- Ukugqugquzelwa kwe-Chain of Thought kuthuthukisa ukucabanga kwe-LLM ngokwenza izinyathelo eziphakathi zicace esikhundleni sokuphoqa izimpendulo ezisheshayo.
- Izinhlobo ezifana ne-zero-shot, i-few-shot, i-Auto-CoT, i-self-consistency kanye ne-Tree-of-Thoughts zishintsha ukunemba, izindleko kanye nomzamo wokusebenzisa.
- I-CoT inamandla kakhulu ezinhlelweni ezisebenzisa amathuluzi ngendlela ehlelekile lapho ukucabanga okusobala kuthuthukisa ukuthembeka kanye nokulungiswa kwamaphutha.
- Ukusetshenziswa kokukhiqiza kwe-CoT kudinga ukubonwa, ukuhlolwa kanye nokwenza ngcono okusheshayo ukuze kulinganiswe ikhwalithi ngokumelene nokubambezeleka kanye nezindleko zamathokheni.
Uchungechunge Lokucabanga Okushukumisayo (i-CoT) luye lwasuka ekubeni yilukuluku lokucwaninga lwaba ngelinye lamathuluzi awusizo kakhulu onjiniyela abanawo ukuze benze amamodeli amakhulu olimi acabange ngempela, esikhundleni sokuqagela igama elilandelayo elingenzeka kakhulu. Ngokucela ngokusobala imodeli ukuthi ichaze izinyathelo zayo eziphakathi, uvula ukusebenza okungcono kakhulu emisebenzini yezibalo, i-logic kanye nokwenza izinqumo, ngenkathi uthola nomzila ocacile ongawulungisa futhi uwuhlole.
Uma wakha izinhlelo zokusebenza ezisebenzisa i-LLM, amanxusa noma abashayeli bezindiza abangontanga futhi usaqhubeka nokukhipha izixwayiso zesinyathelo esisodwa kuphela, ushiya ikhwalithi eningi etafuleni. Kulo mhlahlandlela ogxile kubathuthukisi sizochaza ukuthi iyini i-Chain of Thought, ukuthi kungani isebenza, izinhlobo eziyinhloko (i-zero-shot, i-few-shot, i-Auto-CoT, i-self-consistency, i-Tree-of-Thoughts, i-least-to-most, i-multimodal), ukuthi iqhathaniswa kanjani ne-prompt chaining, kanye nendlela yokuyihlanganisa nokuyiqapha ezinhlelweni zangempela kusetshenziswa amathuluzi esimanje.
Kusukela ekuphenduleni okuqondile kuya ekucabangeni okucacile
Iningi leziphakamiso abantu abazithumela ku-LLM “ziwukuzama okukodwa”: uma ubuza umbuzo, imodeli ikhipha impendulo, akukho mibuzo ebuziwe, akukho sizathu esiboniswayo. Ngento efana nokuthi “Isibhakabhaka sinombala onjani?” kulungile lokho: imodeli imane ibuyisele ukuthi “Isibhakabhaka siluhlaza okwesibhakabhaka.” Akukho sakhiwo esibonakalayo, akukho logic ephakathi, umusho wokugcina nje ozwakala ulungile.
Isikhuthazo Somcabango siguqula leli phethini ngokutshela imodeli ukuthi ilandise ngempela izinyathelo zokucabanga ezilandelayo. Buza ukuthi “Kungani isibhakabhaka sibukeka siluhlaza okwesibhakabhaka? Cabanga isinyathelo ngesinyathelo.” bese imodeli ingase ichaze umqondo “oluhlaza okwesibhakabhaka”, ikhulume ngendlela ukukhanya kwelanga okuxhumana ngayo nomoya, ikhulume ngokusabalala kweRayleigh, bese isho kuphela ukuthi amaza amafushane aluhlaza okwesibhakabhaka asakazeke kuzo zonke izinhlangothi, ngakho isibhakabhaka sibonakala siluhlaza okwesibhakabhaka kithi.
