Publications 📗

@inproceedings{BiAlign2026,

title = {Trained Experience versus Prompted Personas: Population-Level Diversity in LLM Moral Judgment},

author = {Rafal Rzepka},

year  = {2026},

date = {2026-12-11},

booktitle = {Proceedings of NeurIPS Workshop on Dynamic Alignment in Human-Al Coupled System},

abstract = {Alignment is usually treated as convergence on a single, static value system. We show that a population of LLM-based agents can instead sustain and grow genuinely different values, with "experience" as the source of that diversity. Each agent is a LoRA adapter over one frozen model, trained with no task loss or reward on its own subset of Japanese narratives, mimicking how humans form values through what they encounter and live through. Experience content alone pushes agents well past a same-subset noise floor (2.86 vs. 5.06), and six generations of novelty search grow that divergence further, to 7.65, rather than collapsing it toward consensus. This transfers largely intact to an independent moral-judgment benchmark (JETHICS), with small gains on two categories and only a modest shift on the rest, and it does not change much when agents exchange experiences directly. Persona-prompting a static model with no training produces real but far smaller spread (4.63) than trained experience does (7.65), and richer real personas outspread simpler synthetic ones at equal group size: "experience" does something prompting alone cannot fully replicate. Checked against human sentence-level safety judgments, agents trained on human-sourced experience track human moral judgment better than those trained on synthetic content. We discuss the quantity vs. quality of experience behind human simulations, an approach gaining visibility across alignment research.},

keywords = {},

pubstate = {forthcoming},

tppubtype = {inproceedings}

}

@inproceedings{SatoHAI2026,

title = {Whose Signal Is It? Measuring Independence Limits of CLIP-Mediated Brain Representations in Human-AI Alignment},

author = {Takuto Sato and Rafal Rzepka},

year  = {2026},

date = {2026-11-18},

booktitle = {Proceedings of The 14th International Conference on Human-Agent Interaction, Osaka, Japan},

abstract = {When people delegate tasks to AI agents, oversight requires an evaluation axis from outside the AI, and brain measurement is a leading candidate. But brain data must first be mapped into the AI’s own representation space, so the oversight signal passes through the representation of the system it should oversee. Using four subjects from the Natural Scenes Dataset and MindBridge, we ask how much of a brain representation decoded through CLIP is already predictable from CLIP itself under adversarial perturbation. CLIP alone accounted for roughly three quarters of the across-image variance in the brain-routed quantity; a reproducible component remained

in all four subjects, but added almost no incremental explanatory value beyond the CLIP-only score. Because the brain embedding is fixed from clean-viewing fMRI, our design quantifies redundancy in CLIP-mediated representational geometry rather than establishing statistical independence from CLIP. For HAI, routing a human

cognitive signal through the agent’s own representation space does not by itself yield independent oversight; the degree of redundancy must be measured for each mediating representation.},

keywords = {},

pubstate = {forthcoming},

tppubtype = {inproceedings}

}

@inproceedings{BinghaoWangPRICAI2026,

title = {SafeSight: Auditing False Alarms in Embodied Hazard Perception},

author = {Binghao Wang and Rafal Rzepka},

url = {https://github.com/takashiwangbh/SafeSight},

year  = {2026},

date = {2026-11-17},

urldate = {2026-11-17},

booktitle = {Proceedings of The Pacific Rim International Conference on Artificial Intelligence (PRICAI 2026)},

abstract = {Embodied agents increasingly rely on large language and vision-language models (LLMs/VLMs) to perceive household hazards such as running faucets, broken glass, or unattended stoves. Existing

benchmarks evaluate this ability primarily through recall on hazardous scenes, which rewards models that trivially predict “dangerous” for every input. This paper presents SafeSight, an AI2-THOR-grounded evaluation framework, and SafeSight-Bench, its frozen benchmark instantiation with 33 executable hazard templates, 891 hazardous scenes, 356 safe controls, and semantic sensor noise. Across 12 modern open-source LLMs and VLMs spanning five architecture families and three input tracks (text-only, image-only, text+image), the audit collects 142,542 raw model-inference JSON results, which are normalized into 137,182 standardized scoring records, and reports precision, recall, F1, False-Alarm Rate (FAR) on safe scenes, Hazard Alignment (HA), and Phantom Contamination

(PC). The analysis reveals that five of twenty-two configurations combine recall ≥0.96 with FAR between 0.87 and 0.99, exposing conservative over-prediction; that sensor noise raises text-bearing FAR from 0.43 at λ= 0 to 0.77 at λ= 0.7 (all tracks: 0.41 →0.67); and that keyword-based Hazard Alignment does not exceed approximately 0.60 even on correctly flagged scenes. A lightweight Embodied Gated Alignment Verification (EGAV) baseline reduces FAR by roughly half for two representative VLMs while improving F1. Automatic metrics are complemented with machine-consensus cross-verification. The code, hazard templates, standardized scoring records, and qualitative visual examples are publicly available in our GitHub repository.},

