---
title: "When Discourse Pressures Conflict: Information Structure in"
description: "Vision-language models (VLMs) are usually judged on whether they get visual content right. 46 references."
authors: ["Marcell Fekete", "Johannes Bjerva", "Tamas Kaldi"]
year: 2026
venue: "arXiv:2605.28346"
source: "https://arxiv.org/abs/2605.28346"
doi: "10.48550/arxiv.2605.28346"
arxiv: "2605.28346"
licence: "CC BY-NC-SA 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by-nc-sa/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/fekete-2026-discourse-pressures-conflict-information-structure/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/fekete-2026-discourse-pressures-conflict-information-structure/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/48/) / When Discourse Pressures Conflict: Information Structure in Vision-Language Model Outputs

# When Discourse Pressures Conflict: Information Structure in Vision-Language Model Outputs

**Authors:** Marcell Fekete, [Johannes Bjerva](https://md.research.elamur.ai/free/en/authors/johannes-bjerva/) and Tamas Kaldi · **Year:** 2026 · **Venue:** arXiv:2605.28346 · **Licence:** [CC BY-NC-SA 4.0 (arXiv)](http://creativecommons.org/licenses/by-nc-sa/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2605.28346) · [PDF](https://arxiv.org/pdf/2605.28346) · [DOI 10.48550/arxiv.2605.28346](https://doi.org/10.48550/arxiv.2605.28346) · [arXiv 2605.28346](https://arxiv.org/abs/2605.28346) · [OpenAlex](https://openalex.org/W7162628935)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
Vision-language models (VLMs) are usually judged on whether they get visual content right. This paper asks a different question: do they phrase that content in a way that fits the discourse? It focuses on information structure (IS), the split between discourse-old Topics and discourse-new Foci. English mostly marks IS through intonation, which is invisible in writing. So the authors use Hungarian, where Topic sits sentence-initially and Focus sits right before the verb, displacing the verb modifier (igekötő). They adapt a psycholinguistic visual question answering task. Participants see an image of a transitive event, and a wh-question makes either the subject or the object discourse-old. The answers of six VLMs, both proprietary and open-weight, are compared with those of Hungarian speakers. The models do produce IS-appropriate constructions. However, when discourse status, grammatical role and definiteness pull in different directions, humans vary their strategies while the models collapse onto a few fixed templates, a pattern reminiscent of mode collapse. The authors conclude that evaluation should check the distribution of linguistic form, not just content accuracy.

## Key points
- The authors describe this as the first systematic study of IS realisation in VLM outputs. They found no existing IS benchmark for either LLMs or VLMs, only a narrow English probing study (Cuneo et al., 2025).
- Responses are coded by Hungarian word order (SVIO, OSVI, OIVS, SIVO, SOVI, OVIS). Each order is classified as topicalisation, preverbal Focus, post-verbal Focus, neutral, or infelicitous under each focus condition. The most direct written cue for Focus is a displaced verb modifier.
- The VLMs topicalise subjects far more strongly than humans do. They also show a subject–object asymmetry in Focus definiteness that is absent from the human data.
- The stimuli are semantically reversible transitive events with animate human characters (20 male, 4 female) and verbs carrying verb modifiers. Two expert annotators coded the responses, and the anonymised dataset is promised upon publication.
- The human sample is young and mostly female, and only one language is tested. The authors flag both as limits on generality.

## Interesting ideas
- Make a hidden phenomenon observable by testing it in a language that marks it overtly. Here Hungarian word order exposes Topic and Focus, which English hides in prosody, so a covert property becomes checkable in plain text.
- Let the question decide which referent counts as given, then score only the form of the answer. This separates getting the content right from packaging it right.
- Compare the spread of a system's outputs with the spread of human outputs, not just its accuracy. Collapse onto a single template is a failure even when every individual answer is correct.
- Cross several interacting pressures (discourse status, grammatical role, definiteness) so they sometimes conflict. This reveals whether a system follows the real cue or falls back on a prior such as "subject = topic".
- Use semantically reversible events, such as a king and a ghost acting on each other, so that world plausibility cannot reveal who did what to whom.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (frames): prag could tag Topic and Focus from the overt English devices that do exist, such as clefts, fronting and passives, and validate this tagging on cases where the question context fixes what is given.
- **I2** (back): the generator answering a wh-question could keep the given referent as subject or pronoun and put the new one in final or clefted position, and a test set could check this.
- **I3** (learn): when LLM teacher tiers generate silver data, measure how varied their constructions are against human treebank distributions to catch template collapse before training.
- **I4** (learn): negative-control evaluations could set role priors against contextual cues to expose shortcut heuristics in the parser or the frame builder.
- **I5** (answer): fable and tale QA tests could include reversible-role questions, forcing readers to rely on the UD tree and frames rather than plausibility guessing.

