{"id":{"repo_id":"vilnius","oai_identifier":"oai:vu.lt:elaba:308332492"},"canonical_url":"https://search.ndltd.org/etd/vilnius/oai:vu.lt:elaba:308332492","repository":{"repo_id":"vilnius","name":"Vilnius University","base_url":"https://epublications.vu.lt/oai"},"display":{"title":"Analysis of algorithms to reduce hallucinations in vision language models /","abstract":"Vision-Language Models (VLMs) exhibit impressive multimodal capabilities but remain susceptible to hallucinations — fluent yet visually ungrounded or factually incorrect outputs. This thesis analyzes a white-box hallucination detection framework, Multi-Token Reliability Estimation (MTRE), applied to the LLaVA-v1.5-7B architecture, which exploits internal logit dynamics to detect unreliability during generation. Experiments were conducted on the MAD-Bench and POPE benchmarks, investigating the impact of hyperparameters: generation temperature and the token logit sequence length. Results show that in-domain performance peaks at an ultra-narrow window of 2 logits (AUROC of 97.67%), improves substantially as the sequence is expanded to 20 logits. Simultaneous cross-dataset and cross-prompt transfer causes complete classifier collapse. High decoding temperatures introduce arithmetic overflow that is recoverable via sanitization, though boundary calibration remains limited under small training sets. The findings suggest that effective hallucination detection benefits from wider token horizons and is highly sensitive to both prompt format and dataset shifts.","abstract_html":"Vision-Language Models (VLMs) exhibit impressive multimodal capabilities but remain susceptible to hallucinations — fluent yet visually ungrounded or factually incorrect outputs. This thesis analyzes a white-box hallucination detection framework, Multi-Token Reliability Estimation (MTRE), applied to the LLaVA-v1.5-7B architecture, which exploits internal logit dynamics to detect unreliability during generation. Experiments were conducted on the MAD-Bench and POPE benchmarks, investigating the impact of hyperparameters: generation temperature and the token logit sequence length. Results show that in-domain performance peaks at an ultra-narrow window of 2 logits (AUROC of 97.67%), improves substantially as the sequence is expanded to 20 logits. Simultaneous cross-dataset and cross-prompt transfer causes complete classifier collapse. High decoding temperatures introduce arithmetic overflow that is recoverable via sanitization, though boundary calibration remains limited under small training sets. The findings suggest that effective hallucination detection benefits from wider token horizons and is highly sensitive to both prompt format and dataset shifts.","abstract_has_math":false,"creators":["Bruškytė, Brigita,"],"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":["vision-language models, hallucination detection, multi-token reliability estimation, prompt format sensitivity, white-box methods"],"languages":["eng"],"rights":["info:eu-repo/semantics/openAccess"],"rights_urls":[],"identifier_entries":[],"person_identifier_entries":[]},"links":{"outbound_url":"https://repository.vu.lt/VU:ELABAETD308332492&prefLang=en_US","outbound_label":"Repository record","outbound_source":"dc:identifier","record_url":"https://repository.vu.lt/VU:ELABAETD308332492&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%3A308332492","prefix":"oai_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:creator","label":"Author","values":["Bruškytė, Brigita,"]}]},{"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:308332492/308332492.pdf"]},{"key":"dc:type","label":"Dc Type","values":["info:eu-repo/semantics/bachelorThesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["vision-language models, hallucination detection, multi-token reliability estimation, prompt format sensitivity, white-box methods"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["eng"]},{"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:ELABAETD308332492&prefLang=en_US"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Vision-Language Models (VLMs) exhibit impressive multimodal capabilities but remain susceptible to hallucinations — fluent yet visually ungrounded or factually incorrect outputs. This thesis analyzes a white-box hallucination detection framework, Multi-Token Reliability Estimation (MTRE), applied to the LLaVA-v1.5-7B architecture, which exploits internal logit dynamics to detect unreliability during generation. Experiments were conducted on the MAD-Bench and POPE benchmarks, investigating the impact of hyperparameters: generation temperature and the token logit sequence length. Results show that in-domain performance peaks at an ultra-narrow window of 2 logits (AUROC of 97.67%), improves substantially as the sequence is expanded to 20 logits. Simultaneous cross-dataset and cross-prompt transfer causes complete classifier collapse. High decoding temperatures introduce arithmetic overflow that is recoverable via sanitization, though boundary calibration remains limited under small training sets. The findings suggest that effective hallucination detection benefits from wider token horizons and is highly sensitive to both prompt format and dataset shifts."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Analysis of algorithms to reduce hallucinations in vision language models /","Haliucinacijų mažinimo algoritmų tyrimas regos kalbos modeliams."]}]}],"canonical_facts":{"dc:creator":["Bruškytė, Brigita,"],"dc:date":["2026"],"dc:description":["Vision-Language Models (VLMs) exhibit impressive multimodal capabilities but remain susceptible to hallucinations — fluent yet visually ungrounded or factually incorrect outputs. This thesis analyzes a white-box hallucination detection framework, Multi-Token Reliability Estimation (MTRE), applied to the LLaVA-v1.5-7B architecture, which exploits internal logit dynamics to detect unreliability during generation. Experiments were conducted on the MAD-Bench and POPE benchmarks, investigating the impact of hyperparameters: generation temperature and the token logit sequence length. Results show that in-domain performance peaks at an ultra-narrow window of 2 logits (AUROC of 97.67%), improves substantially as the sequence is expanded to 20 logits. Simultaneous cross-dataset and cross-prompt transfer causes complete classifier collapse. High decoding temperatures introduce arithmetic overflow that is recoverable via sanitization, though boundary calibration remains limited under small training sets. The findings suggest that effective hallucination detection benefits from wider token horizons and is highly sensitive to both prompt format and dataset shifts."],"dc:format":["application/pdf"],"dc:identifier":["https://repository.vu.lt/VU:ELABAETD308332492&prefLang=en_US"],"dc:language":["eng"],"dc:publisher":["Institutional Repository of Vilnius University"],"dc:relation":["https://epublications.vu.lt/object/elaba:308332492/308332492.pdf"],"dc:rights":["info:eu-repo/semantics/openAccess"],"dc:subject":["vision-language models, hallucination detection, multi-token reliability estimation, prompt format sensitivity, white-box methods"],"dc:title":["Analysis of algorithms to reduce hallucinations in vision language models /","Haliucinacijų mažinimo algoritmų tyrimas regos kalbos modeliams."],"dc:type":["info:eu-repo/semantics/bachelorThesis"]},"updated_at":"2026-11-19T10:12:52Z"}