Ngobuchwepheshe, awushintshi isisindo semodeli noma uyinike ulwazi olusha; ushintsha ifomethi yokubala oyicelayo ukuba ikwenze. Esikhundleni sokucindezela ukuhlaziya, ukucabanga, ukubala kanye nokuphendula kube yiphasi elilodwa eliya phambili, uvumela ukuthi kusakaze uchungechunge lwemicabango ephakathi eyakhela esiphethweni.
Empeleni, lokhu kungaba lula njengokwengeza umyalelo onjengokuthi “bonisa ukucabanga kwakho isinyathelo ngesinyathelo” noma “masixazulule lokhu ngendlela ehlelekile” kuze kube sekupheleni kwesicelo sakho. Lokho kwengeza okuncane kukhuthaza imodeli ukuthi yembule uchungechunge lwezimo eziphakathi eziholela emphumeleni wokugcina, kunokuba igxumele ngqo empendulweni ezwakala sengathi inengqondo.
I-CoT yenza nokubona kube lula kakhulu. Uma imodeli inephutha, uvame ukukhomba isinyathelo esiqondile lapho ukucabanga kwayo kwaphuma khona, esikhundleni sokubheka inombolo engalungile eyimfihlakalo noma isinqumo esingalungile ngaphandle kwencazelo.
Igebe phakathi kokufanisa amaphethini nokucabanga kwangempela
Ama-LLM anekhono elingakholeki lokufanisa amaphethini ngoba empeleni ayimishini emikhulu yokucabanga eqeqeshwe ngobuningi obukhulu bombhalo. Buza, “Yini esindayo, iphawundi lezimpaphe noma iphawundi lomthofu?” futhi imodeli yesimanje ibone lo mbuzo onzima izikhathi eziningi; iphendula ngokuqiniseka ukuthi unesisindo esifanayo.
Kodwa uma ubuza umbuzo odinga imisebenzi eminingana exhunyiwe, ukusebenza kungancipha ngokushesha. Isibonelo esijwayelekile: “Uma kuthatha imishini emi-5 nemizuzu emi-5 ukwenza amawijethi ama-5, kungathatha isikhathi esingakanani imishini eyi-100 ukwenza amawijethi ayi-100?” Amamodeli amaningi azoyixazulula impendulo enembile kodwa engalungile ngaphandle kokuthi iqondiswe ngokucophelela.
Inkinga eyinhloko ngokuvamile akukhona ukuntuleka kolwazi kodwa ukuntuleka kwesakhiwo. Ukucabanga ngezinyathelo eziningi ngokungaqondile kudinga imodeli ukuthi ihlanganise imisebenzi eminingi ngokulandelana: ukuqonda umbhalo, ukuhlonza okubuzwayo, ukuhlanganisa ubudlelwano noma amafomula afanele, ukwenza izibalo nokubhala impendulo. Uma ufuna impendulo esheshayo, ucela ukuthi icindezele wonke lowo msele ube yi-shot eyodwa.
Uchungechunge Lokucabanga lunikeza imodeli “isikhala sokucabanga” ngokuguqula lolo chungechunge olungacacile lube umbhalo ocacile. Ucwaningo oluvela ku-Google nabanye lubonise ukuthi uma ucela amamodeli ukuthi “abonise umsebenzi wawo”, ukunemba kwezibalo, ukucabanga okunengqondo, kanye nemisebenzi yokuphatha ngokufanekisa kugxuma kakhulu uma kuqhathaniswa nokuphendula okuqondile.
Ukuhlolwa okukodwa okuhlaba umxhwele kakhulu: lapho abacwaningi bebuza imibuzo yezibalo yesikole samabanga aphansi se-GPT‑3, bathola ngaphansi kuka-20% wayo ngezimpendulo ezicacile. Lapho beshintsha nje isikhuthazo sokucela ukucabanga okuphakathi, ukunemba kwadlula ku-50%, futhi ukubeka ukuzivumelanisa nezimo phezulu kwashukumisela kwaze kwaba maphakathi nawo-70s. Izisindo ezifanayo, imodeli efanayo—indlela ehlakaniphile yokubuza umbuzo.
