---
title: "Better Scores, Worse Grounding — Park 2026"
description: "Dialogue fact verification (DFV) checks whether a response is supported by evidence, but the response often only makes sense once its pronouns are…"
authors: ["Hyunkyung Park", "Arkaitz Zubiaga"]
year: 2026
venue: "Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue"
source: "https://aclanthology.org/2026.sigdial-1.51/"
doi: "10.18653/v1/2026.sigdial-1.51"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/park-2026-better-scores-worse-grounding-hidden/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/park-2026-better-scores-worse-grounding-hidden/)

[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/2/) / Better Scores, Worse Grounding: Hidden Regressions after Fine-Tuning in Dialogue Fact Verification

# Better Scores, Worse Grounding: Hidden Regressions after Fine-Tuning in Dialogue Fact Verification

**Authors:** Hyunkyung Park, [Arkaitz Zubiaga](https://md.research.elamur.ai/free/en/authors/arkaitz-zubiaga/) · **Year:** 2026 · **Venue:** Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.sigdial-1.51/) · [PDF](https://aclanthology.org/2026.sigdial-1.51.pdf) · [DOI 10.18653/v1/2026.sigdial-1.51](https://doi.org/10.18653/v1/2026.sigdial-1.51) · [OpenAlex](https://openalex.org/W7214030851)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14)

## Summary
Dialogue fact verification (DFV) checks whether a response is supported by evidence, but the response often only makes sense once its pronouns are resolved against earlier turns. The authors ask whether fine-tuning can raise benchmark scores while quietly damaging this kind of contextual grounding. They kept the original joint input intact, with evidence as premise and dialogue context plus response as hypothesis, instead of rewriting responses into standalone claims. Using referent-annotated audit files constructed from DialFact and FaithDial, they evaluated six encoder-only verifiers, comparing each model's behaviour prior to and following fine-tuning matched to the corresponding data source. The diagnostic has two parts. First, examples are grouped by how their predictions change, contrasting negative flips (correct before, wrong after) with stable-correct cases. Second, the Premise-Preference Score (PPS) uses control-adjusted masking to estimate whether gold-label support relies more on evidence-side or context-side referent cues. Macro-F1 rose for every model and evaluation set. Yet the newly broken cases leaned more on premise cues. This shows that aggregate metrics can mask grounding regressions, so evaluations should compare model behaviour before and after fine-tuning.

## Key points
- **Audit resources:**
  - DF-Coref-Valid and DF-Coref-Test, derived from DialFact, which serves as the in-domain source.
  - FD-Random and FD-Topic, derived from FaithDial, which provides out-of-domain corroboration.
  - Records annotate response pronouns, such as "It", and their referent mentions in context.
- **Panel:** six encoder-only verifiers by three evaluation sets, giving 18 comparisons. Fine-tuning improved Macro-F1 in all of them.
- **Main finding:** negative-flip (NF) cases showed stronger premise-side sensitivity under PPS than positive no-flip (PNF) cases in 17 of 18 comparisons. This held for all six models on DF-Coref-Test.
- **Scope:** the audit deliberately covers a narrow, conservative slice. It includes only response-side pronouns whose referent is recoverable from context and that have a surface-matched competing cue in the evidence. For example, on DF-Coref-Test, 305 jointly diagnosable examples per model were available for PPS analysis.
- **Ablation:** a PPS-stratified selection ablation, with tertiles defined on DF-Coref-Valid, is reported only as exploratory. Its results are given as NF recovery and PNF harm against an oracle upper bound.

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

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

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

```bibtex
@inproceedings{park2026better,
  title = {{Better Scores, Worse Grounding: Hidden Regressions after Fine-Tuning in Dialogue Fact Verification}},
  author = {Park, Hyunkyung and Zubiaga, Arkaitz},
  booktitle = {Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue},
  year = {2026},
  doi = {10.18653/v1/2026.sigdial-1.51},
  url = {https://doi.org/10.18653/v1/2026.sigdial-1.51}
}
```

- [← previous Best-Worst Scaling of Hype in Biomedical Research: Building an Intensity Lexicon of Promotional Adjectives](https://md.research.elamur.ai/free/en/papers/millar-2026-best-worst-scaling-hype-biomedical/)
- [next → Beyond BLEU: Linguistic Invisibility and Interactional Repair Sequence in End-to-End Sign Language Translation](https://md.research.elamur.ai/free/en/papers/wang-2026-beyond-bleu-linguistic-invisibility-interactional/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/2/)

