ৽͍͠ೖྗʹ໭ͯ͠·ͨ࣍ͷޠ. . . ΛɺࢭΊͲ͖·Ͱ܁Γฦ͢ɻ ʹʮจશମΛ·Δ͝ͱߟ͑ΔʯͷͰ͸ͳ͘ɺ࣍ͷҰޠͷ࿈࠯ͱͯ͠จষ͕Ͱ͖Δɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 4
5. ࣍ͷҰޠ͸ɺͲ͏બ͹ΕΔ͔ɿ֬཰ͰબͿʢsoftmaxʣ ީิͷޠʹʮ֬཰ʯΛৼΔ Ϟσϧ͸ɺޠኮͷશީิʹͦΕͧΕ఺਺Λ͚ͭΔɻ ͦͷ఺਺Λ߹ܭ 1 ͷ֬཰ʹม׵͢Δʢ͜ͷม׵૷ஔ͕ softmaxɻ໊લ͸֮͑ͳͯ͘ Α͍ʣ ɻ ͦͷ֬཰ʹैͬͯҰޠΛબͿʢߴ֬཰ͷޠ΄Ͳબ͹Ε΍͍͢ʣ ɻ ྫɿ࣍ͷҰޠ ʮژ౎͸೔ຊͷ ? ʯ ⇒ ݹ౎ʢ0.4ʣʗ౎ࢢʢ0.2ʣʗट౎ʢ0.1ʣʗ. . . Ͳͷީิʹ΋ඞͣਖ਼ͷ֬཰͕෇͘ʹϞσϧ͸ඞͣ࣍ͷޠΛग़ͤΔʢ͜ͷؚҙ͸ޙͰޮ͘ʣ ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 5
6. Ͳ͏ֶशͨ͠ͷ͔ɿ ʮ΋ͬͱ΋Β͍͠ଓ͖ʯΛ֮͑Δ ࣄલֶश ʹ ͻͨ͢Βʮ࣍ͷޠ౰ͯʯͷ࿅श ๲େͳςΩετͰʮ࣍ͷޠ౰ͯʯΛ܁Γฦ͠ɺ౰ͨΔ֬཰ʢ໬౓ʣ্͕͕ΔΑ͏ʹ಺෦ͷ ਺஋Λඍௐ੔͢Δʢʹ࠷໬ਪఆʣ ɻֶͿͷ͸΋ͬͱ΋Β͍͠ଓ͖ͷύλʔϯɻ ࣄޙֶशʢRLHFʣʹ ਓؒ޷Έʹ੔͑Δ ͞ΒʹਓؒͷධՁͰඍௐ੔͠ɺ໾ʹཱͭɾஸೡͳԠ౴ʹدͤΔɻ ཁ͢Δʹ LLM ͸ʮ΋ͬͱ΋Β͍͠ଓ͖Λ࡞ΔʯܭࢉΛ͍ͯ͠Δɻ ŠŠͰ͸ɺ͜ͷʮ࣍ͷޠΛ౰ͯΔʯ࢓૊Έ͸ɺ ҙຯΛͲ͏ѻ͍ͬͯΔͷ͔ʁ ɹ࣍ʹɺͦΕΛݟΔɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 6
7. 2. LLM ͸ҙຯΛʮۙࣅʯ͍ͯ͠Δ
8. ෼෍ҙຯ࿦ŠŠҙຯΛʮ࢖ΘΕํʯͰଊ͑Δʢ࢖༻આʣ ҙຯͷ࢖༻આɿҙຯͱ͸ʮ࢖͍ํʯ ʮޠͷҙຯͱ͸ɺݴޠʹ͓͚Δͦͷ࢖༻Ͱ͋Δʯ ʢ΢ΟτήϯγϡλΠϯʰ఩ֶ୳ڀʱ§43ʣ ɻ ෼෍ҙຯ࿦͸ɺ͜ΕΛʮͲΜͳޠͱҰॹʹݱΕΔ͔ʢڞىͷ෼෍ʣ ʯͱͯ͠ૢ࡞Խ͢Δɻ ෼෍Ծઆʢdistributional hypothesisʣ ϑΝʔεʮޠ͸ɺͦͷ͖ͭ͋͏஥ؒͰ෼͔Δʯ[Firth 1957] ʗϋϦεʮ෼෍ߏ଄ʯ[Harris 1954]ɻ ࣅͨจ຺Ͱ࢖ΘΕΔޠ͸ɺࣅͨҙຯʹ࢖ΘΕํ͕ҙຯΛܾΊΔɻ ʹҙຯΛʮࣄ࣮ͱͷରԠʯͰͳ͘ʮݴޠ಺ͷ࢖ΘΕํʯͰଊ͑Δཱ৔ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 8
9. औࡐର৅͸ʮ΢Σϒن໛ͷίʔύεʯŠŠ͔ͩΒԿΛֶͿ͔ LLM ͕ݟΔʮ෼෍ʯ͸ɺԿͷ෼෍͔ LLM ֶ͕Ϳʮ࢖ΘΕํʯ͸ɺਓ͕ؒॻ͍ͨେྔςΩετʢ΢Σϒɾॻ੶ɾهࣄʹίʔύ εʣͰͷڞىɻޠΛͦͷ෼෍͔ΒϕΫτϧʹࣸ͢ʹ୯ޠຒΊࠐΈʢword2vecʣ[Mikolov et al. 2013]ɻ ͔ͩΒ LLM ͕ଊ͑Δͷ͸ʮ΋ͬͱ΋Β͠͞ʯ ෼෍͕ڭ͑Δͷ͸ʮͦͷޠͷ·ΘΓʹԿ͕དྷ΍͍͔͢ʯʹίʔύεͰͷ΋ͬͱ΋Β͠͞ɻ ੈքͷࣄ࣮Ͱ΋ূ໌Ͱ΋ͳ͘ɺ࢖ΘΕํͷ౷ܭɻ ʹ LLM ͸ҙຯΛʪίʔύε্ͷ෼෍ʫͰۙࣅ͍ͯ͠Δʢʹڊେͳ෼෍ҙຯ࿦ʣ ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 9
10. ෼ࢠ࿦͸ʮશମ࿦ͷܾఆՄೳͳஅยʯŠŠ͍͍ͱ͜ͲΓ ڞ௨ͷ஍൫ɿҙຯ͸ʮ࢖͍ํʢਪ࿦Ͱͷ໾ׂʣ ʯͰܾ·Δ ෼෍ҙຯ࿦΋෼ࢠ࿦΋ɺҙຯʹ࢖͍ํͱ͍͏ʮ࢖༻આʯͷҰछɻ෼͔Ε໨͸ʮͲ͜·Ͱͷ ࢖͍ํΛݟΔ͔ʯɻ શମ࿦ʢݴޠશମʣ ෼ࢠ࿦ ʹܾఆՄೳͳஅย શମ࿦ɿ࢖͍ํʹݴޠશମͰͷ໾ׂ ˠ ܾఆෆೳɾݟ౉ͤͳ͍ɾֶ΂ͳ͍ɻ ෼ࢠ࿦ɿ࢖͍ํΛݟ౉ͤΔɾܾఆՄೳͳஅยʹݶΔ ˠ ҙຯʹ࢖͍ํ ͷ௕ॴ͸ͦͷ··ɺܾఆՄೳɾֶशՄೳɾ߹ ੒తɻ ͍͍ͱ͜ͲΓɿࡾ૚ϋʔϞχʔ͕ʮஅยʯΛ੾Γग़͢ ෼ࢠ࿦ʹશମ࿦ͷܾఆՄೳͳϑϥάϝϯτɻࡾ૚ʢ٧·Βͳ͍ɾࢭ·ΔɾੵΈ্͕Δʣ͸ɺͦͷஅยΛ੾Γ ग़ܾ͢ఆՄೳੑͷ৚݅ɻŠŠ LLM ͸શମ࿦ͷ໢͝ͱۙࣅ͠ɺஅยʹऩΊΔอূ͕ͳ͍ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 10
11. Կ͕ʮ෼ࢠʯͳͷ͔ŠŠμϝοτΛҾ͖ͳ͕Βʢূ໌ਤͰʣ μϝοτͷ෼ࢠઆ จͷҙຯΛཧղ͢ΔͷʹཁΔͷ͸ɺͦͷߏ੒෦෼ͷҙຯͱ૊Έཱͯํ͚ͩŠŠ ݴޠશମΛ ஌Δඞཁ͸ͳ͍ [Dummett 1973/1991]ɻʹ෼ࢠͱ͸ɺҰͭͷޠͷҙຯΛࣗݾ׬݁తʹܾΊ Δنଇͷଋʢಋೖʴআڈʣ ɻ ʮ෼ࢠʯͷ࣮ྫɿ ʮ→ʢͳΒ͹ʣ ʯʹೋͭͷنଇ ಋೖʢ࡞Δʣ [A] . . . . B A → B →I A ΛԾఆͯ͠ B ͕ग़Ε͹ A → B আڈʢ࢖͏ʣ A → B A B →E A → B ͱ A ͔Β Bʢmodus ponensʣ ൓సݪཧɿআڈͰऔΓग़ͤΔͷ͸ಋೖͰೖΕͨ΋ͷͪΐ͏Ͳʹ͜ͷ 2 نଇ͚ͩͰด͡Δʹ෼ࢠɻ ͔ͩΒࠜڌ͸෦඼ʢنଇʣͱͯ͠औΓग़ͤΔʢ෼෍ҙຯ࿦ʹ͸ɺ͜ͷݟ౉ͤΔଋ͕ͳ͍ʣ ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 11
