Kwenzekeni
I-HackerNoon ibika ukuthi unjiniyela wesofthiwe u-Prajwal S. Venkateshmurthy wakha i-TeXFix-Bench, ibhentshimakhi yokuhlola ukuthi izinhlelo ze-AI ziyayilungisa yini imibhalo ye-LaTeX, iTypst ne-Markdown ephukile ngaphandle kokushintsha incazelo yazo. Umbiko uthi impumelelo yokuhlanganisa iyodwa ikhiqize amazinga adukisayo.
I-HackerNoon ibika ukuthi i-Venkateshmurthy idale i-TeXFix-Bench ngemuva kokuphetha ngokuthi "ukwenza lokhu kuhlanganisa" isilinganiso esinganele sokulungisa idokhumenti. Ibhentshimakhi icela amamodeli ukuthi abuyisele ifayela lomthombo elilungisiwe eliphelele, ngaphandle kokukhomba iphutha noma ukunikeza amathuluzi, bese ihlola umphumela endaweni. Umgomo obikiwe uwukuhlukanisa idokhumenti emane yakha kuleyo egcina okuqukethwe kwedokhumenti yokuqala.
NgokukaHackerNoon, ibhentshimakhi iqukethe imisebenzi yokulungisa elawulwayo eyi-10,437 kuyo yonke iLaTeX, iTypst neMarkdown. Imisebenzi yenziwe nge-taxonomy esekelwe emaphutheni e-LaTeX aqinisekisiwe angu-168 aqoqwe kwa-TeX Stack Exchange, ukuzibophezela kwe-GitHub kanye nemibhalo yephakheji. Umbiko uthi umtapo wezincwadi wokuguqula i-DocMut onemiphumela unama-opharetha e-syntax-aware angama-48 kuwo wonke amafomethi amathathu futhi usebenzisa imbewu enqumayo, amasango enjini kanye nesheke lokubonisa umehluko.
I-HackerNoon ibika ukuthi ukuqhathanisa okuyinhloko kusebenzise amamodeli ayisikhombisa kuwo wonke ama-matrix wesibonelo esingu-6,613, okukhiqiza imizamo eyinhloko engu-46,291. Kubandakanya nemizamo eyengeziwe erekhodiwe, iphakheji yocwaningo iqukethe izicelo ezingama-48,651. Ukuhlola kubale izimpendulo ezingenalutho, imiphumela encishisiwe, ukuphela kwesikhathi, ukwehluleka kwezokuthutha kanye nemikhawulo yezilinganiso njengokuhluleka. Imiphumela iye yahlolwa kabusha endaweni nge-Tectonic 0.17.0 ye-LaTeX, i-Typst 0.15.1 kanye ne-Pandoc 3.10.1 ye-Markdown, nombhalo we-PDF okhishiwe osetshenziselwa ukuthola amaphuzu okubuyisela.
Umbiko uthi imodeli enezinga eliphezulu lokuhlanganisa okunemibandela, i-Qwen3.7-Max engu-94.4%, ibe nesilinganiso sokuphela konyaka esingu-56.7% ngoba sibuyise impendulo esebenzisekayo kuphela u-60.1% wesikhathi. I-Grok-4.3 kubikwa ukuthi ihlanganise u-84.2% wayo yonke imizamo, kuyilapho i-GLM-5.2 ihlanganise u-64.9% ekupheleni ukuze iphele naphezu kwezinga elinemibandela elingu-93.8%. I-HackerNoon iphinde ibike ukuthi u-13.6% kuya ku-18.5% wokulungiswa kokuhlanganiswa kwashintsha ngokubonakalayo idokhumenti, nokuthi i-Qwen3.7-Max inenani eliphakeme kakhulu elibikiwe lesilinganiso sokubuyisela kuyilapho i-Llama-4 i-Maverick inerekhodi elibuthakathaka kakhulu lokubuyisela.
Imininingwane yomthombo: hackernoon.com ↗
Kungani kubalulekile
Ibhentshimakhi ibhekana nemodi yokwehluleka okusebenzayo kumathuluzi okubhala nekhodi e-AI: idokhumenti ingase ihlanganiswe ngempumelelo ngemva kokubhala kabusha okunolaka okususa noma okushintsha okuqukethwe buthule. Indlela yayo ingasiza amaqembu omkhiqizo ukuthi ahlole amasistimu okulungisa amadokhumenti esebenzisa ukwethembeka nokulondolozwa okubonakalayo komsebenzisi, kunomaka olulodwa oluluhlaza.
Ukuthola okuphakathi kuka-HackerNoon ukuthi ukuhlanganisa nokulungisa okuthembekile kuyimisebenzi ehlukene. Isistimu ingabuyisela idokhumenti encane ehlala ihlanganiswa ngenkathi ilahla iphepha lomsebenzisi, noma ingase ibhale kabusha isethulo, isitayela sezincwadi zezincwadi, ithebula noma isigaba ngezindlela ezingabonakali ku-PDF ewumphumela. Umbiko uthi u-3.7% wabafundi abamukelwe bebehlehlisiwe ncamashi, okusho ukuthi uhlelo luxazulule isigameko ngokulungisa iphutha elijovwe kunokuba libonise ukulungiswa okuvamile.
Imiphumela ibalulekile kumathuluzi okubhala e-AI ngoba abasebenzisi bahlangabezana nokungalethwa njengokwehluleka. I-HackerNoon ibika ukusabalala kwamaphoyinti angu-27.5 phakathi kwamazinga okuhlanganisa angcono kakhulu namabi kakhulu futhi iphikisa ngokuthi umehluko omkhulu uqhamuke ekunikezeni ukwethembeka kunekhono lemodeli. Lowo mehluko ubalulekile ekusebenzeni: ukuthuthukisa ukutholakala komhlinzeki, imikhawulo yokuqedela kanye nomzila kungase kusize abasebenzisi ngaphezu kokushintsha imodeli eyisisekelo, kuyilapho ukunemba kokuhlanganisa kuphela kungafihla lokho kwehluleka kwesevisi.