Izinhlobo eziyinhloko zokukhuthaza uchungechunge lwemicabango
Abathuthukisi bathuthukise izinhlobo ezimbalwa ze-CoT ukuze balinganisele ukunemba, izindleko kanye nobunzima bokusebenzisa. Uzobona izinhlobo ezifana ne-zero-shot CoT, i-few-shot CoT, i-Automatic CoT (i-Auto-CoT), ukuzivumelanisa nezimo, i-Tree-of-Thoughts kanye nokusikisela okuncane kakhulu, ngayinye ifanele izimo ezihlukile kancane.
Uchungechunge Lokucabanga Olungenaso Isibhamu
I-Zero-shot CoT iyindlela elula kakhulu: awunikezi izibonelo, umane nje ulandela imiyalelo yokucabanga. Imisho efana nokuthi “Ake sicabange isinyathelo ngesinyathelo”, “Xazulula lokhu ngokucophelela, isinyathelo ngasinye” noma “Chaza ukucabanga kwakho ngaphambi kokuphendula” yizinto ezaziwayo ezivuselela ukuziphatha kokucabanga okufundiwe kwemodeli.
Ngokwemvelo, lokhu kulungiswa okulula kungaba nomthelela omkhulu. Kuma-benchmarks ezibalo, umsebenzi wokuqala ubonise ukukhuphuka kokunemba kusuka cishe ku-10% kuya ngaphezu kuka-40% ngokungeza nje umyalelo wesinyathelo ngesinyathelo. Uthola intuthuko enkulu ekhwalithini yokucabanga ngaphandle kokwakha noma ukugcina umtapo wolwazi oyisibonelo.
I-Zero-shot CoT iyakhanya uma ufuna ukunqoba okusheshayo emisebenzini yokucabanga evamile futhi ukhathalela ukubambezeleka kanye nezindleko. Izicelo zihlala zifushane, ngakho-ke ukhokhela amathokheni ambalwa futhi unciphise ukwakheka komongo, kuyilapho usathola ukuhunyushwa okunamandla nokunemba.
Okubi ukuthi imodeli kufanele isungule indlela yayo yokucabanga, engase ibe nenkulumo ephindaphindayo, engaguquki kuzo zonke izizinda, noma ngezinye izikhathi engenangqondo ngisho noma impendulo yokugcina ibukeka kahle. Ezindaweni ezikhethekile—ezezimali, ezokwelapha, ezomthetho, izinqumo ezibalulekile ngokuphepha—lokhu akwanele.
Uchungechunge Lwemicabango Oluncane
I-CoT ethatha imizuzwana embalwa ithatha indlela egxile kakhulu emibonweni: ubonisa amabhangqa emibuzo nezimpendulo zesibonelo lapho izimpendulo zifaka khona izinyathelo zokucabanga ezicacile. Ngemva kwemiboniso embalwa enjalo, ufaka umbuzo wakho wangempela bese uvumela imodeli ukuthi ilingise iphethini.
Le ndlela inamandla kakhulu uma isakhiwo sokucabanga okunengqondo sibaluleke ngempela. Ukuze uthole ithuluzi lokuhlaziya ezezimali, ungafaka izibonelo ezichaza ukubalwa kwemali engenayo, amazinga esaphulelo kanye nokulungiswa kwengozi. Ukuze uthole i-bot ye-medical triage, ungafaka izihlahla zezinqumo zezokwelapha: izimpawu, umlando, izimpawu ezibomvu, umehluko, bese kuba nezincomo.
Umehluko ukuthi i-CoT encane idinga umzamo omkhulu wobunjiniyela ngokushesha. Kumelwe uklame izibonelo ezihlanzekile nezihlukahlukene, uqiniseke ukuthi ukucabanga kwazo kulungile futhi kuyamelela, futhi uzigcine zinolwazi olusha njengoba imikhawulo yomkhiqizo noma yesizinda sakho ishintsha. Izicelo ezinde zisho namathokheni amaningi, izindleko eziphakeme kanye nokubambezeleka okwengeziwe ngocingo ngalunye.
