Elamur Research⌕
☰
home / papers / 2026 / Success and Cost Elicit Convention Formation for Efficient Communication

Success and Cost Elicit Convention Formation for Efficient Communication

Paper · 2026 · Saujas Vaduguru, Yilun Hua, Yoav Artzi et al.

Summary

Large multimodal models (LMMs) tend to keep producing equally long, detailed descriptions throughout an interaction. People, by contrast, settle on shorter shared expressions (ad hoc conventions) as common ground builds up. This paper trains LMM speakers to show the same adaptive behavior, using no extra human-produced interaction data.

The method works in three steps:

  • Simulation: speaker and listener models play repeated reference games, with several candidate descriptions sampled at each trial.
  • Preference pairs: candidates are compared on a utility combining communicative success (did the listener pick the target?) and cost (word count).
  • Training: these pairs drive DPO-style preference optimization of the speaker.

The trained speakers were evaluated with human listeners on MS COCO photographs and tangram images. Over the course of games, their messages became considerably shorter while accuracy rose, and people responded faster to them than to a non-adapting baseline. Ablations show that success and cost are both required to elicit conventions.

The work matters because it shows that generic, task-independent rewards learned in model-only simulation can produce efficiency that people can follow.

Key points

  • COCO games with humans: message length fell by 41% and communicative success rose by 16% over the interaction. For tangrams, length fell by 29% and success rose by 27%.
  • Human response speed: by an image's 4th repetition, people answered 52% faster (COCO) and 35% faster (tangrams) than with a non-adapting model.
  • Ablations:
    • SUCCESS-only training raised accuracy but made messages about 7% longer.
    • COST-only training gave short messages with near-chance accuracy, measured with a GPT-5-mini simulated listener.
  • Models: the approach was applied to Gemma- and Pixtral-based speakers. With the Pixtral speaker and the simulated listener, message length dropped 13.4% (8.95 → 7.74 words) by repetition 4.
  • Word choice: spaCy POS tagging shows the adapted speaker drops function words and relies increasingly on nouns, verbs and adjectives. This mirrors human patterns reported by Hawkins et al. (2020b).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Carnegie Mellon UniversityPittsburgh · United States598 works · 191,225 citations🎓Cornell UniversityIthaca · United States134 works · 65,159 citations🎓Cornell TechNew York · United States9 works · 2,342 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 worksSocial Sciences · 1,954 worksCultural Studies · 20 works

Cite

Cite as Saujas Vaduguru, Yilun Hua, Yoav Artzi and Daniel Fried. 2026. Success and Cost Elicit Convention Formation for Efficient Communication. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1946

@inproceedings{vaduguru2026success,
  title = {{Success and Cost Elicit Convention Formation for Efficient Communication}},
  author = {Saujas Vaduguru, Yilun Hua, Yoav Artzi and Daniel Fried},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1946},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1946}
}

