Inkampani i-OpenAI imemezele namuhla (31 kuJulayi) ukuthi ngokuthuthukisa i-framework, ikwazile ukukhuphula ukusebenza kwemodeli i-GPT-5.6 Sol ku-ARC-AGI-3 ngo-188%.
Qaphela: I-ARC-AGI-3 yisivivinyo esisha sokucabanga ngokusebenzisana esethulwe yithimba le-ARC Prize. Siyisilinganiso esiqinile nesisezingeni eliphezulu kunazo zonke esamukelwa emhlabeni wonke njengendlela yokukala ubuhlakani bokwenziwa obujwayelekile (AGI).
Lesi sivivinyo siqeda ngokuphelele indlela yakudala yokuhlola i-AI ngokuthi “iphendule imibuzo” nolwazi ngendlela emile, bese siyifaka ngqo endaweni “yomdlalo” engajwayelekile ngokuphelele, ukuze kuhlolwe ukuthi inawo yini amandla okuf探索, ukufunda nokuzivumelanisa nesimo ngesikhathi sangempela njengomuntu.
Ngaphambi kokulungiswa nokwenziwa ngcono, i-GPT-5.6 Sol ithole u-7.8% esivivinyweni se-ARC-AGI-3, kanti imodeli i-GPT-5.5 ithole u-0.4% kuphela.
Ngemva kokuhlaziya kabusha, i-OpenAI ithole ukuthi i-framework esemthethweni yochungechunge lwamamodeli e-GPT-5.6 inezindlela ezi-2 ezithinta ukusebenza kwe-GPT-5.6 Sol:
Eyokuqala, ukucabanga okuyimfihlo kuyasuswa ngemva kwesenzo ngasinye somdlalo. Lokhu kusho ukuthi ngemva kokwenza isenzo esi-1, i-GPT-5.6 Sol kufanele iphinde iwuqonde umdlalo. Imodeli isakwazi ukubona amarekhodi ezenzo zangaphambilini namanothi amafushane, kodwa ayikwazi ukubona izinhlelo, ukuqonda kanye nenqubo yokucabanga okuholele kulezo zenzo.
Eyesibili, i-framework esemthethweni isebenzisa iwindi lokususa okuqhubekayo. Njengoba umlando uba mude, izenzo zokuqala ziyanyamalala emongweni. Ngamanye amazwi, i-GPT-5.6 Sol ayikwazi nje kuphela ukugcina ukucabanga kwangaphambilini, kodwa futhi ilahlekelwa yinkumbulo yezenzo zangaphambilini.
Ukuze ibhekane nalezi zinkinga, i-OpenAI iphakamise izixazululo ezimbili: ukugcinwa kokucabanga (inference preservation) kanye nokucindezelwa komongo (context compression).
Mayelana nokugcinwa kokucabanga, i-OpenAI iphinde yakha i-framework yokuhlola i-ARC-AGI-3 isebenzisa i-Responses API. Ku-GPT-5.6, kwanele ukudlulisa i-ID yempendulo yangaphambilini ukuze ukucabanga kugcinwe ngokuzenzakalelayo ekubizweni kwamathuluzi nasekusebenzelaneni kwemijikelezo eminingi.
I-OpenAI ithe ngemva kokunika amandla ukugcinwa kokucabanga, i-GPT-5.6 ayisadingi ukuphinda ichaze umdlalo emzuliswaneni ngamunye. Lokhu kunciphisa isikhathi esichithwa ekucabangeni ngaphambi kwesenzo ngasinye. Njengoba imodeli igcina imicabango yangaphambilini, iba ngcono ekufundeni ngokuhamba kwesikhathi futhi isebenzise amasu ahambisana kangcono.
Mayelana nokucindezelwa komongo, i-framework esemthethweni ye-ARC-AGI-3 ilahla imilayezo emidala kakhulu lapho umongo wengxoxo weqa izinhlamvu ezingu-175000. I-OpenAI ithe ukususa okuqhubekayo kunezinkinga ezi-2: imodeli ilahlekelwa ukubuka nezenzo zokuqala; futhi ingxenye enkulu yesikhathi somsebenzi imodeli isebenza inewindi lomongo eligcwele kakhulu, okungase kwehlise ukusebenza kancane.
Ngemva kokunika amandla ukucindezelwa komongo, i-GPT-5.6 Sol ikwazi kangcono ukugcina ulwazi olufundiwe lomdlalo ngamunye emisebenzini emide, futhi ithole amaphuzu aphezulu ngokukhiqiza ama-token ambalwa.
Nokho, ngemva kokusebenzisa i-harness esemthethweni, i-GPT-5.6 Sol ithole u-13.3% ku-ARC-AGI-3. Lapho kunikwa amandla ukugcinwa kokucabanga (inference preservation) nokucindezelwa komongo (context compression), i-GPT-5.6 Sol yehlise inani lama-token akhishwayo laba ingxenye yesithupha, ngesikhathi amaphuzu enyuka afinyelela ku-38.3%, okuwukukhula ngo-188.42%.
i-harness, evame ukubizwa ngokuthi ubunjiniyela bokulawula, ibhekisela ku-architecture ephelele yesistimu eboshelwe ngaphandle kwemodeli enkulu, enikeza indawo yokuyisebenzisela, ukuhlanganiswa kwamathuluzi, imithetho nemikhawulo, kanye nendlela yokuthola impendulo.
Uma siqhathanisa imodeli ye-AI enamandla nehhashi lasendle elinamandla kodwa elilula ukuphuma ekulawulweni, i-harness iyizintambo zalo, isihlalo kanye nomzila womgibeli. Ayishintshi i-architecture yemodeli uqobo, kodwa inquma ukuthi imodeli ingalawulwa kanjani ngokuphepha nangokwethembeka ukuze iqede imisebenzi eyinkimbinkimbi.