*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

- [Max Planck Institute for Psycholinguistics](https://md.research.elamur.ai/free/en/labs/max-planck-institute-for-psycholinguistics/) — Research institute, Nijmegen, Netherlands · 6 works · 11,996 citations
- [Aalborg University](https://md.research.elamur.ai/free/en/universities/aalborg-university/) — University, Aalborg, Denmark · 12 works · 743 citations
- [Eötvös Loránd University](https://md.research.elamur.ai/free/en/universities/eotvos-lorand-university/) — University, Budapest, Hungary · 5 works · 256 citations

Also on the byline: [Hungarian University of Sports Science](https://ror.org/01zh80k81).

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 2,869 works

🧭 Topics (4): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,320 works · [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,263 works · [🖼️ Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) · 657 works · [🧵 Discourse analysis](https://md.research.elamur.ai/free/en/topics/discourse-analysis/) · 466 works

## Cite

Cite as: Marcell Fekete, Johannes Bjerva and Tamas Kaldi. 2026. When Discourse Pressures Conflict: Information Structure in Vision-Language Model Outputs. arXiv:2605.28346.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{fekete2026when,
  title = {{When Discourse Pressures Conflict: Information Structure in Vision-Language Model Outputs}},
  author = {Marcell Fekete, Johannes Bjerva and Tamas Kaldi},
  year = {2026},
  doi = {10.48550/arxiv.2605.28346},
  eprint = {2605.28346},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2605.28346}
}
```

- [← previous When depth is redundant: Efficient transformer-based speech anti-spoofing](https://md.research.elamur.ai/free/en/papers/tran-2026-when-depth-is-redundant-efficient/)
- [next → When Does Auxiliary Modality Matter in Solving Geometric Problems? A Comprehensive Study of Textual, Formal, and Visual Modalities](https://md.research.elamur.ai/free/en/papers/namgoong-2026-when-does-auxiliary-modality-matter/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/48/)

## References

46 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Judith Aissen. 1999. Markedness and subject choice in optimality theory. Natural Language & Linguistic Theory, 17(4):673–711. [↗](https://doi.org/10.1023/a:1006335629372)
2. Jennifer E. Arnold, Elsi Kaiser, Jason M. Kahn, and Lucy K. Kim. 2013. Information structure: linguistic, cognitive, and processing approaches. WIREs Cognitive Science, 4(4):403–413. [↗](https://doi.org/10.1002/wcs.1234)
3. Harald Cramér. 1946. Mathematical Methods of Statistics. Princeton University Press, Princeton. [find](https://scholar.google.com/scholar?q=Harald+Cram%C3%A9r%2E+1946%2E+Mathematical+Methods+of+Statistics%2E+Princeton+University+Press%2C+Princeton%2E)
4. Nicole Cuneo, Eleanor Graves, Supantho Rakshit, and Adele E. Goldberg. 2025. [For GPT-4 as with humans: Information structure predicts acceptability of long-distance dependencies](https://md.research.elamur.ai/free/en/papers/cuneo-2025-gpt-4-humans-information-structure/). Preprint, arXiv:2505.09005. [↗](https://arxiv.org/abs/2505.09005)
5. Delphine Dahan, Michael K. Tanenhaus, and Craig G. Chambers. 2002. Accent and reference resolution in spoken-language comprehension. Journal of Memory and Language, 47(2):292–314. [↗](https://doi.org/10.1016/s0749-596x(02)00001-3)
6. Dorottya Demszky. 2021. The role of verb semantics in Hungarian verb-object order. Proceedings of the Linguistic Society of America, 6(1):54–68. [↗](https://doi.org/10.3765/plsa.v6i1.4941)
7. Katalin É. Kiss, editor. 1995. Discourse Configurational Languages. Oxford University Press. [find](https://scholar.google.com/scholar?q=Katalin+%C3%89%2E+Kiss%2C+editor%2E+1995%2E+Discourse+Configurational+Languages%2E+Oxford+University+Press%2E)