## References

22 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Abbas Ghaddar, David Alfonso-Hermelo, Philippe Langlais, Mehdi Rezagholizadeh, Boxing Chen, and Prasanna Parthasarathi. 2024. Charp: Conversation history awareness probing for knowledge-grounded dialogue systems. In Findings of the Association for Computational Linguistics: ACL 2024, pages 1534– 155… [find](https://scholar.google.com/scholar?q=Abbas+Ghaddar%2C+David+Alfonso%2DHermelo%2C+Philippe+Langlais%2C+Mehdi+Rezagholizadeh%2C+Boxing+Chen%2C+and+Prasanna+Parthasarathi%2E+2024%2E+Charp%3A+Conversation+history+awareness+probing+for+knowledge%2Dgrounded+dialo)
2. Max Glockner, Ieva Staliūnaitė, James Thorne, Gisela Vallejo, Andreas Vlachos, and Iryna Gurevych. 2024. AmbiFC: Fact-checking ambiguous claims with evidence. Transactions of the Association for Computational Linguistics, 12:1–18. [find](https://scholar.google.com/scholar?q=Max+Glockner%2C+Ieva+Staliu%CC%84naite%CC%87%2C+James+Thorne%2C+Gisela+Vallejo%2C+Andreas+Vlachos%2C+and+Iryna+Gurevych%2E+2024%2E+AmbiFC%3A+Fact%2Dchecking+ambiguous+claims+with+evidence%2E+Transactions+of+the+Association+for+Com)
3. Sachin Goyal, Christina Baek, J. Zico Kolter, and Aditi Raghunathan. 2025. Context-parametric inversion: Why instruction finetuning can worsen context reliance. Preprint, arXiv:2410.10796. — [arXiv:2410.10796](https://arxiv.org/abs/2410.10796) [find](https://scholar.google.com/scholar?q=Sachin+Goyal%2C+Christina+Baek%2C+J%2E+Zico+Kolter%2C+and+Aditi+Raghunathan%2E+2025%2E+Context%2Dparametric+inversion%3A+Why+instruction+finetuning+can+worsen+context+reliance%2E+Preprint%2C+arXiv%3A2410%2E10796%2E+%E2%80%94+%5BarXiv%3A24)
4. Anisha Gunjal and Greg Durrett. 2024. Molecular facts: Desiderata for decontextualization in LLM fact verification. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3751–3768, Miami, Florida, USA. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Anisha+Gunjal+and+Greg+Durrett%2E+2024%2E+Molecular+facts%3A+Desiderata+for+decontextualization+in+LLM+fact+verification%2E+In+Findings+of+the+Association+for+Computational+Linguistics%3A+EMNLP+2024%2C+pages+3751)
5. Prakhar Gupta, Chien-Sheng Wu, Wenhao Liu, and Caiming Xiong. 2022. DialFact: A benchmark for fact-checking in dialogue. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3785–3801, Dublin, Ireland. Association for Computational… [find](https://scholar.google.com/scholar?q=Prakhar+Gupta%2C+Chien%2DSheng+Wu%2C+Wenhao+Liu%2C+and+Caiming+Xiong%2E+2022%2E+DialFact%3A+A+benchmark+for+fact%2Dchecking+in+dialogue%2E+In+Proceedings+of+the+60th+Annual+Meeting+of+the+Association+for+Computational+)
6. Hongyu Li, Liang Ding, Meng Fang, and Dacheng Tao. 2024. Revisiting catastrophic forgetting in large language model tuning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 4297–4308, Miami, Florida, USA. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Hongyu+Li%2C+Liang+Ding%2C+Meng+Fang%2C+and+Dacheng+Tao%2E+2024%2E+Revisiting+catastrophic+forgetting+in+large+language+model+tuning%2E+In+Findings+of+the+Association+for+Computational+Linguistics%3A+EMNLP+2024%2C+pa)
7. Elita Lobo, Chirag Agarwal, and Himabindu Lakkaraju. 2025. On the impact of fine-tuning on chain-of-thought reasoning. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), p… [find](https://scholar.google.com/scholar?q=Elita+Lobo%2C+Chirag+Agarwal%2C+and+Himabindu+Lakkaraju%2E+2025%2E+On+the+impact+of+fine%2Dtuning+on+chain%2Dof%2Dthought+reasoning%2E+In+Proceedings+of+the+2025+Conference+of+the+Nations+of+the+Americas+Chapter+of+t)
