--- title: "Lexical Conditioning of Model's Distribution through" description: "The authors propose Uncertainty-Gated Lexical Decoding (UGLD) to ease this tension. 9 references." authors: ["Michele Papucci", "Giulia Venturi", "Felice Dell'Orletta"] year: 2026 venue: "Proceedings of the Joint Workshop on Readability and Text Simplification (READIxTSAR) @ LREC 2026" source: "https://aclanthology.org/2026.readi-1.7/" doi: "10.63317/5hgiz64xsj6s" licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)" licence_url: "https://creativecommons.org/licenses/by-nc/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/papucci-2026-lexical-conditioning-model-s-distribution/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/papucci-2026-lexical-conditioning-model-s-distribution/) [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/34/) / Lexical Conditioning of Model's Distribution through Uncertainty-gated Soft-Mixing of Probabilities # Lexical Conditioning of Model's Distribution through Uncertainty-gated Soft-Mixing of Probabilities **Authors:** [Michele Papucci](https://md.research.elamur.ai/free/en/authors/michele-papucci/), [Giulia Venturi](https://md.research.elamur.ai/free/en/authors/giulia-venturi/) and [Felice Dell'Orletta](https://md.research.elamur.ai/free/en/authors/felice-dell-orletta/) · **Year:** 2026 · **Venue:** [Proceedings of the Joint Workshop on Readability and Text Simplification (READIxTSAR) @ LREC 2026](https://md.research.elamur.ai/free/en/venues/readi/) ([Joint Workshop on Readability and Text Simplification (READIxTSAR) @ LREC 2026](https://md.research.elamur.ai/free/en/venues/readi/2026/)) · **Publisher:** [European Language Resources Association (ELRA)](https://md.research.elamur.ai/free/en/publishers/european-language-resources-association-elra/) · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.readi-1.7/) · [PDF](https://aclanthology.org/2026.readi-1.7.pdf) · [DOI 10.63317/5hgiz64xsj6s](https://doi.org/10.63317/5hgiz64xsj6s) · [OpenAlex](https://openalex.org/W7166531214) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary Decoding-time steering of LLM output often trades fluency for control: the harder the push toward or away from certain words, the worse the text tends to read. The authors propose Uncertainty-Gated Lexical Decoding (UGLD) to ease this tension. It applies lexical intervention mainly when the model is uncertain about the next token, measured by the Shannon entropy of its predictive distribution. A sigmoid gate turns that entropy into an intervention weight. The method has two variants. UGLD-t blends the model's distribution with a prior concentrated on a "green" word list. UGLD-a subtracts scaled penalties from the logits of "red" tokens. Neither variant needs extra training, classifiers or auxiliary models, which makes the approach attractive for low-resource languages. Experiments in Italian used ANITA 8B and Qwen 3 4B on paraphrasing and free continuation, in two scenarios: - **Simple Vocabulary Conditioning:** steering output toward words from the New Basic Italian Vocabulary. - **Jargon Reduction:** steering output away from medical terminology. Automatic coverage scores improved over standard decoding, and human judges found fluency largely intact. The authors see these two directions as building blocks for future controlled text simplification. ## Key points - **Gating mechanism:** the gate is computed as σ((H(p) − τ)/s). - τ is the entropy threshold and s controls how sharply the gate opens. - The effective strength is the gate value times a maximum, α_max for mixing (UGLD-t) or λ_max for logit penalties (UGLD-a). - **Simple vocabulary data:** 100 sentences sampled from Wikipedia, with green tokens drawn from the NBIV (De Mauro and Chiari, 2016). - The NBIV contains about 7,000 words; only content words in the "Fundamental Words" class were used. - These lemmas expand to 98,430 word forms, which map to 4,686 tokens for ANITA and 4,654 for Qwen. - **Jargon Reduction:** the setup is domain-general in principle but was instantiated on the medical domain via token penalties. - **Evaluation:** a lexical coverage metric was tracked across several intervention strengths (Soft, Medium, Strong), alongside human ratings of fluency. - **Availability:** the code is released on GitHub (michelepapucci/ugld) and installable from PyPI as `ugld`. *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 - [University of Pisa](https://md.research.elamur.ai/free/en/universities/university-of-pisa/) — University, Pisa, Italy · 25 works · 3,842 citations - [Institute for Computational Linguistics “A. Zampolli”](https://md.research.elamur.ai/free/en/labs/institute-for-computational-linguistics-a-zampolli/) — Research institute, Pisa, Italy · 14 works · 1,028 citations 📚 Shelves (4): [🖥️ 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 ## Cite Cite as: Michele Papucci, Giulia Venturi and Felice Dell'Orletta. 