Teburin Abubuwan Ciki
1 Gabatarwa
Juyin halittar Manyan Samfuran Harshe (LLMs) da Samfuran Harshe Masu Nau'i-nau'i (MLLMs) sun kawo sauyi ga iyawar tunanin AI, duk da haka akwai manyan kalubale a fagen son fahimtar harshe na halitta da ingantaccen lissafi. Tsarin Agent na AI na yanzu ya dogara sosai akan hanyoyin tunani na waje kamar Sarkar Tunani (CoT) da Maimaita Tunani (IoT), waɗanda ke haifar da manyan farashi na token kuma sun gaji iyakokin LLM.
Tsarin Binciken Tunani (INoT) da muka gabatar yana magance waɗannan iyakokin ta hanyar ba da damar yin tunani a cikin LLM kanta ta hanyar tunani na tattaunawa na shirye-shirye, yana rage maimaita waje da kuma haɗarin lissafi masu alaƙa.
7.95%
Matsakaicin Ingantaccen Aiki
58.3%
Rage Farashin Token
6
Ma'auni da aka Ƙididdige
2 Ƙirar Tsarin INoT
2.1 Gurfanar da Lambar LLM-Karatu
Babban ƙirƙira na INoT yana cikin ƙirar gurfanar da lambar LLM-Karatu, wanda ke canza tunanin harshe na halitta zuwa tsarin aiwatarwa na shirye-shirye. Ba kamar ƙirar gurfanar da na gargajiya wanda ke dogaro da bambance-bambancen harshe ba, INoT yana amfani da samfuran lamba masu tsari waɗanda LLMs za su iya fassara da aiwatarwa kai tsaye.
2.2 Tsarin Ƙin Kai
INoT yana aiwatar da binciken ciki inda LLM ke kimanta tsarin tunaninta ba tare da madogaran tabbaci na waje ba. Wannan tsarin suka na ciki yana rage buƙatar hulɗar wakili da yawa ko tabbaci na waje mai maimaitawa.
3 Aiwar Fasaha
3.1 Tushen Lissafi
Tsarin INoT yana inganta tsarin tunani ta hanyar ƙirar ƙirar yiwuwar. Idan aka ba da shigarwar $x$ da kuma fitarwa da ake so $y$, hanyoyin gargajiya suna lissafin:
$P(y|x) = \prod_{t=1}^{T} P(y_t|x, y_{ INoT yana inganta wannan ta hanyar tunani na ciki: $P_{INoT}(y|x) = \prod_{t=1}^{T} P(y_t|x, y_{ inda $R_t$ ke wakiltar yanayin tunani na ciki a mataki na $t$, wanda aka lissafa kamar haka: $R_t = f_{reflect}(x, y_{ Aikin tunani $f_{reflect}$ yana aiki a cikin sararin ɓoyayyen LLM, yana rage yawan amfani da token na waje yayin kiyaye amincin tunani. Duk da yake PDF bai ba da lamba bayyanna ba, ana iya fassara tsarin INoT ta wannan tsarin pseudocode: An kimanta INoT a cikin ma'auni shida da suka haɗa da tunanin lissafi, ayyukan shirye-shirye, da amsa tambayoyi masu nau'i-nau'i. Tsarin ya sami matsakaicin ingantaccen aiki na 7.95% idan aka kwatanta da hanyoyin tushe da suka haɗa da CoT, IoT, da ProgCo. Mafi girman nasarar da INoT ya samu shine rage farashin token da kashi 58.3% idan aka kwatanta da mafi kyawun hanyar tushe. Wannan ribar ingantaccen ta samo asali ne daga shigar da tsarin tunani ciki, yana kawar da buƙatar zagayawa na tabbaci na waje da yawa. INoT ba wani ƙarin ci gaba ne kawai ba—canji ne na asali kan yadda muke tunkarar tunanin LLM. Tsarin ya yi nasarar ƙalubalantar akidar da ta mamaye cewa tunani mai sarƙaƙiya yana buƙatar madogaran tabbaci na waje da