MemHarness – uhlaka lokwakha kabusha inkumbulo oluqalwe yi-Shanghai AI Lab nabanye Uyasebenza

I-MemHarness iwuhlaka lokwakha kabusha inkumbulo lwama-LLM Agent, olwethulwe ngokuhlanganyela yi-Shanghai AI Lab namanyuvesi ahlanganisa i-Zhejiang University, i-Fudan University kanye ne-Shanghai Jiao Tong University. Ama-Agent amaningi athuthukiswe ngenkumbulo ajwayele ukufaka ngokuqondile okuhlangenwe nakho kwangaphambilini okutholiwe kumongo; uma okuhlangenwe nakho okudala kungahambisani nesimo samanje, kungadala ukudluliselwa okungekuhle. Igqugquzelwa indlela inkumbulo yabantu eyakhiwa ngayo kabusha, i-MemHarness ifaka izinyathelo ezicacile zokuhlaziya nokwakha kabusha phakathi kokusesha inkumbulo nokwenza isenzo. Ihlola okuhlangenwe nakho kwangaphambilini ngokomongo wamanje, ikugcine, ikubhale kabusha noma ikulahle, futhi iqeqeshwe nge-GRPO kusukela ekuqaleni kuye ekugcineni ngaphandle kwedatha eyengeziwe enelebula eyenziwe abantu.

Ukusesha inkumbulo: I-Agent ikhiqiza umbuzo ngokusekelwe kulokho ekubonayo manje, bese ibuyisa okuhlangenwe nakho kwangaphambilini okuhlobene kakhulu kanye nezimo zakho zemvelaphi kusizindalwazi senkumbulo samavekhtha i-Milvus.
Ukuhlaziya nokwakha kabusha: Imodeli yenqubomgomo eyodwa iqhathanisa isimo semvelaphi yenkumbulo nesimo samanje, ihlole ukufaneleka kwaso, bese ibhala kabusha okuhlangenwe nakho kube ulwazi oluqondiswe esimweni, noma inqume ukuthi akufanele kusetshenziswe bese iyakulahla.
Ukukhiqiza izenzo: Ngokusekelwe olwazini olwakhiwe kabusha noma ekucabangeni okuzimele, ikhiqiza izenzo zemvelo ezingenziwa.
Ukubuyisela nokususa okuhlangenwe nakho: Ngemva komjikelezo ngamunye wokuxhumana, umzila ufingqwa ube okuhlangenwe nakho bese ubhalwa kusizindalwazi senkumbulo. Kwenziwa ukususa impinda ngokomqondo, bese inkumbulo enenani eliphansi isuswa ngezikhathi ezithile ngokusekelwe ekusizeni kwayo.
Ukuqeqeshwa kusukela ekuqaleni kuye ekugcineni: I-GRPO ithuthukisa ngokuhlanganyela ukusesha, ukwakha kabusha nokukhiqizwa kwezenzo, ngaphandle kwedatha ezimele enelebula yesigaba sokwakha kabusha.

Landela i-WeChat uphendule ngokuthi “开源” ukuze ujoyine iqembu lokuxoxa ngamaphrojekthi e-AI anomthombo ovulekile
Inqubo yokuthatha izinqumo enezigaba ezinhlanu: I-MemHarness ihlukanisa ukuthatha izinqumo okuqondiswa inkumbulo kube izigaba ezinhlanu ezilandelanayo: ukubuka imvelo, ukusesha okuhlangenwe nakho, ukuhlaziya inkumbulo, ukwakha kabusha inkumbulo ngokomongo kanye nokukhiqiza isenzo. Lokhu kulingisa inqubo yabantu yokusesha, ukuhlola nokwakha kabusha, esikhundleni sendlela yakudala lapho i-Agent idlala inkumbulo futhi ngokuqondile.
Indlela yokwakha kabusha inkumbulo ngokomongo: Imodeli yenqubomgomo ihlanganisa incazelo yomsebenzi, umlando wamanje, okuhlangenwe nakho okuseshiwe kanye nesimo semvelaphi yakho kube umongo wokwakha kabusha. Ngokuqhathanisa isimo semvelaphi yangaphambilini nesimo samanje, ibona umehluko, igcine ulwazi oludlulayo, ibhale kabusha okuqukethwe okungahambisani, noma ikhiphe umaka othi ukuze yenqabe leyo nkumbulo bese ibuyela ekucabangeni okuzimele.
Ukwakhiwa kwemodeli yenqubomgomo eyodwa: Izigaba ezintathu zokukhiqiza umbuzo wokusesha, ukuhlaziya nokwakha kabusha inkumbulo kanye nokukhiqiza isenzo esingenziwa zabelana ngamapharamitha emodeli yenqubomgomo efanayo. Lokhu kususa isidingo sokuqeqesha inethiwekhi yenani noma idatha egadiwe ngokwehlukile yemojuli yokwakha kabusha, futhi kuhlanganisa ukusetshenziswa kwenkumbulo nokucabanga kokuthatha izinqumo ngaphakathi kwemodeli eyodwa.
Ukuqeqeshwa kwe-GRPO kusukela ekuqaleni kuye ekugcineni: Ngenxa yokuthi ulwazi lokuqondisa ukwakha kabusha alunamalebula ayiqiniso, i-MemHarness isebenzisa i-GRPO ukuthuthukisa lonke uchungechunge lokusesha, ukwakha kabusha nokwenza isenzo kusukela ekuqaleni kuye ekugcineni. Umvuzo wakhiwa yimiphumela yomsebenzi eyivelakancane nemivuzo yefomethi. Inzuzo ejwayelekile yeqembu isetshenziselwa ukwabela ikhredithi yezinga lomzila kuwo wonke ama-token, ngesikhathi esisodwa kuqeqeshwa amakhono okucabanga, okwakha kabusha nawokukhiqiza izenzo.