Noma kunjalo, uma isizinda sibucayi noma siyinkimbinkimbi, i-CoT enezithombe ezimbalwa ivame ukwenza kahle kakhulu kune-zero-shot futhi ivame ukuba yisiqalo osifunayo ekukhiqizweni. Uthola ukulawula okwengeziwe ngesitayela kanye nokujula kokucabanga, futhi ungaqondisa imodeli kude nezindlela zokucabanga ezibuthakathaka noma ezingafanele.
Uchungechunge Lokucabanga Oluzenzakalelayo (i-Auto-CoT)
Ukwenza ngezandla izibonelo ezinhle ze-CoT akwenzi kahle, ngakho abacwaningi baphakamise i-Automatic Chain of Thought (Auto-CoT) ukuze kulayishwe ingxenye enkulu yalowo msebenzi kumodeli. Umqondo uwukukhiqiza ngokuzenzakalelayo izintambo zokucabanga ezahlukahlukene ongazisebenzisa kabusha njengemiboniso.
I-Auto-CoT ivame ukuvela ngezigaba ezimbili:
- Ukuhlanganiswa kwemibuzo: Uthatha isethi yedatha yezinkinga, uzifake (isibonelo usebenzisa i-transformer yemisho), bese uzihlanganisa ukuze imibuzo efanayo iphele ndawonye.
- Isampula yokubonisa: kusuka eqenjini ngalinye, ukhetha umbuzo omele bese ucela i-LLM ukuthi idale uchungechunge lokucabanga olune-CoT engenalutho, ngokuvamile isebenzisa ama-heuristics alula njengokuthi "imibuzo emifushane enezinyathelo zokucabanga ezi-~5".
Umphumela uba umtapo wezibonelo ze-CoT ezikhiqizwe ngokuzenzakalelayo, ezihlukahlukene ngokunengqondo ngaphandle kokubhala ngesandla. Uma kufika umbuzo omusha, ungathola noma usampula imiboniso efanele kule layibrari bese uyifaka esicelweni njengezibonelo ezimbalwa ze-CoT.
Ngisho noma amanye amaketanga akhiqizwa ngokuzenzakalela azoqukatha amaphutha amancane, ukuhlukahluka kanye nokubuyisa kuvame ukunciphisa umthelela wanoma yisiphi isibonelo esisodwa esinamaphutha. Empeleni, i-Auto-CoT ivame ukushaya i-CoT engahluziwe kanye ne-CoT engenalwazi emigomeni yokucabanga, kuyilapho isindisa isikhathi esiningi somuntu.
Ukuzivumelanisa nezimo ngezindlela eziningi zokucabanga
Ukuzivumelanisa nezimo kuyisandiso esithuthukisiwe esihweba ngokubala ukuze sithole ukuthembeka. Esikhundleni sokucela imodeli ukuthi inikeze uchungechunge olulodwa lokucabanga kanye nempendulo, uthatha amasampula amaningana azimele (ngokugudluza izinga lokushisa noma amapharamitha okuthatha amasampula), bese uhlanganisa izimpendulo zokugcina ngokuvota kweningi.
Umbono uwukuthi kunezindlela eziningi zokucabanga ezinengqondo eziholela empendulweni efanayo efanele, kodwa izindlela ezinephutha zivame ukuhlukahluka. Isibonelo, u-“15 − 3 + 8” angabalwa njengo-“12 + 8”, noma “15 + 8 = 23, bese ususa u-3”, noma “hlola kwesobunxele kuya kwesokudla”. Konke kukhiqiza u-20, kodwa uchungechunge oluphukile lungagcina luku-21. Uma usebenzisa amasampula amaningana, impendulo engalungile ivame ukudlulwa yivoti.
