CRL-Prompt: Contrastive and Reinforcement Learning for Soft Prompt Tuning for Text Classification

Danila Lapokin, Andrey Savchenko


Abstract
Prompt choice is crucial in adapting language models to text classification tasks, particularly under low-resource conditions. Manual prompt engineering is time-consuming, non-scalable, and brittle, while current auto-prompting techniques are still far from maturity. This paper presents a two-stage method for prompt learning of frozen language models, CRL-Prompt, based on soft prompt initialization followed by contrastive and reinforcement-based refinement. An experimental study demonstrates that our approach achieves consistent improvements in accuracy over baseline prompt tuning strategies, with gains of up to 2.2% while training fewer than 0.25% of model parameters.
Anthology ID:
2026.acl-srw.123
Volume:
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026)
Month:
July
Year:
2026
Address:
San Diego, California, United States
Editors:
Santosh T.Y.S.S., Juan Diego Rodriguez, Ona de Gibert
Venue:
ACL
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
1391–1398
Language:
URL:
https://preview.aclanthology.org/ingest-acl/2026.acl-srw.123/
DOI:
Bibkey:
Cite (ACL):
Danila Lapokin and Andrey Savchenko. 2026. CRL-Prompt: Contrastive and Reinforcement Learning for Soft Prompt Tuning for Text Classification. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), pages 1391–1398, San Diego, California, United States. Association for Computational Linguistics.
Cite (Informal):
CRL-Prompt: Contrastive and Reinforcement Learning for Soft Prompt Tuning for Text Classification (Lapokin & Savchenko, ACL 2026)
Copy Citation:
PDF:
https://preview.aclanthology.org/ingest-acl/2026.acl-srw.123.pdf