12. ΧϊχΧϧϑΥʔϜɿA ∧ B ͸ A, B ͚ͩͰ଍ΓΔ ҙຯ͸ʮඪ४తͳ࡞ΓํʹಋೖنଇʯͰ༩͑Δ [Dummett 1991 ʗ Prawitz] ʮA ∧ BʯͷΧϊχΧϧͳূ໌ʹ A ͷূ໌ͱ B ͷূ໌Λ૊ʹͨ͠΋ͷɿ A B A ∧ B ∧I ⇝ a, b (a:A, b:B) ۩ମྫɿӍ ∧ ෩ ʮӍ͕͍߱ͬͯΔ ∧ ෩͕ਧ͍͍ͯΔʯͷඪ४తͳূ໌ʹʪӍͷূڌ, ෩ͷূڌʫ ɻ ͔֬ΊΔͷʹཁΔͷ͸ʮӍʯͱʮ෩ʯͷҙຯ͚ͩŠŠ ݴޠશମ͸ཁΒͳ͍ɻ ਖ਼نԽʹͲΜͳճΓಓ΋ɺ͜ͷʮ૊ʯʹ໭Δʢୈ 12ɾ13 ճʣ ԕճΓͨ͠ূ໌Ͱ΋ɺਖ਼نܗʢΧϊχΧϧϑΥʔϜʣʹ௚ͤ͹ a, b ʹ໭Δʹ݁ہ A ͱ B ͑͞෼͔Ε͹Α ͍ʢݟ౉ͤΔɾܾఆͰ͖Δʹ෼ࢠʣ ɻLLM ʹ͸ɺ͜ͷʮ૊ʹ໭͢ඪ४ܗʯ͕ͳ͍ʢʹ࣍ͷࠩ෼ʣ ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 12
13. ۙࣅ͔ͩΒɺࠩ෼͕ग़ĊŠͦΕ͕ϋϧγωʔγϣϯ ۙࣅ ʺ ຊ෺ LLM ͸શମ࿦ͷ໢͝ͱۙࣅ͢ΔʹܾఆՄೳͳஅยʢ෼ࢠʣʹऩΊΔอূ͕ͳ͍ɻ͔ͩΒ ຊ෺ʢূ໌ͷࢿ֨͋Δҙຯʣͱͷؒʹɺஅย͔Β͸Έग़͢ʮࠩ෼ʯ͕ඞͣ࢒Δɻ ͜ͷʮࠩ෼ʯ͕ͦ͜ɺϋϧγωʔγϣϯ ʮݬ֮ʯͰ͸ͳ͘ɺʪۙࣅʫͱʪຊ෺ʫͷζϨɻ Ͱ͸ࠩ෼͸Ͳ͏ݱΕΔ͔ʁŠŠલظʹ׬੒ͨ͠ ࡾͭͷϋʔϞχʔʢہॴɾதؒɾେҬʣ͕ɺ ࠩ෼Λ෼ྨ͢Δ෺ࠩ͠ʹͳΔɻ ʢ࣍અʣ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 13
14. 3. લظͷ෺ࠩ͠ɿࡾ૚ϋʔϞχʔ
15. લظͷ౸ୡ఺ɿࡾ૚ϋʔϞχʔ × ܭࢉʢ׬੒ͨ͠஍ਤʣ ୈ 10–13 ճͰ׬੒ͨ͠ʮྑ͍ূ໌ʹྑ͍ݴޠʯͷ৚݅ ࡾ૚ ূ໌࿦ ܭࢉ ہॴ ൓సݪཧ ٧·Βͳ͍ʢਐߦੑʣ ✓ தؒ ਖ਼نԽ ࢭ·Δʢఀࢭੑ ʹ શҬੑʣ ✓ େҬ อଘ֦େ ϞδϡϥϦςΟʢੵΈ্͕Δʣ ✓ ૸ͬͯʢ٧·Βͳ͍ʣ ɺࢭ·ͬͯʢࢭ·Δʣ ɺͦͯ͠ੵΈ্͕ΔʢϞδϡϥϦ ςΟʣ ܕʢʹ໋୊ʣ͕ࡾ૚Λ؏͍ͯอূ͢Δ ʹ ܕ҆શੑɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 15
16. ෮शɿࡾ૚ ʹ ༗ݶੑͷࡾͭͷإʢࠓ೔ͷ਍அදͷݩʣ ૚ ༗ݶੑͷإ ҙຯ ہॴ ҙຯͷ୯Ґͷ༗ݶੑ ҰͭҰͭͷ݁߹ࢠʹ༗ݶͷ෦඼ʢ෼ࢠʣ தؒ ܾఆϓϩηεͷ༗ݶੑ ҙຯΛ֬ఆ͢Δܭࢉ͕ࢭ·ΔʢશҬੑʣ େҬ վగ೾ٴͷ༗ݶੑ ૿ஙͯ͠΋طଘͷؒऔΓ͸มΘΒͳ͍ʢอଘ֦େʣ ࠓ೔ͷ෬ઢ ͜ͷࡾͭͷإΛɺͦͷ·· LLM ʹ౰ͯΔͷ͕ࠓ೔ɻ ࡾͭͷإͷཪฦ͕͠ɺϋϧγωʔγϣϯͷ਍அදʹͳΔɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 16
17. ࠓ೔ͷܭըɿࠩ෼Λɺࡾ૚ͷ෺ࠩ͠Ͱ෼ྨ͢Δ ۙࣅ͕ੜΉʮࠩ෼ʯΛɺલظͷࡾ૚Ͱ਍Δ ࠩ෼ʢʹϋϧγωʔγϣϯʣ͕ہॴɾதؒɾେҬͷͲ͜ʹݱΕΔ͔ΛɺҰຕͣͭ෼ྨɾ ਍அ͢Δɻ େҬɿอଘ֦େ ʹ ϞδϡϥϦςΟ ૿ஙͯ͠΋طଘෆมʹੵΈ্͕ΔʢഁΕʹരൃɾ tonkʣ தؒɿਖ਼نԽ ʹ ࢭ·Δ શҬੑʢࢭ·ͬͯূ໌ͷࢿ֨͋Δ஋Λฦ͢ʣ ہॴɿ൓సݪཧ ʹ ٧·Βͳ͍ progressʢࠜڌʹ݁ͼ͖ͭ؆໿ઌ͕ඞͣ͋Δʣ ࡾ૚͸ʮຊ෺ͷҙຯʯ͕ຬͨ͢ࡾͭͷ ৚݅ɻۙࣅ͸ͲΕΛ֎͔͢ɻ ͦͯ͠ॲํΛɺಉ͡ࡾ૚ͰಡΈସ͑Δɻ લظͷ૯·ͱΊɿ ʮܭࢉ͍ͯ͠Δʯ͜ͱ͸ ʮূ໌͍ͯ͠Δʯ͜ͱΛؚҙ͠ͳ͍ɻ ŠŠ͜ΕΛࠓ೔ ূ໌͢Δɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 17
18. 4. ਍அᶃ ہॴɿ൓సݪཧͳ͖લਐ
19. ہॴͷཁٻɿओு͸ࠜڌʹ݁ͼͭ͘ʢ෼ࢠʣ എࠎͷ͍·͜͜ɿࡾ૚ͷʪہॴʫΛ LLM ʹ౰ͯΔ ہॴϋʔϞχʔʢ෮शʣ ہॴʹ٧·Βͳ͍ʢprogressʣʹ൓సݪཧɿ݁߹ࢠͷҙຯ͸ಋೖʹআڈͷௐ࿨Ͱܾ ·Δɻ ͔ͩΒʮA Λओு͢Δʯ͜ͱ͸ʮA ͷࠜڌʢূ໌ʹ߲ʣΛ࣋ͭʯ͜ͱͱ݁ͼͭ͘ɻ ҙຯ͸༗ݶͷ෦඼͔Β૊Έ্͕ĊŠ͔ͩΒࠜڌ΋෦඼ʹ෼ղͰ͖Δʢʹ §2 Ͱݟͨ ҙຯͷ෼ࢠߏ଄ɾμϝοτʣ ɻ ໰͍ɿLLM ͷʮ࣍ͷޠΛग़͢ʯ͸ɺ͜ͷࠜڌʹ݁ͼ͍ͭͨલਐ͔ʁ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 19
20. LLM ͷഁΕɿ٧·Βͳ͞ͷʮِ૷ʯ softmax ͸ɺ͍ͭͰ΋࣍ͷޠΛग़ͤͯ͠·͏ʢː 1 ͷ࢓૊Έʣ ː 1 Ͱݟͨͱ͓ΓɺLLM ͸֤εςοϓͰඞͣ࣍ͷޠΛબͿʢ໧Εͳ͍ʣ ɻ ʹҰݟʮ٧·Βͳ͍ʯɻߦ͖ࢭ·Γ͕ܾͯ͠ͳ͍ɻ ͕ͩɺͦΕ͸൓సݪཧʹࢧ͑ΒΕ͍ͯͳ͍ ࿦ཧͷʮ٧·Βͳ͍ʯ͸ࠜڌʢ؆໿نଇʣ͕͋Δ͔ΒਐΊΔɻ LLM ͷʮ٧·Βͳ͍ʯ͸ࠜڌ͕ͳͯ͘΋ਐΊΔʹࠜڌͳ͖લਐɻ ιʔεࢀরဃ཭ʢsource-reference divergenceʣʹओு͕ιʔεʹࠜڌʹґڌ͠ͳ͍ɻ [Ji et al. 2023, ACM Comput. Surv.ɿ఻هจͷ 62%͕ιʔεඇґڌ৘ใΛؚΉʢWIKIBIOʣ] Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 20
