VeriLoop E2 Uyasebenza

I-VeriLoop E2 iyimodeli engu-27B yokuqeqeshwa kwangemuva, ekhishwe njengomthombo ovulekile yi-Robotics Laboratory yase-Shenzhen International Graduate School, Tsinghua University. Isekelwe ku-Qwen3.8-27B futhi iqondiswe ekucabangeni ngekhodi, izibalo ne-physics, inelayisensi ye-Apache-2.0 kanye nomongo womdabu ongu-262K. Indlela eyinhloko ye-VGR ye-VeriLoop E2 ihlukanisa “iziphakamiso zemodeli” “ekuqinisekisweni kwangaphandle/nasekuphikeleni”, futhi isimo somuntu ozokhethwa sithunyelwa kuphela uma izixhumanisi zobufakazi ezivikelekile zingenakho ukuwohloka futhi okungenani kukhona ukuthuthuka okukodwa okuqinile; uma kungenjalo, kubuyiselwa emuva. Imodeli ayikwazi ukuzifakazela ukuthi iphumelele.

Imodeli yokuqeqeshwa kwangemva evulekile: imodeli ye-27B esekelwe ku-Qwen3.8-27B, enesisindo nesihlukanisi samathokheni esikhishwe ngaphansi kwe-Apache-2.0, umongo wendabuko wamathokheni angu-262,144, futhi ingasebenza nge-vLLM 0.17.0 ngendlela ehambisana ne-OpenAI.
Amakhono ekhodi nama-ejenti: ihlanganisa imisebenzi efana nokulungiswa kwesoftware ezingeni lenqolobane kanye nokusebenzisa i-terminal/amathuluzi; amaphuzu ashicilelwe afaka i-SWE-bench Pro 76.2%, i-Terminal-Bench 2.1 88.8%, ne-DeepSWE v1.1 64.6%.
Ukucabanga kwezibalo nesayensi: iqondiswe ezibalweni zokuncintisana nasemibuzweni yesayensi enzima, futhi kubikwe i-AIME 2026 98.3%, i-ArXivMath 100%, i-GPQA Diamond 93.9%, ne-Apex 2025 89.6%.
Umboniso wesayensi ovalwe ngobufakazi: kushicilelwe isitifiketi esiqinile esinobukhulu obunomkhawulo sesilinganiso sama-zero e-Riemann ζ esingu-67.350003708785593%, kanye nohlaka lwe-asymptotic graviton tomography T1–T9 ne-suite yezitifiketi esebenzisekayo ephumelele ngo-6/6.
Inkontileka yobunjiniyela ehlolisisekayo: kushicilelwe i-Golden Rules eziyi-14, ezibophezela imisebenzi yesikhathi eside ukuthi ihlale inomkhawulo, ihlolwe futhi ilandeleke; futhi ivimbela ngokusobala ukwenza impumelelo mbumbulu, ukuhlehlisa isabelomali sezinsiza, noma ukweqa izivumelwano zasendaweni.
Imikhawulo ecacile: i-VeriLoop Harness yenguqulo yokukhiqiza, ukuhlelwa kwangasese nedatha yokuqeqesha akufakiwe; izinto zesayensi nelayisensi yemodeli kuhlukene, futhi imiphumela ayikwazi ukuhunyushwa ngokuthi i-Riemann hypothesis noma indida yolwazi lwemigodi emnyama isixazululiwe.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI anomthombo ovulekile
Indlela evamile ye-VGR: I-VeriLoop-Governed Recurrence imodela ukuxazulula umsebenzi owodwa njengokudluliselwa kwezimo eziphakanyisiwe ezizungeze isimo esigciniwe: imodeli isebenza njenge-proposer efundile ekhiqiza ukususwa okuphakanyisiwe kweziguquguquko ezifihlekile noma ama-artifact aphakanyisiwe, i-decoder engaguquki ikuguqulela kube umphumela, bese i-verifier yangaphandle ibuyisa izinga lobufakazi obuvikelwe ngaphansi kwenkontileka engaguquki; kuphela lapho izinga eliphakanyisiwe libusa kancane izinga lamanje futhi okungenani ikoordineyithi eyodwa ithuthuka ngokuqinile lapho kwenziwa i-atomic commit, kungenjalo i-bundle yamanje igcinwa ngokuphelele, bese ukususwa okunqatshiwe kubhalwa kuphela kumemori yokuhlola.
I-partial order evikelwe kanye ne-atomic commit: Inkontileka ye-verifier ichaza uhlu lwezibopho oluhlelwe O=(o1​,…,om​) kanye nevektha yezinga r∈N0m​, lapho inani elincane kwikoor­dineyithi lisho ukwephulwa okusele okumbalwa; i-predicate yokuthumela χt​=1{rt′​≺P​rt​} idinga ukuthi akukho koordineyithi ewohlokayo futhi okungenani eyodwa ithuthuke ngokuqinile, ngakho-ke ukuthuthuka kwesikolo esihlanganisiwe akunakufihla ukuhlehla kwephuzu elilodwa, futhi amakhandidethi angaqhathaniseki noma ahlehlela emuva ayenqatshwa njalo. Iyunithi yokuthumela yi-CertifiedBundle ebopha i-artifact, isimo esifihlekile, imiphumela yokuqinisekisa kanye nesifinyezo; lapho yenqatshwa, yonke i-bundle igcinwa injengoba injalo, okuqinisekisa ukubuyiselwa emuva nokudlalwa kabusha, kanti ukwehla kwe-integer potential kuveza ukuthi inani lama-commit aqinile lilinganiselwe.
Impendulo yobufakazi kanye nokuhlola okunqatshiwe: Ngemva kokuzama ngakunye, i-evidence compiler Γ ibala izici ezilinganiselwe ezibonwa imodeli kusukela “kumphumela wamanje—umphumela wekhandidethi—isinqumo sokuthumela” ngaphambi kokushintsha into yokuqinisekisa yamanje, njenge-compression yezinga nesamba, izilinganiso zokuthuthuka/ukonakala/ukuma, kanye nezinkomba zokuthumela nezinga elingu-zero; lezi zisetshenziselwa kuphela ukwenza isiphakamiso esilandelayo sibe ngaphansi kwezimo, futhi azinalo igunya lokuthumela. Ukususwa okunqatshiwe kungena kumemori yokuhlola Et​, kungathonya ukusesha okulandelayo kodwa akunakubhala phezu kwesimo esivikelwe, ngaleyo ndlela imikhondo yokwehluleka iba “nolwazi kodwa ayihlali”.
I-interface yokugadwa kwangemva kokuqeqeshwa: Ngenxa yokuthi amalebula e-VGR achazwa ekudlulisweni kwezimo ezingeni lomsebenzi, i-E2 isebenzisa amarekhodi ahlelekile angu-1,841,831 ukwenza ukuqeqeshwa kwangemva kokuqeqeshwa: izenzo zenqubekelaphambili eqinile eziqinisekisiwe zinikeza ukugadwa kokulungisa, amakhandidethi angenakuthunyelwa anikeza izibonelo ezingezinhle zokuqhathanisa, kanti ama-artifact ezinga elingu-zero anikeza izinhloso zokugcina zokukhiqiza; ukulahleka kwakhiwa ngokukala itemu le-logistic lokuthandwayo ngamabili, i-action NLL kanye ne-generation NLL, kusetshenziswa i-QLoRA-4/bfloat16, i-gradient checkpointing, umzila ngokwendawo yekhono kanye ne-protected-core replay. Umbiko ugcizelela ukuthi lokhu kwenza ukufunda kuthuthukise “ukusatshalaliswa kweziphakamiso”.