8. Mamta and Oana Cocarascu. 2025. FactEval: Evaluating the robustness of fact verification systems in the era of large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1:… [find](https://scholar.google.com/scholar?q=Mamta+and+Oana+Cocarascu%2E+2025%2E+FactEval%3A+Evaluating+the+robustness+of+fact+verification+systems+in+the+era+of+large+language+models%2E+In+Proceedings+of+the+2025+Conference+of+the+Nations+of+the+Americ)
9. Kawshik Manikantan, Makarand Tapaswi, Vineet Gandhi, and Shubham Toshniwal. 2025. IdentifyMe: A challenging long-context mention resolution benchmark for LLMs. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Languag… [find](https://scholar.google.com/scholar?q=Kawshik+Manikantan%2C+Makarand+Tapaswi%2C+Vineet+Gandhi%2C+and+Shubham+Toshniwal%2E+2025%2E+IdentifyMe%3A+A+challenging+long%2Dcontext+mention+resolution+benchmark+for+LLMs%2E+In+Proceedings+of+the+2025+Conference+of)
10. Dasha Metropolitansky and Jonathan Larson. 2025. Towards effective extraction and evaluation of factual claims. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6996–7045, Vienna, Austria. Association for Computational Linguist… [find](https://scholar.google.com/scholar?q=Dasha+Metropolitansky+and+Jonathan+Larson%2E+2025%2E+Towards+effective+extraction+and+evaluation+of+factual+claims%2E+In+Proceedings+of+the+63rd+Annual+Meeting+of+the+Association+for+Computational+Linguisti)
11. Hyunkyung Park and Arkaitz Zubiaga. 2026. Bicongate: Consistency-gated de-colloquialisation for dialogue fact-checking. In Proceedings of the Ninth Fact Extraction and VERification Workshop (FEVER), pages 59–73. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Hyunkyung+Park+and+Arkaitz+Zubiaga%2E+2026%2E+Bicongate%3A+Consistency%2Dgated+de%2Dcolloquialisation+for+dialogue+fact%2Dchecking%2E+In+Proceedings+of+the+Ninth+Fact+Extraction+and+VERification+Workshop+%28FEVER%29%2C+p)
12. Gorjan Radevski, Kiril Gashteovski, Shahbaz Syed, Christopher Malon, Sebastien Nicolas, Chia-Chien Hung, Timo Sztyler, Verena Heußer, Wiem Ben Rim, Masafumi Enomoto, Kunihiro Takeoka, Masafumi Oyamada, Goran Glavaš, and Carolin Lawrence. 2025. On synthesizing data for context attribution in question… [find](https://scholar.google.com/scholar?q=Gorjan+Radevski%2C+Kiril+Gashteovski%2C+Shahbaz+Syed%2C+Christopher+Malon%2C+Sebastien+Nicolas%2C+Chia%2DChien+Hung%2C+Timo+Sztyler%2C+Verena+Heu%C3%9Fer%2C+Wiem+Ben+Rim%2C+Masafumi+Enomoto%2C+Kunihiro+Takeoka%2C+Masafumi+Oyamada)
13. Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. 2023. Measuring attribution in natural language generation models. Computational Linguistics, 49(4):777–840. [find](https://scholar.google.com/scholar?q=Hannah+Rashkin%2C+Vitaly+Nikolaev%2C+Matthew+Lamm%2C+Lora+Aroyo%2C+Michael+Collins%2C+Dipanjan+Das%2C+Slav+Petrov%2C+Gaurav+Singh+Tomar%2C+Iulia+Turc%2C+and+David+Reitter%2E+2023%2E+Measuring+attribution+in+natural+languag) [our card](https://md.research.elamur.ai/free/en/papers/rashkin-2023-measuring-attribution-natural-language-generation/)
14. Evgeniia Razumovskaia, Ivan Vulić, Pavle Marković, Tomasz Cichy, Qian Zheng, Tsung-Hsien Wen, and Paweł Budzianowski. 2024. Dial BeInfo for faithfulness: Improving factuality of information-seeking dialogue via behavioural fine-tuning. In Findings of the Association for Computational Linguistics:… [find](https://scholar.google.com/scholar?q=Evgeniia+Razumovskaia%2C+Ivan+Vulic%CC%81%2C+Pavle+Markovic%CC%81%2C+Tomasz+Cichy%2C+Qian+Zheng%2C+Tsung%2DHsien+Wen%2C+and+Pawe%C5%82+Budzianowski%2E+2024%2E+Dial+BeInfo+for+faithfulness%3A+Improving+factuality+of+information%2Dseeking+)
15. Ieva Staliunaite and Andreas Vlachos. 2025. Dis2Dis: Explaining ambiguity in fact-checking. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 246–267, Albuquerque, New Mexico. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Ieva+Staliunaite+and+Andreas+Vlachos%2E+2025%2E+Dis2Dis%3A+Explaining+ambiguity+in+fact%2Dchecking%2E+In+Findings+of+the+Association+for+Computational+Linguistics%3A+NAACL+2025%2C+pages+246%E2%80%93267%2C+Albuquerque%2C+New+Me)