21. ࠜڌΛऔΓग़ͤͳ͍ཧ༝ɿҙຯͷ෼ࢠߏ଄͕ͳ͍ LLM ͷҙຯ͸ʮ෦඼ʯʹ෼͔Ε͍ͯͳ͍ §2 Ͱݟͨͱ͓Γɺ࿦ཧͷࠜڌ͸෦඼ʢنଇʣͱͯ͠औΓग़ͤΔʢ෼ࢠߏ଄͕͋Δʣ ɻ ͕ͩ LLM ͷҙຯ͸෼ࢄදݱʹશମ࿦తŠŠޠͷҙຯ͕໢ͷશମʹ༹͚͍ͯͯɺ ෦඼ ʹ෼͔Ε͍ͯͳ͍ɻ ͔ͩΒʮ͜ͷओுͷࠜڌʯΛ෦඼ͱͯ͠औΓग़ͤͳ͍ʹ෼ࢠߏ଄͕ͳ͍ɻ ͦ΋ͦ΋ࠜڌ͕ଘࡏ͠ͳ͍৔߹͢Β͋Δ ֶशσʔλʹҰ౓͔͠ݱΕͳ͍ࣄ࣮ʢsingleton ʗ arbitrary factsʣ͸ɺࠜڌ͕౷ܭతʹཱ֬ ͠ͳ͍ɻ[Kalai & Vempala 2024, STOC] Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 21
22. ہॴͷ਍அɿޠኮࡏݿͰݴ͍௚͢ ہॴ͸ʮ٧·ΔʯͰഁΕΔͷͰ͸ͳ͍ɻ ʮ٧·Βͳ͞ͷِ૷ʯͰഁΕΔ ʹ߲͕ͳ͍ओுʗܕͷ͔ͭͳ͍߲ΛɺฏવͱલʹਐΊͯ͠·͏ ୈ 12 ճͰݟͨʮࢭ·Βͳ͍߲ʯͱରΛͳ͢ɺࠓ೔ͷҰͭ໨ͷ঱ঢ়ɻ ࣍͸ɺ͜Ε͕தؒ૚ͰͲ͏ݟ͑Δ͔ŠŠ͕͜͜ࠓ೔ͷ৺ଁɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 22
23. 5. ਍அᶄ தؒɿࢭ·Δ͕ূ໌Ͱͳ͍
24. தؒͷཁٻɿࢭ·ͬͯɺূ໌ͷࢿ֨͋Δ஋Λฦ͢ എࠎͷ͍·͜͜ɿࡾ૚ͷʪதؒʫΛ౰ͯĊŠ͕͜͜ࠓ೔ͷ৺ଁ தؒϋʔϞχʔʢ෮शɾୈ 12–13 ճʣ தؒʹࢭ·Δʢਖ਼نԽʣʹશҬੑʢtotalʣ ɻ શҬੑͱ͸ʮඞͣ౴͑Δʯ͜ͱͰ͸ͳ͍ɻ ࢭ·ͬͯɺূ໌ͷࢿ֨͋Δ஋Λฦ͢͜ͱʢBHKʣ ɻ ࢭ·Βͳ͍ܭࢉʢΩɾρʣ͸ূ໌ͷࢿ֨Λ࣋ͨͳ͍ɻ ໰͍ɿLLM ͷग़ྗ͸ʮࢭ·ͬͯূ໌Λฦ͢ʯͷ͔ɺ ʮࠩ͠ग़ͨ͠;Γʯ͔ʁ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 24
25. LLM ͷഁΕɿࠩ͠ग़ͨ͠;Γʢ໿ଋखܗͷෆཤߦʣ ࿦ཧɿࢭ·߲ͬͯΛฦ͢ [x : A] . . . . M : B λx.M : A → B →I ஋ λx.M ʹূ໌ͷࢿ֨͋Δ߲ɻ ܕʢ໋୊ʣʹ߲͕൐͍ͬͯΔɻ LLMɿओுΛࢭΊͯฦ͕͢. . . ʮA → B ͕੒ΓཱͭʯͱจࣈྻΛฦ͢ɻ ͕ͩɺͦΕΛࢧ͑Δ߲ʢূ໌ʣΛ࣋ͨͳ͍ɻ ʹܕ͚ͩݴͬͯɺ߲͕ͳ͍ɻ ࢭ·ͬͯ͸͍Δɻ͕ͩࠩ͠ग़ͨ͠ͷ͸ʮূ໌ʯͰͳ͘ʮূ໌ͷ;Γʯ ʹӬԕʹཤߦ͞Εͳ͍໿ଋखܗʢୈ 13 ճ຤ඌͷޠኮʣ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 25
26. Ͱ͸ LLM ͸Ͳ͏ʮࢭ·Δʯͷ͔ŠŠਖ਼نԽͰ͸ͳ͘ଧͪ੾Γ ໰͍ɿ౴͑ʹਖ਼نܗʹʮ౸ୡ͔ͨ͠Βʯࢭ·Δͷ͔ʁ ূ໌ͷਖ਼نԽ͸಺ࡏతʹࢭ·Δʢܕ͕͋Ε͹ਖ਼نԽఆཧɾୈ 12 ճʣ ɻLLM ͸ҧ͏ŠŠࢭ ·Δͷ͸ɺ֎͔Β՝͞Εͨଧͪ੾Γͷ૚ɻ 1. ༧ࢉɿ࠷େτʔΫϯɾਪ࿦τʔΫϯɾ࣌ؒɾܭࢉྔɻ ਚ͖ͨΒଧͪ੾Γɻ 2. ֶशͨ͠ EOS ܏޲ɿ௕͘ͳΔͱʮ·ͱΊͯ݁࿦Λ ॻ͖ऴ͑ΔʯΛ౷ܭతʹֶशɻ 3. ൓෮ɾҟৗݕ஌ʢج൫ଆʣ ɿಉޠ൓෮ɾλΠϜΞ΢τ Λػցతʹଧͪ੾Γʗ࠶ࢼߦɻ ӡ༻ʢٖࣅίʔυʣ while not finished: think() if budget exhausted: print best answer() ʹͲΕ΋ਖ਼نԽఆཧͷ಺ࡏతอূͰ͸ͳ͘֎ࡏతͳଧͪ੾ΓɻҙຯΛཧղͨ͠؂ಜऀͰ͸ͳ͍ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 26
27. ֩৺ɿਪ࿦ͷਖ਼نԽͱɺग़ྗͷਖ਼نԽ͸ʮ෼཭ʯ͍ͯ͠Δ ূ໌࿦ɿਪ࿦ͦͷ΋ͷ͕ਖ਼نԽ ԕճΓΛফ͠੾ͬͯਖ਼نܗʹଥ౰ͳূ໌ʹࢸΔɻ಺ࡏ తʹอূʢਖ਼نԽఆཧɾୈ 12 ճʣ ɻ LLMɿग़ྗʢจষʣ͚ͩਖ਼نԽ ಺෦ਪ࿦͸ਖ਼نԽ͠ͳ͍͜ͱ͕͋ΔʢԕճΓ͕૿͑ Δɾಉ͡มܗΛ൓෮ɾ୳ࡧࢬ͕໎૸ʣ ɻ͕ͩʮͨ͠ ͕ͬͯʯͰจষΛด͡Δͷ͸ֶशࡁΈͷߦಈɻ LLM ͸ূ໌Λਖ਼نԽͰ͖ͳͯ͘΋ɺ ূ໌ʮΒ͍͠จॻʯΛਖ਼نԽͯ͠ग़ͤΔ ୳ࡧঢ়ଶ S −→ quote(S)ʢ΋ͬͱ΋Β͍͠֎෦දݱʣŠŠ ਖ਼͍͠ূ໌ͷਖ਼نܗͱ͸ݶΒͳ͍ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 27
28. ਖ਼نԽ vs LLMŠŠͦͯ͠ʮ٧·ͬͨͱ͖΄Ͳಊʑʯ ূ໌࿦ʢਖ਼نԽʣ LLM ؐݩྻ͕ਖ਼نܗ΁౸ୡ ୳ࡧɾੜ੒͕౴͑΁౸ୡ ڧਖ਼نԽʹͲͷॱংͰ΋ඞͣਖ਼ن ܗʢܕ͕อূʣ ͦͷ಺ࡏతอূ͸ͳ͍ʢ༧ࢉʹΑΔ֎ࡏతͳଧͪ੾ΓͷΈʣ ਖ਼نܗʹଥ౰ͳূ໌ ʢূ໌ͷࢿ֨ʣ ʮ݁࿦Β͍͠จষʯʹਖ਼نܗͱ͸ݶΒͳ͍ ൃࢄ͢Δؐݩྻ ൓෮ɾಊʑ८Γɾ୳ࡧࢬͷ໎૸ ͔ͩΒʮ٧·ͬͨͱ͖΄Ͳɺ࠷ޙ͚ͩಊʑͱͨ͠ޡ౴ʯ ಺෦୳ࡧ͕ະ׬Ͱ΋ճ౴ϑΣʔζ΁ҠΔʢreasoning ܥ͸୳ࡧͱճ౴͕ผϑΣʔζʣ ɻʹʮ಺෦Ͱղܾͨ͠ʯͱ ʮճ౴ʹҠͬͨʯ͸ผɻ࠷ޙ͸୳ࡧ͔Β࡞จ΁ͷඇৗޱΛ࢖͏ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 28