Kopisha inqolobane bese udala indawo: Sebenzisa git clone https://github.com/KnowledgeXLab/MemHarness.git bese udala indawo ye-Conda ene-python==3.12, ufake ngokulandelana i-vLLM, i-Flash Attention 2 kanye ne-MemHarness uqobo.
Sebenzisa insizakalo yokushumeka: Qalisa vllm serve BAAI/bge-m3 --port 8001 ukuze uqale imodeli yokushumeka i-BGE-M3, enikeza isevisi yokufaka amakhodi ama-vekhtha kusizindalwazi senkumbulo i-Milvus.
Faka izindawo okuqondiswe kuzo: Ngokuya ngezidingo zomsebenzi, faka indawo yokuxhumana ye-ALFWorld noma ye-WebShop, bese ulanda amafayela omdlalo kanye nezitholi eziqeqeshwe kusengaphambili.
I-SFT yokuqala ebandayo (uyazikhethela): Qalisa scripts/build_*_coldstart_data.py ukuze wakhe idatha yokuqala ebandayo, bese usebenzisa scripts/cold_start_sft.sh ukuze uqondanise imodeli nefomethi yokuxhumana kanye nefomethi yesifinyezo senkumbulo.
Ukuqeqeshwa kwe-GRPO kusukela ekuqaleni kuye ekugcineni: Qalisa run_scripts/train_alfworld.sh noma run_scripts/train_webshop.sh. Uhlaka luzoqalisa ngokuzenzakalela isizindalwazi samavekhtha enkumbulo, lwenze ukusesha kwe-Agent, lubuyisele okuhlangenwe nakho futhi lususe inkumbulo enenani eliphansi, bese luqedela ukuqeqeshwa kwe-GRPO.

Ukwakha kabusha kudlula ukudlala kabusha: Izinyathelo ezicacile zokuhlaziya nokwakha kabusha ziguqula izingcezu zenkumbulo ezingaguquki zibe iziqondiso eziqondaniswe nesimo futhi ezizwelayo kumongo, kugwenywe ukudluliselwa okungekuhle.
Imodeli encane idlula enkulu: Imodeli enamapharamitha angu-7B idlula i-Gemini-2.5-Pro ngo-23.1% ku-ALFWorld nango-39.7% ku-WebShop.
Ukuqina okuphezulu ku-OOD: Ezimweni ezingaphandle kokusatshalaliswa, izinga lokuphumelela elijwayelekile lifinyelela ku-85.9%, liphakeme kakhulu kuno-76.3% wokudlalwa kabusha kwenkumbulo kwasekuqaleni.
Ukuthuthukiswa kokucabanga okungaqondile: Ngisho noma inkumbulo ivaliwe ngesikhathi sokuhlolwa, inhloso yokuqeqesha yokwakha kabusha ikhuphula izinga lokuphumelela lenqubomgomo eyisisekelo lisuka ku-76.4% liye ku-83.0%, ithuthukise ngokuyisisekelo ikhono lokucabanga langaphakathi le-Agent.
Akudingeki amalebula engeziwe: Ikhono lokwakha kabusha livela ngokwemvelo ngokuqeqeshwa kwe-GRPO kusukela ekuqaleni kuye ekugcineni, ngaphandle kokuncika kudatha yokugada yokwakha kabusha ebhalwe abantu.

Inqolobane ye-GitHub: https://github.com/KnowledgeXLab/MemHarness
Inqolobane yamamodeli e-HuggingFace: https://huggingface.co/KnowledgeXLab/MemHarness
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2607.28272

Imisebenzi yasekhaya yamarobhothi ahlakaniphile: Ezindaweni zasekhaya ezifana ne-ALFWorld, i-Agent ingakha kabusha okuhlangenwe nakho kwangaphambilini kokuthatha izinto nokuhlanza, ijwayelane nezakhiwo zamakamelo ezahlukene ukuze iqedele imisebenzi yasekhaya eyinkimbinkimbi.
Ukuthenga ku-inthanethi ngokuzenzakalela: Ezinkundleni ze-e-commerce ezifana ne-WebShop, yakha kabusha okuhlangenwe nakho kokuthenga kwangaphambilini kube amasu amanje okusesha nokuhlunga imikhiqizo, ithuthukise izinga lokuphumelela lokuthenga okuqondiswe emgomweni.
Izingxoxo zesevisi yamakhasimende ezihlakaniphile: I-Agent yesevisi yamakhasimende isesha amathikithi omlando afanayo bese yakha kabusha isixazululo, igweme ukusebenzisa impendulo endala ngokuqondile bese iphendula ngokungahambisani nombuzo.
Ukusiza ekuthuthukisweni kwekhodi: I-Agent yokuhlela yakha kabusha okuhlangenwe nakho kwangaphambilini kokulungisa iziphazamisi, ikuvumelanise nomongo wekhodi yamanje, futhi inikeze iziphakamiso eziqondile zokulungisa esikhundleni sokukopisha uhlelo oludala.
Ukuhlelwa kokuhlolwa kocwaningo: I-Agent yokuhlola yakha kabusha amapharamitha nemiphumela yokuhlolwa kwangaphambilini ibe iziphakamiso zokumiswa ezifanele kakhulu zezimo zokuhlolwa zamanje, inciphise izindleko zokuzama nokwenza amaphutha.