Chii chaitika
Bepa rearXiv rinounza FLARE, kana Formulation-Level Automated Reformulation Evaluation, nzira yekuona yakasanganiswa-integer linear programming reformulations. Inogadzirisa tsananguro inovaka yekuvandudza muLean uye inoshandisa LLM-based agent kuburitsa humbowo hunogona kutariswa nemuchina uchipesana nereferensi yekumisikidza.
Mixed-integer linear programming chishandiso chepakati chekubatanidza optimization, uye bepa rinoti rinoshandiswa mukuwanda kwezvishandiso zvepasirese. Kugadzira magadzirirwo ayo ari computational anobudirira kwakaoma. Mamiriro makuru emitauro anogona kubatsira kuburitsa kana kusimbisa iwo maumbirwo, asi dhizaini inoita seine musoro inogona kutadza kuchengetedza dambudziko rekugadzirisa. Muchigadziro ichocho, kunetseka hakusi kungoti kana modhi yakatsanangurwa inogona kuitiswa, asi kuti inomiririra basa rakafanana rekugadzirisa muzviitiko zviri kutariswa. Bepa rinoisa izvi semubvunzo wekuchengetedza zvinoreva padanho rekuumba.
Vanyori vanocherekedza kuongororwa, pane kugadzira chizvarwa chega, sechinhu chepakati chinodiwa kune yakavimbika otomatiki. FLARE inogadzirisa dambudziko iri nekuunza tsananguro inovaka yeyakasanganiswa-integer linear programming reformulation inogona kugadzirwa muLean, mubatsiri wehumbowo. Iyo sisitimu inosanganisa LLM-yakavakirwa mumiriri neLean kuti aone gadziriso yakarongwa ichipesana nereferensi yekugadzira. Kana FLARE ikabvuma kuvandudzwa, bepa rinoti rinoburitsa chitupa chinotariswa nemuchina. Chitupa ichocho chakagadzirirwa kuita kuti mhedzisiro ionekwe nehurongwa hwepamutemo pane kungovimba nenhamba yekuedza kana tsananguro yeLLM. Iko kugadziridzwa saka kunopa marongero ayo tsamba yakarongwa inogona kutaurwa nekutariswa. Chinangwa chayo mukutsanangurwa kwekufamba kwebasa ndechekufumura hukama hunogoneka pakati pemagadzirirwo maviri aya.
Vanyori vanoongorora FLARE paFormulationBench, ine makumi maviri matambudziko uye 109 maumbirwo. Vanoshuma kuti FLARE inowana 100% kunyatsoita pane iyo 's NP-yakaoma subset. Iyo bepa zvakare inosuma FLARE-NL, inotsanangurwa seyekukurumidza uye yakachipa LLM proxy yemamiriro ezvinhu apo vimbiso yepamutemo isingadikanwi. Benchmark ndiyo marongero echiyero chakashumwa, saka mhedzisiro inofanirwa kuverengwa pamwe chete nerondedzero yedataset uye chiyero chekuongorora. Pepa rinoshandisa muongorori kuratidza nzira inodiwa kushandiswa.
Zvinoenderana nekwakabva, FLARE-NL inofananidza kurongeka kweFLARE pakuongorora asi haiburitse chitupa. Kwainobva hakutauri kuti mangani matambudziko ndeeNP-yakaoma subset, tsvaga iwo chaiwo mabhesi, kana kutsanangura chero kugadzirwa kwekugadzirwa. Izvi zvinodiwa zvinosiya miganhu yeratidziro yakashumwa yakavhurika. Vanoitawo mutsauko pakati pehurongwa hwepamutemo uye proxy yakakosha pakududzira mhedzisiro.
Nei zvichikosha
Iro basa rinogadzirisa dambudziko rekuvimbika mukuyedza kugadzirisa otomatiki modhiyo nemamodheru emitauro. Nhamba dzebvunzo dzinogona kusaratidza kuti kuumbwa kunoshanda kune zvakajairika dambudziko zviitiko; Zvitupa zveFLARE zvakagadzirirwa kupa humbowo hwakasimba kana kurongeka kuchivimbisa nyaya.
