20.500.12556/RUL-189579 Dekodiranje jezikovnih modelov kot iskanje regulariziranega Nashevega ravnovesja Language model decoding as regularized Nash equilibrium search Veliki jezikovni modeli na isto vprašanje pogosto odgovorijo različno, odvisno od tega, ali jih vprašamo generativno ali diskriminativno. To neskladje lahko naslovimo z igro soglasja; signalno igro med generatorjem in diskriminatorjem, katere regularizirano ravnovesje uporabimo za rangiranje kandidatnih odgovorov. V diplomski nalogi igro soglasja formalno zapišemo kot ekstenzivno igro z nepopolno informacijo, izračunamo njeno strateško obliko in množico čistih Nashevih ravnovesij ter dokažemo karakterizacijo njenih šibkih zaporednih ravnovesij, s čimer pokažemo, da neregularizirana igra ne izloči nesmiselnih rešitev. Podamo podrobno izpeljavo algoritma piKL-Hedge in pravila posodabljanja strategij z regularizacijo s Kullback–Leiblerjevo divergenco. Metodo po našem vedenju prvič preizkusimo na nalogi naravno–jezikovnega sklepanja in prvič na slovenščini, in sicer na množici SI-NLI z modeli GaMS pri pozivih brez dodatnih primerov, z enim in s tremi primeri. Large language models often answer the same question differently depending on whether they are queried generatively or discriminatively. This inconsistency can be addressed through the consensus game; a signaling game between a generator and a discriminator whose regularized equilibrium is used to rank candidate answers. In this thesis we formally describe the consensus game as an extensive-form game with imperfect information, compute its strategic form and its set of pure Nash equilibria, and prove a characterization of its weak sequential equilibria, showing that the unregularized game does not rule out implausible solutions. We give a detailed derivation of the piKL-Hedge algorithm and of the strategy update rule regularized by Kullback–Leibler divergence. To the best of our knowledge, we are the first to evaluate the method on natural language inference and the first to do so for Slovene, using the SI-NLI dataset with the GaMS models under zero-shot, one-shot and three-shot prompts. teorija iger ekstenzivne igre z nepopolno informacijo Nashevo ravnovesje igra soglasja dekodiranje jezikovnih modelov game theory extensive-form games with imperfect information Nash equilibrium consensus game language model decoding true false false Slovenski jezik Angleški jezik Diplomsko delo/naloga 2026-10-09 11:55:07 2026-10-09 11:55:15 2026-10-29 12:33:13 0000-00-00 00:00:00 2026 0 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 38722 303083011 Brodar_Aljaz_-_Dekodiranje_jezikovnih_modelov_kot_iskanje_regulariziranega_Nashevega_ravnovesja.pdf Brodar_Aljaz_-_Dekodiranje_jezikovnih_modelov_kot_iskanje_regulariziranega_Nashevega_ravnovesja.pdf 1 7CFCE853641707F623F987B49B1EE76C 5ccaac05f10bf7748ab3411ae54324f1a2591eab7eb2dd783e3851e8a7cb2b06 60d88f10-c3c7-11f1-8bc5-0050569b8976 https://repozitorij.uni-lj.si/Dokument.php?lang=slv&id=250230 Fakulteta za računalništvo in informatiko 0 0 0