Umbiko uphinde uphakamise ukuthi ifomethi yedokhumenti nohlobo lwamaphutha kuthinta kakhulu ukusebenza. I-HackerNoon ithi impumelelo yokuhlanganisa ukuphela kuye ekupheleni ibicishe ibe ngu-74.2% ku-LaTeX, u-60.3% weTypst no-90.2% we-Markdown. Izinkinga eziphathelene nesakhiwo nokuncika, okuhlanganisa ukwehliswa kokungenisa kwe-Typst, izimfuneko ze-LaTeX shell-scape kanye nezibalo ezingavaliwe, kubikwa ukuthi zibonakale ziqinile kunendlela yokubhala imiyalo yendawo. Lokhu okutholakele kungasiza onjiniyela baklame ukuxilonga okuqondiwe futhi babuyekeze ukugeleza komsebenzi.
I-HackerNoon ibika ukuthi amaphutha okwenziwa asekelwe ku-taxonomy abengamaphesenti angu-5.6 kuya kwangu-9.2 anzima kunokuguqulwa okusekelwe ephethinini kuyo yonke imindeni emithathu yamamodeli. Ocwaningweni oluhlukile, imodeli eqinile etholakalayo kubikwa ukuthi ilungise u-67.0% wokuphahlazeka kwangempela kwabantu okungalinganiseka okungama-88, uma kuqhathaniswa no-81.3% kusethi eqinile yokwenziwa kanye no-90.5% ezinguqukweni ezisekelwe iphethini. I-athikili yethula lokhu njengobufakazi bokuthi ukuhlola okokwenziwa okusekelwe phansi kungase kube okwangempela kunokuhlelwa kwe-ad hoc, hhayi njengesilinganiso senani labantu sokusebenza komhlaba wangempela.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Umbiko uthi umsebenzi wesikhathi esizayo uzohlola ukulungisa kuzo zonke izinjini eziningi ze-TeX, wengeze amamodeli avaliwe nokuxilongwa, futhi wakhe ithrekhi yephutha lomhlaba wangempela onelayisensi, ephindaphindekayo. Okutholwe yibhentshimakhi kuhlala kuyilokho okubikwe ngumbhali we-HackerNoon futhi akuzange kuqinisekiswe ngokuzimela lapha.
Ukulandelela okubaluleke kakhulu ukuthi izinga elibikiwe liyasinda yini ekuhlolweni okubanzi. I-HackerNoon ithi umsebenzi wamanje awusunguli izinga lemodeli yendawo yonke yayo yonke i-LaTeX, futhi ukuhlola kwayo okubonakalayo kunomkhawulo ngoba ukubuyisela kukalwe ngombhalo okhishiwe esikhundleni sokubonakala okuphelele noma ukulingana kwesakhiwo. Ngakho-ke ukulungisa kungalondoloza umbhalo kuyilapho kushintsha ukwakheka kwekhasi, ukufometha noma isethulo ngezindlela ezilandelanayo.
Umbiko uthi ukuhlolwa okuzayo kuzohlola ukuthi ukulungisa okusebenza ngaphansi kwe-Tectonic nakho kuyasebenza ngaphansi kwe-pdfLaTeX, i-XeLaTeX ne-LuaLaTeX. Lokho kubalulekile ngoba umehluko wenjini ungathinta ukuphatheka. I-HackerNoon iphinda ihlonze amamodeli avaliwe emngceleni kanye nesimo sokuxilonga ngokushesha njengesingekho kuphaneli yamanje, ngakho-ke imiphumela ebikiwe ye-zero-shot akufanele ithathwe njengesilinganiso esiphelele salokho okungenziwa amathuluzi okuthengisa asizayo.
Omunye umbuzo ovulekile ukuthi ingabe ibhentshimark ingathuthukisa ithrekhi yomhlaba wangempela ekwazi ukukhiqizwa kabusha. I-HackerNoon ithi ithrekhi yayo yomhlaba wangempela eqandisiwe okwamanje ayinazo izimo ezamukelwayo ngoba umbhali ubedinga ukugunyazwa okuqinisekisiwe, ukuvela kwegama kanye nokukhiqizwa kabusha. Lowo mkhawulo unengqondo: isethi yokwenziwa ine-oracle ecacile, kodwa okwamanje ayikabonisi ukuthi amaphutha okugunyaza avela kaningi kangakanani noma ukuthi amadokhumenti abasebenzisi aziphatha kanjani ekuhambeni komsebenzi okubukhoma.
Izinga elisebenzayo elihlongozwe umbiko ukushicilela ukulethwa, ukuhlanganiswa, ukubuyisela, ukuhlela ubuncane kanye nokukhiqiza kabusha ngokuhlukene, ngamadenomineyina ashiwo. Lezo zinyathelo zingase zibe nolwazi oluningi kunezinga lokuphasa elilodwa, kodwa i-HackerNoon ayiqambi ngokuzimela ukuthi imiphi imikhawulo okufanele ibuse ukuthunyelwa. I-athikili ithi ngokuqondile ukuthi awukho umkhawulo owodwa wokufana nombhalo ongaqinisekisa ukulungiswa okulungile, ngakho ukubuyekezwa komuntu nokuhlola okubanzi kwe-semantic kuhlala kungaxazululiwe.