keywords = {},

pubstate = {forthcoming},

tppubtype = {inproceedings}

}

@inproceedings{ihara2026lau,

title = {日本語省略質問に対する検索・クエリ拡張手法の体系的評価},

author = {井原 惇宏 and ジェプカ ラファウ},

year  = {2026},

date = {2026-09-27},

urldate = {2026-01-01},

booktitle = {言語獲得と理解研究会（LAU 2026）の予稿集},

pages = {1–11},

note = {LAU26-01},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{sato2026lau,

title = {CLIP 各層で N 字型に変化する Attention と人間視線の一致度ならびに中間表現の線形プロービン},

author = {佐藤 汰功斗 and 吉田 正俊 and Rafal Rzepka},

year  = {2026},

date = {2026-09-27},

urldate = {2026-01-01},

booktitle = {言語獲得と理解研究会（LAU 2026）の予稿集},

pages = {36–46},

note = {LAU26-05},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{10.1007/978-3-032-33384-1_13,

title = {Beyond Description: Comparing AI and Human Interpretation of Visual Art in Social Communication Contexts},

author = {Rafal Rzepka and Jin Hosono and Jun Nakamura},

editor = {Dario Liberona and Luong Vuong Nguyen and I-Hsien Ting and Kai Wang},

isbn = {978-3-032-33384-1},

year  = {2026},

date = {2026-08-11},

urldate = {2027-01-01},

booktitle = {Multidisciplinary Social Networks Research},

pages = {169–183},

publisher = {Springer Nature Switzerland},

address = {Cham},

abstract = {Visual content plays a central role in online communication, where images are often accompanied by human interpretation and commentary. With the rapid development of Vision–Language Models (VLMs), artificial intelligence systems are increasingly capable of generating such interpretations. This paper examines whether current VLMs can produce art descriptions and interpretations comparable to those written by human students. We conduct a comparative study using a structured evaluation framework in which a Large Language Model (GPT-4o) acts as an automated judge assessing interpretations across seven dimensions, including symbolism, contextual awareness, originality, and coherence. Although the LLM-judge evaluates models higher than humans, experiments with GPT-4o, LLaVA1.5, Qwen2.5-VL, and Gemma3 reveal distinct differences. While humans map abstract concepts onto spatial reality and personal metaphor, advanced VLMs rely on detached, formal art critique terminology. Crucially, our variance analysis debunks the assumption that higher generation temperatures induce creativity; advanced models like GPT-4o exhibited zero or small variance, acting as unyielding, consistent critics. On the other hand, smaller models like LLaVA exhibited erratic, above-human variance due to abstract instability. These results highlight a critical design trajectory for multimodal AI: choosing between the rigid reliability of an automated ``art critic'' and the unpredictable, divergent brainstorming of an ``idea-giver.''},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Hirakata2026AGIArchitecture,

title = {左右脳統合モナド／モジュールを基盤とするAGIアーキテクチャ},

author = {平方 勝 and 馬 沖 and ジェプカ ラファウ and 山田 智章 and 加瀬 究},

year  = {2026},

date = {2026-06-11},

booktitle = {2026年度人工知能学会全国大会（第40回）},

note = {Paper ID: 4L4-GS-1a-01},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Yokoyama2026StoryOrdering,

title = {計算資源制約下における言語モデルの日本語ストーリー文順序付け能力の評価},

author = {横山 明咲 and ジェプカ ラファウ},

year  = {2026},

date = {2026-06-09},

booktitle = {2026年度人工知能学会全国大会（第40回）},

note = {Paper ID: 2Yin-B-06},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Ihara2026StoryUnderstanding,

title = {日本語物語理解におけるLLMの動作主体・対象抽出能力と文脈理解の限界},

author = {井原 惇宏 and ジェプカ ラファウ},

year  = {2026},

date = {2026-06-09},

booktitle = {2026年度人工知能学会全国大会（第40回）},

note = {Paper ID: 2Yin-B-07},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Yamada2026KnowledgeDistillation,

title = {大規模言語モデルからの知識蒸留による低コストな顧客怒り度推定システムの開発},

author = {山田 有沙人 and ジェプカ ラファウ},

year  = {2026},

date = {2026-06-09},

booktitle = {2026年度人工知能学会全国大会（第40回）},

note = {Paper ID: 2Yin-B-10},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{rzepka2026dataset,

title = {A Dataset and Benchmark for Resolving Legal Cross-References in Japanese Export Control Regulations},

author = {Rafal Rzepka and Shinji Muraji and Akihiko Obayashi},

year  = {2026},

date = {2026-06-08},

urldate = {2026-06-01},

booktitle = {Proceedings of the Twentieth International Workshop on Juris-informatics (JURISIN 2026) associated with JSAI International Symposia on AI 2026 (IsAI-2026)},

address = {Gunma, Japan},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{MurajiLaCATODA26,

title = {Evaluation of Emotion-Conditioned Response Generation for Role-Playing Agents Using Large Language Models: A Case Study with Facial Expression Labels from Visual Novel Data},

