Buyela Ezindabeni
UkuqambaAI Understanding ukwaziswa

Ibhentshimakhi ye-ChemDIRT ithola izinguquko zokusebenza kwekhemistri LLM ngemiyalo nokumelwa kwamangqamuzana

Ibhentshimakhi entsha ye-arXiv ihlola amamodeli ezilimi amakhulu agxile kukhemistri kuyo yonke imiyalo ehlukahlukene, izethulo zamangqamuzana nezigaba zemisebenzi eyisishiyagalombili, kubika ukuzwela okukhulu okusheshayo, ukuncika kokumeleleka kanye nokusebenza okungalingani.

5 min readRead the primary source
Source-page capture accompanying ChemDIRT benchmark finds chemistry LLM performance changes with prompts and molecular representations
Idokhumenti yomthombo oyinhlokoUmthombo urekhodiwe
Umshicileli
arxiv.org
Isixhumanisi somthombo
arxiv.orghttps://arxiv.org/abs/2608.21504
Uhlobo lomthombo
Idokhumenti eyisisekelo — isimemezelo esisemthethweni, iphepha, ukugcwalisa, noma ikhasi lomuntu wokuqala esilifunda ngokuqondile.
UmongoQonda lokhu ngemizuzwana engama-60

Qala lapha

Imigomo ebalulekile

Imodeli Yolimi Olukhulu (LLM)
Imodeli yolimi eqeqeshwe ku-massive text corpora ukuze ikhiqize futhi ihlaziye umbhalo.
Ibhentshimakhi
Ukuhlolwa okujwayelekile noma isethi yedatha esetshenziselwa ukukala nokuqhathanisa ukusebenza kwemodeli.
Ukuqina
Ikhono lemodeli lokugcina ukusebenza ngaphansi komsindo, amashifu, noma okokufaka okuphikisayo.
ZihloleImibuzo Ecacisiwe yamamodeli e-AI

Kwenzekeni

Abacwaningi bethule i-ChemDIRT, ibhentshimakhi eklanyelwe ukuhlola ukuthi amamodeli ezilimi amakhulu angabonisana ngokungaguquki mayelana nekhemistri lapho amagama enkinga noma indlela ulwazi lwamakhemikhali olumelelwa ngayo lushintsha. Leli phepha lithi lihlola ukunemba nokungaguquguquki kukho konke ukuhlukahluka okulawulwayo emiyalweni nasezethulweni zamangqamuzana, okuhlanganisa izigaba eziyisishiyagalombili zemisebenzi yamakhemikhali. Ababhali bayo babika ukuzwela okukhulu okusheshayo, ukumelela ukumelela nokusebenza okungalingani kuyo yonke imindeni yemisebenzi phakathi kwesethi ehlukahlukene yamamodeli wemithombo evulekile nevaliwe.

I-ChemDIRT imele i-Diversified Instruction, Representation, kanye ne-Task . Ababhali bayethula njengohlaka lokuhlola lwamamodeli wezilimi amakhulu agxile kumakhemikhali, ukusetshenziswa kwawo kuzilungiselelo zesayensi kunwetshiwe. Umthombo ubeka inkinga njengomkhawulo wamabhentshimakhi ekhemistri akhona: amaningi ahlola isethi ewumngcingo yemisebenzi futhi asebenzise izinhlobo ezilinganiselwe zokwakhiwa kwenkinga noma ukumelwa kwamakhemikhali. Ngokombono wababhali, lokho kunganikeza isithombe esingaphelele sekhono lemodeli lokucabanga ngokungaguquki.

Ibhentshimakhi ishintsha okokufaka okubili okungathinta kakhulu impendulo yemodeli yolimi: umyalo osetshenziselwa ukudala inkinga kanye nokumelela okusetshenziselwa ukuveza ulwazi lwamakhemikhali. I-abstract ayicacisi izinguquko zamagama eziqondile noma izethulo ezifakiwe. Ithi i-ChemDIRT ikala kokubili ukusebenza nokungaguquguquki ngaphansi kokuphazamiseka okulawulwayo, kunokuthembela kuphela kumphumela owodwa osuka kufomethi eyodwa egxilile.

Leli phepha lithi ukuhlola kuhlanganisa izigaba eziyisishiyagalombili zemisebenzi yamakhemikhali futhi kufaka phakathi isethi ehlukahlukene yama-LLM emithombo evulekile nevaliwe. Umthombo onikeziwe awuyisho imindeni yomsebenzi noma amamodeli, futhi awunikezi izibalo zesethi yedatha, izibalo zokunemba, amaphuzu afanayo noma imiphumela eyisisekelo. Ngakho-ke isekela isimangalo sokuthi abacwaningi benze ukuhlola okubanzi, okuhlukahlukene, kodwa hhayi ukuqhathanisa okuningiliziwe kwezinhlelo ngazinye.