8. Katalin É. Kiss. 1998. Identificational focus versus information focus. Language, 74(2):245–273. [↗](https://doi.org/10.1353/lan.1998.0211)
9. Katalin É. Kiss. 2002. The Syntax of Hungarian, 1 edition. Cambridge University Press. [find](https://scholar.google.com/scholar?q=Katalin+%C3%89%2E+Kiss%2E+2002%2E+The+Syntax+of+Hungarian%2C+1+edition%2E+Cambridge+University+Press%2E)
10. Bradley Efron and Robert J. Tibshirani. 1993. An Introduction to the Bootstrap. Number 57 in Monographs on Statistics and Applied Probability. Chapman & Hall, New York. [↗](https://doi.org/10.1007/978-1-4899-4541-9)
11. Nomi Erteschik-Shir. 2013. Information structure and (in)definiteness. In Crosslinguistic Studies on Noun Phrase Structure and Reference, pages 23–51. Brill, Leiden. [↗](https://doi.org/10.1163/9789004261440_003)
12. Carol A. Fowler and Jonathan Housum. 1987. Talkers’ signaling of “new” and “old” words in speech and listeners’ perception and use of the distinction. Journal of Memory and Language, 26(5):489–504. [↗](https://doi.org/10.1016/0749-596x(87)90136-7)
13. Talmy Givón. 1983. Topic continuity in discourse: The functional domain of switch reference. In John Haiman and Pamela Munro, editors, Switch Reference and Universal Grammar, pages 51–82. John Benjamins, Amsterdam. [↗](https://doi.org/10.1075/tsl.2.06giv)
14. Geoffrey Haig and Stefan Schnell. 2016. The discourse basis of ergativity revisited. Language, 92(3):591– 618. [↗](https://doi.org/10.1353/lan.2016.0049)
15. Eva Hajičová. 2019. A Plea for Information Structure as a Part of Meaning Representation. In Proceedings of the First International Workshop on Designing Meaning Representations, pages 66–72, Florence, Italy. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Eva+Hajic%CC%8Cov%C3%A1%2E+2019%2E+A+Plea+for+Information+Structure+as+a+Part+of+Meaning+Representation%2E+In+Proceedings+of+the+First+International+Workshop+on+Designing+Meaning+Representations%2C+pages+66%E2%80%9372%2C+Florenc)
16. Eva Hajičová, Petr Sgall, and Hana Skoumalova. 1995. An Automatic Procedure for Topic-Focus Identification. Computational Linguistics, 21(1):81–94. [find](https://scholar.google.com/scholar?q=Eva+Hajic%CC%8Cov%C3%A1%2C+Petr+Sgall%2C+and+Hana+Skoumalova%2E+1995%2E+An+Automatic+Procedure+for+Topic%2DFocus+Identification%2E+Computational+Linguistics%2C+21%281%29%3A81%E2%80%9394%2E)
17. Jorge E. Hankamer. 1971. Constraints on deletion in syntax. Ph.D. thesis, Yale University. PhD dissertation. [find](https://scholar.google.com/scholar?q=Jorge+E%2E+Hankamer%2E+1971%2E+Constraints+on+deletion+in+syntax%2E+Ph%2ED%2E+thesis%2C+Yale+University%2E+PhD+dissertation%2E)
18. Irene Heim. 1982. The semantics of definite and indefinite noun phrases. Ph.D. thesis, University of Massachusetts, Amherst. PhD dissertation. [↗](https://doi.org/10.5281/zenodo.8100764)
19. Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. [The state and fate of linguistic diversity and inclusion in the NLP world](https://md.research.elamur.ai/free/en/papers/joshi-2020-state-fate-linguistic-diversity-inclusion/). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6282–6293, Online. Association for Computational Linguistics.
20. Jaap Jumelet, Leonie Weissweiler, Joakim Nivre, and Arianna Bisazza. 2026. MultiBLiMP 1.0: A massively multilingual benchmark of linguistic minimal pairs. Transactions of the Association for Computational Linguistics, 14:193–216. Tamás Káldi and Ágnes Lukács. in preparation. [↗](https://doi.org/10.1162/tacl.a.600)
21. Tamás Káldi and Anna Babarczy. 2021. Linguistic focus guides attention during the encoding and refreshing of working memory content. Journal of Memory and Language, 116:104187. [↗](https://doi.org/10.1016/j.jml.2020.104187)