2026. Lexical Conditioning of Model's Distribution through Uncertainty-gated Soft-Mixing of Probabilities. Proceedings of the Joint Workshop on Readability and Text Simplification (READIxTSAR) @ LREC 2026. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{papucci2026lexical, title = {{Lexical Conditioning of Model's Distribution through Uncertainty-gated Soft-Mixing of Probabilities}}, author = {Michele Papucci, Giulia Venturi and Felice Dell'Orletta}, booktitle = {Proceedings of the Joint Workshop on Readability and Text Simplification (READIxTSAR) @ LREC 2026}, year = {2026}, doi = {10.63317/5hgiz64xsj6s}, url = {https://doi.org/10.63317/5hgiz64xsj6s} } ``` - [← previous Lexical and Syntactic Diversity: Still Lost in Machine Translation?](https://md.research.elamur.ai/free/en/papers/volkart-2026-lexical-syntactic-diversity-still-lost/) - [next → Lexical Innovation in Business Colour Idioms: Evidence from Large Language Models in Five Languages](https://md.research.elamur.ai/free/en/papers/oleskeviciene-2026-lexical-innovation-business-colour-idioms/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/34/) ## References 9 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Souvik Das, Lifeng Jin, Linfeng Song, Haitao Mi, Baolin Peng, and Dong Yu. 2025. Entropy guided Marco Polignano, Pierpaolo Basile, and Giovanni Semeraro. 2024. Advanced natural-based interaction for the italian language: Llamantino-3-anita. [find](https://scholar.google.com/scholar?q=Souvik+Das%2C+Lifeng+Jin%2C+Linfeng+Song%2C+Haitao+Mi%2C+Baolin+Peng%2C+and+Dong+Yu%2E+2025%2E+Entropy+guided+Marco+Polignano%2C+Pierpaolo+Basile%2C+and+Giovanni+Semeraro%2E+2024%2E+Advanced+natural%2Dbased+interaction+for+t) 2. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8233–8248, Singapore. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Proceedings+of+the+2023+Conference+on+Empirical+Methods+in+Natural+Language+Processing%2C+pages+8233%E2%80%938248%2C+Singapore%2E+Association+for+Computational+Linguistics%2E) 3. Matt Post and David Vilar. 2018. [Fast lexically constrained decoding with dynamic beam allocation for neural machine translation](https://md.research.elamur.ai/free/en/papers/post-2018-fast-lexically-constrained-decoding-dynamic/). In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers),… 4. Wangchunshu Zhou, Yuchen Eleanor Jiang, Ethan Wilcox, Ryan Cotterell, and Mrinmaya Sachan. 2023. Controlled text generation with natural language instructions. In Proceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org. [find](https://scholar.google.com/scholar?q=Wangchunshu+Zhou%2C+Yuchen+Eleanor+Jiang%2C+Ethan+Wilcox%2C+Ryan+Cotterell%2C+and+Mrinmaya+Sachan%2E+2023%2E+Controlled+text+generation+with+natural+language+instructions%2E+In+Proceedings+of+the+40th+International) 5. Seungmin Shin, Dooyoung Kim, and Youngjoong Ko. 2025. ECO decoding: Entropy-based control for controllability and fluency in controllable dialogue generation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 28297–28309, Suzhou, China. Association for… [find](https://scholar.google.com/scholar?q=Seungmin+Shin%2C+Dooyoung+Kim%2C+and+Youngjoong+Ko%2E+2025%2E+ECO+decoding%3A+Entropy%2Dbased+control+for+controllability+and+fluency+in+controllable+dialogue+generation%2E+In+Proceedings+of+the+2025+Conference+on+) 6. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jin… — [Qwen3 Technical Report](https://md.research.elamur.ai/free/en/papers/yang-2025-qwen3-technical-report/) 7. Kevin Yang and Dan Klein. 2021. [FUDGE: Controlled text generation with future discriminators](https://md.research.elamur.ai/free/en/papers/yang-2021-fudge-controlled-text-generation-future/). In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3511–3535, Online. Association for Computational Ling… 8. Hanqing Zhang, Haolin Song, Shaoyu Li, Ming Zhou, and Dawei Song. 2023. A survey of controllable text generation using transformer-based pre-trained language models. ACM Comput. Surv., 56(3). [find](https://scholar.google.com/scholar?q=Hanqing+Zhang%2C+Haolin+Song%2C+Shaoyu+Li%2C+Ming+Zhou%2C+and+Dawei+Song%2E+2023%2E+A+survey+of+controllable+text+generation+using+transformer%2Dbased+pre%2Dtrained+language+models%2E+ACM+Comput%2E+Surv%2E%2C+56%283%29%2E) 9. Tianqi Zhong, Quan Wang, Jingxuan Han, Yongdong Zhang, and Zhendong Mao. 2023. Air-decoding: Attribute distribution reconstruction for decoding-time controllable text generation. In [find](https://scholar.google.com/scholar?q=Tianqi+Zhong%2C+Quan+Wang%2C+Jingxuan+Han%2C+Yongdong+Zhang%2C+and+Zhendong+Mao%2E+2023%2E+Air%2Ddecoding%3A+Attribute+distribution+reconstruction+for+decoding%2Dtime+controllable+text+generation%2E+In) *Updated 1 November 2026*