Umphumela obikiwe uqondisa kunenombolo emthonjeni otholakalayo. Ababhali bathi bathole ukuzwela okukhulu emizamweni, ukuncika ekumeleleni kwamangqamuzana nokusebenza okungalingani kuyo yonke imindeni yomsebenzi. Ngamagama asebenzayo, i-abstract ithi umphumela wemodeli kufomethi yebhentshimakhi yekhemistri eyodwa akufanele uthathwe ngokuzenzakalelayo njengobufakazi bokucabanga kwamakhemikhali okuzinzile kuwo wonke amanye amafomethi. Umthombo awuqinisekisi ukuthi kungani amamodeli athile ayezwela noma noma iyiphi isistimu yayisebenza kahle kakhulu kunezinye.

Imininingwane yomthombo: arxiv.org ↗

Kungani kubalulekile

Amabhentshimakhi ekhemistri asebenzisa ifomethi eyodwa engaguquki angenza imodeli ibonakale inekhono kakhulu kunalokho esetshenziswayo. Umnikelo omaphakathi we-ChemDIRT, ngokomthombo, uwukukala ukuqina ngaphansi kokuhlukahluka uhlelo lwamakhemikhali olungahlangabezana nalo ngaphandle kokuhlolwa okukodwa okujwayelekile. Lokho kungase kunikeze abacwaningi nabasebenzisi isisekelo esingcono sokwahlulela ukuthi ama-LLM ekhemistri akhiqiza imiphumela ezinzile noma ancike kakhulu emagameni nakwifomethi yokufaka.

Ukubaluleka okuyinhloko kuyi-methodological. I-LLM yekhemistri ingaphendula ngendlela efanele lapho inkinga yethulwa ngendlela eyodwa evamile kuyilapho ikhiqiza impendulo ehlukile noma engalungile ngemva kokuguqulwa kwamagama noma ushintsho endleleni i-molecule efakwe ngayo ikhodi. Uma lokho kuziphatha kungalinganiswa, ibhentshimark ingase ivuze ukujwayelana kwefomethi kanye nokucabanga kwamakhemikhali adlulisekayo. Umklamo we-ChemDIRT uqondise ngqo lelo gebe ngokuphatha ukungaguquguquki njengento yokuhlola eduze nokunemba.

Lokhu kubalulekile kubacwaningi abaqhathanisa amasistimu. Isilinganiso esisodwa sokulinganisa singafihla amandla okungalingani: imodeli ingase yenze kahle kweminye imisebenzi yekhemistri futhi ingasebenzi kahle kweminye, noma yenze ngokuqinile kuphela izethulo ezithile. Umbiko womthombo wokusebenza okungalingani kuyo yonke imindeni yemisebenzi uphakamisa ukuthi amaphuzu ahlanganisiwe kufanele ahunyushwe ngokucophelela. Ukuhlolwa okuhlukahlukene kungasiza onjiniyela bamamodeli ukukhomba lapho kudingeka khona ukuqeqeshwa okwengeziwe, ukuphatha ukumelela noma ukuvikela, nakuba i-abstract ingabonisi ukuthi yikuphi ukungenelela okungabhekana nobuthakathaka obuboniwe.

Lolu daba lubalulekile nakubantu abacabangela umsebenzi wesayensi osizwa yi-AI. Amamodeli ekhemistri angase asetshenziselwe ukuhlela ulwazi, ukuphendula imibuzo yobuchwepheshe noma izinqumo zocwaningo ezisekelayo, kodwa umthombo awubonisi ukuthi ukusebenza kwe-ChemDIRT kubikezela impumelelo endaweni yaselabhorethri noma yokukhiqiza. Ukuqina ekuphazamisekeni kwebhentshimakhi kuwubufakazi obuwusizo mayelana nekhwalithi yokuhlola; akubona nje ubufakazi bokuthi imodeli iphephile ezinqumweni zesayensi ezingagadiwe.

Ngenkambu ye-AI ebanzi, iphepha libonisa ukuthi kungani ukuhlola okuqondene nesizinda kungenakuncishiswa kube izikolo ezijwayelekile zemodeli yolimi. I-Chemistry ihlanganisa izethulo ezikhethekile nezinhlobo zemisebenzi ezingase zidalule izindlela zokuhluleka ezifihlwe izivivinyo ezijwayelekile zokuphendula imibuzo. Umthombo usekela isiphetho esincane sokuthi i-ChemDIRT inikeza indlela ehlukahlukene yokuhlola ukuziphatha kwe-chemistry-LLM. Ayiqinisekisi ukuthi ibhentshimakhi iqinisekile noma ukuthi okutholakele kusebenza kuzo zonke izizinda zesayensi.