22. Tamás Káldi, Lili Madarász, and Anna Babarczy. 2020. Contextual triggers of the Hungarian pre-verbal focus structure – A guided production study. In Veronika Hegedűs and Irene Vogel, editors, Approaches to Hungarian: Volume 16: Papers from the 2017 Budapest Conference, volume 16, pages 73–96. John Benjamins Publishing Company. [↗](https://doi.org/10.1075/atoh.16.04kal)
23. Tamás Káldi, Ágnes Szőllősi, and Anna Babarczy. 2021. Hungarian structural focus: Accessibility to focused elements and their alternatives in working memory and delayed recognition memory. Frontiers in Psychology, 12. [↗](https://doi.org/10.3389/fpsyg.2021.514886)
24. István Kenesei. 2006. Focus as identification. In Valéria Molnár and Susanne Winkler, editors, The Architecture of Focus, pages 137–168. Mouton de Gruyter, Berlin. [↗](https://doi.org/10.1515/9783110922011.137)
25. Robert Kirk, Ishita Mediratta, Christoforos Nalmpantis, Jelena Luketina, Eric Hambro, Edward Grefenstette, and Roberta Raileanu. 2024. [Understanding the effects of RLHF on LLM generalisation and diversity](https://md.research.elamur.ai/free/en/papers/kirk-2023-understanding-effects-rlhf-llm-generalisation/). In The Twelfth International Conference on Learning Representations. [↗](https://doi.org/10.48550/arxiv.2310.06452)
26. Manfred Krifka. 2008. Basic notions of information structure. Acta Linguistica Hungarica, 55(3–4):243– 276. [↗](https://doi.org/10.1556/aling.55.2008.3-4.2)
27. Susumu Kuno. 1972. Functional sentence perspective: A case study from Japanese and English. Linguistic Inquiry, 3(3):269–320. [find](https://scholar.google.com/scholar?q=Susumu+Kuno%2E+1972%2E+Functional+sentence+perspective%3A+A+case+study+from+Japanese+and+English%2E+Linguistic+Inquiry%2C+3%283%29%3A269%E2%80%93320%2E)
28. Knud Lambrecht. 1994. Information Structure and Sentence Form: Topic, Focus, and the Mental Representations of Discourse Referents, volume 71 of Cambridge Studies in Linguistics. Cambridge University Press, Cambridge. [find](https://scholar.google.com/scholar?q=Knud+Lambrecht%2E+1994%2E+Information+Structure+and+Sentence+Form%3A+Topic%2C+Focus%2C+and+the+Mental+Representations+of+Discourse+Referents%2C+volume+71+of+Cambridge+Studies+in+Linguistics%2E+Cambridge+University+)
29. Bolei Ma, Yuting Li, Wei Zhou, Ziwei Gong, Yang Janet Liu, Katja Jasinskaja, Annemarie Friedrich, Julia Hirschberg, Frauke Kreuter, and Barbara Plank. 2025. [Pragmatics in the era of large language models: A survey on datasets, evaluation, opportunities and challenges](https://md.research.elamur.ai/free/en/papers/ma-2025-pragmatics-era-large-language-models/). In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8679–8696, Vienna, Austria. Association for Computational Linguistics.
30. Brian MacWhinney and Csaba Pléh. 1988. The processing of restrictive relative clauses in hungarian. Cognition, 29(1):95–141. [↗](https://doi.org/10.1016/0010-0277(88)90034-0)
31. Tim P. Morris, Ian R. White, and Michael J. Crowther. 2019. Using simulation studies to evaluate statistical methods. Statistics in Medicine, 38(11):2074–2102. [↗](https://doi.org/10.1002/sim.8086)
32. Csaba Oravecz, Tamás Váradi, and Bálint Sass. 2014. The Hungarian Gigaword corpus. In Proceedings of LREC 2014, page 5. [find](https://scholar.google.com/scholar?q=Csaba+Oravecz%2C+Tam%C3%A1s+V%C3%A1radi%2C+and+B%C3%A1lint+Sass%2E+2014%2E+The+Hungarian+Gigaword+corpus%2E+In+Proceedings+of+LREC+2014%2C+page+5%2E)
33. György Orosz, Gergő Szabó, Péter Berkecz, Zsolt Szántó, and Richárd Farkas. 2023. Advancing Hungarian Text Processing with HuSpaCy: Efficient and Accurate NLP Pipelines. In Text, Speech, and Dialogue, pages 58–69, Cham. Springer Nature Switzerland. [find](https://scholar.google.com/scholar?q=Gy%C3%B6rgy+Orosz%2C+Gergo%CC%8B+Szab%C3%B3%2C+P%C3%A9ter+Berkecz%2C+Zsolt+Sz%C3%A1nt%C3%B3%2C+and+Rich%C3%A1rd+Farkas%2E+2023%2E+Advancing+Hungarian+Text+Processing+with+HuSpaCy%3A+Efficient+and+Accurate+NLP+Pipelines%2E+In+Text%2C+Speech%2C+and+Dialogue)