Thola imodeli: Yidede kusuka ku-Hugging Face ngokusebenzisa i-tsinghua-sigs-robot-lab/VeriLoop-E2, bese ugcina i-tokenizer, i-chat template kanye nokucushwa okuza nokhosombe.
Lungisa indawo: Faka i-Python 3.12.x, i-PyTorch 2.10.0 + CUDA 12.9, i-Transformers 4.57.6 ne-Triton 3.6.0, bese usebenzisa python -m pip install "vllm==0.17.0".
Qalisa isevisi: Sebenzisa VLLM_USE_MODELSCOPE=true vllm serve --served-model-name veriloop-e2 --dtype bfloat16 --model-impl vllm --language-model-only --max-model-len 131072 --max-num-seqs 16 --gpu-memory-utilization 0.92 --generation-config vllm --host 127.0.0.1 --port 8001.
Yenza ukuhlola okulula kuqala: Thumela isicelo esincane esine-temperature: 0 ku-http://127.0.0.1:8001/v1/chat/completions, uqinisekise ukuthi impendulo efana no-8001_OK iyabuya ngaphambi kokuxhuma ithrafikhi yangempela.
Shayela ngokuvumelana nephrothokholi ye-OpenAI: Sebenzisa i-/v1/chat/completions ene-model="veriloop-e2"; ohlangothini lwe-Python dala iklayenti usebenzisa i-OpenAI(base_url="http://127.0.0.1:8001/v1", api_key="EMPTY").
Phatha ithempulethi njengengxenye yephrothokholi: Ungawashintshi amasepharetha ezindima, isintaksi sokubiza amathuluzi, incazelo yokuma noma ifomethi yomlando wokucabanga; noma izisindo zingashintshi, lokhu kungashintsha ukusetshenziswa kwamathuluzi nokuziphatha kwesistimu.
Hlela izinsiza zomongo omude: Umkhawulo wendabuko we-tokenizer oshicilelwe ngu-262,144 tokens, kodwa isevisi yokuqinisekisa esesidlangalaleni isebenzisa ama-tokens angu-131,072; amawindi amade adinga ukuthi uhlole ngokwakho inkumbulo ye-GPU ne-KV cache.
Hlukanisa izimvume zemodeli neze-Harness: I-E2 iphethe kuphela iziphakamiso, ukukhishwa kwemiqondo, ukuxilonga nokulungisa; ukwamukelwa kobufakazi, ukuhlolwa kokunqunywa, ukuthumela/ukuhlehlisa, ukumisa nokuphikelela kufanele kusetshenziswe i-verifier/Harness yakho.