yawa. Ta hanyar matsar da tunani cikin samfurin, marubutan sun gano wata muhimmiyar rashin inganci a cikin gine-ginen wakilin AI na yanzu. Binciken ya bi sahihan ci gaba na ma'ana: Hanyoyin na yanzu → Rashin inganci da aka gano → Hasashen tunani na ciki → Aiwarwa → Tabbaci. Sarkar ta daɗe saboda tana magance wata takuraƙƙeniyar takura (farashin token) yayin inganta aiki, yana haifar da wani yanayi mai nasara da wuya a cikin ingantaccen AI. Abubuwan Haske: Rage token na 58.3% yana da girma—kwatankwacin ribar ingancin da aka gani a cikin ƙwararrun ingantaccen kamar ingantaccen gine-ginen Transformer na asali akan RNNs. Ƙarfin tsarin a cikin ma'auni da yawa yana nuna ingantaccen haɗakarwa. Iyakoki: Hanyar tana ɗauka cewa LLMs suna da isasshen ƙarfin wakilci na ciki don ingantaccen tunani na kai. Kamar yadda aka lura a cikin takardar CycleGAN ta asali, ƙuntatawa na gine-gine na iya iyakance irin waɗannan hanyoyin ingantaccen ciki. Bugu da ƙari, hanyar na iya fuskantar wahala tare da ayyukan da ke buƙatar sabon tunani da gaske fiye da rarraba horon samfurin. Wannan binciken ya kamata ya haifar da sake kimanta gine-ginen tsarin tunani a cikin masana'antu nan da nan. Kamfanonin da ke gina wakilan AI ya kamata su ba da fifiko ga hanyoyin tunani na ciki akan madogaran tabbaci na waje. Sakamakon ya nuna cewa ya kamata injiniyan gurfanar da ya karkata zuwa tsarin shirye-shirye maimakon bambance-bambancen harshe na halitta. Kamar yadda binciken DeepMind akan ingantaccen tushen samfura ya nuna, tunani na ciki sau da yawa yana fi tabbaci na waje lokacin da aka tsara shi yadda ya kamata. Tsarin INoT yana buɗe hanyoyi masu ban sha'awa da yawa don ci gaba na gaba: Aikin gaba ya kamata ya bincika hanyoyin haɗaka waɗanda ke haɗa tunani na ciki na INoT tare da zaɓin tabbaci na waje don mafi kyawun aiki a cikin nau'ikan ayyuka daban-daban.3.2 Aiwar Lamba
class INoTReasoner:
def __init__(self, llm_model):
self.llm = llm_model
self.reflection_states = []
def reason_with_introspection(self, query):
# Fage na tunani na farko
initial_response = self.llm.generate(query)
# Lokacin tunani na ciki
reflection_prompt = self._build_reflection_prompt(query, initial_response)
reflection = self.llm.generate(reflection_prompt)
# Haɗaɗɗen amsa ta ƙarshe
final_prompt = self._integrate_reflection(query, initial_response, reflection)
return self.llm.generate(final_prompt)
def _build_reflection_prompt(self, query, response):
return f"""Bincika tunanin mai zuwa don yuwuwar ingantawa:
Tambaya: {query}
Amsa na Yanzu: {response}
Gano gibi na ma'ana da kuma ba da shawarwarin ingantawa:"""4 Sakamakon Gwaji
4.1 Ma'aunin Aiki
4.2 Ingantaccen Token
Mahimman Bayanai
5 Bincike Mai mahimmanci
Hangen Nesa na Manazin Masana'antu
Yanke zuwa Ga Tabbawa
Sarkar Ma'ana
Abubuwan Haske da Iyakoki
Bayyanannun Ayyuka
6 Aikace-aikacen Gaba
7 Nassoshi