Mupiro wepakati webepa inzira yekuyedza kana AI-yakagadzirwa optimization yekugadzira inochengetedza dambudziko rainomirira kumiririra. Vanyori vanoti nzira dziripo dzinoongorora maumbirwo enhamba uye hadzifungi nezvematambudziko akajairwa. Musiyano iwoyo unokosha nekuti kupasa bvunzo dzenhamba dzakasarudzwa hakuiti, pachezvako, kuratidza chokwadi chechiitiko chimwe nechimwe chakakodzera. Humbowo hwepamutemo hunogamuchirwa naLean hunogona kupa hwaro hwakasimba hwekuvimba kana modhi yekuvandudza ichishandiswa kutungamira sarudzo dzakakosha. Kukosha kunoshanda kwemusiyano iwoyo kunotsamira pane chiri kuvandudzwa uye kuti kuvimba kwakadini kunodiwa mumuenzaniso unobuda. Nharo yekwakabva ndeye humbowo chiyero chekugadzirisa, pane chirevo chekuti basa remuenzaniso rinoda mubatsiri wehumbowo.
Chitupa chinoshandura basa rinogona kutamba neLLM mukufambiswa kwebasa. Panzvimbo pekubata gadziriso yakarongwa yemuenzaniso kana tsananguro inoperekedza sehumbowo hwekupedzisira, mushandisi anogona kuda kuti chikumbiro chishandurwe kuita chirevo chinogona kutariswa nemubatsiri. Izvi zvinogadzira kupatsanurwa kwakajeka pakati pechizvarwa nekusimbisa: iyo LLM inogona kutsvaga kana kuvaka reformulation, uku Lean achitarisa kana chirevo chepamutemo chinotevera. Kupatsanurwa kwevashandi uku hakubvisi kudiwa kwekutsanangudza chikumbiro nemazvo kana kupa chirevo chakakodzera. Izvo, zvisinei, inozivisa apo cheki yepamutemo inotarisirwa kugara mukuita kunotsanangurwa nebepa.
Kwauri kunopa izvi sechinhu chinogonesa otomatiki inovimbika, kwete seumboo hwekuti mamodheru emitauro akazvimiririra anovimbisa kurongeka kwemasvomhu. Mhedzisiro yacho inovimbisa asi inoratidzwa nepadiki. Iyo 100% nhamba ndiyo mhedzisiro yevanyori paFormulationBench, kwete chiyero chakajairwa chekufunga kweLLM kana kuvimbika kwehumbowo pakugadzirisa. Mamiriro ekuenzanisa naizvozvo akakosha pakududzira chikamu. Inoratidza uko vanyori vakayera kuita, nepo izvo zvisipo zvichidzikamisa kuenzanisa uye kuwedzera kwakawedzera.
The ine 20 matambudziko uye 109 maumbirwo, uye tsime haapi ruzivo pamusoro sei mumiririri nyaya idzodzo, zvakaoma sei asiri-NP-akaoma nyaya akanga, kana kuti zvikanganiso zvakaitika. Iro bepa ndeye arXiv kuendesa, uye kwainopa hakuna humbowo hwekuongorora nevezera, kudzokorora kwekunze, kutorwa kwemushandisi, kana zvakagadziridzwa mhedzisiro mukushanda kwekugadzirisa. Makapu iwayo ane chekuita nesimba uye kukura kwehumbowo, kwete musiyano uripo pakati pekuyedzwa kwenhamba nekusimbisa zviri pamutemo zvinotsanangurwa nebepa. Humwe humbowo hwaizodiwa usati watora mhedziso nezve mashandisiro enguva dzose.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Mhedzisiro yakashumwa inouya kubva padiki bhenji rematambudziko makumi maviri uye 109 maumbirwo, uye tsime haripe saizi yeNP-yakaoma subset kana kudonongodza nzira dzekukwikwidza. Kudzokorora pamatambudziko akakura uye akasiyana-siyana ekugadzirisa, pamwe nehumbowo nezve mutengo, kukurumidza uye kubatanidzwa mune chaiyo yekumisikidza workflows, inozoona kuti nzira yacho inobatsira zvakadii.
