Ucwaningo olusha olwenziwe yi-Microsoft Research luveze ukuthi ubuhlakani bokwenziwa obukhiqizayo, uma busetshenziselwa ukuhlela nokubhala kabusha imibhalo yebhizinisi isikhathi eside, buvame ukungenisa amaphutha nokunciphisa ikhwalithi yolwazi. Lolu cwaningo, olubizwa nge-DELEGATE-52, luhlole ukusebenza kwezinhlelo ezinkulu zolimi (LLMs) emisebenzini elandelanayo yokufunda, ukuhumusha nokuguqula imibhalo eyinkimbinkimbi. Imiphumela ikhombisa ukuthi nakuba lezi zinsimbi zihlaba umxhwele emisebenzini emifushane, zingakwazi ukususa ulwazi olubalulekile, zishintshe ulwazi olulungile, futhi zidale ukuphambuka okukhulayo uma zisebenza ngaphandle kwesandla somuntu esiqaphile. Lokhu kuphakamisa izingozi ezinkulu emkhakheni wezebhizinisi lapho ukunemba kungukhiye.
I-DELEGATE-52: Indlela entsha yokuhlola
I-benchmark i-DELEGATE-52 yakhelwe ukulingisa imisebenzi yangempela yochwepheshe, ihlanganisa izindawo eziningi zolwazi. Ngokungafani nezivivinyo ezijwayelekile ezigxile emibuzweni eyedwa, lesi sivivinyo esisha silinganisa ukuthi kwenzekani lapho uhlelo lwe-IA lunikezwa inkululeko yokwenza imisebenzi emide — njengokubhala imibiko, ukudala izethulo, nokufingqa okuqukethwe — ngezinyathelo eziningi. Abacwaningi babona ukuthi izinkinga ziba nkulu njengoba inani lokuxhumana okwenziwa ubuhlakani bokwenziwa ngaphakathi kwedokhumenti eyodwa likhula. Lokhu kwenzeka ngoba amaphutha amancane, nakuba engabonakali esigabeni ngasinye, ayakwazi ukunqwabelana ngokuhamba kwesikhathi. Ngakho, ukwethemba i-IA ngaphandle kokuqapha kungaholela emiphumeleni emibi kakhulu.
Ukunqwabelana kwamaphutha okubangwa ukuhlela okuphindaphindiwe
Enye yezinto eziyinhloko ezitholwe ucwaningo ukuncipha okuhamba kancane kwekhwalithi yedokhumenti, okubizwa ngokuthi i-degradation yedokhumenti. Lokhu ukulahlekelwa okuncane kokunemba njengoba idokhumenti ihlelwa izikhathi eziningi yi-IA. Ulwazi oluguqulwe kancane esigabeni esisodwa lungenziwa njengolulungile ezigabeni zakamuva, okudala ukuphambuka okukhulayo. Lesi simo sikhumbuza umthelela wokudluliselwa kwemiyalezo phakathi kwabantu, lapho izinguquko ezincane ezihlangene zikhiqiza umphumela ohluke kakhulu kowokuqala. Ngokusho kocwaningo, le ndlela yokuziphatha ibonwe kumamodeli amaningi athuthukile atholakalayo emakethe.
Inselelo yemibhalo emide nezinhlelo eziyinkimbinkimbi
Izinhlelo ezinkulu zolimi zisebenza ngokubikezela ukuthi amagama athile angavela kanjani ngokulandelana, ngaphakathi komongo othile. Nakuba le ndlela ikhiqiza imibhalo eyinkimbinkimbi, ayiqinisekisi ukuqonda okuphelele kwencazelo yolwazi. Lapho idokhumenti ihlelwa kaningi, imodeli idinga ukunquma ukuthi yini okufanele igcinwe, isuswe noma iguqulwe — futhi ezimweni eziningi, ulwazi olubalulekile lufinyezwa ngokweqile, luhunyushwe ngendlela engafanele, noma luthathelwe indawo okuqukethwe okubukeka kulungile kodwa kungelona iqiniso. Imibhalo emide imelela inselelo eyengeziwe, ngoba idinga ukuthi uhlelo lubheke inani elikhulu lomongo ngesikhathi esisodwa.
I-Python ikhombisa ukusebenza okungcono
Phakathi kwezindawo ezihloliwe, uhlelo lwe-Python lubonise ukusebenza okuhle kakhulu. Abacwaningi baqaphele ukuthi imisebenzi yokudala nokuguqula ikhodi inezici ezisiza ukuhlolwa okuzenzakalelayo: amaphutha angatholakala ngezivivinyo, izihlanganisi, nezihloli. Lokhu akwenzeki emibhalweni ejwayelekile. Lokhu kusiza ukuchaza impumelelo enkulu yokuzenzakalela kwe-IA ekuthuthukisweni kwesofthiwe. Kodwa noma kunjalo, ochwepheshe baxwayisa ukuthi amakhodi akhiqizwe ubuhlakani bokwenziwa adinga ukubuyekezwa ngobuchwepheshe ngaphambi kokuthi asetshenziswe.
Iqhaza elingenakushintshwa lomuntu ekuhloleni
Isiphetho esiyinhloko se-DELEGATE-52 ukuthi ukwenganyelwa komuntu kuhlala kubalulekile. Amamodeli amanje, nakuba ethuthukile, awanakuqonda kwangempela komongo, izinhloso, noma imiphumela ehlobene nolwazi abalusebenzisa. Ochwepheshe abanolwazi banendima ebalulekile ekuqinisekiseni amaqiniso, ekuhlaziyeni okujulile, ekutholeni ukungahambelani, nasekuqinisekiseni imiphumela. Empeleni, ukuhlanganisa ubuhlakani bokwenziwa nokwenganyelwa komuntu kuvame ukunikeza imiphumela engcono kunanoma iyiphi indlela esetshenziswa yodwa. Ucwaningo luqinisa ukuthi emisebenzini ebalulekile njengemibiko yezezimali, izinkontileka zomthetho, nocwaningo lwesayensi, i-IA kufanele ibe ithuluzi lokusekela, hhayi esikhundleni sabantu.
Naphezu kwemikhawulo yamanje, ochwepheshe bakholelwa ukuthi ama-agent e-IA azoqhubeka nokukhula ngokushesha. Izakhiwo ezintsha, amawindi omongo amakhulu, ukuhlanganiswa nezizinda zedatha zangaphandle, kanye nezindlela eziphambili zokuqinisekisa zinganciphisa kakhulu izinkinga ezibonwa namuhla. Abaningi bavela ukuthi ikusasa lokuzenzakalela lizoncika ekwakhiweni kwezinhlelo ezikwazi ukuqinisekisa izimpendulo zazo ngokuqhubekayo — mhlawumbe ngama-agent amaningi asebenza ndawonye kanye nokuqinisekiswa okuzimele. Indlela ethembisayo, ngokusho kocwaningo, ukubambisana phakathi kwabantu nemishini, kuhlanganisa isivinini sekhompyutha nokwahlulela komuntu.