Kuma-benchmark afana ne-GSM8K, ukuhlanganisa ukuhambisana kwe-CoT kunikeze ukuthuthukiswa kwamaphesenti amabili ekunembileni. Okusobala ukuthi manje wenza izingcingo eziningi ze-LLM ngombuzo womsebenzisi ngamunye, okwandisa kokubili i-latency kanye ne-token spend ngenani lakho lesampula.
Lokho kwenza ukuzivumelanisa nezimo kulungele kangcono imithwalo yemfanelo edinga ukuxazululwa: ukubalwa kwezimali, ukucabanga okusemthethweni, ukwesekwa kwezinqumo zezokwelapha, ukuhlolwa kokuphepha. Kwi-bot yokuxoxa evamile, ukubala okwengeziwe akuvamile ukuphuma, kodwa ku-ejenti ebaluleke kakhulu, ukuthembeka okwengeziwe kungafaneleka yonke imizuzwana.
I-Tree-of-Thoughts: ukuhluma esikhundleni sokucabanga okuqondile
I-Tree-of-Thoughts (ToT) yelula uchungechunge lwemicabango oluvela ochungechungeni olulodwa luye esihlahleni sokusesha esihlanganisa amagatsha phezu kwemicabango engaba khona. Esikhundleni sokulandela indlela eyodwa yokucabanga kusukela ekuqaleni kuze kube sekupheleni, uhlelo luhlola izinketho eziningana esinyathelweni ngasinye, luthena amagatsha abuthakathaka bese luqhubeka nokwehla kwanamandla kakhulu.
Lokhu kusondelene nendlela ongazixazulula ngayo izinkinga zokuhlanganisa noma zamasu ekhanda lakho. Ucabanga ngezinyathelo ezimbalwa ezingenzeka, uzihlole kancane, uzilahle lezo ezibonakala zingenakwenzeka, bese uqhubeka nokwandisa iziqondiso ezithembisayo uze ufinyelele esixazululweni esiqinile.
Ngokwemigomo yokusetshenziswa, i-ToT ivame ukuhlanganisa izingcingo eziningi ze-LLM. Ekujuleni ngakunye kwesihlahla, imodeli iphakamisa izinyathelo ezilandelayo; umlawuli uhlola izimo ezingaphelele, mhlawumbe esebenzisa enye i-LLM noma amaphuzu e-heuristic, bese ekhetha ukuthi yimaphi amagatsha okufanele awandise. Amademo ocwaningo asebenzise i-ToT ukubhekana nemidlalo yephazili, imisebenzi yokuhlela kanye nemibono yokudala ngemiphumela engcono kakhulu kune-CoT ecacile.
Ukushintshana kwentengo kuba yizindleko: ungase udinge izingcingo eziningi ngenkinga eyodwa. Yingakho i-ToT ingcono kakhulu ukugcinelwa izindawo lapho ukuhlola okuphelele kubaluleke kakhulu kunesivinini—umklamo oyinkimbinkimbi, ama-ejenti okudlala imidlalo, noma ukucabanga lapho ukujula nokuhlukahluka kuyimigomo.
Isikhuthazo esincane kakhulu
Ukunxusa okuncane kakhulu kungenye isu elithuthukisiwe elihlukanisa inkinga eyinkimbinkimbi ibe yizinkinga ezincane ezilula ezisingathwa ngokulandelana. Okokuqala, ucela imodeli ukuthi ibone umsebenzi omncane kakhulu engayixazulula; okulandelayo, ubuyisela leso sixazululo bese ucela ingxenye elandelayo eyinkimbinkimbi kakhulu; njalo njalo kuze kube yilapho inkinga ephelele isixazululiwe.
Le ndlela isebenza kahle kakhulu ekucabangeni kokwakheka. Cabanga ngemibuzo yesakhiwo sedatha esihlanganisiwe, i-algebra enezinyathelo eziningi, noma ukukhiqizwa kwekhodi ngezici eziyinkimbinkimbi lapho ingxenye ngayinye incike emiphumeleni yangaphambilini. Ngokuphoqa ukubola okuhlanzekile, unciphisa umthwalo wokuqonda kumodeli esinyathelweni ngasinye futhi wenza umkhondo wokucabanga wonke ube lula ukuwuhlola.