29. ਍அᶄͷݟऔΓਤɿ ʮࢭ·Δʯ ɾ͕ɾ ʮূ໌Ͱͳ͍ʯ தؒͷഁΕ͸ɺೋޠͰͰ͖͍ͯΔ ਍அᶄʹʮࢭ·Δʯ ɾ͕ɾ ʮূ໌Ͱͳ͍ʯɻ͜ͷೋޠΛɺ͜͜·Ͱͱ͔͜͜ΒͰผʑʹݟΔɻ ͜͜·Ͱʹʮࢭ·Δʯ͸Ͳ͏ى͖͔ͨ ʢN1–N3ʣ ਖ਼نԽͰ͸ͳ͘֎͔Βͷଧͪ੾Γʴग़ྗ͚ͩͷਖ਼ن Խʹ͍ͭͰ΋ʮ΋ͬͱ΋Β͘͠ด͡Δʯ ʢquote(S)ʣ ɻ ͔͜͜Βʹʮূ໌Ͱͳ͍ʯ͸ͳ͔ͥ ࢭΊͯࠩ͠ग़ͨ͠౴͑͸ɺͳͥਖ਼͍͠ͱݶΒͣɾ֎Ε ͕ආ͚ΒΕͳ͍ͷ͔ɻࣝผͱੜ੒ → Kalai ͷԼք → ධՁ͕غݖΛേ͢Δ → ઇͩΔ·ɻ ܧ͗໨ɿ ʮ΋ͬͱ΋Β͘͠ด͡Δʯ͜ͱ͸ʮਖ਼͍͠ʯ͜ͱΛอূ͠ͳ͍ʢഎࠎʹۙࣅͷࠩ෼ʣ ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 29
30. ʮࣝผʯͱʮੜ੒ʯŠŠͳͥੜ੒ͷ΄͏͕೉͍͔͠ എࠎͷ͍·͜͜ɿ࣍εϥΠυͷෆ౳ࣜΛಡΉͨΊͷݴ༿Λઌʹ༻ҙ͢Δ ೋͭͷ՝୊Λ۠ผ͢Δ ࣝผʢIs-It-Validʣ ɿࠩ͠ग़͞Εͨจʮ͜Ε͸ਖ਼͍͠ʯΛ Yes/No Ͱ൑ఆ͢Δɻ ੜ੒ɿਖ਼͍͠จΛࣗ෼Ͱ࡞ͬͯࠩ͠ग़͢ɻ ྫɿ͋Δਓͷ஀ੜ೔ ࣝผɿ ʮ஀ੜ೔͸ 3/4 ͩʯͱݴΘΕͯɺਖ਼͍͔͠൑ఆ͢Δɻ ੜ੒ɿ஀ੜ೔Λࣗ෼Ͱ౰ͯͯݴ͏ɻ ֶशσʔλʹҰ౓͔͠ग़ͳ͍ࣄ࣮ʢsingletonʣͳΒɺਖ਼͍͠೔෇Λِ෺͔Βݟ෼͚ΒΕͳ͍ʹࣝผʹࣦഊ ͢Δɻ·ͯࣗ͠෼Ͱ౰ͯΔੜ੒͸΋ͬͱແཧɻ ԡ͑͞ΔҰ఺ɿੜ੒Ͱ͖Δ ⇒ ࣝผͰ͖Δʢ࡞ΕΔͳΒྑ͠ѱ͠͸෼͔Δ͸ͣʣ ɻ ରۮɿࣝผͰ͖ͳ͍ ⇒ ੜ੒Ͱ͖ͳ͍ɻੜ੒ͷ೉͠͞͸ࣝผͷ೉͠͞ʹԼ͔ΒറΒΕΔɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 30
31. ͳͥආ͚ΒΕͳ͍͔ɿݬ֮ ʹ ೋ஋෼ྨͷޡΓ ੜ੒ͷޡΓ͸ɺࣝผͷޡΓʹ຋༁Ͱ͖Δ લεϥΠυͷʮੜ੒ ⇒ ࣝผʯΛఆྔԽ͢Δͱʢ[Kalai, Nachum, Vempala, Zhang 2025, Nature/OpenAI]ʣ ɿ ʢੜ੒ͷޡΓ཰ʣ ≳ 2×ʢࣝผͷޡΓ཰ʣ ಡΈํɿ≳ ͸ʮগͳ͘ͱ΋͓͓Αͦʯ ɻੜ੒ͷޡΓ͸ࣝผͷޡΓͷ໿ 2 ഒΛԼճΕͳ͍ɻ͔ͩΒࣝผ͕ຊ࣭ తʹ೉͍͠ʢʹ arbitrary factsʣͱɺੜ੒ͷޡΓ΋ඞͣҾ͖ͣΒΕͯग़Δɻ ͳͥʮ2 ഒʯ͔ʗσʔλ͕ޡΓθϩͰ΋ Is-It-Valid ͸ʮຊ෺ͷਖ਼͍͠จʯͱʮ΋ͬͱ΋Βِ͍͠෺ʯΛ൒ʑʹࠞͥͨೋ୒ɻੜ੒ͷޡΓʹ͜ͷೋ୒ ͷऔΓҧ͑ʹ຋༁͞Εɺߏ੒্΄΅ 2 ഒͰޮ͘ɻ ʹͨͱֶ͑शσʔλ͕׬શʹਖ਼ͯ͘͠΋ɺ࠷খԽ͢Δ໨తؔ਺ͦͷ΋ͷ͕ޡΓΛੜΉɻ ʹʮࠩ͠ग़ͨ͠;Γʯ͸ઃܭϛεͰͳ͘ɺ΋ͬͱ΋Β͠͞Λ࠷దԽͨ݁͠Ռɻ ˞Լք͕ޮ͘ͷ͸ arbitrary factsʢਅֶِ͕शσʔλ͔Βܾ·Βͳ͍ࣄ࣮ʣ ɻମܥతࣄ࣮ɾݕࡧซ༻Ͱ͸ผɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 31
32. શҬੑͷ෗ഊɿධՁ͕ʮΘ͔Βͳ͍ʯΛേ͢Δ ࠷໬ਪఆ͸ʮ΋ͬͱ΋Β͠͞ʯΛ࠷దԽ͢Δʢː 1 ͷֶशͷؼ݁ʣ ː 1 Ͱݟͨͱ͓Γֶश্͕͛Δͷ͸໬౓ʹ΋ͬͱ΋Β͠͞Ͱɺࣄ࣮తෆ੔߹Λേ͠ͳ͍ɻ [Ji et al. 2023] ࠷దԽ͞ΕΔͷ͸ਅཧੑͰ͸ͳ͘΋ͬͱ΋Β͠͞ɻ 0-1 ࠾఺͕ɺ౰ͯͣͬΆ͏Λ࠷దղʹ͢Δ େ൒ͷϕϯνϚʔΫ͸ਖ਼ޡೋ஋Ͱɺغݖʢ ʮΘ͔Βͳ͍ʯ ʣʹಘ఺Λ༩͑ͳ͍ɻˠ IDK Α Γ౰ͯͣͬΆ͏͕࠷దɻ[Kalai et al. 2025] ʹʮඞͣ౴͑ΔʯِͷશҬੑʹڲਖ਼͞ΕΔʢغݖΛേ͢Δ epidemicʣ ɻ શҬੑͷຊٛ͸ʮ౴͑ΔͳΒূ໌Λ൐͏ʯ ɻLLM ͸ʮͱʹ͔͘౴͑ΔʯʹତͪΔɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 32
33. ࢭ·ͬͯ΋ਖ਼نܗͰͳ͍ɿઇͩΔ·ࣜͷޡΓ Hallucination Snowballing [Zhang et al. 2023, ACL] ࣗݾճؼσίʔσΟϯά͸ࠨˠӈͷ֬ఆΛఫճͰ͖ͳ͍ɻॳظͷ 1 ޡτʔΫϯ͕Ҏ߱શମ ͷೖྗͱͳΓɺҰ؏ੑҡ࣋ͷͨΊޡΓΛੵΉɻ ୯ಠͳΒݟൈ͚ΔͷʹɺྲྀΕͷதͰ͸ॏͶΔ GPT-4 ͸ޡओுΛ୯ಠఏࣔͳΒ 87%ݟൈ͚Δͷʹɺ௚લͷࣗ෼ͷग़ྗͱͷ੔߹Λ༏ઌ͠ ͯޡΓΛॏͶΔɻ ʹܭࢉ͸ࢭ·Δ͕ɺͦΕ͸ਖ਼نԽʢ੔͑ͯূ໌ʹ͢ΔܭࢉʣͰ͸ͳ͘๫૸ɻ ୈ 12 ճͷ Ωʢࢭ·Βͳ͍߲ʣͷྡʹ͋Δ঱ঢ়ɿࢭ·͕ͬͨਖ਼نܗͰͳ͍߲ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 33
34. ΋͏Ұஈɿ ʮ΋ͬͱ΋Β͠͞ʯ͸ɺਅཧʹ͸ಧ͔ͳ͍ ֶश͕࠷దԽ͢Δͷ͸ʮ෼෍ʯͰ͋ͬͯʮਅཧʯͰ͸ͳ͍ ࣄલֶश͸ਓؒͷςΩετͷ෼෍Λ࠶ݱ͢Δ͜ͱʢʹີ౓ਪఆʣ ɻ ໨ඪ͸ʮਓ͕ؒॻ͖ͦ͏ͳจʯͰ͋ͬͯɺ ʮਖ਼͍͠จʯͰ͸ͳ͍ɻ ͔ͩΒʮੈؒͷΑ͋͘Δޡղʯ΋஧࣮ʹ࠶ݱ͢Δ ΋ͬͱ΋Β͍͠ޡΓʢimitative falsehoodʣ·ͰਅࣅΔɻ ͦΕΛଌΔͷ͕ TruthfulQA [Lin et al. 2022]ŠŠਓؒͷࢥ͍ࠐΈΛͲΕ͚ͩ໛฿͢Δ͔ɻ ई౓͕ҧ͏ɿશମ࿦తͳʮ΋ͬͱ΋Β͠͞ʯ vs ෼ࢠతͳʮূ໌ʯ ΋ͬͱ΋Β͠͞ʹίʔύεશମ΁ͷద߹ʢશମ࿦తʣʗূ໌ͷࢿ֨ʹ෦඼͔ΒہॴنଇͰ૊ΊΔʢ෼ࢠతʣ ɻ ผͷ෺͔ࠩͩ͠Βɺ΋ͬͱ΋Β͠͞Λ͍͘Β্͛ͯ΋ূ໌ʹ͸ಧ͔ͳ͍ɻŠŠଛࣦθϩͷʮ׬ᘳͳʯ LM Ͱ ΋ݬ֮͸ফ͑ͳ͍ʢୈ 2 ճʮS1 ʹ΋ͬͱ΋Β͠͞ͷ໛฿ʯͷਖ਼ମʣ ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 34
