Komawa Labarai
Bidi'aAI Understanding takaitaccen bayani

Werewolf benchmark ya gano LLMs na iya wuce gona da iri ga amintattun masu zargin

Alamar ma'auni na saitin LLM mai buɗewa 40 ya gano cewa zarge-zarge na iya canza imanin ƙirar koda lokacin da mai ƙara ya daidaita tare da bangaren adawa.

4 min readRead the primary source
Source-provided image accompanying Werewolf benchmark finds LLMs can overvalue trusted accusers
Takardun tushe na farkoAn rubuta tushen tushe
Mawallafi
arxiv.org
Tushen hanyar haɗin gwiwa
arxiv.orghttps://arxiv.org/abs/2609.12446
Nau'in tushe
Takardun farko - sanarwar hukuma, takarda, yin rajista, ko shafi na farko da muka karanta kai tsaye.
MaganaFahimtar wannan a cikin daƙiƙa 60

Fara a nan

Mabuɗin sharuddan

Alamar alama
Daidaitaccen gwaji ko saitin bayanai da aka yi amfani da shi don aunawa da kwatanta aikin ƙira.
Babban Samfurin Harshe (LLM)
Samfurin harshe da aka horar akan babban haɗin gwiwar rubutu don samarwa da tantance rubutu.
Bayani
Alamomin da aka ƙara ɗan adam ko metadata da aka yi amfani da su don horarwa ko kimanta ƙirar koyon inji.
Gwada kankaAI Model An Bayyana Tambayoyi

Me ya faru

Masu bincike sun gabatar da ma'auni na Werewolf wanda ke auna yadda lura da LLMs ke sabunta imaninsu bayan sun karɓi saƙon tuhuma da zargi. Fiye da bayanan bayanan 1,224 da 40 buɗaɗɗen ƙirar ƙira, ƙira mafi girma galibi ana gano kerkeci na gaskiya daga tarihin wasan amma sun kasance suna tasiri sosai ta zarge-zarge da kuma fahimtar amincin mai zargi.

Takardar ta ba da shawarar kimanta sauye-sauyen imani maimakon dogaro da sakamakon ƙarshe na wasan ragi na zamantakewa. A cikin saitin sa, samfurin gefen ƙauye yana karɓar saƙonnin wasa, gami da zato da zargi, kuma masu bincike suna auna yadda imaninsa ke canzawa bayan kowane saƙo.

Rahotann da aka zayyana suna haifar da buɗaɗɗen nauyi LLM jeri 40 da saƙon da aka rubuta 1,224. Manyan samfura sun yi aiki mafi kyau wajen bambance kerkeci daga ƙauye ta amfani da cikakken tarihin wasan. Duk da haka, har yanzu zarge-zarge na ƙara zato ga wanda ake tuhuma da kuma rage zato ga wanda ake tuhuma, musamman lokacin da aka riga an amince da wanda ake tuhuma.

Tsarin da aka ruwaito ya ci gaba har ma lokacin da amintaccen wanda ake zargi ya kasance mai haɗin kai. Manyan samfura sun fi iya tsayayya da zarge-zarge daga masu zargin da suka rigaya ba su amince da su ba, amma samfuran har zuwa sigogin biliyan 120 har yanzu suna gwagwarmaya don haɗa abubuwan da ake zargi da amincin tushen sa. Madogarar ba ta bayar da cikakken makin kowane samfuri, rashin tabbas na ƙididdiga, ko kwafi mai zaman kansa a cikin rubutun da aka kawo.

Bayanan tushe: arxiv.org ↗

Me ya sa yake da mahimmanci

Binciken ya gano takamaiman rauni a cikin dabarun sadarwa: ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila za su raba amincin mai magana da shaidar da ke ƙunshe a da'awar mai magana. Wannan yana da mahimmanci ga wakilan LLM da ke aiki a cikin saituna inda saƙonni za su iya zama zaɓi, yaudara, ko abokan gaba. Sakamakon shine ma'auni da binciken bincike, ba shaida cewa duk tsarin AI da aka tura suna yin wannan hanya ba ko kuma kimantawa ta wuce Werewolf.

Ga masu binciken da ke kimanta wakilan LLM, sakamakon ya nuna cewa nasarar wasan ƙarshe na iya ɓoye mahimman rauni a cikin tunani na tsaka-tsaki da sabunta imani. Samfurin zai iya cimma matsaya mai ma'ana yayin da yake da kiba wanda ya ba da da'awa maimakon tantance da'awar tare da samuwan shaidar.

Ma'anar aiki tana iyakance amma mai amfani: kimantawa na wakilai waɗanda ke sadarwa ko yanke shawara daga rahotanni da yawa na iya buƙatar auna canje-canjen imani bayan saƙon mutum ɗaya, gami da lamuran da sahihancin magana ke yaudara. Binciken bai tabbatar da cewa halayen iri ɗaya na faruwa a cikin samfuran da aka tura ko a cikin saitunan da ba na wasa ba.

Interactive Mechanism

Ingantacciyar hanyar sadarwa: Yadda A zahiri yake Aiki

Bincika fasahar da ke bayan wannan ci gaban ta hanyar mu'amala.

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.
Duba ra'ayi na hulɗa+10 Points
AI Models Explained Quiz

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

Abin kallo na gaba

Ana samun maƙasudin maƙasudi da lambar ta hanyar gidan yanar gizon aikin, amma tushen ba ya bayyana lasisi, cikakkun bayanai na baƙi, ko samfuran da aka kimanta suna samuwa don amfanin jama'a. Ya kamata ƙarin aiki ya gwada ko binciken yana canjawa zuwa wasu ayyukan sadarwa, ko horarwa yana inganta tunani-abun ciki, da kuma yawan sakamakon ya dogara da ƙirar wasa da zaɓin bayanin.

Marubutan sun ce alamar alama da lambar suna samuwa a shafin aikin da aka haɗa. Tushen da aka kawo ba ya rubuta buƙatun samun dama, lasisi, farashi, ginshiƙai masu goyan baya, ko kuma maƙasudin ya haɗa da fitattun samfura fiye da bayanan bayanan.

Muhimman abubuwan da ba a san su ba sun haɗa da yadda aka ƙirƙira saƙon da bayanin, yadda aka kafa amana, ko sakamakon ƙididdiga ya yi ƙarfi a cikin kowane nau'in mutum ɗaya, da kuma ingantattun tunanin tarihin wasan kwaikwayo mafi girma na fassara zuwa mafi kyawun juriya ga magudi.

Maimaitawa a cikin sauran ayyukan sadarwa na dabarun zai taimaka tantance ko wannan takamaiman tasirin Werewolf ne ko iyakancewa mai faɗi cikin yadda wakilan LLM ke haɗa amincin tushe tare da abun ciki na zargi.

Jagorori masu alaƙa & tambayoyin tambayoyi

AI Model ya bayyanaWakilan AIƊa'a ta AIGwada abin da kuka sani - gwada gwajin AI kyautaNemo kalmar AI a cikin ƙamus ɗin muBi samfurin AI na sakin tracker
An sami wannan yana da amfani?