Uchungechunge Lwemicabango ezinhlelweni ezisebenzisa ubuchwepheshe kanye namathuluzi
I-CoT iba yigugu nakakhulu uma usuqala ukwakha ama-ejenti athatha izinyathelo, abize amathuluzi futhi ahlele izinyathelo eziningi. Esikhundleni sokuphendula umbuzo owodwa bese umisa, lezi zinhlelo zijikeleza ngemijikelezo yokucabanga, yokwenza kanye nokubuka, zivuselela izinhlelo zazo ngolwazi ngalunye olusha.
Cabanga nge-ejenti yokusekela ephethe: “Ngi-ode ijezi elibomvu ngoLwesibili olwedlule kodwa ngathola eliluhlaza okwesibhakabhaka. Ngingalibuyisela?” Indlela yokuziphatha efanelekile ingaba: ukuqonda inkinga, ukuthola i-oda, ukuhlola inqubomgomo yokubuyisa, ukuhlola iwindi lokubuyisa, ukunquma ukufaneleka bese ekugcineni uqala ukubuyisa.
Ngokusebenzisa iseluleko esilula, umenzeli angase agxumele ku-“Impela, nansi ilebula” noma “Cha, asikwazi ukukwenza lokho” ngokusekelwe ekufaneni kwephethini okusheshayo, eqa ukuhlola okubalulekile. Nge-Chain of Thought, uyikhuthaza ukuthi ilandise into efana nale: “Ngizoqala ngibheke i-oda lakho kusukela ngoLwesibili olwedlule, bese ngiqinisekisa ukuthi into kanye nombala awufani, bese ngihlola ukuthi usesikhathini sezinsuku ezingu-30, bese ngiqala ukugeleza kokubuya uma kufaneleka.”
Lokhu kuseduze nephethini ye-ReAct (Reason + Act): i-ejenti iyashintshana phakathi kokucabanga kwangaphakathi (“Ngidinga ukubuza nge-oda le-API”) kanye nezenzo zangaphandle (ukwenza ucingo lwe-API), bese ihlanganisa okubonwayo esinyathelweni esilandelayo sokucabanga. Ingxenye ngayinye "yomcabango" iba yingxenye yomkhondo ongawubhala phansi, uwulungise bese uwuhlaziya.
Kumasistimu e-ejensi, i-CoT akuyona nje into enhle ukuba nayo; ngokuvamile iyisivikelo sakho esiyinhloko sokuthembeka, ukucaca kanye nokuphepha. Uma into iphuka—ithuluzi elingalungile, ipharamitha engalungile, ukuhumusha okungalungile—ungabona ngempela ukuthi i-ejenti iphambukile kuphi bese ulungisa isixwayiso, amathuluzi noma inqubomgomo esikhundleni sokuqagela ebumnyameni.
Ukuxhumanisa ngokushesha vs uchungechunge lokucabanga
Ukuxhumanisa ngokushesha kanye nochungechunge lokucabanga kokubili kusiza ngemisebenzi eyinkimbinkimbi, kodwa kusebenza emazingeni ahlukene. Ngokuhlanganisa ngokushesha, uhlukanisa umsebenzi omkhulu phakathi kwezimpendulo eziningi ezihlukene, uhambisa umphumela wesinye uye kwesilandelayo. Nge-CoT, ufaka yonke inqubo yokucabanga ngaphakathi kokushintshana okukodwa kwempendulo ngokushesha.
Isibonelo sokuxhumanisa ngokushesha: ukuhlaziya incwadi ngezinyathelo ezintathu—isinyathelo sokuqala sesifinyezo sendaba, isinyathelo sesibili sokuhlaziya itimu kusetshenziswa leso sifinyezo, isinyathelo sesithathu sokubuyekeza kokugcina kusetshenziswa zombili. Isinyathelo ngasinye siwubizo lwe-LLM oluhlukile olunemiyalelo yalo.