35. 6. ਍அᶅ େҬɿੵΈ্͕Βͳ͍
36. େҬͷཁٻɿ૿ஙͯ͠΋طଘ͕յΕͳ͍ʢੵΈ্͕Δʣ എࠎͷ͍·͜͜ɿࡾ૚ͷʪେҬʫΛ౰ͯΔ େҬϋʔϞχʔʢ෮शɾୈ 13 ճː 3ɾୈ 2 ճԾ૝ܭࢉػʣ େҬʹอଘ֦େʹϞδϡϥϦςΟɿ৽͍݁͠߹ࢠΛ଍ͯ͠΋طଘͷਪ࿦͸ෆมɻ ͔ͩΒଞਓͷิ୊Λ҆৺ͯ͠आΓΒΕɺ஌͕ࣝੈ୅Λ௒͑ͯੵΈ্͕Δɻ ഁΕʹരൃʢtonkɾୈ 11 ճʣʹ૿ங͕طଘΛԚછ͢Δɻ ໰͍ɿLLM ͸ʮֶͿͨͼʹյΕͣʹੵΈ্͕Δʯͷ͔ʁ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 36
37. LLM ͷഁΕɿ࠶ֶशͰطଘ͕յΕʮ͏Δʯ ʢഁ໓త๨٫ʣ ഁ໓త๨٫ʢcatastrophic forgettingʣͱ͸ χϡʔϥϧωοτΛ৽͍͠λεΫʹஞֶ࣍शͤ͞Δͱɺطशͷੑೳ͕ٸམ͠͏Δݱ৅ɻ෼ ࢄදݱ͕ॏͳΓ߹͏ͨΊʹى͖Δʢ҆ఆੑʖՄ઼ੑͷδϨϯϚʣ ɻ จݙʢྫͱͯ͠ʣ ൃݟ [McCloskey & Cohen 1989 ʗ Ratcliff 1990] ʗ૯આ [French 1999, TICS] ʗޯ഑ϕʔε໢Ͱ࣮ূ [Goodfellow et al. 2013]ɻ LLM Ͱ΋ܧଓతϑΝΠϯνϡʔχϯάͰ؍ଌ [Luo et al. 2023, 1B–7B]ʢ͜ͷൣғͰ͸ن໛େ΄Ͳਂࠁʣ ɻ ʴ௥ैʢsycophancyʣ ࿩ऀɾલఏΛ૿ங͢Δͱ݁࿦͕ಈ͘ʹඇอଘతɾෆ҆ఆ [Sharma et al. 2023, Anthropic]ɻŠŠͲͪΒ΋ʮ ଍ ͢ͱطଘ͕յΕΔʗಈ͘ʯํ޲ͷ঱ঢ়ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 37
38. ਖ਼͘͠͸ʮඞͣյΕΔʯͰ͸ͳ͘ʮյΕͳ͍อূ͕ͳ͍ʯ ๨٫͸ෆՄආͰ͸ͳ͍ŠŠ؇࿨Ͱ͖Δ ͔ͭͯʮഁ໓త๨٫͸ෆՄආͱ޿͘৴͡ΒΕ͍ͯͨʯ͕ɺͦΕ͸෴ͬͨɻ[Kirkpatrick et al. 2017, PNAS] EWCʢॏཁͳॏΈͷֶशΛ཈͑Δʣ ɾϦϓϨΠɾύϥϝʔλִ཭ɾҰൠࢦࣔσʔλͷࠞ߹ ͳͲͰେ͖͘؇࿨Ͱ͖Δɻ ͕ͩɺͲͷख๏΋ʮඇ๨٫ʯΛอূ͠ͳ͍ EWC ͸ιϑτͳਖ਼ଇԽͰɺλεΫྻ͕௕͍ͱྼԽ͢Δ [Kemker et al. 2018]ɻޮՌ͸ن໛ɾλ εΫ਺ͳͲͷ৚݅࣍ୈɻܧଓֶशͱ͸ɺ๨٫Λௐ੔͢ΔӦΈͰ͋ͬͯɺճආ͕ࣗಈอূ͞ ΕΔ࢓૊ΈͰ͸ͳ͍ [De Lange 2021 ʗ Wang 2024]ɻ ࿦ཧ͸อଘ֦େΛ಺ࡏతʹอূ͢Δʢ૿ஙͯ͠΋طଘ͸յΕͳ͍ʣ ɻ LLM ࠶ֶशʹ͸ͦͷอূ͕ͳ͍ŠŠ͔ͩΒʮ ੵΈ্͕Δʯͱݴ͍੾Εͳ͍ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 38
39. ਍அͷ·ͱΊɿࡾͭͷഁΕ͸ɺ͕ࠜҰͭ େҬͷഁΕɿੵΈ্͕Δอূ͕ͳ͍ ഁ໓త๨٫ʹ࠶ֶशͰطଘ͕յΕ͏Δ தؒͷഁΕɿࢭ·Δ͕ূ໌Ͱͳ͍ ࠩ͠ग़ͨ͠;Γʹ໿ଋखܗͷෆཤߦ ہॴͷഁΕɿ൓సݪཧͳ͖લਐ ٧·Βͳ͞ͷِ૷ʹࠜڌͳ࣍͘ͷޠΛग़͢ ࠜ͸ͻͱͭɿҙຯͷ෼ࢠߏ଄ͷܽ೗ ෼ࢠ͕ͳ͍ͱ. . . ʢہॴʣࠜڌʹ݁΂ͳ͍ ʢதؒʣ஋ΛݕূͰ͖ͳ͍ ʢେҬʣ෦඼Λ଍ͤͳ͍ ʹʮ༗ݶੑͷࡾͭͷإʯ ʢS3ʣͷ ͪΐ͏Ͳཪฦ͠ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 39
40. 7. ॲํɿࡾ૚ͷཁٻΛճ෮͢Δ
41. ॲํͷಡΈํɿͲͷ૚ͷཁٻΛิڧ͢Δ͔ എࠎͷ͍·͜͜ɿഁΕͨ૚ΛɺͲ͏ิڧ͢Δ͔ ରࡦΛࡾ૚ͰಡΈସ͑Δ ఏҊ͞Ε͍ͯΔରࡦ͸ଟ͍͕ɺຊतۀͷ෺ࠩ͠ͰݟΔͱ ʮͲͷ૚ͷཁٻΛิڧ͢Δ͔ʯͰ੔ཧͰ͖ΔɻҰͭͷख๏͕ෳ਺૚ʹޮ͘ɻ ˞ ݸʑͷख๏ͷ༗ޮੑ͸ݚڀ్্ʢஅఆ͸͠ͳ͍ʣ ɻ͜͜Ͱ͸Ґஔ͚ͮΛࣔ͢ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 41
42. ہॴΛิڧɿओுΛࠜڌʹ෼ࢠʹ݁ͼ͚ͭΔ ൓సݪཧΛʮ֎஫ʯ͢Δ RAG ʗݕࡧ֦ுɿ֎෦ιʔεʹओுΛ݁ͼ͚ͭΔʹιʔεࢀরဃ཭ΛຒΊΔɻ FActScore[Min et al. 2023]ɿओுΛ atomic fact ʹׂ֤ͬͯʑݕূʹ෼ࢠʹ෼ղɻ ه߸తཪ͚ͮɾ֎෦πʔϧɾܗࣜݕূɿओுʹܕΛ͚ͭΔʢہॴʴதؒʣ ɻ ͍ͣΕ΋ʮࠜڌͳ͖લਐʯΛʮࠜڌʹ݁ͼ͍ͭͨલਐʯ΁໭ͦ͏ͱ͢ΔࢼΈɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 42
43. தؒΛิڧɿશҬੑΛਖ਼௚ʹ͢Δʢغݖʣ ཤߦͰ͖ͳ͍໿ଋखܗΛɺ੾Βͳ͍ غݖʢabstentionʣ ɾ ʮΘ͔Βͳ͍ʯ܇࿅ɾෆ࣮֬ੑਪఆɿ ஋Λฦͤͳ͍ͱ͖͸ฦ͞ͳ͍ʹʮূ໌ͷࢿ͕֨ͳ͍ͳΒࠩ͠ग़͞ͳ͍ʯ ʢBHKʣ ɻ ධՁվֵɿغݖʹಘ఺Λ༩͑ΔʢKalai et al. ͷॲํʣʹதؒͷཁٻΛධՁʹ૊ΈࠐΉɻ ࢭ·ͬͨ஋͕ຊ෺͔ɺ͔֬ΊΔ self-consistency ʗ SelfCheckGPT[Manakul et al. 2023]ɿෳ਺ੜ੒ͷҰ؏ੑͰࢿ֨Λۙ ࣅݕূɻ DoLa[Chuang et al. 2023]ɿࣄ࣮ੑํ޲ʹσίʔσΟϯάΛิਖ਼ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 43