16. Yufei Tao and Ameeta Agrawal. 2026. No-worse context-aware decoding: Preventing neutral regression in context-conditioned generation. Preprint, arXiv:2604.16686. [find](https://scholar.google.com/scholar?q=Yufei+Tao+and+Ameeta+Agrawal%2E+2026%2E+No%2Dworse+context%2Daware+decoding%3A+Preventing+neutral+regression+in+context%2Dconditioned+generation%2E+Preprint%2C+arXiv%3A2604%2E16686%2E)
17. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Vo… [find](https://scholar.google.com/scholar?q=James+Thorne%2C+Andreas+Vlachos%2C+Christos+Christodoulopoulos%2C+and+Arpit+Mittal%2E+2018%2E+FEVER%3A+a+large%2Dscale+dataset+for+fact+extraction+and+VERification%2E+In+Proceedings+of+the+2018+Conference+of+the+Nort) [our card](https://md.research.elamur.ai/free/en/papers/thorne-2018-fever-large-scale-dataset-fact/)
18. Zineddine Tighidet, Jiali Mei, Benjamin Piwowarski, and Patrick Gallinari. 2024. Probing language models on their knowledge source. In Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 604–614, Miami, Florida, US. Association for Computational Lin… [find](https://scholar.google.com/scholar?q=Zineddine+Tighidet%2C+Jiali+Mei%2C+Benjamin+Piwowarski%2C+and+Patrick+Gallinari%2E+2024%2E+Probing+language+models+on+their+knowledge+source%2E+In+Proceedings+of+the+7th+BlackboxNLP+Workshop%3A+Analyzing+and+Interp)
19. Miriam Wanner, Benjamin Van Durme, and Mark Dredze. 2025. DnDScore: Decontextualization and decomposition for factuality verification in long-form text generation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 23609–23626, Suzhou, China. Association… [find](https://scholar.google.com/scholar?q=Miriam+Wanner%2C+Benjamin+Van+Durme%2C+and+Mark+Dredze%2E+2025%2E+DnDScore%3A+Decontextualization+and+decomposition+for+factuality+verification+in+long%2Dform+text+generation%2E+In+Proceedings+of+the+2025+Conferenc)
20. Yumo Xu, Peng Qi, Jifan Chen, Kunlun Liu, Rujun Han, Lan Liu, Bonan Min, Vittorio Castelli, Arshit Gupta, and Zhiguo Wang. 2025. CiteEval: Principle-driven citation evaluation for source attribution. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1… [find](https://scholar.google.com/scholar?q=Yumo+Xu%2C+Peng+Qi%2C+Jifan+Chen%2C+Kunlun+Liu%2C+Rujun+Han%2C+Lan+Liu%2C+Bonan+Min%2C+Vittorio+Castelli%2C+Arshit+Gupta%2C+and+Zhiguo+Wang%2E+2025%2E+CiteEval%3A+Principle%2Ddriven+citation+evaluation+for+source+attribution%2E+)
21. Junjie Ye, Yuming Yang, Yang Nan, Shuo Li, Qi Zhang, Tao Gui, Xuanjing Huang, Peng Wang, Zhongchao Shi, and Jianping Fan. 2025. Analyzing the effects of supervised fine-tuning on model knowledge from token and parameter levels. In Proceedings of the 2025 Conference on Empirical Methods in Natural La… [find](https://scholar.google.com/scholar?q=Junjie+Ye%2C+Yuming+Yang%2C+Yang+Nan%2C+Shuo+Li%2C+Qi+Zhang%2C+Tao+Gui%2C+Xuanjing+Huang%2C+Peng+Wang%2C+Zhongchao+Shi%2C+and+Jianping+Fan%2E+2025%2E+Analyzing+the+effects+of+supervised+fine%2Dtuning+on+model+knowledge+from+)
22. Yingming Zheng, Hanqi Li, Kai Yu, and Lu Chen. 2025. When long helps short: How context length in supervised fine-tuning affects behavior of large language models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 10293–10308, Suzhou, China. Association… [find](https://scholar.google.com/scholar?q=Yingming+Zheng%2C+Hanqi+Li%2C+Kai+Yu%2C+and+Lu+Chen%2E+2025%2E+When+long+helps+short%3A+How+context+length+in+supervised+fine%2Dtuning+affects+behavior+of+large+language+models%2E+In+Proceedings+of+the+2025+Conferenc)