Isibonelo Sochungechunge Lokucabanga lomsebenzi ofanayo: ngaphakathi kwesikhuthazo esisodwa uthi, “Okokuqala ufingqa indaba, bese ukhomba izingqikithi ezinkulu, bese uphetha ngombono omfushane obucayi. Cabanga ngesigaba ngasinye isinyathelo ngesinyathelo.” Imodeli bese idala umugqa wayo omncane wemicabango kanye nempendulo yokugcina ngesithombe esisodwa.
Empeleni, izinhlelo zangempela zivame ukuhlanganisa zombili: sebenzisa i-CoT ngaphakathi kwesinyathelo ngasinye esihlanganisiwe ukuze uthuthukise ukucabanga, futhi uhlanganise izinkomba eziningana ezingeziwe ze-CoT ukuze uhlele imisebenzi emide yokusebenza. Umehluko omkhulu ukuthi ukuxhumanisa okusheshayo kwakheka ukuhamba komsebenzi kwe-macro kuzo zonke izingcingo eziningi, kuyilapho uchungechunge lwemicabango lwakha ukucabanga okuncane ngaphakathi kocingo ngalunye.
Uchungechunge Lwemicabango Ehlukahlukene
Njengoba amamodeli amaningi evuthwa, i-Chain of Thought ayisagcini embhalweni omsulwa. I-Multimodal CoT ivumela uhlelo ukuthi lucabange ndawonye ngombhalo, izithombe kanye nokunye okufakiwe okufana nomsindo noma amathebula, kuyilapho lusaxoxa ngezinyathelo zalo zangaphakathi.
Thatha isithombe sebhishi eligcwele abantu bese ubuza umbuzo othi “Ingabe le ndawo ibukeka ithandwa yizivakashi njengamanje?” Imodeli ye-CoT enezindlela eziningi ingaphawula ngokusobala inani lezambulela, ubuningi babantu, indawo yokupaka ematasa kanye nezinkomba zesikhathi sosuku noma izithunzi, bese iphikisana ngokuthi zonke lezo zimpawu ezibonakalayo zikhomba ukuthandwa okuphezulu kwamanje.
Ngokwenza ukucabanga okubonakalayo kucace, awugcini nje ngokuthola ukunemba okungcono kodwa futhi uthola nezinqumo eziqondakala kalula. Abasebenzisi bangabona ukuthi yiziphi izakhi zesithombe imodeli egxile kuzo, futhi ungabona izindlela zokwehluleka njengokufaka imininingwane eminingi engabalulekile.
Ukuthuthukisa Uchungechunge Lokucabanga Ngezinga Eliphezulu
Uma usushintshele kuma-demo ambalwa uye kuthrafikhi yangempela, iqiniso elingcolile liyafika: Ukusebenza kahle kwe-CoT kuncike kakhulu emsebenzini, izibuyekezo zemodeli kanye nomhlahlandlela wokufuduka, indlela yokubhala kanye nezibonelo ezithile ozinikeza zona. Ukucabanga okubhalwe kahle kusengaholela ezimpendulweni ezingalungile, futhi izintambo zokucabanga ezibanzi zingashisa amathokheni ngaphandle kokwengeza inani elikhulu.
Ukuze i-CoT isebenze ekukhiqizeni, udinga i-feedback loop elandelela ubukhulu obuningana ngesikhathi esisodwa:
- Ukunemba kokugcina: Ingabe impendulo yomodeli ihambisana neqiniso elilindelekile noma isahlulelo somuntu?
- Ikhwalithi yokucabanga: ingabe izinyathelo eziphakathi ziyasebenza, ziyahambisana ngokunengqondo futhi zihambisana nemikhawulo yesizinda?
- Ukuvumelana: Ingabe imibuzo efanayo iveza ukucabanga nezimpendulo ezifanayo phakathi nokugijima kanye nesikhathi?
