Dellu ci xibaar yi
YeesalAI Understanding

Preprint dafay méngale ni xeetu làkk yu mag yi di jëlee ngëmu noon yi ci jeu économique

Benn preprint bu bees dafa wax ni xeetu làkk yu mag yi dañu wane màndarga yuñ mëna natt ci mentalisation ci ñaari jeu économique, waaye seen performance ak seeni pexe wuute nañu ci xeetu famiy ak dayo. Auteur yi neena ñu GPT-5 agents yi dañu ànd ak noon yu gëna xarañ ba noppi gëna am doole ci mbooloo mi ñuy méngale ak nit ñi ci...

5 min readRead the primary source
Primary-source image accompanying Preprint compares how large language models infer opponents’ beliefs in economic games
Këyitu xët bu njëkkSource biñ enregistre
Siiwalkat
arxiv.org
Lëkkalekaayu cosaan
arxiv.orghttps://arxiv.org/abs/2608.26291
Xeetu balluwaay
Këyitu njëkk - ab yëgle ofisel, këyit, dosiye, wala xëtu pàrti bu njëkk bi ñuy jàng ci saasi.
KontekstXam lii ci 60 seconde

Tambalil fii

Term yu am solo

Etalonnage
Naka la poñ yi model bi am ci wóolu boppam méngoo ak probabilite yu dëggu yi.
Référence
Test buñ yamale wala ensemble done yuñ jëfandikoo ngir natt ak méngale liggéeyu model bi.
laaj
Tegtal yiñ dugal ak muy tekki biñ jox xeetu generatif bi.
Nattal sa boppModèlu IA leeral quiz

Lu xew

Gëstukat yi jàngat nañu ndax xeeti làkk yu mag yi mën nañu jëfandikoo ay leeral ci ngëm ak mébetu yeneen joueur ngir jël yenn dogal. Natt nañu 2,099 ndawu liggéey yu bawoo ci ñeenti famiy model - DeepSeek, GPT-4.1, GPT-5 ak Gemini 2.0 Flash - ci ñaari match koom-koom ak ay noon yu am niveau yu wuute ci xarañte, ba noppi ñu méngale li ñu am ak nit ñi.

Preprint bi, ñu jox ko arXiv ci Aug. 26, dafay jàngat mentalization, bi bindkat yi wax ni dafay jël gëm-gëmi ñeneen ñi ak seen yéene ngir teg tànneefi bopp. Laaj bi gëstu bi di laaj dafa gëna gàtt ndax model yi am nañu xel mu nit wala xam-xam bu nekk: dafay laaj ndax seeni dogal dañuy wane siñaale yu jeffin ak yu ordinatër yu méngoo ak jëfandikoo gis-gis yu yeneen agents 'states mental.

Gëstukat yi jëfandikoo nañu ñaari jeu économique ak modeling cognitif informatique ngir jàngat pexe yu nëbbu yi ci ginaaw dogal model. Natt nañu agent yu benn-benn yu bawoo ci ñeenti famiy model-DeepSeek, GPT-4.1, GPT-5 ak Gemini 2.0 Flash-ci noon yi ñu wax ci abstract bi ni dañu am xarañte yu wuute. Tolluwaayu misaal bi tollu ci 2,099 ndawu liggéey. Benn méngale bu wuute dafa amaale 251 nit ñu ci bokk.

Gëstu bi dafa jàngat itam ab pexe buy yee nit ñi ngir génne xalaat yu gëna jub ci pexe. Sunu sukkandikoo ci abstract bi, ñaaxtu pexe dafay gëna yokk performance, ndigam dayo njariñ bi dafa wuute ci ñaari jeu yi. Kon bindkat yi dañu wax ni amna njeexital ci njariñ yi ak ci pexem xalaat yiñ jëlee, du ñu am yokkute buñ yamale ci bépp jekkal.

Resultaa bi gëna leer ci model bi mooy GPT-5. Auteur yi neena ñu GPT-5 agents yi dañu defar xóotaayu seen xalaat ci anam wu yomb ndax noon yi dañu gëna xarañ ba noppi gëna am performance yu gëna baax ci nit ñi bokk ci jeu yi. Résumé bi waxul poñ yi ci lim yi, marge différence bi, dàntite yi wala version yi ci sistem yi ñu natt ginaaw etiket yiñ lim, wala ni ñu jëlee nit ñi bokk ci jàngat bi.

Ay leeral ci cosaan: arxiv.org ↗

Lu tax mu am solo

Gëstu bi mungi tontu ap laaj bu am solo bu jëm ci sistem yu juntuwaay yu bees yi di jaxasoo ak nit ñi: ndax jeffin ju niru ak doxalinu xel mi munna jappale pexe ngir mëna ànd ak nit ñi. Resultaa yi dañuy wane ni model yi mën nañu wane xeeti mentalizing yu wuute, yuñ mëna natt, ci noonu lañuy wane lu tax modeli làkk yu nuróo yi waru ñu leen jàppee ni ay model cognitif yuñ mëna weccoo.