44. େҬΛิڧʴϝλɿ໰୊ΛʮԼքͷޮ͔ͳ͍ྖҬʯ΁Ҡ͢ ੵΈ্͕ΔΛճ෮͢Δ RAGʢ࠶ܝʣɿ஌ࣝΛ֎෦Խʹߋ৽͕طଘΛյ͞ͳ͍ʢඇഁյత૿ஙʣʹೋ౛ɻ ܧଓֶशɾ๨٫ରࡦʗπʔϧҕৡɿମܥతࣄ࣮͸ه߸ܥʹ೚ͤΔɻ ϝλɿෆՄආੑ͸ʮݶఆ͖ͭʯ͔ͩΒͦ͜ॲํ͕ޮ͘ ౷ܭతෆՄආੑͷԼք͸ arbitrary facts ݶఆʢRAGɾࢉज़ʹ͸ඇద༻ʣ ɻ ରࡦͱ͸ɺ໰୊ΛԼքͷޮ͔ͳ͍ྖҬ΁Ҡ͢ʹࡾ૚ͷཁٻΛ֎ૠ͢ΔӦΈɻ ධՁث TruthfulQA[Lin et al. 2022]ɾHaluEval[Li et al. 2023]ɾFActScore ʹͲͷ૚͕ഁΕ͔ͨଌΔ਍அثɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 44
45. ࢀߟɿϋϧγωʔγϣϯͷඪ४తͳ෼ྨʢαʔϕΠΑΓʣ ֶք͸ʮԿͱ৯͍ҧ͏͔ʯͰ੾Δ ෼ྨ࣠ʢ୅දαʔϕΠʣ ԼҐܕ ࣄ࣮ੑʗ஧࣮ੑ [Huang et al. 2023/25] ࣄ࣮ੑʹໃ६ɾ፻଄ʗ஧࣮ੑʹࢦࣔɾจ຺ɾ࿦ཧͷෆ੔߹ ಺ࡏʗ֎ࡏ [Ji et al. 2023] ಺ࡏʹιʔεͱໃ६ʗ֎ࡏʹιʔεͰݕূෆೳ ࡾ छ ͷ ি ಥ [Zhang 2023, Siren’s Song] ೖྗিಥʗจ຺িಥʗੈք஌ࣝিಥ ຊतۀͷࡾ૚͸ɺ͜Εͱʮ௚ަʯ͢Δ ඪ४෼ྨʹԿͱ৯͍ҧ͏͔ɻຊतۀʹͲͷܭࢉత৚͕݅յΕ͔ͨʢہॴɾதؒɾେҬʣ ɻྫɿ࿦ཧతෆ੔߹˺ ہॴ/தؒʢূ໌ͱͯ͠૊Ίͳ͍ʣʗ፻଄ (fabrication) ˺தؒʢࢿ֨ͳࠩ͘͠ग़͢ʣʗจ຺ෆ੔߹˺ہॴʢೖ ྗʹࠜڌ΁ͷ஧࣮ͷഁΕʣ ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 45
46. ࢀߟɿରࡦΧλϩάΛࡾ૚ʹ഑ΔʢαʔϕΠΑΓɾྫʣ ૚ʢຊतۀʣ ୅දతͳରࡦख๏ʢҰ࣍จݙʣ ہॴʢࠜڌʹ݁Ϳʣ RAG [Lewis 2021] ʗ FActScore [Min 2023] ʗه߸తݕূɾ֎෦πʔϧ தؒʢશҬੑɾغݖʣ R-Tuning ʹغݖ [Zhang 2023] ʗ SelfCheckGPT [Manakul 2023] ʗ DoLa [Chuang 2023] ʗ CoVe [Dhuliawala 2023] ʗ ITI [Li 2023] େҬʢඇഁյߋ৽ʣ RAGʢ஌ࣝͷ֎෦Խʣʗܧଓֶशɾ๨٫ରࡦʗπʔϧҕৡ ؇࿨αʔϕΠɿTonmoy et al. 2024ʢ32 ख๏௒ɾϓϩϯϓτ޻ֶʗϞσϧ։ൃʣ[2401.01313]ɺJi et al. 2023 ʢσʔλؔ࿈ʗϞσϧɾਪ࿦ʣ ɻ ˞ख๏ͷ༗ޮੑ͸ݚڀ్্ɻ͜͜Ͱ͸ʮͲͷ૚ͷཁٻΛิڧ͢Δ͔ʯͱ͍͏Ґஔ͚ͮͷྫΛࣔ͢ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 46
47. 8. Φνɿݬ֮͸ޡশŠŠܭࢉ͸ূ໌Ͱ͸ͳ͍
48. ʮݬ֮ʯ͸ޡশŠŠΧϦʔɾϋϫʔυͰଊ͑௚͢ ͦ΋ͦ΋ʮݬ֮ʯͱ͍͏໊લ͕ѱ͍ LLM ͸ײ֮Λ࣋ͨͳ͍ɻޡΓ͸ܹࢗͷෆࡏͰͳ͘σʔλͱ໨తؔ਺ʹ༝དྷ͢Δɻ[Maleki et al. 2024, IEEE CAI][Ostergaard & Nielbo 2023] ຊतۀͷݴ͍׵͑ ʮݬ֮ʯͰ͸ͳ͘ɺ ʮূ໌ͷࢿ֨ͷͳ͍߲ΛɺશҬੑΛ૷ͬͯࠩ͠ग़͢ߏ଄తੑ࣭ʯɻ ໊લΛΧϦʔɾϋϫʔυͷޠኮʹม͑ΔͱɺॲํʢͲͷ૚Λิڧ͢Δ͔ʣ͕ݟ͑Δɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 48
49. લظͷҰߦɿܭࢉ͸ɺূ໌Ͱ͸ͳ͍ ʮܭࢉ͍ͯ͠Δʯ͜ͱ͸ɺ ʮূ໌͍ͯ͠Δʯ͜ͱΛؚҙ͠ͳ͍ ࡾ૚ͷ෺ࠩ͠ͰଌΔͱ LLM ͸૸ΔʢΉ͠Ζ૸Γ͗͢Δʹ٧·Βͳ͞ͷِ૷ʣ ɻ ͕ͩࢭ·ͬͯূ໌Λࠩ͠ग़͢͜ͱʢશҬੑʣ΋ɺ յΕͣʹੵΈ্͕Δ͜ͱʢϞδϡϥϦςΟʣ΋ɺอূ͠ͳ͍ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 49
50. ͳͥਓؒ͸ʮγεςϜ2ʯΛҭͯͨͷ͔ʢୈ2ճͷճऩʣ γεςϜ 1 ʹ LLM ͕໛฿͍ͯ͠Δଆ ଎͍ɾ௚ײɾࣗಈɾੜಘతʢϕΠζతʣ ɻLLM ʹ ͜ͷʮγεςϜ 1ʯͷ໛฿ʹ΋ͬͱ΋Β ͠͞ͷ໛฿ʢୈ 2 ճʣ ɻŠŠ͕ͩ γεςϜ 1 Ͱ͸ূ໌͸Ͱ͖ͳ͍ɻ ͔ͩΒਓؒ͸ʮγεςϜ 2ʯΛҭͯͨ γεςϜ 2 ʹ ೴ͷ্ͷԾ૝ܭࢉػʢσωοτʣʹݴޠͱ͍͏ OSʢΫϥʔΫʣ͕ಈ͔͢૚ɻ ୈ 2 ճͷҰ൪େࣄͳࣄ࣮ Ծ૝ܭࢉػ͸࠷ॳ͔Β೴಺ʹͳ͔ͬͨɻਓ͸੠ʹग़ͯ͠ਪ࿦ͨ͠ʹܭࢉ͸֎෦ʹ͋Γɺ΍͕ͯ಺Խͨ͠ɻγ εςϜ 2 ͷىݯ͸ࣾձతɾݴޠత࣮ફŠŠ͔ͩΒ ϓϩάϥϚϒϧʢֶ΂Δɾ௚ͤΔʣ ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 50