- Ukusebenza kahle kwethokheni: Zingaki amathokheni owasebenzisayo ngombuzo ngamunye, futhi ingabe uthola ikhwalithi eyanele ngokubuyisela?
Ukuhlola indawo ngesandla ezibonelweni ezimbalwa akukwenzi kube nzima uma usunezinhlobo eziningi ezisheshayo kanye namakhulu ezimo zokuhlola. Udinga ingqalasizinda engakwazi ukuhumusha izikhuthazo, iqhube ukuhlolwa okuhlelekile futhi ibone ngeso lengqondo imikhondo yokucabanga ngezinga.
Amathuluzi okubona akhelwe ngenhloso ama-LLM asiza lapha ngokuthwebula umkhondo ogcwele—okusheshayo, imodeli, ukucabanga kwe-CoT, izingcingo zamathuluzi, umphumela wokugcina—wesicelo ngasinye. Amapulatifomu afana ne-Opik, isibonelo, akuvumela ukuthi ubhalise futhi uhlole izintambo ze-CoT ngokuningiliziwe, uqhathanise izinguqulo ezahlukene ze-prompt, futhi usebenzise ngisho nokusethwa kwe-LLM-as-a-judge ukuthola amaphuzu ngokuzenzakalelayo kokubili izimpendulo zokugcina kanye nekhwalithi yokucabanga.
Ngaleyo datha esandleni, ungathuthukisa kancane kancane ukusethwa kwe-CoT yakho: ukulungisa amagama, ukushintshanisa i-zero-shot ne-shot-shot embalwa, ukulungisa noma ukuvuselela izibonelo nge-Auto-CoT, noma ukwethula ukuzivumelanisa kuphela lapho kuhambisa khona inaliti. Ezinye izinhlaka zihlangana ngisho nemitapo yolwazi yokwenza ngcono njenge-DSPy noma usesho lokuziphendukela kwemvelo ukuze zithuthukise izikhuthazo ezingcono ngokuqhubekayo ngokusekelwe kumamethrikhi okuhlola.
Khumbula ukuthi i-Chain of Thought cishe njalo ibiza okungaphezu kokuphendula ngokuqondile: umbhalo wokucabanga wodwa ungakhulisa ukusetshenziswa kwethokheni ngokuphindwe kabili kuya kane, ukuzivumelanisa nezimo kuphindaphinda lokho ngenani lamasampula, futhi i-Tree-of-Thoughts ingaba yi-oda elikhulu kakhulu futhi. Yingakho ufuna ukuqapha okucacile, ukuze wazi kahle ukuthi leso sabelomali esengeziwe sithela kuphi.
Kumaqembu amaningi, isu elisebenzayo lihlelwe ngokwezigaba: okuzenzakalelayo kube yi-light zero-shot noma i-shot short few-shot CoT, ikhuphukela ekubeni yi-self-consistency noma i-ToT kuphela uma imibuzo iphawulwe njengenani eliphezulu, ukungaqondakali okuphezulu noma ingozi ephezulu. Ukuqaphela nokuhlola yikona okwenza lolu hlobo lwesu eliguquguqukayo lube nokwenzeka.
Njengoba uzama i-CoT ezinhlelweni zakho zokusebenza—kungakhathaliseki ukuthi usebenzisa izixwayiso ezisheshayo ze-zero-shot, izibonelo ezikhethwe kakhulu ze-shot ezimbalwa, imitapo yolwazi ye-Auto-CoT ezenzakalelayo noma ukuzivumelanisa nezimo okunezibonelo eziningi—isihluthulelo ukuphatha ukucabanga kwemodeli njengomkhiqizo weklasi eliphezulu. Kwenze kucace, kubhale phansi, kubhale phansi bese uphinda-phinda kukho, futhi uzovula ukuziphatha okuthembekile kakhulu, okuchazekayo nokunamandla okuvela kumamodeli afanayo ayisisekelo kunokuba ungakwenza ngezimpendulo ezicacile.
