{"id":{"repo_id":"vilnius","oai_identifier":"oai:vu.lt:elaba:308334486"},"canonical_url":"https://search.ndltd.org/etd/vilnius/oai:vu.lt:elaba:308334486","repository":{"repo_id":"vilnius","name":"Vilnius University","base_url":"https://epublications.vu.lt/oai"},"display":{"title":"Didžiųjų kalbos modelių loginio mąstymo gebėjimų tobulinimas lietuvių kalbai taikant sustiprinimo mokymą: metodologija, duomenų paruošimas ir palyginamoji analizė /","abstract":"This thesis investigates the possibilities of enhancing the logical and mathematical reasoning capabilities of small­scale Large Language Models (LLMs) using Reinforcement Learning (RL) methods. During the study, a Lithuanian ”Knights and Knaves” logical task dataset comprising 5000 unique examples was algorithmically synthesized and categorized by distinct difficulty levels. The baseline Qwen3­1.7B model was trained using three different algorithms: GRPO, DAPO, and LogicRL Reinforce++. The primary objective of the experiment was to incentivize the model to generate accurate, deep, and coherent reasoning chains to successfully solve complex logical tasks. The experimental results revealed that the 1.7 billion parameter model trained with the Logic­RL Reinforce++ method achieved the highest performance. This model attained a 13.3% accuracy rate on Olympiadlevel AIME benchmarks, outperforming both its baseline model (4.4%) and the significantly larger Qwen3­4B (8.9%) and Qwen3­8B (6.7%) models. The model trained with the GRPO method also demonstrated superiority over the baseline version (6.7%) and secured an intermediate position among the investigated algorithms. Conversely, the application of the DAPO algorithm failed due to a reward hacking problem that restricted the generated reasoning chains. The conducted research proves that targeted reinforcement learning successfully enhances the logical reasoning capabilities of small models, and the developed sequential reasoning chain generation ensures model versatility across tasks of varying difficulty.","abstract_html":"This thesis investigates the possibilities of enhancing the logical and mathematical reasoning capabilities of small­scale Large Language Models (LLMs) using Reinforcement Learning (RL) methods. During the study, a Lithuanian ”Knights and Knaves” logical task dataset comprising 5000 unique examples was algorithmically synthesized and categorized by distinct difficulty levels. The baseline Qwen3­1.7B model was trained using three different algorithms: GRPO, DAPO, and LogicRL Reinforce++. The primary objective of the experiment was to incentivize the model to generate accurate, deep, and coherent reasoning chains to successfully solve complex logical tasks. The experimental results revealed that the 1.7 billion parameter model trained with the Logic­RL Reinforce++ method achieved the highest performance. This model attained a 13.3% accuracy rate on Olympiadlevel AIME benchmarks, outperforming both its baseline model (4.4%) and the significantly larger Qwen3­4B (8.9%) and Qwen3­8B (6.7%) models. The model trained with the GRPO method also demonstrated superiority over the baseline version (6.7%) and secured an intermediate position among the investigated algorithms. Conversely, the application of the DAPO algorithm failed due to a reward hacking problem that restricted the generated reasoning chains. The conducted research proves that targeted reinforcement learning successfully enhances the logical reasoning capabilities of small models, and the developed sequential reasoning chain generation ensures model versatility across tasks of varying difficulty.","abstract_has_math":false,"creators":["Girdauskas, Žygimantas,"],"institution":"Institutional Repository of Vilnius University","degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026","date_published":"2026","updated_at":"2026-11-19T10:12:52Z","subjects":[],"languages":["lit"],"rights":["info:eu-repo/semantics/openAccess"],"rights_urls":[],"identifier_entries":[],"person_identifier_entries":[]},"links":{"outbound_url":"https://repository.vu.lt/VU:ELABAETD308334486&prefLang=en_US","outbound_label":"Repository record","outbound_source":"dc:identifier","record_url":"https://repository.vu.lt/VU:ELABAETD308334486&prefLang=en_US","record_label":"Repository record","record_kind":"record","record_source":"dc:identifier","download_url":null,"download_label":null,"download_source":null,"access_status":"open","embargo_until":null,"access_note":null},"source_record":{"url":"https://epublications.vu.lt/oai?verb=GetRecord&metadataPrefix=oai_dc&identifier=oai%3Avu.lt%3Aelaba%3A308334486","prefix":"oai_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:creator","label":"Author","values":["Girdauskas, Žygimantas,"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2026"]},{"key":"dc:publisher","label":"Institution","values":["Institutional Repository of Vilnius