34. György Orosz, Zsolt Szántó, Péter Berkecz, Gergő Szabó, and Richárd Farkas. 2022. HuSpaCy: an industrial-strength Hungarian natural language processing toolkit. In XVIII. Magyar Számítógépes Nyelvészeti Konferencia, pages 59–73. [↗](https://doi.org/10.48550/arxiv.2201.01956)
35. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. [Training language models to follow instructions with human feedback](https://md.research.elamur.ai/free/en/papers/ouyang-2022-training-language-models-follow-instructions/). In Advances in Neural Information Processing Systems, volume 35, pages 27730–27744.
36. Tanya Reinhart. 1981. Pragmatics and linguistics: an analysis of sentence topics. Philosophica, 27:53–94. [↗](https://doi.org/10.21825/philosophica.82606)
37. Timo B. Roettger, Tim Mahrt, and Jennifer Cole. 2019. Mapping prosody onto meaning – the case of information structure in American English. Language, Cognition and Neuroscience, 34(7):841–860. [↗](https://doi.org/10.1080/23273798.2019.1587482)
38. Alison J. S. Sanford, Jessie Price, and Anthony J. Sanford. 2009. Enhancement and suppression effects resulting from information structuring in sentences. Memory & Cognition, 37(6):880–888. [↗](https://doi.org/10.3758/mc.37.6.880)
39. Alison J. S. Sanford, Anthony J. Sanford, Jo Molle, and Catherine Emmott. 2006. Shallow processing and attention capture in written and spoken discourse. Discourse Processes, 42(2):109–130. [↗](https://doi.org/10.1207/s15326950dp4202_2)
40. Taiga Someya and Yohei Oseki. 2023. JBLiMP: Japanese benchmark of linguistic minimal pairs. In Findings of the Association for Computational Linguistics: EACL 2023, pages 1581–1594, Dubrovnik, Croatia. Association for Computational Linguistics. [↗](https://doi.org/10.48448/qn5t-y408)
41. Ekaterina Taktasheva, Maxim Bazhukov, Kirill Koncha, Alena Fenogenova, Ekaterina Artemova, and Vladislav Mikhailov. 2024. RuBLiMP: Russian benchmark of linguistic minimal pairs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 9268–9299, Miami, Florida, USA. Association for Computational Linguistics. [↗](https://doi.org/10.48550/arxiv.2406.19232)
42. Patricia Ward and Patrick Sturt. 2007. Linguistic focus and memory: An eye movement study. Memory & Cognition, 35(1):73–86. [↗](https://doi.org/10.3758/bf03195944)
43. Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng-Fu Wang, and Samuel R. Bowman. 2020. BLiMP: A benchmark of linguistic minimal pairs for English. In Proceedings of the Society for Computation in Linguistics 2020, pages 409–410, New York, New York. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Alex+Warstadt%2C+Alicia+Parrish%2C+Haokun+Liu%2C+Anhad+Mohananey%2C+Wei+Peng%2C+Sheng%2DFu+Wang%2C+and+Samuel+R%2E+Bowman%2E+2020%2E+BLiMP%3A+A+benchmark+of+linguistic+minimal+pairs+for+English%2E+In+Proceedings+of+the+Socie)
44. Thomas Wasow. 1997. Remarks on grammatical weight. Language Variation and Change, 9(1):81–105. [↗](https://doi.org/10.1017/s0954394500001800)
45. Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. [Finetuned language models are zero-shot learners](https://md.research.elamur.ai/free/en/papers/wei-2021-finetuned-language-models-zero-shot/). In International Conference on Learning Representations. [↗](https://doi.org/10.48550/arxiv.2109.01652)
46. Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael R Tomz, Christopher D Manning, and Weiyan Shi. 2025. Verbalized sampling: How to mitigate mode collapse and unlock llm diversity. arXiv preprint arXiv:2510.01171. [↗](https://arxiv.org/abs/2510.01171)

*Updated 1 November 2026*