Workshop Program (preliminary)

Saturday, October 17th, 2026, 09:00–12:30, Wachtkammer 3e Klasse, Utrecht, the Netherlands / online https://2026.inlgmeeting.org/

All times are in Central European Summer Time (CEST), UTC+2.

All talks in Sessions 1–3 are peer reviewed papers accepted to the workshop. Long papers are given 15 minutes and Short papers 10 minutes, including Q&A. Presenters are marked with an asterisk (*). The program is preliminary and subject to change.

09:00–09:15 Opening and contest results

  • Opening: what is AIWolf? (Yoshinobu Kano)
  • AIWolfDial 2026: results of the 2026 International AIWolf Contest, Natural Language Division

09:15–09:35 Contest team presentations (6): non-archival AIWolf agent descriptions, 3 minutes each

  • Team kataken14f (Atsuro Hamanaka, Ariana Sheikhi, Daiki Asakura, and Ryosuke Sueyoshi*)
  • Team CamelliaDragons (Momoka Kato*, Kaito Okuno, Syuji Otsuki, Keita Kasuya, and Toma Inagaki)
  • Team CanisLupus (Yu Sugawara*)
  • Team gotsumori-NLP (Tatsuya Mikami*)
  • Team SunameLLi (Satoko Natsubori*, Hiroki Nakanishi, and Soichiro Mikami)
  • Team wool_100% (Kento Fukuda*)

09:35–10:15 Session 1 — Self-improvement and learning: peer reviewed papers

  • 09:35 [Long] Self-Improvement of AIWolf Agents Based on Self-Play Logs (Taku Umeda*, Takao Endo, Jumpei Hato, and Kotaro Otomura) (Team AGKAI)
  • 09:50 [Long] Do Cross-Game Lessons Appear in Later Dialogue? A Controlled Measurement of Lesson Realization in Werewolf Dialogue Agents (Kota Munaoka* and Yoshinobu Kano) (Team Oxymoron)
  • 10:05 [Short] The moon is new each night: enhancing reasoning with self-reflection in the game of Werewolf (Alicja Kasicka*, Łukasz Krajewski, Bartosz Przybył, Natalia Krawczyk, Michał Szczerbak, and Justyna Gromada) (Team carrot)

10:15–10:35 Coffee break

10:35–11:20 Session 2 — Role-specific strategy and consistency: peer reviewed papers

  • 10:35 [Long] Power Play Instructions in LLM Werewolf: Cross-Request Execution and Game Outcomes (Shoma Yato*, Ikumu Komatsu, Haruka Itakura, and Daisuke Katagami) (Team yatolab)
  • 10:50 [Short] An AIWolf Agent for Consistent Dialogue with Role-Specific Strategies and Utterance Regeneration (Yuki Kubo*, Taiga Sano, Eiji Kumakawa, and Tae Sato) (Team NTT-HAI)
  • 11:00 [Short] Improving Dialogues in LLM-Based AI Werewolf Agents with Few-Shot Discussion Examples (Yuya Harada* and Yoshinobu Kano) (Team yharada)
  • 11:10 [Short] Keeping a Public Story Straight: Memory Interventions for LLM Agents in AIWolf (Ryan Lee*) (Team ryanleeai)

11:20–12:20 Session 3 — Timing, persona, evaluation, and robustness: peer reviewed papers

  • 11:20 [Long] Beyond What to Say: Proactive Utterance Timing with Inner Thoughts for Werewolf Agents (Yoshiki Tanaka*, Shuya Uchiyama, Daichi Hamaguchi, and Michimasa Inaba) (Team UEC-IL)
  • 11:35 [Long] Preserving Character in Four Letters: Personality Information Compression in AIWolf Agents Using MBTI-like Labels (Ryosuke Sueyoshi, Kentaro Hoshi*, Miyu Ueki, and Chaeyoung Sung) (Team noisy_watcher)
  • 11:50 [Long] Fine-Grained Absolute Evaluation of Utterance Naturalness in Japanese AIWolf Dialogue: An Agreement Analysis between Human Annotators and LLM Judges (Chaeyoung Sung*, Ryosuke Sueyoshi, Kentaro Hoshi, Miyu Ueki, and Masato Taki) (Team noisy_watcher)
  • 12:05 [Long] What Breaks Prompt-Specified Settings in Multi-Turn Dialogue Attacks: Attack Type, Burst-Injection Granularity, and Role Boundaries (Yuki Shimoda* and Yoshinobu Kano) (Team yshimoda)

12:20–12:30 Closing and general discussion

(Lunch is served at the venue after the workshop.)