51. ΦνɿLLM ͷதʹʮγεςϜ2ʯˊŠͦΕ͕࿦ཧ ଍Γͳ͍ͷ͸ɺγεςϜ 2 LLM ʹ γεςϜ 1ʢ΋ͬͱ΋Β͠͞ʣ ɻ͚͍ܽͯΔͷ͸γεςϜ 2 ʹ ূ໌Λ૸ΒͤΔԾ ૝ܭࢉػʹࠓ೔ͷࡾ૚ϋʔϞχʔʢہॴͷنଇɾఀࢭɾϞδϡϥϦςΟʣ ɻ ࠓ೔ͷʮॲํʯ͸ɺͦͷޙ෇͚ͩͬͨ ֎෦πʔϧɾه߸తݕূɾCoVeɾRAGɾغݖŠŠͲΕ΋ʮ γεςϜ 1 ͷ্ʹɺγεςϜ 2 Λޙ෇͚͢ΔʯࢼΈɻ ͔ͩΒɺ࿦ཧ͕ཁΔ ࿦ཧͦ͜ɺͦͷʮγεςϜ 2ʯͷ࢓༷ॻɻ ଎౓ͷ୅ঈɿγεςϜ 1 ͸଎͍҆͘ʗγεςϜ 2 ͸஗͘ߴՁʢҰา͔ͣͭ֬ΊΔʣ ɻ͕ͩূ໌ͷࢿ֨͸ɺͦ ͜ʹ͔͠ͳ͍ɻ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 51
52. ࿦ཧͱ͸ɺ૸ͬͯɾࢭ·ͬͯɾੵΈ্͕ΔӦΈɻ ͦͷࡾ૚ͷ෺ࠩ͠͸ɺAI ΛଌΔ෺ࠩ͠Ͱ΋͋Γ·ͨ͠ɻ ૸ͬͯɺࢭ·ͬͯɺੵΈ্͕Δɻ ൒೥ؒɺ͓͔ͭΕ͞·Ͱͨ͠ɻ
53. 9. ԋशʢ࠷ऴճʣ
54. ԋशʢ࠷ऴճɾܰΊʣ ԋशʢղ౴ʹ͔͔ͬͨ࣌ؒ΋ॻ͘͜ͱʣ ͋ͳ͕ͨ࠷ۙݟ͔͚ͨʢ͋Δ͍͸ࣗ෼Ͱ LLM ʹग़ͤͨ͞ʣϋϧγωʔγϣϯͷ࣮ྫΛҰ ͭڍ͛ɺͦΕ͕ࡾ૚ͷͲΕͷഁΕ͔ŠŠ ہॴʢࠜڌͳ͖લਐʣʗதؒʢূ໌ͷࢿ֨ͳࠩ͘͠ग़͢ʣʗେҬʢطଘΛյ͢ɾ௥ैʣ ŠŠͷͲΕʹ౰ͨΔ͔Λ਍அ͠ɺ ҰͭॲํΛࡾ૚ͷݴ༿ͰఏҊͤΑɻ ώϯτ ʮ΋ͬͱ΋Β͍͕ࠜ͠ڌΛग़ͤͳ͍ʯʹہॴɻ ʮࣗ৴ຬʑʹஅఆ͢Δ͕ཪ͚͕ͮͳ͍ʯʹதؒʢશҬੑͷِ૷ʣ ɻ ʮલͷ࿩ͱໃ६͢Δʗࢦఠʹܴ߹ͯ͠ҙݟ͕มΘΔʯʹେҬɻ ʢఏग़ํ๏ɾక੾͸ KULMS ͷࢦࣔʹै͍ͬͯͩ͘͞ɻ ʣ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 54
55. 10. ࢀরจݙ
56. ࢀরจݙʢ1/4ʣҙຯͷཧ࿦ Wittgenstein, L. (1953). Philosophical Investigations. Blackwell.ʢː 43ʮҙຯͱ͸࢖༻ʯ ʣ Firth, J. R. (1957). “A synopsis of linguistic theory, 1930–1955.” Studies in Linguistic Analysis, 1–32. Blackwell. Harris, Z. S. (1954). “Distributional structure.” Word 10(2–3): 146–162. Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). “Efficient Estimation of Word Representations in Vector Space.” arXiv:1301.3781.ʢword2vecʣ Dummett, M. (1973). Frege: Philosophy of Language. Duckworth. Dummett, M. (1991). The Logical Basis of Metaphysics. Harvard Univ. Press. Prawitz, D. (1965). Natural Deduction: A Proof-Theoretical Study. Almqvist & Wiksell.ʢਖ਼نԽɾূ໌࿦తҙຯ࿦ʣ Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 56
57. ࢀরจݙʢ2/4ʣϋϧγωʔγϣϯɿ෼ྨɾݪҼɾෆՄආੑ Ji, Z., et al. (2023). “Survey of Hallucination in Natural Language Generation.” ACM Computing Surveys 55(12): 1–38. Huang, L., et al. (2023/2025). “A Survey on Hallucination in Large Language Models.” ACM Trans. Information Systems. arXiv:2311.05232. Zhang, Y., et al. (2023). “Siren’s Song in the AI Ocean: A Survey on Hallucination in LLMs.” arXiv:2309.01219. Zhang, M., et al. (2023). “How Language Model Hallucinations Can Snowball.” arXiv:2305.13534ʢICML 2024ʣ. Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025). “Why Language Models Hallucinate.” arXiv:2509.04664ʢOpenAIʣ. Kalai, A. T., & Vempala, S. S. (2024). “Calibrated Language Models Must Hallucinate.” STOC 2024. arXiv:2311.14648. Sharma, M., et al. (2023). “Towards Understanding Sycophancy in Language Models.” Anthropic. arXiv:2310.13548. Maleki, N., et al. (2024). “AI Hallucinations: A Misnomer Worth Clarifying.” IEEE CAI. arXiv:2401.06796. Østergaard, S. D., & Nielbo, K. L. (2023). “False Responses From AI Models Are Not Hallucinations.” Schizophrenia Bulletin 49(5). Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 57