Ukuhlukaniswa kokukhiqiza nokuqinisekisa: imodeli inesibopho kuphela sokuphakamisa iziphakamiso, ukuxilonga nokulungisa, kuyilapho i-Harness yangaphandle iyodwa iphethe ukwamukela, ukubuyisela emuva, ukumisa nokuphikelela, ngaleyo ndlela igweme ukuthi imodeli iziqinisekisele impumelelo.
Imithetho yokuthumela evimbela ukubuyela emuva: i-VGR ithumela kuphela lapho “izixhumanisi zobufakazi ezivikelwe zingenakho ukuwohloka futhi okungenani kukhona ukuthuthuka okuqinile okukodwa”; ukuthuthuka okuhlanganisiwe akukwazi ukufihla ukuhlehla kwephuzu elilodwa.
Ukubuyisela emuva kwe-athomu nokwehluleka okuphephile: ukwenqatshwa kwekhandidethi kugcina i-CertifiedBundle ephelele; imikhondo yokwehluleka ingasetshenziswa kabusha kodwa ayibhali phezu kwesimo esiqinisekisiwe.
Izimpawu zokuqondisa zivela ekuqinisekisweni kwangempela: ukuqeqeshwa kwangemva kokwakhiwa kusebenzisa inqubekelaphambili eqinile, izibonelo ezingathunyelwa neziphetho zokugcina ze-zero-rank ukwakha izinhloso ze-preference/NLL.
Ikhono eliqinile elihlanganisiwe: ngosayizi we-27B, ihlanganisa ama-agent ekhodi, ukusebenzisa i-terminal, ukucabanga kwezibalo nesayensi, ibike imiphumela efana ne-SWE-bench Pro 76.2%, AIME 2026 98.3% kanye ne-GPQA Diamond 93.9%.
Umongo omude nesevisi evulekile: umongo wama-tokens angu-262,144 ngokomdabu, izisindo ze-Apache-2.0, kanye nendlela yesevisi ehambisana ne-OpenAI eqinisekiswe nge-vLLM 0.17.0.
Iyahlolwa futhi iyacwaningeka, hhayi ukubika amaphuzu kuphela: iphakheji yobufakazi bokuhlola ibopha kumsebenzi ngamunye ubunikazi, okukhiphayo, amarekhodi omdabu okwenza nokuhlola, impumelelo nokwehluleka, kanye nemethadatha ephelele yobuqotho.

Inqolobane yamamodeli ye-HuggingFace: https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2
Iphepha lobuchwepheshe: https://openreview.net/pdf?id=P6FIQILHwX

Ukulungiswa kwekhodi ezingeni lenqolobane: kusetshenziselwa ukukhiqiza ama-patch emisebenzini efana ne-SWE, ukuhlonza ukuhlehla, bese ukuhlolwa/umhloli kunquma ukuthi agcinwe yini; kulungele izimo zokulungisa lapho “izinguquko kufanele ziqinisekiswe”.
Ama-agent etheminali namathuluzi: ezindaweni ezifana ne-Terminal-Bench, asebenzisa imiyalo, abize amathuluzi, athathe isinyathelo esilandelayo ngokusekelwe ebufakazini, futhi asebenzise ubufakazi bokuphumelela/ukwehluleka ukulawula ukuqhubeka.
Ukucabanga ngezibalo nokuxazulula imincintiswano: kugxile ekudonseni iziphetho ezinzima okufana ne-AIME/ArXivMath; ngemva kokusampula izikhathi eziningi, kugcinwa izimpendulo eziphindaphindekayo ngokusebenzisa umhloli wangaphandle, kuncishiswe inhlanhla yokubheka impendulo eyodwa kuphela.
Ukubala kwesayensi nokuhlola i-physics yethiyori: ocwaningweni lwe-Riemann ζ, i-quantum field theory, ama-graviton e-asymptotic nezinye izindawo, kusiza ukuphakamisa izakhiwo nokuthola izikhala ebufakazini.