Interactive Mechanism

I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela

Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
I-Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

Ongakubuka ngokulandelayo

Iphepha liyi-arXiv, futhi umthombo onikeziwe uqukethe kuphela i-abstract yawo. Ayiwahlonzi amamodeli ahloliwe, izigaba zomsebenzi, izethulo zamangqamuzana, osayizi bedathasethi, imiphumela yezinombolo noma izindlela zokuqhathanisa. Leyo mininingwane iyadingeka ukuze kuhlolwe amandla kanye nokujwayelekile kokutholakele. Ukucutshungulwa kokulandelela kufanele kuhlole ukuthi i-ChemDIRT iyakwazi ukukhiqizwa kabusha, ukuthi ukuphazamiseka kwayo kubonisa ukuhamba komsebenzi wangempela wekhemistri, nokuthi amamodeli asebenza ngokungaguquki kubhentshimakhi nawo enza ngendlela ethembekile yini elabhorethri, imisebenzi yomtholampilo noma yezimboni.

Okokuqala okungaziwa ukwakheka kwebhentshimakhi. Ikhasi elinikeziwe le-arXiv linikeza isihloko kanye nokungabonakali kodwa hhayi izindlela ezigcwele zephepha, ngakho-ke abafundi abakwazi ukunquma ukuthi yiziphi izigaba zomsebenzi wamakhemikhali eziyisishiyagalombili ezasetshenziswa, ukuthi izethulo zamangqamuzana zakhethwa kanjani, noma ukuthi ukuhluka kwemiyalelo kwakhiwe kanjani. Lezo zinketho zizoba nomthelela ekutheni ukuhlola kulinganisa ukuqina kwangempela noma ikakhulukazi ukuzwela izinguquko zokufometha kokwenziwa.

Okwesibili okungaziwa isilinganiso nokuqhathaniswa kokuhlola. Umthombo uthi ababhali balinganise amamodeli wemithombo evulekile nevaliwe, kodwa awuwahlonzi noma awusho ukuthi mangaki amasistimu ahloliwe. Ayinikezi futhi amasayizi omthelela wezinombolo, izilinganiso zokungaqiniseki noma ukuhlolwa kwezibalo. Ngaphandle kwaleyo mininingwane, ukuzwela okusheshayo “okukhulu” nokuncika kokumelela akukwazi kukalwa ngokuzimele kuqhathaniswa nomehluko wemodeli kuya kwemodeli, ubunzima bomsebenzi noma ukungcoliswa kwedatha okungenzeka.

Umbuzo wesithathu ungowokuba semthethweni kwangaphandle. Imodeli ehlala ifana kukho konke ukuhlukahluka okulawulwayo kwe-ChemDIRT isengenza izincomo ezingalungile ngokwekhemikhali noma ezingaphephile kuzilungiselelo ezingamelwe ibhentshimakhi. Umsebenzi wesikhathi esizayo kufanele uhlole ukuthi ingabe amaphuzu ebhentshimakhi ayahambisana yini nezinqumo zochwepheshe, imiphumela eqinisekiswe ngokuhlolwa noma ukusebenza kokugeleza komsebenzi wekhemistri wangempela. Ukuphindaphinda okuzimele nakho kungasiza ekunqumeni ukuthi amaphethini abikiwe ayaqhubeka yini kuzo zonke izinguqulo namasethi edatha.

Okokugcina, bheka ukuthi ingabe i-ChemDIRT iba insiza yokuhlola okwabelwana ngayo noma ihlala iyisiphakamiso sephepha elilodwa. Umthombo awusho ukuthi idatha yebhentshimakhi, ikhodi noma izintambo zokuhlola ziyatholakala yini esidlangalaleni. Lokho okweqiwe kukhawulela ukuqinisekiswa okusheshayo nokutholwa okungokoqobo. Kuze kube yilapho kuhlolwa iphepha eligcwele nezinto ezisekelayo, isiphetho esivikeleke kakhulu ukuthi i-ChemDIRT ikhomba inkinga yokuhlola enengqondo futhi ibike ubufakazi bokungazinzi, kuyilapho ubukhulu kanye nemiphumela yezwe langempela yalokho kungazinzi kusazosungulwa.

Imihlahlandlela ehlobene nemibuzo

Amamodeli e-AI AchaziweChatGPT ne-LLMsUkuqeqeshwa kwe-AIUkuziphatha kwe-AIHlola okwaziyo — zama imibuzo ye-AI yamahhalaBheka igama le-AI kuhlu lwethu lwamagamaLandela i-tracker yokukhishwa kwemodeli ye-AI
Uthole lokhu kuwusizo?