Mubvunzo wekutanga ndewekuti kurongeka kwakataurwa kunobata kunze kweFormulationBench. Inobatsira yekutevera ongororo yaizoda kuyedza zvakanyanya matambudziko, akasiyana masitayipi ekugadzira uye makirasi akakura eakasanganiswa-integer mutsara zvirongwa. Vanofanirawo kuzivisa ukuru uye kuumbwa kweNP-yakaoma subset, nzira dzekuenzanisa chaiyo, zviitiko zvekukundikana uye chiyero chekupindira kwevanhu kunodiwa. Pasina iwo ruzivo, bhenji mhedzisiro inogadza ratidziro inovimbisa asi kwete kuvimbika kwakakura. Kuedza kwakadai kuchaita kuti zvive nyore kusiyanisa mashandiro pane zvakashumwa muunganidzwa kubva pakuita pane yakakura modhi mamiriro anokurudzira basa. Kwaizoitawo kuti kururama kwakashumwa kuve nyore kududzira.
FLARE-NL inobvumidza kuongororwa kwakapatsanurwa nekuti yakagadzirirwa zviitiko apo vimbiso yepamutemo isingakodzeri. Iyo sosi inoitsanangura inokurumidza uye yakachipa kupfuura FLARE uye inoti inoenderana nekurongeka kweFLARE, asi zvakare inotaura pachena kuti haiburitse chitupa. Vashandisi vanozoda humbowo nezve kumhanya uye musiyano wemitengo, kangani iyo proxy isingabvumirani neyakajeka verification pamatambudziko akaoma kana asina kujairika, uye kana kusavapo kwechitupa kuri kutengeserana kunogamuchirwa. Kwakabva hakupi zviyero izvozvo kana mitemo yesarudzo. Kuenzanisa kwakakodzera saka hakusi pakati pekumhanya maviri kana mitengo, asi pakati pehumbowo imwe neimwe modhi inoita kuti iwanikwe kumushandisi. Kunobva kunosiya iyo sarudzo yekushandisa isina kugadziriswa.
Kuendesa kunoshanda kunozobvawo pane zvinopfuura kuongorora humbowo. Chinyorwa chebepa hachitauri kuti kuumbwa kwereferenzi kunosarudzwa sei, humbowo husina kukwana kana hukakundikana hunobatwa sei, zviwanikwa zvekombuta zvinodikanwa, kana kuti nzira yacho inoshanda nesoftware iripo. Iyi mibvunzo inonyanya kukosha kune chero mafambiro ebasa umo magadzirirwo anogadziridzwa, kuturikirwa kana kuongororwa kakawanda. Rondedzero iripo hairatidze kuti nzira yacho yaizoita sei mumamiriro ezvinhu aya.
Basa remangwana rinofanirwa kujekesa kana zvitupa zvinoramba zvichidzoreka sezvo maumbirwo achikura zvakanyanya uye kana sisitimu ichigona kuona zvikanganiso zvinomuka isati yagadzirwa. Kusvika panguva iyoyo, FLARE inonzwisiswa zvakanyanya senzira yekutsvagisa yeAI-yakabatsirwa verification, ine zvinokurudzira mabhenji mhedzisiro asi zvine muganho pane izvo zvinogona kutaurwa pamusoro pekushandiswa kwepasirese. Humbowo hwazvino hunotsigira kutarisisa nzira uye kune chinangwa chayo chekusimbisa, uku ichisiya mibvunzo yekutumirwa kuti iongororwe gare gare. Ndiwo muganho wemhedziso dzinotsigirwa nekwakabva sekutsanangurwa kwazvinoitwa.