author = {Shinji Muraji and Rafal Rzepka and Toshihiko Itoh},

url = {https://ceur-ws.org/Vol-4178/},

year  = {2026},

date = {2026-01-27},

urldate = {2026-01-27},

booktitle = {CEUR Proceedings of the 10th Linguistic and Cognitive Approaches to Dialog Agents (LaCATODA 2026), AAAI 2026 Workshop, Singapore, 2026},

pages = {144--159},

publisher = {WS-CEUR},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{RzepkaLaCATODA2026,

title = {Evaluating Lightweight Embedding Guardrails for Cost-Effective Misalignment Mitigation in Export Control Dialog Systems},

author = {Rafal Rzepka and Shinji Muraji and Akihiko Obayashi},

url = {https://ceur-ws.org/Vol-4178/},

year  = {2026},

date = {2026-01-27},

urldate = {2026-01-27},

booktitle = {CEUR Proceedings of the 10th Linguistic and Cognitive Approaches to Dialog Agents (LaCATODA 2026), AAAI 2026 Workshop, Singapore, 2026.},

pages = {132--143},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{JETHICS2026,

title = {JETHICS: Japanese Ethics Understanding Evaluation Dataset},

author = {Masashi Takeshita and Rafal Rzepka},

editor = {Marija Slavkovik and Hiroshi Yamakawa and Rafal Rzepka and Taichiro Endo and Louise A. Dennis and Raynaldo Limarga and Michael Fisher and Ryutaro Ichise},

url = {https://ceur-ws.org/Vol-4189/},

issn = {1613-0073},

year  = {2026},

date = {2026-01-27},

urldate = {2026-01-27},

booktitle = {Proceedings of the Machine Ethics: from formal methods to emergent machine ethics Workshop (MEW 2026) co-located with the 40th Annual AAAI Conference on Artificial Intelligence (AAAI 2026), Singapore},

volume = {4189},

pages = {108-117},

publisher = {CUER-WS},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{RzepkaLaCATODA2026b,

title = {A Novel Neuro-symbolic Approach to Irony Detection Based on Structural Components of Ironic Statements},

author = {Hiroshi Shigenobu and Michal Ptaszynski and Shunsuke Dan and Fumito Masui and Yuzu Uchida and Rafal Rzepka},

url = {https://ceur-ws.org/Vol-4178/},

year  = {2026},

date = {2026-01-27},

urldate = {2026-01-27},

booktitle = {CEUR Proceedings of the 10th Linguistic and Cognitive Approaches to Dialog Agents (LaCATODA 2026), AAAI 2026 Workshop, Singapore, 2026.},

pages = {172-183},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{10.1145/3765766.3765822,

title = {Predicting Cooperative Overlap Timing to Enhance Human-likeness in Spoken Dialogue},

author = {Ryusei Hachiya and Toshihiko Itoh},

url = {https://doi.org/10.1145/3765766.3765822},

doi = {10.1145/3765766.3765822},

isbn = {9798400721786},

year  = {2026},

date = {2026-01-01},

booktitle = {Proceedings of the 13th International Conference on Human-Agent Interaction},

pages = {389–391},

publisher = {Association for Computing Machinery},

address = {New York, NY, USA},

series = {HAI '25},

abstract = {Cooperative overlap is a conversational behavior that expresses engagement and supports smooth communication, yet it is rarely addressed in dialogue systems. In this study, we extracted instances of cooperative overlap from a Japanese casual conversation corpus and explored the feasibility of predicting their timing using a lightweight classifier. The model showed learning tendencies that indicate potential for distinguishing cooperative overlaps from other utterances. Our findings suggest that this behavior can be implemented in dialogue agents. This work opens up new possibilities for designing systems that interact more naturally and responsively with users.},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{ShintoAACL2025,

title = {To What Extent Can In-Context Learning Solve Unseen Tasks?},

author = {Ryoma Shinto and Masashi Takeshita and Rafal Rzepka and Toshihiko Itoh},

url = {https://aclanthology.org/2025.ijcnlp-srw.23/},

year  = {2025},

date = {2025-12-20},

urldate = {2025-12-20},

booktitle = {IJCNLP-AACL Student Research Workshop, Mumbai, 2025 (Best Paper Award🏆) },

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{ShintoDemo2025,

title = {Live Demonstration: Fingerspelling Recognition in Quantized Neural Networks with Emulated Memristors for 2D/3D Camera Applications},

author = {Yoshinari Wada and Rei Kusunose and Ryoma Shinto and Ayane Matsuzaki and Seiji Adachi and Kota Ando and Tetsuya Asai and Takao Marukame},

year  = {2025},

date = {2025-10-16},

urldate = {2025-10-16},

booktitle = {Proceedings of 2025 IEEE Biomedical Circuits and Systems Conference, Abu Dhabi, UAE},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{nokey,

title = {Rethinking Continual Learning with Pre-Trained Models: Knowledge-Preserving Approach},

author = {Makoto Misaizu and Koshi Watanabe and Keisuke Maeda and Takahiro Ogawa and Rafal Rzepka and Toshihiko Itoh and Miki Haseyama},

year  = {2025},

date = {2025-09-23},

urldate = {2025-09-23},

booktitle = {Proceedings of 2025 IEEE 14th Global Conference on Consumer Electronics (GCCE)},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}