58. ࢀরจݙʢ3/4ʣରࡦɾධՁϕϯνϚʔΫ Lewis, P., et al. (2020). “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS 2020. arXiv:2005.11401. ʢRAGʣ Chuang, Y.-S., et al. (2023). “DoLa: Decoding by Contrasting Layers.” arXiv:2309.03883ʢICLR 2024ʣ. Dhuliawala, S., et al. (2023). “Chain-of-Verification Reduces Hallucination in LLMs.” arXiv:2309.11495.ʢCoVeʣ Manakul, P., Liusie, A., & Gales, M. (2023). “SelfCheckGPT.” EMNLP 2023. arXiv:2303.08896. Zhang, H., et al. (2023). “R-Tuning: Instructing LLMs to Say ‘I Don’t Know’.” arXiv:2311.09677ʢNAACL 2024ʣ. Li, K., et al. (2023). “Inference-Time Intervention.” NeurIPS 2023. arXiv:2306.03341.ʢITIʣ Tonmoy, S. M. T. I., et al. (2024). “A Comprehensive Survey of Hallucination Mitigation Techniques in LLMs.” arXiv:2401.01313. Lin, S., Hilton, J., & Evans, O. (2022). “TruthfulQA.” ACL 2022. arXiv:2109.07958. Li, J., et al. (2023). “HaluEval.” EMNLP 2023. arXiv:2305.11747. Min, S., et al. (2023). “FActScore.” EMNLP 2023. arXiv:2305.14251. Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 58
59. ࢀরจݙʢ4/4ʣഁ໓త๨٫ɾܧଓֶश McCloskey, M., & Cohen, N. J. (1989). “Catastrophic Interference in Connectionist Networks.” Psychology of Learning and Motivation 24: 109–165. Ratcliff, R. (1990). “Connectionist Models of Recognition Memory.” Psychological Review 97(2): 285–308. French, R. M. (1999). “Catastrophic Forgetting in Connectionist Networks.” Trends in Cognitive Sciences 3(4): 128–135. Goodfellow, I. J., et al. (2013). “An Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks.” arXiv:1312.6211ʢICLR 2014ʣ. Kirkpatrick, J., et al. (2017). “Overcoming Catastrophic Forgetting in Neural Networks.” PNAS 114(13): 3521–3526.ʢEWCʣ Kemker, R., et al. (2018). “Measuring Catastrophic Forgetting in Neural Networks.” AAAI 2018. arXiv:1708.02072. Luo, Y., et al. (2023). “An Empirical Study of Catastrophic Forgetting in LLMs During Continual Fine-tuning.” arXiv:2308.08747 ʢIEEE/ACM TASLP 2025ʣ. De Lange, M., et al. (2021). “A Continual Learning Survey.” IEEE TPAMI 44(7). arXiv:1909.08383. Wang, L., et al. (2024). “A Comprehensive Survey of Continual Learning.” IEEE TPAMI. arXiv:2302.00487. Պֶ఩ֶՊֶ࢙ (ԋश) ࿦ཧֶʢલظʣ ୲౰ɿ໼ా෦ढ़հ 59