Mentalisation amna solo ci sistem IA yi wara tontu nit ñi, concurrent yi wala yeneen agent losisel yi. Sistem bu mëna xayma li beneen pàrti xam, gëm wala xalaat mën na jël yenn dogal yu wuute ak sistem bu méngoo ak motif surface yi ci biir mbind. Li gëstu bi jàppale mooy wane mën-mën bii muy luñu mëna natt ci tànneef ak ci xeetu ordinatër, te baña yam ci tontu ci waxtaan ci laaji theorie-of-mind.

wuute yiñ wax ci diggante ñiy joxe xeetu làkk yi ak dayo yi dañu am solo ndax dañuy werante xalaat biy wax ni ab resultaa bu am doole ci benn test buy xalaat ci askan wi dafay fësal xeetu làkk yu mag yépp. Abstract bi dafa wax ni sistem yi dañu wane "siñaale yu leer ci jeffin ak ordinatër" ci mentalizing, waaye siñaale yooyu wuute nañu lool ci diggante ñiy joxe ak dayo model yi. Ci jàngat bu mën jëfandikoo, lii dafay wane ni dañu wara natt ay xeetu jaxasoo ci anam wu amul benn werante, moo gën ñuy wéeru ci benn etiketu xarañteg général.

Resultaa biy laaj amna lu gëna néew waaye lu am njariñ. Laaj benn model mu bokk ci xalaat strategique mën na yokk jikkoom ci yenn jekkal, waaye jafe-jafe yi tolloowul ci ñaari jeu yi dañuy wane ni ñaaxtu du yeesal kàttan universel. Gaawaayu performance mën na aju ci liggéey bi, noon bi ak ni ñuy jëlee xalaatu model bi. Source bi firndewul ni model yi 'processus yi ci biir dañu gëna niru ak nit; Dafay xamle coppite yi am ci doxalin ak ci pexe yuñ jëlee ci ordinatër.

Resultaa biñ joxe ci GPT-5 mën na am solo ci gëstu ci waxtaan, koordinasioŋ ak sistem yu bari, waaye waru ñu ko jàppee ni firnde ni GPT-5 moo gëna baax ci nit ñi ci xalaat ci askan wi. Tegtale bi dafa yam ci ñaari jeu économique, te balluwaay bi joxewul benn firnde ci diggante bis bu nekk, dogal yu am solo, njuuj njaaj, wuute ci aada wala jaxasoo ci àdduna dëgg. Du firndeel itam ni tekki ci jeu yii mingi aju ci jeffin ju wóor wala ju wóor ci bitim njàng mi.

Interactive Mechanism

Mekanism buy weccoo xalaat: naka lay doxee

Saytu xarala yu bees yi ci ginaaw yokkute bii ci anam wu weccoo xalaat.

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.
Saytu konsept buy weccoo xalaat+10 Points
AI Models Explained Quiz

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

Li nga wara seetaan ci topp

Këyit bi ab preprint arXiv la, te xëtu source bi dafay joxe ab abstract moo gën poñ yi ci niveau liggéey bi, ñakk wóor ci lim yi, ay laaj wala ay detay yu leer ci njuréef yiñ soxla ngir jàngat li ñu gis ci anam wu jege. Liggéey bi ci topp dafa wara natt ndax pexe yi ñu wax dañu ëpp pomi ñuy saytu, ak ndax dañuy wane xalaat yu dëgër, duñu jàng ay motif yu méngoo ak ay yuñ tànn.

Laaj bi gëna am solo mooy ndax këyit bi dafay joxe leeral yu doy ci njuréef yi ngir jàngat méngale bi. Leeral yu am solo yi ñàkk ci balluwaay bi bokkuna ci sàrti jeu yi, niveau sophistication yiñ jox noon yi, laaj yi gëna jubal, limu jéem yi ci agent bu nekk, jekkal misaal yi, doxalinu poñ yi ak xayma yu wóorul. Detay yooyu ñooy wane ndax resultaa bi dëgër na wala dafa méngoo ak jëmmal bi.

Baamtu ci jàngat model independent yi dina tax ñu saytu ndax li ñu gis mingi wéy ci yeesali model yi ak ci anam yu wuute ngir mëna dem. Boroom bi dafay xàmmee GPT-4.1, GPT-5 ak Gemini 2.0 Flash, waaye leeralul jekkal jëfandikoo gi, tegtali sistem bi wala bisu jëfandikoo gi. Ndax anam wi ñuy doxalee mën na soppiku ci laaj, jël misaal ak tabb produit, jàngat yi ci topp dañu wara bind variable yooyu ba noppi jëfandikoo protocole evaluation yuñ njëkka bind fépp fuñu ko mënee.

Gëstukat yi dañu wara xoolaat itam ndax xalaat yu bari yi ñuy jëlee ci liggéey yi ñu xamul dañuy wax luy waaja am. Benn model mën na def lu baax ci benn jeu ndax done yi mu tàggat wala structure bi dafay jàppale pexe buñ miin, te amul benn representation buñu mëna toxal ci yeneen xel. Extension yu am njariñ ñooy jeux yu bees, groupe yu wuute ak nit ñi, noon yuy doxalee ci anam wuñu ko mëna xamul, ak jekkal yu siñaal sosiaal yi mattwul wala ñuy réerloo nit.

Liggéeyukaay bi ba leegi kenn xamul. Bii preprint du natt benn produit buñ dugal, dafay digal ñu jëfandikoo sistem IA ci yenn dogal yi ci topp wala natt loraange yi. Laata ñuy jëfandikoo li ñu gis ci waxtaan, njàngale, jàppale wérgi-yaramu xel wala yeneen jëfandikoo yu jëm ci nit, jàngat yi dina ñu soxla firnde ci wàllu wóor, etalonnage, yoon ci askan wi, baña nangu njuuj njaaj ak anam yu ñàkka mëna dem sudee xalaati model bi ci beneen nit juumu ñu.

Gid ak quiz yu ci méngoo

Model IA leeral nañu koChatGPT & LLMsJikko yu AIËllëgu AINatt li nga xam — natt quiz IA bu amul faydaSeetal benn baat IA ci sunu glossaireToppal toppukaayu génne xeetu IA
Gis nga lii am njariñ?