University"]},{"key":"dc:relation","label":"Dc Relation","values":["https://epublications.vu.lt/object/elaba:308334486/308334486.pdf"]},{"key":"dc:type","label":"Dc Type","values":["info:eu-repo/semantics/bachelorThesis"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["lit"]},{"key":"dc:rights","label":"Dc Rights","values":["info:eu-repo/semantics/openAccess"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://repository.vu.lt/VU:ELABAETD308334486&prefLang=en_US"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["This thesis investigates the possibilities of enhancing the logical and mathematical reasoning capabilities of small­scale Large Language Models (LLMs) using Reinforcement Learning (RL) methods. During the study, a Lithuanian ”Knights and Knaves” logical task dataset comprising 5000 unique examples was algorithmically synthesized and categorized by distinct difficulty levels. The baseline Qwen3­1.7B model was trained using three different algorithms: GRPO, DAPO, and LogicRL Reinforce++. The primary objective of the experiment was to incentivize the model to generate accurate, deep, and coherent reasoning chains to successfully solve complex logical tasks. The experimental results revealed that the 1.7 billion parameter model trained with the Logic­RL Reinforce++ method achieved the highest performance. This model attained a 13.3% accuracy rate on Olympiadlevel AIME benchmarks, outperforming both its baseline model (4.4%) and the significantly larger Qwen3­4B (8.9%) and Qwen3­8B (6.7%) models. The model trained with the GRPO method also demonstrated superiority over the baseline version (6.7%) and secured an intermediate position among the investigated algorithms. Conversely, the application of the DAPO algorithm failed due to a reward hacking problem that restricted the generated reasoning chains. The conducted research proves that targeted reinforcement learning successfully enhances the logical reasoning capabilities of small models, and the developed sequential reasoning chain generation ensures model versatility across tasks of varying difficulty."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Didžiųjų kalbos modelių loginio mąstymo gebėjimų tobulinimas lietuvių kalbai taikant sustiprinimo mokymą: metodologija, duomenų paruošimas ir palyginamoji analizė /","Improving reasoning capabilities of large language models on lithuanian data using reinforcement learning: methodology, data preparation and comparative analysis."]}]}],"canonical_facts":{"dc:creator":["Girdauskas, Žygimantas,"],"dc:date":["2026"],"dc:description":["This thesis investigates the possibilities of enhancing the logical and mathematical reasoning capabilities of small­scale Large Language Models (LLMs) using Reinforcement Learning (RL) methods. During the study, a Lithuanian ”Knights and Knaves” logical task dataset comprising 5000 unique examples was algorithmically synthesized and categorized by distinct difficulty levels. The baseline Qwen3­1.7B model was trained using three different algorithms: GRPO, DAPO, and LogicRL Reinforce++. The primary objective of the experiment was to incentivize the model to generate accurate, deep, and coherent reasoning chains to successfully solve complex logical tasks. The experimental results revealed that the 1.7 billion parameter model trained with the Logic­RL Reinforce++ method achieved the highest performance. This model attained a 13.3% accuracy rate on Olympiadlevel AIME benchmarks, outperforming both its baseline model (4.4%) and the significantly larger Qwen3­4B (8.9%) and Qwen3­8B (6.7%) models. The model trained with the GRPO method also demonstrated superiority over the baseline version (6.7%) and secured an intermediate position among the investigated algorithms. Conversely, the application of the DAPO algorithm failed due to a reward hacking problem that restricted the generated reasoning chains. The conducted research proves that targeted reinforcement learning successfully enhances the logical reasoning capabilities of small models, and the developed sequential reasoning chain generation ensures model versatility across tasks of varying difficulty."],"dc:format":["application/pdf"],"dc:identifier":["https://repository.vu.lt/VU:ELABAETD308334486&prefLang=en_US"],"dc:language":["lit"],"dc:publisher":["Institutional Repository of Vilnius University"],"dc:relation":["https://epublications.vu.lt/object/elaba:308334486/308334486.pdf"],"dc:rights":["info:eu-repo/semantics/openAccess"],"dc:title":["Didžiųjų kalbos modelių loginio mąstymo gebėjimų tobulinimas lietuvių kalbai taikant sustiprinimo mokymą: metodologija, duomenų paruošimas ir palyginamoji analizė /","Improving reasoning capabilities of large language models on lithuanian data using reinforcement learning: methodology, data preparation and comparative analysis."],"dc:type":["info:eu-repo/semantics/bachelorThesis"]},"updated_at":"2026-11-19T10:12:52Z"}