{"id":"W4412862721","doi":"10.5120/ijca2025925482","title":"Semantic Jailbreaks and RLHF Limitations in LLMs: A Taxonomy, Failure Trace, and Mitigation Strategy","year":2025,"lang":"en","type":"article","venue":"International Journal of Computer Applications","topic":"Digital and Cyber Forensics","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"Trinity College","funders":"","keywords":"Computer science; TRACE (psycholinguistics); Taxonomy (biology); Data science; Ecology; Linguistics; Biology","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001334344,0.00009605133,0.0001341592,0.0003532943,0.00004622661,0.0003780287,0.0004793586,0.00004142725,0.000001050276],"category_scores_gemma":[0.000009915319,0.00009210446,0.00004321903,0.0002833974,0.00007438363,0.0007357747,0.0001413517,0.0001419309,0.000002707159],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003725584,"about_ca_system_score_gemma":0.0001035925,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000006581905,"about_ca_topic_score_gemma":0.00005500673,"domain_scores_codex":[0.9991209,0.00002306615,0.0003965737,0.0001725597,0.0001905752,0.00009631448],"domain_scores_gemma":[0.9991403,0.0001601135,0.0001726873,0.0001262391,0.0003394512,0.00006126775],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.000005177885,0.0001435687,0.003405847,0.0000124853,0.00008024595,0.00001136763,0.0002356131,0.0007856827,0.00008839188,0.4670891,0.001189456,0.526953],"study_design_scores_gemma":[0.002522174,0.0002110151,0.272859,0.0003935339,0.00004629983,0.0006388084,0.0003020827,0.1448505,0.0004269571,0.4339517,0.143338,0.0004598971],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.03858851,0.000217122,0.9513129,0.006889973,0.0001848081,0.0002071439,0.000005849475,0.00002079951,0.002572906],"genre_scores_gemma":[0.9557891,0.00007454637,0.04363665,0.0002506974,0.00009838521,0.00003566158,0.000004085117,0.000003442674,0.0001074636],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9172006,"threshold_uncertainty_score":0.3755911,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01773420956557482,"score_gpt":0.2427508446670002,"score_spread":0.2250166351014254,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}