References

32 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, Albert Q. Jiang, Kartik Khandelwal, Timothée Lacroix, Guillaume Lample, Diego Las Casas,… — Pixtral 12B
  2. Leon Bergen, R. Levy, and Noah D. Goodman. 2016. Pragmatic reasoning through semantic inference. Semantics and Pragmatics, 9. find
  3. Kranti Chalamalasetti, Jana Götze, Sherzod Hakimov, Brielen Madureira, Philipp Sadler, and David Schlangen. 2023. clembench: Using game play to evaluate chat-optimized language models as conversational agents. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,… find
  4. Herbert H. Clark and Deanna Wilkes-Gibbs. 1986. Referring as a collaborative process. Cognition, 22(1):1– 39. find
  5. Yann Dubois, Percy Liang, and Tatsunori Hashimoto. 2024. Length-controlled alpacaeval: A simple debiasing of automatic evaluators. In First Conference on Language Modeling. find
  6. Michael C. Frank and Noah D. Goodman. 2012. Predicting pragmatic reasoning in language games. Science, 336(6084):998–998.
  7. Gemma Team. 2025. Gemma 3 technical report. Preprint, arXiv:2503.19786. find
  8. Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. 2024. A general theoretical paradigm to understand learning from human preferences. In Proceedings of The 27th International Conference on Artificial Intelligence and Statistic…
  9. Herbert P Grice. 1975. Logic and conversation. In Speech acts, pages 41–58. Brill. find
  10. Mustafa Omer Gul and Yoav Artzi. 2024. Cogen: Learning from feedback with coupled comprehension and generation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 12966–12982, Miami, Florida, USA. Association for Computational Linguistics. find
  11. Janosch Haber, Tim Baumgärtner, Ece Takmaz, Lieke Gelderloos, Elia Bruni, and Raquel Fernández. 2019. The PhotoBook dataset: Building common ground through visually-grounded dialogue. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1895–1910, Florenc…
  12. Robert Hawkins, Minae Kwon, Dorsa Sadigh, and Noah Goodman. 2020a. Continual adaptation for efficient machine communication. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 408–419, Online. Association for Computational Linguistics.
  13. Robert D. Hawkins, Michael C. Frank, and Noah D. Goodman. 2020b. Characterizing the dynamics of learning in repeated reference games. Cognitive Science, 44(6):e12845.
  14. Robert D. Hawkins, Michael Franke, Michael C. Frank, Adele E. Goldberg, Kenny Smith, Thomas L. Griffiths, and Noah D. Goodman. 2023. From partners to populations: A hierarchical bayesian account of coordination and convention. Psychological Review, 130(4):977–1016. find
  15. Matthew Honnibal, Ines Montani, Sofie Van Lan-deghem, and Adriane Boyd. 2020. spaCy: Industrial-strength Natural Language Processing in Python. find
  16. Nicola Horst, Davide Mazzaccara, Antonia Schmidt, Michael Sullivan, Filippo Momentè, Luca Franceschetti, Philipp Sadler, Sherzod Hakimov, Alberto Testoni, Raffaella Bernardi, Raquel Fernández, Alexander Koller, Oliver Lemon, David Schlangen, Mario Giulianelli, and Alessandro Suglia. 2025. Playpen: A… — [arXiv:2504.08590](https://arxiv.org/abs/2504.08590) find
  17. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  18. Yilun Hua and Yoav Artzi. 2024. Talk less, interact better: Evaluating in-context conversational adaptation in multimodal LLMs. In First Conference on Language Modeling.
  19. Yilun Hua, Evan Wang, and Yoav Artzi. 2025. Post-training for efficient communication via convention formation. In Second Conference on Language Modeling. find
  20. Anya Ji, Noriyuki Kojima, Noah Rush, Alane Suhr, Wai Keen Vong, Robert D. Hawkins, and Yoav Artzi. 2022. Abstract visual reasoning with tangram shapes. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 582– 601, Abu Dhabi, United Arab Emirates. Associat… find
  21. Ksenia Konyushkova, Christos Kaplanis, Serkan Cabi, and Misha Denil. 2025. Vision-language model dialog games for self-improvement. Preprint, arXiv:2502.02740. — [arXiv:2502.02740](https://arxiv.org/abs/2502.02740) find
  22. Robert M. Krauss and Sidney Weinheimer. 1964. Changes in reference phrases as a function of frequency of usage in social interaction: a preliminary study. Psychonomic Science, 1:113–114. find
  23. Angeliki Lazaridou, Anna Potapenko, and Olivier Tieleman. 2020. Multi-agent communication meets natural language: Synergies between functional and structural language learning. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7663–7674, Online. Associ… find
  24. Mina Lee, Tatsunori B. Hashimoto, and Percy Liang. 2019. Learning autocomplete systems as a communication game. Preprint, arXiv:1911.06964. find
  25. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision – ECCV 2014, pages 740–755, Cham. Springer International Publishing.
  26. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Thirty-seventh Conference on Neural Information Processing Systems.
  27. Prasann Singhal, Tanya Goyal, Jiacheng Xu, and Greg Durrett. 2024. A long way to go: Investigating length correlations in RLHF. In First Conference on Language Modeling.
  28. Ece Takmaz, Mario Giulianelli, Sandro Pezzelle, Arabella Sinclair, and Raquel Fernández. 2020. Refer, Reuse, Reduce: Generating Subsequent References in Visual and Conversational Contexts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4350–4… find
  29. Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. 2025. Siglip 2: Multilingual vision-language encoders with improv…
  30. Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec. 2020. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl. ↗
  31. Tianhao Wu, Weizhe Yuan, Olga Golovneva, Jing Xu, Yuandong Tian, Jiantao Jiao, Jason E. Weston, and Sainbayar Sukhbaatar. 2024. Meta-rewarding language models: Self-improving alignment with llm-as-a-meta-judge. ArXiv, abs/2407.19594. as a string that was part of the template to generate prompts. We…
  32. Hao Zhu, Graham Neubig, and Yonatan Bisk. 2021. Few-shot language coordination by modeling theory of mind. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 12901–12911. PMLR. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/vaduguru-2026-success-cost-elicit-convention-formation/